Using Remote Sensing and GIS to Monitor and Predict Urban Growth-Case Study in Alachua County, Florida

MISSING IMAGE

Material Information

Title:
Using Remote Sensing and GIS to Monitor and Predict Urban Growth-Case Study in Alachua County, Florida
Physical Description:
1 online resource (291 p.)
Language:
english
Creator:
Guo, Yong Hong
Publisher:
University of Florida
Place of Publication:
Gainesville, Fla.
Publication Date:

Thesis/Dissertation Information

Degree:
Doctorate ( Ph.D.)
Degree Grantor:
University of Florida
Degree Disciplines:
Design, Construction, and Planning Doctorate, Design, Construction and Planning
Committee Chair:
Zwick, Paul D
Committee Members:
Bejleri, Ilir
Carr, Margaret H
Binford, Michael W

Subjects

Subjects / Keywords:
classification -- growth -- lulc -- mlr
Design, Construction and Planning -- Dissertations, Academic -- UF
Urban and Regional Planning -- Dissertations, Academic -- UF
Genre:
Design, Construction, and Planning Doctorate thesis, Ph.D.
Electronic Thesis or Dissertation
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )

Notes

Abstract:
Alachua County experienced remarkable urban growth in the extent of the urban area during the past three decades, largely due to population growth. This dissertation simulates and predicts future urban growth in the county based on population growth. First, this dissertation classifies urban land uses and land covers (LULCs) based on eleven classes equivalent to the U.S. Geological Survey (USGS) Level III standard using remote sensing Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) data, as well as the state-of-the-art classification and regression tree (CART) method. The CART method is then compared to the Vegetation-Impervious Surface-Soil (V-I-S) and conventional supervised classification methods based on a pilot area. Next, this dissertation classifies 1982, 1994, and 2003 LULC classes for Alachua County as a whole. Second, a multinomial logistic regression (MLR) model is employed to simulate urban LULC development in 2003 based on the urban LULC classified for 1982, 1994, and 2003, respectively. Accordingly, a number of probability maps are created for the four classified urban uses, which include single-family, multi-family, commercial-institutional-transportation, and industrial-warehouses. Third, an urban LULC allocation process is undertaken for Alachua County, based on five scenarios for 2020 and 2030. The five scenarios include the business as usual (BAU) scenario, infill development, increased density development, redevelopment, and the conservation scenario. This research has wide application in both the public and the private sectors of urban planning and additional areas in which the ability to describe the spatial extent of urban areas and their future expansion is of interest.
Statement of Responsibility:
by Yong Hong Guo.
General Note:
In the series University of Florida Digital Collections.
General Note:
Includes vita.
Bibliography:
Includes bibliographical references.
General Note:
Description based on online resource; title from PDF title page.
General Note:
This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Thesis:
Thesis (Ph.D.)--University of Florida, 2012.
General Note:
Adviser: Zwick, Paul D.

Record Information

Source Institution:
UFRGP
Rights Management:
Applicable rights reserved.
Classification:
lcc - LD1780 2012
System ID:
UFE0043914:00001


This item has the following downloads:


Full Text

PAGE 1

1 USING REMOTE SENSING AND GIS TO MONITOR AND PREDICT URBAN GROWTH CASE STUDY IN ALACHU A COUNTY, FLORIDA By YONG HONG GUO A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2012

PAGE 2

2 2012 Yong Hong Guo

PAGE 3

3 To my beloved wife Zheng Song and my parents Guo Zhizheng and Xu Jie In memory of my father Mr. Guo Zhizheng

PAGE 4

4 ACKNOWLEDGMENTS I would like to thank my wife, Zheng Song, for her full consideration and wholeheartedness in supporting my Ph.D. study. Because of her vigorous and cooperative support, my Ph.D. study was conducted smoothly and efficiently. Also, I would like to thank my parents, Mr. Guo Zhizheng and Mrs. Xu Jie, for their moral and financial support throughout my entire Ph.D. I would not be able to finish this Ph.D. without their timely and enthusiastic encouragement and assistance. In addition, I would like to thank Dr. Paul Zwick, my committee chair, for his support and advice during my study. Dr. Zwick provided valuable comments and guidance for this research; he also financially assisted me for the research. Dr. Zwick provided me with heuristic discussions, effective technical comments and input for the writing of this dissertation; without his technical support and financial assistance, this dissertation could not have been fulfilled. I am also thankful for my committee members, Dr. Michael Binford, Dr. Ilir Bejleri, and Prof. Margaret Carr for their valuable comments and guidance for this dissertation. Finally, I am also grateful for Mr. Chad Riding at Caltrans, Calif., Mr. Thomas Cole and Ms. Lyndsay Brown at the Reading and Writing Center at the University of Florid a who helped me edit the manuscripts of this dissertation.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ ............... 4 LIST OF TABLES ................................ ................................ ................................ ........................... 8 LIST OF FIGURES ................................ ................................ ................................ ....................... 14 LIST OF ABBR EVIATIONS ................................ ................................ ................................ ........ 17 ABSTRACT ................................ ................................ ................................ ................................ ... 19 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .................. 21 Brief Review of Urban Growth in the U.S. ................................ ................................ ............ 22 Alternatives to Decentralized Urban Development ................................ ................................ 24 Research Questions ................................ ................................ ................................ ................. 28 Research Applications ................................ ................................ ................................ ............ 29 2 URBAN LAND USE AND LAND COVER CLASSIFICATION ................................ ........ 30 Chapter Overview ................................ ................................ ................................ ................... 30 Literature Review ................................ ................................ ................................ ................... 30 Supervised and Unsupervised Classification Methods ................................ .................... 30 V I S Model ................................ ................................ ................................ ..................... 34 CART Method ................................ ................................ ................................ ................. 40 Spectra l Characteristics of Urban LULC ................................ ................................ ................ 43 Research Question ................................ ................................ ................................ .................. 45 Methodologies ................................ ................................ ................................ ........................ 45 Classification System ................................ ................................ ................................ ...... 45 Data Inventory and Pilot Area ................................ ................................ ......................... 47 CART Method ................................ ................................ ................................ ................. 48 ENVI 4.4 RuleGen 1.02 ................................ ................................ ................................ .. 50 IDRISI Andes ................................ ................................ ................................ .................. 52 V I S Method ................................ ................................ ................................ ................... 53 Supervised Metho d ................................ ................................ ................................ .......... 54 Pilot Area Results ................................ ................................ ................................ ................... 54 CART Method ................................ ................................ ................................ ................. 54 V I S Method ................................ ................................ ................................ ................... 57 Supervised Metho d ................................ ................................ ................................ .......... 58 Pilot Area Summaries of CART Method, V I S Method, and Supervised Method ........ 59 County Wide Urban LULC Classifications Using the Preferred CART Methodology ......... 59

PAGE 6

6 3 URBAN MODEL BUILDING THE MULTINOMIAL LOGISTIC REGRESSION MODEL ................................ ................................ ................................ ................................ .. 94 Chapter Overview ................................ ................................ ................................ ................... 94 The Accuracy of Urban Growth Modeling ................................ ................................ ............. 95 Relevant Literature Review for Logistic Regression ................................ .............................. 98 Methodologies ................................ ................................ ................................ ...................... 101 Logistic Regression Results ................................ ................................ ................................ .. 107 Model Calibration Single Family ................................ ................................ .................. 109 Model Calibration Multi family ................................ ................................ .................... 110 Model Calibration Commercial Institutional Transportation ................................ ....... 111 Model Calibration Industrial Warehouses ................................ ................................ .... 113 Refined LULC Logistic Regression Model ................................ ................................ ... 114 Single family ................................ ................................ ................................ .......... 115 Multi family ................................ ................................ ................................ ........... 116 Commercial institutional transportation ................................ ................................ 116 Industrial warehouses ................................ ................................ ............................. 117 Sensitivity Analysis ................................ ................................ ................................ ....... 117 Single family ................................ ................................ ................................ .......... 120 Multi family ................................ ................................ ................................ ........... 122 Commercial institutional transportation ................................ ................................ 125 Industrial warehouses ................................ ................................ ............................. 127 2003 Urban LULC Simulation ................................ ................................ ...................... 129 4 URBAN LAND USE ALLOCATIONS ................................ ................................ ............... 176 Chapter Overview ................................ ................................ ................................ ................. 176 Allocation Literature Review ................................ ................................ ............................... 179 Methodology ................................ ................................ ................................ ......................... 190 Five Scenarios ................................ ................................ ................................ ............... 190 BAU scenario ................................ ................................ ................................ ......... 191 Infill scenario ................................ ................................ ................................ .......... 192 Increased density development scenario ................................ ................................ 193 Redevelopment scenario ................................ ................................ ........................ 194 Conservation scenario ................................ ................................ ............................ 195 Forecasting Development Acreages for Five Scenarios ................................ ................ 196 Baseline forecasting method (BAU forecasting and infill forecasting methods) ... 196 Increased density forecasting method ................................ ................................ .... 199 Redevelopment forecasting method ................................ ................................ ....... 201 The Conflict Analysis ................................ ................................ ................................ .... 202 Creation of the masks ................................ ................................ ............................. 204 Collapse of the preference maps ................................ ................................ ............ 205 Creation of the conflict maps ................................ ................................ ................. 206 Final Allocation ................................ ................................ ................................ ............. 209 Timeframes of the final allocation ................................ ................................ ......... 209 Urban uses to be allocated in final allocation ................................ ........................ 209 Final allocation sequence ................................ ................................ ....................... 210

PAGE 7

7 Final allocation cell size and spatial reference ................................ ....................... 210 Final allocation preference map slicing, policy allocation, and cell statistics ....... 211 Final Allocation Results ................................ ................................ ................................ ....... 212 BAU Scenario ................................ ................................ ................................ ................ 212 Infill Scenari o ................................ ................................ ................................ ................ 214 Increased Density Development Scenario ................................ ................................ ..... 216 Redevelopment Scenario ................................ ................................ ............................... 217 Conservation Scenario ................................ ................................ ................................ ... 220 5 CONCLUSIONS AND DISCUSSION ................................ ................................ ................ 280 LIST OF REFERENCES ................................ ................................ ................................ ............. 285 BIOGRAPHICAL SKETCH ................................ ................................ ................................ ....... 291

PAGE 8

8 LIST OF TABLES Table page 2 1 Error matrix (ENVI 4.4 RuleGen 1.02 QUEST algorithm) ................................ ............... 62 2 2 Accuracy totals (ENVI 4.4 RuleGen 1.02 QUEST algorithm) ................................ .......... 63 2 3 KAPPA (K^) statistics (ENVI 4.4 RuleGen 1.02 QUEST algorithm) .............................. 63 2 4 Error matrix (IDRISI Andes Ratio Gain rule) ................................ ................................ ... 64 2 5 Accuracy totals (IDRISI Andes Ratio Gain rule) ................................ .............................. 65 2 6 KAPPA (K^) statistics (IDRISI Andes Ratio Gain rule) ................................ ................... 65 2 7 Error matrix (CART method V I S model) ................................ ................................ ....... 66 2 8 Accuracy totals (CART method V I S model) ................................ ................................ .. 67 2 9 KAPPA (K^) statistics (CART method V I S model) ................................ ....................... 67 2 10 Comparisons of the strengths and weaknesses of urban LULC classification methods .... 68 2 11 Error matrix (1982 classification map) ................................ ................................ .............. 72 2 12 Accuracy totals (1982 classification map) ................................ ................................ ......... 73 2 13 KAPPA (K^) statistics (1982 classification map) ................................ .............................. 73 2 14 Error matrix (1994 classification map ) ................................ ................................ .............. 74 2 15 Accuracy totals (1994 classification map) ................................ ................................ ......... 75 2 16 KAPPA (K^) statistics (1994 classification map) ................................ .............................. 75 2 17 Error matrix (2003 classification map) ................................ ................................ .............. 76 2 18 Accuracy totals (2003 classification map) ................................ ................................ ......... 77 2 19 KAPPA (K^) statistics (2003 classification map) ................................ .............................. 77 2 20 ROI pixels of LULC classe s in 1982, 1994, and 2003 ................................ ...................... 78 3 1 Independent variables for the single family use ................................ .............................. 130 3 2 Independent variables for the multi family use ................................ ............................... 133 3 3 Independent variables for the commercial institutional transportation use ..................... 135

PAGE 9

9 3 4 Independent variables for the industrial warehouses use ................................ ................ 138 3 5 .............................. 139 3 6 Autocorrelation and four dependent variables for the Z score test ................................ .. 140 3 7 Comparison between the original models and the refined models ................................ .. 140 3 8 Independent va riables for refined single family use ................................ ........................ 141 3 9 Independent variables for refined multi family use ................................ ......................... 142 3 10 Independent variables for refined commercial institutional transportation use .............. 143 3 11 Independent variables for refined industrial warehouses use ................................ .......... 145 3 12 2 2 contingency table for the ROC curve ................................ ................................ ..... 146 3 13 Contingency table of change versus non change for observed values for the single family use ................................ ................................ ................................ ......................... 146 3 14 Contingency table of change versus non change for expected values for the single family use ................................ ................................ ................................ ......................... 146 3 15 Pseudo R family use ................................ 146 3 16 Overall accuracy for the single family use ................................ ................................ ...... 146 3 17 ROC analysis and 2 2 contingency table based on 0.2 percent stratified sampling for the single family use ................................ ................................ ................................ .. 147 3 18 Contingency table of change versus non change for observed values for th e multi family use ................................ ................................ ................................ ......................... 147 3 19 Contingency table of change versus non change for expected values for th e multi family use ................................ ................................ ................................ ......................... 147 3 20 Pseudo R family use ................................ 147 3 21 Overall accuracy for the multi family use ................................ ................................ ....... 147 3 22 ROC analysis and 2 2 contingency table based on 1.5 percent stratified sampling for the multi fam ily use ................................ ................................ ................................ ... 148 3 23 Contingency table of change versus non change for observed values for the commercial institutional transportation use ................................ ................................ .... 148 3 24 Contingency table of change versus non change for expected values for the commercial institutional transportation use ................................ ................................ .... 148

PAGE 10

10 3 25 Pseudo R institutional transportation use ................................ ................................ ................................ ............. 148 3 26 Overall accuracy for the commercial institutional transportation use ............................. 148 3 27 ROC analysis and 2 2 contingency table based on 8.5 percent stratified sampling for the commercial institutional transportation use ................................ ......................... 149 3 28 Contingency table of change versus non change for observed values for the industr ial warehouses use ................................ ................................ ................................ 149 3 29 Contingency table of change versus non change for expected values for the industr ial warehouses use ................................ ................................ ................................ 149 3 30 Pseudo R warehouses use ................. 149 3 31 Overall accuracy for the industrial warehouses use ................................ ........................ 149 3 32 ROC analysis and 2 2 contingency table based on 15 percent stratified sampling for the industrial warehouses use ................................ ................................ ..................... 150 3 33 Accuracy assessment for 2003 urban LULC simulation map based on predicted pixels (1) ................................ ................................ ................................ .......................... 151 3 34 Accuracy assessment for 2003 urban LULC simulation map based on predicted percent (2) ................................ ................................ ................................ ........................ 152 3 35 Accuracy assessment for 2003 urban LULC sim ulation map (3) ................................ .... 153 3 36 Accuracy assessment for 2003 urban LULC simulation map (4) ................................ .... 154 3 37 Accuracy assessment for 2003 urban LULC simulation map (5) ................................ .... 154 4 1 The baseline forecasting method for urban development acreages (single family) ........ 223 4 2 The base line forecasting method for urban development acreages (multi family) ......... 223 4 3 The baseline forecasting method for urban developmen t acreages (total residential) ..... 223 4 4 The baseline forecasting method for urban development acreages (commercial institutional t ransportation) ................................ ................................ .............................. 224 4 5 The baseline forecasting method for urban development acreages (industrial warehouses) ................................ ................................ ................................ ...................... 224 4 6 The baseline forecasting method for urban development acreages (total urban) ............ 224 4 7 The baseline forecasting method for urban acreage changes ................................ ........... 225

PAGE 11

11 4 8 Baseline urban acreage changes with percentages ................................ ........................... 225 4 9 The increased density forecasting method for urban development acreages (single family) ................................ ................................ ................................ .............................. 225 4 10 The increased density forecasting method for urban development acreages (mult i family) ................................ ................................ ................................ .............................. 226 4 11 The increased density forecasting method for urban development acreages (total residential) ................................ ................................ ................................ ........................ 226 4 12 The increased density forecasting method for urban development acreages (commercial institutional transportation) ................................ ................................ ........ 226 4 13 The increased density forecasting method for urban development acreages (industrial warehouses) ................................ ................................ ................................ .... 227 4 14 The increased density forecasting method for urban development acreages (total urban) ................................ ................................ ................................ ............................... 227 4 15 The increased density forecasting method for urban acreage changes ............................ 227 4 16 The redevelopment forecasting method for urban development acreages (single family) ................................ ................................ ................................ .............................. 228 4 17 The redevelopment forecasting method for urban development acreages (multi family) ................................ ................................ ................................ .............................. 228 4 18 The redevelopment forecasting method for urban development acreages (total residential) ................................ ................................ ................................ ........................ 228 4 19 The redevelopment forecasting method for urban development acreages (commercial institutional transportation) ................................ ................................ ........ 229 4 20 The redevelopment forecasting method for urban development acreages (industrial warehouses) ................................ ................................ ................................ ...................... 229 4 21 The redevelopment forecasting method for urban development acreages (total urban) .. 230 4 22 The redevelopment forecasting method for urban acreage changes ................................ 230 4 23 Conflict scores and equivalent descriptions ................................ ................................ ..... 231 4 24 Conflict scores for the BAU scenario ................................ ................................ .............. 237 4 25 Conflict scores for the infill scenario ................................ ................................ ............... 238 4 26 Single family acreage demand and allocation in 2010, 2020, and 2030 for the BAU scenario ................................ ................................ ................................ ............................ 239

PAGE 12

12 4 27 Multi family acreage demand and allocation in 2010, 2020, and 2030 for the BAU scenario ................................ ................................ ................................ ............................ 239 4 28 Commercial institution transportation acreage demand and allocation in 2010, 2020, and 2030 for the BAU scenario ................................ ................................ ....................... 239 4 29 Industrial warehouses acreage demand and allocation in 2010, 2020, and 2030 for the BAU scenario ................................ ................................ ................................ ............. 239 4 30 Infill scenario infill development acreages and conflict development acreages in 2020 and 2030 ................................ ................................ ................................ .................. 240 4 31 Single family acreage demand and allocation in 2010, 2020, and 2030 for the infill scenario ................................ ................................ ................................ ............................ 240 4 32 Multi family acreage demand and allocation in 2010, 2020, and 2030 for the infill scenario ................................ ................................ ................................ ............................ 241 4 33 Commercial institution transportation acreage demand and allocation in 2010, 2020, and 2030 for the infill scenario ................................ ................................ ........................ 241 4 34 Industrial warehouses acreage demand and allocation in 2010, 2020, and 2030 for the infill scenario ................................ ................................ ................................ .............. 241 4 35 Increased density development scenario development acreages and conflict development acreages in 2020 and 2030 ................................ ................................ ......... 242 4 36 Single family acreage demand and allocation in 2010, 2020, and 2030 for the increased density development scenario ................................ ................................ .......... 242 4 37 Multi family acreage demand and allocation in 2010, 2020, and 2030 for the increased density development scenario ................................ ................................ .......... 243 4 38 Commercial institution transportation acreage demand and allocation in 2010, 2020, and 2030 for the increased density development scenario ................................ ............... 243 4 39 Industrial warehouses acreage demand and allocation in 2010, 2020, and 2030 for the increased density development scenario ................................ ................................ .... 243 4 40 Redevelopment scenario development acreages and conflict development acreages in 2020 and 2030 ................................ ................................ ................................ .................. 244 4 41 Single family acreage demand and allocation in 2010, 2020, and 2030 for the redevelopment scenario ................................ ................................ ................................ ... 244 4 42 Multi family acreage demand and allocation in 2010, 2020, and 2030 for the redevelopment scenario ................................ ................................ ................................ ... 244

PAGE 13

13 4 43 Commercial institution transportation acreage demand and allocation in 2010, 2020, and 2030 for the redevelopment scenario ................................ ................................ ........ 245 4 44 Industrial warehouses acreage demand and allocation in 2010, 2020, and 2030 for the redevelopment scenario ................................ ................................ .............................. 245 4 45 Conservation scenario development acreages and conflict development acreages in 2020 and 2030 ................................ ................................ ................................ .................. 245 4 46 Single family acreage demand and allocation in 2010, 2020, and 2030 for the conservation scenario ................................ ................................ ................................ ....... 246 4 47 Multi family acreage demand and allocation in 2010, 2020, and 2030 for the conservation scenario ................................ ................................ ................................ ....... 246 4 48 Commercial institution transportation acreage demand and allocation in 2010, 2020, and 2030 for the conservation scenario ................................ ................................ ........... 246 4 49 Industrial warehouses acreage demand and allocation in 2010, 2020, and 2030 for the conservation scenario ................................ ................................ ................................ 246

PAGE 14

14 LIST OF FIGURES Figure page 2 1 I S model ................................ ................................ ................................ 79 2 2 V I S model and point T ................................ ................................ ................................ .... 79 2 3 Pilot area location for LULC classification ................................ ................................ ....... 80 2 4 Urban LULC classifications using ENVI 4.4 RuleGen 1.02 QUEST module .................. 81 2 5 Urban LULC classifications using IDRISI Andes gain ratio rule ................................ ..... 82 2 6 Urban LULC classifications using IDRISI Andes entropy rule ................................ ........ 83 2 7 Urban LULC classifications using IDRISI Andes Gini rule ................................ ............. 84 2 8 Or iginal maps for five components ................................ ................................ .................... 85 2 9 Exponential transformed maps for five components ................................ ......................... 86 2 10 Adjusted rule maps for five components. ................................ ................................ .......... 87 2 11 V I S model final classification map ................................ ................................ ................. 88 2 12 V I S model using CART method to clas sify LULC ................................ ........................ 89 2 13 Urban LULC classifications using the supervised method ................................ ................ 90 2 14 1982 classification map for Alachua County ................................ ................................ ..... 91 2 15 1994 classification map for Alachua Count y ................................ ................................ ..... 92 2 16 2003 classification map for Alachua County ................................ ................................ ..... 93 3 1 Predicted 2003 single family probability map ................................ ................................ 155 3 2 2003 single family prediction map ................................ ................................ .................. 156 3 3 Predicted 2003 multi family probability map ................................ ................................ .. 157 3 4 2003 multi family prediction map ................................ ................................ ................... 158 3 5 Predicted 2003 commercial institutional transportation probability map ....................... 159 3 6 2003 commercial institutional transportation prediction map ................................ ......... 160 3 7 Predicted 2003 industrial warehouses probability map ................................ ................... 161

PAGE 15

15 3 8 2003 industrial warehouses prediction map ................................ ................................ .... 162 3 9 Predicted 2003 single family probability map (refined) ................................ .................. 163 3 10 Refined 2003 single family prediction map ................................ ................................ .... 164 3 11 Predicted 2003 multi family probability map (refined) ................................ ................... 165 3 12 Refined predicted 2003 multi family prediction map ................................ ...................... 166 3 13 Predicted 2003 commercial institutional transportation probability map (refined) ........ 167 3 14 Refined 2003 commercial institutional transportation prediction map ........................... 168 3 15 Predicted 2003 industrial warehouses probability map (refined) ................................ .... 16 9 3 16 Refined 2003 industrial warehouses prediction m ap ................................ ....................... 170 3 17 ROC curve for the single family use ................................ ................................ ............... 171 3 18 ROC curve for the multi family use ................................ ................................ ................ 172 3 19 ROC curve for the commercial institutional transportation use ................................ ...... 173 3 20 ROC curve for the industrial warehouses u se ................................ ................................ 174 3 21 2003 Alachua County LULC simulation ................................ ................................ ......... 175 4 1 Areas within urban buffer areas and outside urban buffer area s ................................ ..... 247 4 2 Major roads in Alachua County, incorporated towns and cities, and Urban Cluster area s ................................ ................................ ................................ ................................ 248 4 3 The BAU mas k ................................ ................................ ................................ ................ 249 4 4 The infill development mas k ................................ ................................ ........................... 250 4 5 Five scenarios with their relationship with the urban area and the greenfield areas, where the urban area is within 1,000 meters of the urban buffer areas and the green field area is outside 1,000 meters of the urban buffer areas. ................................ .. 251 4 6 Population growth in Alachua County from 1982 2030. Consider ing the natural se is not included ................... 252 4 7 Ratio of single family population/multi family population from 1982 2030. ................. 253 4 8 BAU preference map for single family development ................................ ...................... 254 4 9 Collapsed map for single family use for the BAU scenario ................................ ............ 255

PAGE 16

16 4 10 Conflict map for the BAU scenario ................................ ................................ ................. 256 4 11 Conflict map for the infill scenario ................................ ................................ .................. 257 4 12 2010 Alachua County LULC Current Plan 1 ................................ ................................ ... 258 4 13 2010 Alachua County LULC Current Plan 2 ................................ ................................ ... 259 4 14 2020 Alachua County LULC Alternative Business As Usual Scenario 1 ....................... 260 4 15 2020 Alachua County LULC Alternative Business As Usual Scenario 2 ....................... 261 4 16 2030 Alachua County LULC Alternative Business As Usual Scenario 1 ....................... 262 4 17 2030 Alachua County LULC Alternative Business As Usual Scenario 2 ....................... 263 4 18 2020 Alachua County LULC Alternative Infill Development Scenario 1 ...................... 264 4 19 2020 Alachua County LULC Alternative Infill Development Scenario 2 ...................... 265 4 20 2030 Alachua County LULC Alternative Infill Development Scenario 1 ...................... 266 4 21 2030 Alachua County LULC Alternative Infill Development Scenario 2 ...................... 267 4 22 2020 Alachua County LULC Alternative Increased Density Development Scenario 1 .. 268 4 23 2020 Alachua County LULC Alternative Increased Density Development Scenario 2 .. 269 4 24 2030 Alachua County LULC Alternative Increased Density Development Scenario 1 .. 270 4 25 2030 Alachua County LULC Alternative Increased Density Development Scenario 2 .. 271 4 26 2020 Alachua County LULC Alternative Redevelopment Scenario 1 ............................ 272 4 27 2020 Alachua County LULC Alternative Redevelopment Scenario 2 ............................ 273 4 28 2030 Alachua County LULC Alternative Redevelopment Scenario 1 ............................ 274 4 29 2030 Alachua County LULC Alternative Redevelopment Scenario 2 ............................ 275 4 30 2020 Alachua County LUL C Alternative Conservation Scenario 1 ................................ 276 4 31 2020 Alachua County LULC Alternative Conservation Scenario 2 ................................ 277 4 32 2030 Alachua County LULC Alternative Conservation Scenario 1 ................................ 278 4 33 2030 Alachua County LULC Alternative Conservation Scenario 2 ................................ 279

PAGE 17

17 LIST OF ABBREVIATION S ABM Agent Based Model AHP Analytic Hierarchy Process AI Artificial Intelligence APA American Planning Association BAU Business As Usual BEBR Bureau of Economic and Business Research CA Cellular Automata CART Classification and Regression Tree CBD Central Business District C I T Commercial Institutional Transportation CLUE Conversion of Land Use and Its Effects CLUE S Conversion of Land Use and its Effects at Small R egional Extent CPU Central Processing Unit CUF California Urban Future CV Cross Validation DN Digital Number DOQQ Digital Orthophoto Quarter Quadrangle EML ERDAS Macro Language ETM+ Enhanced Thematic Mapper Plus GA Genetic Algorithm GIS Geographic Information System GRD Gross Residential Density GUD Gross Urban Density LBCS Land Based Classification Standards

PAGE 18

18 LUCIS Land Use Conflict Identification Strategy LULC Land Use and Land Cover MCM Markov Chain Model MLR Multinomial Logistic Regression MSS M ultispectral Scanner System MUA Multiple Utility A ssignment NDVI Normalized Difference Vegetation Index NLCD National Land Cover Database NN Neural Networks OLS Ordinary Least Squares PC Principal Component QUEST Quick, Unbiased, and Efficient Statistical Tree ROC Relative Operating Characteristic ROI Region of Interest SE Standard Error TIGER Topologically Integrated Geographic Encoding and Referencing TM Thematic Mapper UGB Urban Growth Boundary USGS U.S. Ge ological Survey V I S Vegetation, Impervious Surface, Soil

PAGE 19

19 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy USING REMO TE SENSING AND GIS T O MONITOR AND PREDIC T URBAN GROWTH CASE STUDY IN ALACHU A COUNTY, FLORIDA By Yong Hong Guo May 2012 Chair: Paul D. Zwick Major: D esign, Construction and Planning Alachua County experienced remarkable urban growth in the extent of the urban area during the past three decades, largely due to population growth. This dissertation simulates and predicts future urban growth in the county based on population growth. First, this dissertation classifies urban land uses and land covers (LULCs) based on eleven classes equivalent to the U.S. Geological Survey (USGS) Level III standard using remote sensing Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) data, as well as the state of the art classification a nd regression tree (CART) method. The CART method is then compared to the Vegetation Impervious Surface Soil (V I S) and conventional supervised classification methods based on a pilot area. Next, this dissertation classifies 1982, 1994, and 2003 LULC clas se s for Alachua County as a whole. Second, a multinomial logistic regression (MLR) model is employed to simulate urban LULC development in 2003 based on the urban LULC classified for 1982, 1994, and 2003, respectively. Accordingly, a number of probability maps are created for the four classified urban uses, which include single family, multi family, commercial institutional transportation, and industrial warehouses. Third, an urban LULC allocation process is undertaken for Alachua County, based on five scen arios for 2020 and 2030. The five scenarios include the business as usual (BAU) scenario, infill development, increased density

PAGE 20

20 development, redevelopment, and the conservation scenario. Thi s research has wide application in both the public and the private sectors of urban planning and additional areas in which the ability to describe the spatial extent of urban areas and their future expansion is of interest.

PAGE 21

21 CHAPTER 1 INTRODUCTION The State of Florida experienced an average annual population growth of 3.5 percent over the past four decades (Dewey and Denslow, 2001). Following this trend, Alachua County had remarkable urban growth in the extent of the urban area during the same period, expanding 130 percent between 1990 and 2000. 1 This dissertation inves tigates urban growth in Alachua County over the past three decades and simulates urban development for the future 20 years using remote sensing and Geographic Information System ( GIS ) technologies. Specifically, this dissertation illustrates how urban growth proceeded in Alachua County in 1982, 1994, 2003, and 2010, respectively, and projects how it might proceed in 2020 and 2030. The study first discusses three urban land use and land cover (LULC) classification methods in remote sensing, which include the classification and regression tree (CART) method, the Vegetation Impervious Surface Soil (V I S) method, and the supervised method. The goal of this research is to compare the state of the art CART method with the other two in order to determine the best method for the research. Thus, a pilot area is chosen in this regard. Based on the pilot area, this dissertation classifies eleven urban and natural LULC classe s in Alachua County by using the best method obtained. Next, LULC for Alachua County as a whole is classified. In addition, this research predicts urban growth in the county in 2020 and 2030, respectively, by using the multinomial logistic regression (MLR) model. As a result, fi ve development scenarios are proposed for future LULC simulations in the county: the business as usual (BAU) scenario, the infill scenario, the increased density scenario, the redevelopment scenario, and the conservation scenario. The study provides a fram ework for researchers to apply the best and/or newest methods introduced in this research to other counties or jurisdictions in the State of Florida as well as the rest of the nation. 1 Alachua County had approximately 62 square miles of urbanized areas in 1990, and approximately 81 square miles in 2000 (Data source: http://www.fgdl.org / ).

PAGE 22

22 Brief Review of Urban Growth in the U.S. Urban growth patterns in the U. S. have experienced a dramatic change over the past 150 years. The Industrial Revolution, which began in the early nineteenth century, had a huge impact on American cities, which not only grew rapidly in population, but also in areas. U.S. urban growth in the industrial era was concentrated in cities because a large number of people from rural areas, along with international immigrants, poured into cities to find jobs (Jackson, 1985). In Manhattan Island, New York City, for example, there were about 3.3 mil lion inhabitants in a 22 square mile area during 1900, an average density of 100,000 persons per square mile. In the lower East Side of New York, the density in some wards was several times higher (Levy, 2003). When enough people were concentrated inside c ities, a residential development market was created, and large numbers of private houses were built near major urban centers (Jackson, 1985). Decentralization of the central cities began simultaneously with this time of urban concentration: the Industrial Revolution produced faster transportation tools, such as trains, electric trolley cars, and subways, which provided great impetus for affluent central city residents to flee to the suburbs so as to avoid overcrowding and dirtiness in central cities (Jacks on, 1985; Mumford, 1961). In the U.S., a number of suburban towns were built between 1850 and 1920 such as Riverside, Illinois, and Forest Hill Gardens, New York. However, the number and size of railroad suburban towns remained small because people preferr ed a walking environment, and neighborhoods were planned based on walking distance (Mumford, 1961). If concentration and decentralization were major themes of urban growth in the nineteenth century America, decentralization generally characterized urban A merica in the twentieth century, and this process continues today (Levy, 2003). The technology that fostered this decentralized urban growth was the private automobile. With the mass production of

PAGE 23

23 automobiles beginning in earnest in the 1920s, a large numb er of urban dwellers spread from the central city out to the suburbs, which gave rise to the first great wave of suburbanization in the U.S. (Levy, 2003). American automobile ownership has increased steadily over the years. By 1930, there were about 25 mi llion automobiles; by 1950, the number had risen to 40 million; in 1960, the number stood at 62 million (Levy, 2003); and as of the end of 1999, the number reached 132.4 million (Bureau of Transportation Statistics, 2002). Today, automobile ownership has o utnumbered population increase at national, regional, and local levels; as a result, car ownership increased 383 percent between 1950 and 2000, compared to an 80 percent increase in population growth (The Boston Indicators Project, no date). The construct ion of inter state freeways after World War II further contributed to post war suburbanization (Jackson, 1985; Levy, 2003). Along with cheap gas and the mass production of automobiles, inter state freeways provided lower transport costs that greatly stimul ated decentralized urban growth (Jackson, 1985). In addition, the wide usage of telephones made long distance communication without face to face contact possible, and this further contributed to decentralized urban economic activities and growth (Levy, 200 3). Today, the World Wide Web, email, mobile phones, and fax all make long distance communication easier and more economical (Levy, 2003). National demographic trends echoed the above phenomenon. On the one hand, the metropolitan population was on the ris metropolitan areas increased from 55 percent to 80 percent (Lucy and Phillips, 2006). On the other hand, the urban population declined significantly in metropolitan areas, and suburban population increa sed steadily. Between 1950 and 1970, the suburban population grew

PAGE 24

24 approximately two fold, from 36 million to 74 million (Jackson, 1985). In 1970, suburbanites for the first time outnumbered urban dwellers in the U.S. (Jackson, 1985). During the 1980s, the metropolitan population as of 1990 (Benfield et al., 1999). Ba ltimore and St. Louis were two extreme examples: each lost 31 and 59 percent, respectively, of their central city populations between 1950 and 2000 (Lucy and Phillips, 2006). From the perspective of urban land consumption, statistics show that between 195 0 and 1990, suburbanization led to drastic growth in the size of metropolitan areas, from 208,000 square miles to 585,000 square miles (Squires, 2002). In Los Angeles, for example, the urban area grew an astonishing 300 percent between 1970 and 1990, while the urban population increased by only approximately 45 percent (Benfield et al., 1999). The trend of growth in urban land area surpassing urban population growth is unlikely to change in the near future. Alternatives to Decentralized Urban Development I n light of the recent trend for decentralized urban development, researchers are contemplating alternatives to decentralized urban growth because of its negative externalities. Although outward urban growth possesses both good and bad effects ( Cieslewicz, 2002; Benfield et al., 1999; Lucy and Phillips, 2006; Jargowsky, 2002 ; Squires, 2002 ; Kahn, 2006 ; Wassmer, 2001 ), a number of remedies, as well as new visions, have been outlined by several analysts (Lucy and Phillips, 2006; Squires, 2002). Their proposed remedies to decentralized urban growth are discussed below. decentralized urban sprawl in local jurisdictions. Smart growth principles seek to use existing resources efficiently, and also to allocate them more equitably. Based on this notion, Squires

PAGE 25

25 (2002) outlined eight remedies to offset negative externalities of the decentralized outward urban growth: Reusing existing land and infrastructure resources; Restrictin g development in outlying suburban and exurban areas; Relying less on the automobile by developing a number of transportation alternatives; Concentrating residential and commercial development centrally and along mass transit lines; Devoting more money to area wide revenue sharing and regional investment pools; Constructing more affordable housing and distributing it throughout metropolitan areas; Enforcing more vigorously fair housing laws; and Increasing public and private investments in central cities fo r more balanced development throughout the region. Downs (1994) proposed five components of visions to address the negative outcomes of decentralized outward urban growth, which include ownership of detached, single family homes; ownership of private autom otive vehicles; employment in scattered, low density workplaces, themselves in landscaped settings with free parking; living in small communities with strong local governments; and provision of housing to the urban poor through a trickle down 2 (p.10) ne ighborhood change process (Downs, 1994). In opposition to nearly universal single family detached housing, Downs (1994) suggested the construction of high density residential units in new growth areas, or mixing higher density housing with low density sing le family homes. He proposed a confinement mechanism that was similar to the urban growth boundary (UGB) concept. To reduce near universal private car ownership, Downs (1994) recommended the use of public transit, bicycles, walking, or ride sharing. Furthe r, to address scattered jobs among low 2 Federal government subsidies low income households directly. Will be narrated below.

PAGE 26

26 density areas, Downs (1994) proposed that new jobs should be located in large employment clusters. He thought job density ought to be intensified even though new jobs could be widely scattered. For local governments, functioned elected one or more adjacent counties, or using the state government to coordinate local planning actions. Downs (1994) a lso suggested changing the existing affordable housing standards and providing federal low Duany et al. (2000) proposed a traditional neighborhood concept a s a solution to outward urban growth. The traditional neighborhood concept consisted of six rules: a clear center within a neighborhood; no more than a five minute walk for ordinary daily life; the connection of one location to another via a continuous web and numerous paths; small streets; mixed use blocks; and providing unique sites for civic buildings (Duany et al., 2000). In fact, the above six rules are New Urbanism principles proposed to combat decentralized outward urban growth in the twenty first ce ntury. One example includes Alexandra, Virginia, a town built based on the six traditional neighborhood design principles mentioned above (Duany et al., 2000). Duany et al. (2000) proposed a number of strategies to implement their traditional neighborhood concept. They suggested that development take place within a comprehensive regional plan designed to limit the use of automobiles and preserve open space, rather than focusing on individual buildings. Based on this notion, they believed that new developmen t ought to be placed adjacent to infrastructure, e.g., a transit stop. Because existing and future rail lines are often addressed in regional plans, new neighborhoods and town centers could be planned based upon those transit lines. In addition, mixed uses were suggested based on their

PAGE 27

27 traditional neighborhood concept. Duany et al. (2000) recommended that every residential neighborhood should have a corner store to provide residents with daily needs. The proposed corner stores could limit automobile depende nce because of their location within walking distance. In addition to corner stores, neighborhood scaled shopping centers could be considered for a large population, and they can be next to through traffic. Mixed uses should also include office spaces, whi ch ought to be developed simultaneously with residential development. In addition, civic buildings, such as city halls, libraries, neighborhood elementary schools, etc., need to be constructed inside new communities, and land ought to be reserved for those buildings, which enhance community identity. Moreover, Duany et al. (2000) recommended road network connectivity in a residential neighborhood, in which roads ought to be fully connected in all directions around a residential unit in order to avoid using the same few collector roads (Duany et al., 2000). In general, the remedies to decentralized outward urban growth rest upon reshaping neighborhood planning and design conc epts, increasing housing density, promoting mixed use development, and developing new transportation alternatives. In addition, despite the fact that decentralized outward urban growth is a national trend, its degrees vary locally. Remedies should be under taken based on a particular political boundary, with an accurate assessment of the degree of decentralized outward urban growth in a jurisdiction. As a result, monitoring and/or detecting decentralized outward urban growth at a local level becomes a very i mportant task, and the question of how to monitor and predict it, along with measuring the degree in a jurisdiction, is a serious research question. Currently, there are no universal methods that define and simulate decentralized outward urban growth at th e national as well as local level (Broos and Day, no

PAGE 28

28 date). This dissertation uses remote sensing technologies, combined with the GIS tool, to research decentralized urban growth. Research Questions The research questions of this dissertation aim to model urban growth in Alachua County in the past, at present, and in the future and to develop and test a methodology for predicting decentralized urban growth using satellite imagery and property parcel data Specifically, this dissertation looks into three LULC classification methods using remote sensing technologies the CART method, the V I S method, and the supervised and compares them in a pilot area. The dissertation explores a new method the CART method for urban LULC classifications, and investigates these three methods in detail. The newly obtained method, i.e., the CART method, is later used to create a LULC classification for Alachua County. This study uses the Landsat data Thematic Mapper (TM) and t he Enhanced Thematic Mapper Plus (ETM+) imageries for analysis because satellite data are less costly, easier to obtain than aerial photographic data, and are capable of supporting sub parcel LULC analysis. LULC data are helpful for the analysis of physica l urban form as compared to vector parcel data. Also, this study uses GIS technology as a supplemental tool for the satellite data in order to generate both raster and ancillary data. This dissertation compares Alachua County land use of 1982, 1994, and 2 003, respectively, using the most optimal LULC classification method identified herein. Alachua County has experienced remarkable urban growth over the past decades. Based on the identification of the current urban growth in 2010, f uture land use patterns are predicted for 2020 and 2030 respectively This research contributes to the academic field by identifying a new method for LULC classification and predicting future land use based on the new LULC classification s Few studies

PAGE 29

29 have been conducted on urb an growth that use Landsat images to accurately classify LULC equivalent to U.S. Geological Survey ( USGS ) Level III. This study identifies a methodology of utilizing 30 30 meters Landsat TM and ETM+ imageries for LULC classification s Research Applicati ons This research has wide applications in urban planning and supplementary areas including land use planning; urban design; transportation planning; parks, tourism, and recreation facilities planning; population forecasting; urban economics; water and soi l conservation; crime prevention; environmental and ecosystem protection; natural resources conservation; and real es tate and industrial development It will benefit both the public and the private sectors. Specifically, since the major intent of this rese arch is to provide accurate predictions of future land use, it supports evaluation of local and state land use policy for curbing urban sprawl by allowing for comparisons of land use change over time. The intent of this research is to make future land use patterns tangible. The research also provides additional intuitive technical applications, such as population forecasts, traffic counts, and a methodology for site selection for real estate development. In sum, the applications of the research are monitor ing urban growth, heuristic state and local land use policy making, and intuitive technical applications.

PAGE 30

30 CHAPTER 2 URBAN LAND USE AND L AND COVER CLASSIFICA TION Chapter Overview The difficulty of classifying urban LULC has been noticed by researchers for many years. Because of the spectra l characteristics of urban LULC classes it is often difficult to classify urban uses and achieve high accuracy levels using the conventional approaches such as the supervised and unsup ervised classification methods. Alternative approaches such as the CART method is being discussed (Paul, 2007). This chapter explores a way to apply the CART method in urban LULC classifications to Alachua County by comparing the CART method with two other classification methods. This chapter explores the CART method, the V I S method, and the supervised classification method in a pilot area in Alachua County, and uses the CART method to classify urban LULC classe s for the whole county. This new CART method that is applied makes use of the advantages of the method and yields the highest accuracy level. The overall accuracy of using the CART method reaches more than 90 percent for 1982, 1994, and 2003, respectively, in the whole county. Literature Review Sup ervised and Unsupervised Classification Methods Currently, there are several ways that have been adopted by researchers to classify urban LULC classe s using satellite imageries applying the conventional supervised and unsupervised classification methods. Y ang (2000) used an unsupervised classification method to classify urban LULC classe s for the region of Atlanta, Georgia, based on the Multispectral Scanner System (MSS) and TM images. In his research, Yang (2000) classified urban LULC classe s into six type s, which were high density urban, low density urban, cultivated/exposed land, cropland/grassland, forest, and water. For his research, Yang (2000) utilized the ISODATA

PAGE 31

31 iteration to classify urban LULC classe s into two types, namely, high density urban and low density urban, and sorted out the mixed pixels by using the modal filter so as to identify the pixels with the highest frequency in the histogram of a Digital Number (DN) value, i.e., the pixel value, as the representatives of the most commonly occurri ng class in a patch. To address the issue of urban LULC classe s that have similar spectral characteristics, Yang (2000) used an image interpretation approach to identify four major pairs of urban LULC classe s that have similar spectral signatures, which were low density urban and forest; low density urban and cropland, forests; and high density urban and exposed land. After this identification was completed, Yang (2000) started a reclassification process, in which the leftover erroneous land covers were m anually recoded into the correct land covers. The overall accuracy level was elevated from the initial, about 60 percent, to 87 percent. Finally, Yang (2000) used aerial photography as training data for accuracy assessment. However, Yang (2000) classified urban landscapes based on only two categories, which were the high density urban and low density urban. He did not separate urban LULC classe s further into several detailed categories such as residential, commercial, and industrial. Similar to Yang (2000) Lo and Choi (2004) employed a hybrid supervised method combined with unsupervised approach to mapping urban LULC classe s by using Landsat ETM+ imagery. Their study area was also Atlanta, Georgia. Similar to Yang (2000), they initially utilized the ISODAT A to cluster all likely homogeneous pixels. They subsequently used a fuzzy supervised classification method to address the mixed pixels based on the created clusters. Their LULC classe s were also high density urban use, low density urban use, cultivated/ex posed land, cropland or grassland, forest, and water.

PAGE 32

32 Using the unsupervised classification method, Lo and Choi (2004) generated 60 classes as an optimal number for ISODATA experienced from their previous Atlanta related projects and then labeled the homo genous pixels and left out those unlabeled mixed pixels to supervised fuzzy classification. After this was completed they created a fuzzy subset for those heterogeneous pixels based on a membership function, in which membership grade was assigned to each class. The membership grade was between 0 and 1. This was a fuzzy signature assigning process. To determine the initial membership grade for each LULC class, they utilized DOQQ images to collect training samples, for which three to four training sites were selected. For example, for a typical site the fuzzy membership grade that was assigned for high density urban land use was 0.92 and 0.08 for low density urban land use while 0 was used for other LULC classe s. Then, based on those fuzzy membership grades, Lo and Choi (2004) calculated fuzzy mean and fuzzy covariance matrix parameters for each class, which was significant for supervised fuzzy classification. When the supervised fuzzy classification was finished Lo and Choi (2004) used an overlay to create a union of the pervious ISODATA classified LULC map with the supervised fuzzy classified LULC map in order to formulate a complete LULC map for the study area. The overall accuracy for the hybrid classification method was 91.5 percent (Lo and Choi, 2004). From Lo and Choi (2004), the hybrid classification requires a fine assignment of initial membership grade matrix for each urban LULC class. Although they obtained a high accuracy level for urban classes in their research, like Yang (2000), they did not sep arate each urban LULC further to residential, commercial, and industrial. Whether a high accuracy level can be obtained by applying their method to more detailed urban LULC classe s is unknown.

PAGE 33

33 Jensen and Toll (1982) used the Landsat MSS imagery to detect residential land use development at the urban fringe in Denver, Colorado. For this purpose, Jensen and Toll (1982) proposed a change detection model to identify stages of the residential development with respect to the residential development cycle from p arcel clearing to complete landscaping. Jensen and Toll (1982) first used panchromatic aerial photography to manually identify 10 stages of single family development, which namely were (1) original land cover; (2) area cleared; (3) area cleared, subdivided and paved roads; (4) area cleared, subdivided, paved roads, and building; (5) subdivided, paved roads, buildings, partially landscaped; (6) cleared, subdivided, dirt roads; (7) cleared, subdivided, dirt roads, and buildings; (8) subdivided, dirt roads, b uildings, and partially landscaped; (9) subdivided, dirt roads, buildings, and landscaped; and (10) subdivided, paved roads, buildings, and landscaped (p.630). Then, they utilized a gray tone spatial dependency matrix to extract textural information from t he MSS Band 5 imagery because they thought Band 5 (0.6 0.7 m) enhanced the contrast between vegetated and non vegetated surfaces (Jensen and Toll, 1982). Next, they chose the image differencing method to do the change analysis based on the previously ex tracted texture images because they found that the image differencing method produced lower change detection errors (Jensen and Toll, 1982). After overlapping the Band 5 change image with the panchromatic aerial photography, they found that Band 5 was good at detecting land use changes from original, natural vegetation to partially or fully landscaped residential development while unable to detect the changes between partially developed land and un irrigated rangeland. Therefore, they made a texture change map and overlaid the newly created texture change map onto the previously generated spectral change map and formed a single change map. The absolute accuracy of the new change map hence reached 81 percent.

PAGE 34

34 nd that texture difference provides useful information for the identification of different residential development stages at urban fringes because each stage of development presents different texture characteristics. However, textural information must be u sed in complement with the spectral information such as Band 5. In addition, the identification of the residential development cycle is only one of the aspects in urban LULC identification. Whether Landsat imageries can be applied to identify and classify additional urban LULC classe s and simultaneously yield high accuracy has not been reported by researchers so far. V I S Model Ridd (1995) proposed a V I S model in 1995 for the first time to classify LULC classe s for Salt Lake City. Borrowing ideas from s oil texture compositions, he put vegetation, impervious surface, and soil on vertexes of a triangle (Figure 2 1). Urban land features are determined based on the component percentages to be represented as a percentage on the triangle sides. For example, hi gh density residential use is located on the impervious surface side, which is close to the central business district (CBD), occupying about 60 percent of the impervious surfaces while low density residential occupies about 30 percent of the impervious sur faces that are close to vegetation (Figure 2 1). For practice, Ridd (1995) applied the V I S model to a pilot area in Salt Lake City with data collected from two linear transects radiating from the city center. The linear transect Ridd (1995) applied was a line stretching from the city center to the suburbs, on which equal frames were selected that forms a transect. Then, sample points were collected in each sample frame of a transect. The sample data were tabulated and marked as seven zones, in which hom ogeneous data were found. These zones included CBD, near town old residential, stable mature residential, country club, postwar mid income resi dential, Mt. Olympic Cove, and n atural woodland. Each

PAGE 35

35 zone has its specific V I S compositions, and the V I S com positions for an urban environment are identified from interpreting the diagrams of the compositions of each zone. Ward et al. (2000) applied the V I S model in Queensland, Australia, which was an elements of their research were similar to vegetation index (NDVI), Band 5, and Band 3. Ward et al. (2000) also applied mineral indexes to distinguish rural and agr icultural land uses from urban land uses. For urban impervious surfaces, Ward et al. (2000) used Band 3 to extract the texture image. Phinn et al. (2002) explored the V I S model in the City of Brisbane, Australia, on a sub pixel basis, mimicking the approach of Ward et al. (2000) They first conducted per pixel image classification based on the NDVI, Band 5, and Band 3 to produce 20 classes, which separated vegetated and non vegetated classes. Then, similar to Ward et al. (2000), they classified the unclassifi ed non vegetated surfaces into the soil related classes using mineral indexes. After this is completed, they employed a texture layer to extract urban areas. The five classes output from their V I S process were forest/woodland vegetation; grass/sparse veg etation; cleared areas soil; developed areas impervious; and water bodies. Similar to Ridd (1995), Phinn et al. (2002 ) also extracted the V I S components from aerial tropolitan area and found 1 by 1 km sample frames. Within each sample frame, thirteen random sets were generated, which determined whether a set belong to a certain component: vegetation, impervious surfaces, and soil. In this case, the V I S components of a certain point can be calculated so that various component percentages can be calculated based on the V I S diagram

PAGE 36

36 for each location. Therefore, an area can be identified for their specific use based on the extracted V I S components. Hung (2003) expl ored the V I S model in order to analyze urban LULC classe s on a sub pixel basis for the Salt Lake Valley area. According to Hung (2003), most V I S researchers still used a per pixel based analytic method to explore urban LULC change, in which one pixel r epresented one LULC type and other types were ignored (Hung, 2003). Therefore, Hung (2003) proposed a sub pixel method by using the V I S model, which he believed could better deal with mixed and confused pixels than a conventional pixel by pixel method. Hung (2003) used three different types of satellite images for his research, which were MSS, TM, and ETM+, covering the timeframe from 1972 through 1999 with each sensor covering a particular timeframe. He used a supervised sub pixel classifier to classify the ground LULC classe s. The resulting images included multi channels with each channel indicating a certain proportion of a typical ground cover. According to Hung (2003), the channel with the proportion of a ground cover contained more ground cover information than a single p ixel I S model is presented in Figure 2 2, which is similar to percent impervious surfaces so that urban land use types su ch as residential, commercial, industrial, and recreational can be determined based on the component percentages of each type. Hung (2003) also categorized urban LULC classe s into six types based on the V I S diagram above, which were Vgg (green grass veg etation), Vts (tree/shrub vegetation), Ibr (bright impervious surface), Imd (medium impervious surface), Idk (dark impervious surface), and Sdv (soil/dry vegetation). After conducting proper geometric rectification and co registration processes, Hung (2003 ) defined training sites from remotely sensed images and used them as

PAGE 37

37 end members with each pixel softly classified into six membership grades, which were later translated into component percentages for ground covers and checked by a linear spectral mixtur e model (Hung, 2003). Then, for accuracy assessment purposes, the resulting component percentages of ground covers were compared to another set of component percentages of ground covers derived from aerial photography visual interpretation (Hung, 2003). Hu covers in which the likelihood between the candidate pixel and the six predefined ground covers, such as Vgg, Vts, and so on, was calculated. Each pixel had six likelihood v alues and these likelihoods were then converted to percentages (which equaled to 100 percent) based on a conversion algorithm. Then, a linear spectral mixture model was applied. The linear spectral mixture model is the algorithm that spectral reflectance o f a typical pixel was the sum of all the spectral reflectance of its ground components weighted by the proportion of each component on the ground (Hung, 2003). Hung (2003) chose Band 4 as the checking band because he thought Band 4 was sensitive to vegeta tion. An expert system rule was also applied, which was designed to determine the dominant component in a candidate pixel as well as to determine which component(s) to adjust. An accuracy assessment was processed later in his study. Because MSS, TM, and E TM+ each covered a specific timeframe, accuracy assessments were processed based on different satellite images, which specifically were the 1979 MSS image, the 1987 and 1990 TM images, and the 1999 ETM+ image. Hung (2003) selected two sample areas, namely, Salt Lake City and West Jordan, for accuracy assessments. The component percentages of ground covers derived from the above images were compared to another set of the component percentages of ground cover derived from the visual interpretation from aerial photography (Hung, 2003). The

PAGE 38

38 comparison was processed based on regression analysis, in which correlation coefficients were used to measure accuracy level. In terms of the Salt Lake City sample area, the overall correlation coefficient for MSS 1979 was 0. 883 with some specific correlation coefficient values for each component; the overall correlation coefficient for TM 1990 was 0.75; and the overall correlation coefficient for ETM+ 1999 was 0.755. Similar results were obtained for the West Jordan area. Hun g (2003) concluded that overall the estimated percentages showed a significant relationship with surveyed percentages but had an underestimate of percentages of soil and an overestimate of impervious surfaces for foothills, mountains, and farmlands due to the seasonable agricultural landscape change and soil types (Hung, 2003). Based on the ground component percentages characteristics according to the V I S model, Hung (2003) categorized urban land uses into several general urban features, namely, downtown, light industry, heavy industry, city park, golf course, low density residential, medium density residential, high density residential, the University of Utah campus, foothill, ranch, crop field with vegetation, crop field with mixed vegetation and soil, a nd crop field with soil. For example, downtown had 7 percent vegetation, 3 percent soil, and 90 percent impervious surface. Low density residential had 68 percent vegetation, 6 percent soil, and 26 percent impervious surface, and so on for other urban feat ures. is evident that the V I S model is capable of identifying various urban LULC types in detail, such as residential in various densities, light and heavy industri al, recreation, agricultural, and the CBD. The V I S model can successfully deal with urban LULC issues such as mixed pixels Ridd (1995), Ward et al. (2000), Phinn e t al. (2002 ), and Hung (2003) did not classify the

PAGE 39

39 commercial, institutional, and transportation uses further from their urban uses; for example, they just classified the commercial uses based on the CBD or downtown. The classifications of commercial, inst itutional, and transportation uses in other parts of the study area are missing. Their simple classifications of the CBD are not satisfactory to address the complicated nature of urban LULC classe s because the commercial use, the institutional use, and the transportation use are important urban LULC classe s, which may have similar spectral characteristics to industrial and be intermingled with residential or other uses. Furthermore, wetlands were not addressed by the V I S researchers either. This is due to the fact that the V I S model does not take water into consideration. To make up the inadequacy of the V I S model, Lu and Weng (2004) proposed a model I S model. The Lu Weng Model divides urban landscapes into three components, which are shade, green vegetation, and soil/impervious surface rather than vegetation, impervious surface, and soil. Lu and Weng (2004) argued that the V I S model was not adequate to address water and wetlan ds because it did not have a water component. The Lu Weng model had a shade element, however, which could be used to classify water and wetland. Lu and Weng (2004) utilized their model in Marion County, Indiana. Their final product included the classificat ion of the urban LULC classe s into six categories, which were forest, grass, pasture and agricultural lands, residential, urban, and water. Specifically, they combined commercial and industrial into urban as well as high intensity residential and low inten sity residential into residential. The overall accuracy reached 80 percent. LULC classe s, which covers not only the vegetation covers such as forests, grasses, and pastures

PAGE 40

40 and agricultures, but also non vegetation covers such as water, as well as some urban uses and features such as residential and urban. However, their research is still un able to sort out commercial and industrial from urban uses because commercial and industrial are in similar impervious percentages in their model. In addition, their overall accuracy is still low because the urban landscape components are different in diff erent geographic locations and hence one fixed model cannot fit all the landscape scenarios. For example, in an area that is predominantly covered by vegetation, it is inappropriate to use the shape, green vegetation, and soil/impervious surface model beca use of less impervious surfaces in the area, while in an urban area that is predominantly covered by impervious surfaces, this model is the best to represent each component on the ground, and usually the study area incorporates these two areas together at the same time. In this case, researchers need to adjust their model for each geographic landscape scenario and propose different models for each sub area. This creates technical complexity for the research. CART Method Although efforts have been mad e by using the conventional classification methods as well as the V I S method for urban LULC classe s classifications, research finds that the conventional supervised and unsupervised classification methods using spectral data alone have a 5 to 10 percent lower overall accuracy than methods using ancillary data (Rogan et al., 2003). Therefore, alternative approaches such as the rule based classification methods have received increasing attention as of late (Zambon et al., 2006; Lawrence and Wright, 2001). A knowledge based decision tree model named CART has emerged to classify urban LULC classe s, which showed promise for improving classification accuracy in a number of studies (Zambon et al., 2006). Lewis (2000) compared CART with the multivariate logistic regression model in his clinical research and found that CART had several advantages over the multivariate regression

PAGE 41

41 model. First, CART is non parametric, which can handle highly skewed or multi modal data; compared to multivariate regression models, CART does not require transformation if data are not normally distributed. This is concurred by Huang and Jensen (1997) in remote sensing, in which their research showed clearly the maximum likelihood classifier would be significantly influenced by the distrib ution of data and could not effectively handle a bi or multi modal distribution; because of that, spectral data alone were not able to discriminate among mixed classes. Second, CART can deal with missing data better than multivariate regression models. Fo CART is different from other expert systems because it has an automate knowledge base building program without extensive a priori expert knowledge (Lawrence and Wright, 2001, p.1,141), which Jensen, 1997, p.1,185). As a result, CART requires relatively little input while the multivariate regression model requires extensive input from the analyst, analysis of interim results, and later modification of the method (Lewis, 2000, p.6). CART allows ancillary data to go with the classification, which has advantages over other parametric methods for the elevation of accuracy levels (Lawrence and Wright, 2001). It can incorporate either continuous/categorical or raster/vector an cillary data. Usually GIS data, as well as texture information, can be included into ancillary data (Lawrence and Wright, 2001). Other data such as elevation, slope, soils, road networks, and so on can also be included (Jensen, 2005).

PAGE 42

42 Herold et al. (2003) applied the CART to map impervious surfaces and forest canopies on a sub pixel basis for a study area in the eastern portion of the Chesapeake Bay watershed in Maryland, where the study area was covered by clouds. Like Lewis (2000), Herold et al. (2003) u sed a recursive binary partition process for the training data, which were sampled from the high resolution imagery sources. These training data were representatives of the entire dataset and used in the production rule sets, which enabled the software to et al., 2003, p.1). Because the CART method was able to provide low cost, high quality knowledge bases, the CART software was able to approximate the human learning process The research of Herold et al. (2003) utilized 30 meter ETM+ data from three different seasons between 1999 and 2001. After ETM+ images were geometrically and radiometrically corrected, Herold et al. (2003) selected training samples from the DOQQ pictures of 1 meter reso lution and classified them based on spatial and spectral characteristics of the ETM+ imagery. The sampling process was accomplished by applying a stratified random sampling method e the possibility of accuracy of the dataset as a whole (Herold, et al., 2003). After the high resolution classification was completed, they used CART software to determine a classification rule based on the percent surface values that were identified by the sample sites as well as the relationship between various data layers. The CART software was developed with the C language and the ERDAS Macro Language (EML). During the process, validity data were also made in a random selection manner. As a result, ea ch of the target bands had 50,000 training pixels and 50,000 validity pixels (Herold et al., 2003). Then, the training data and validity data were input into the CART software, in which regression trees and a

PAGE 43

43 production rule set were formulated. The output predicted values were evaluated based on the training data and validity data. The final outputs were a series of maps showing the impervious surfaces and tree canopies free of clouds. The accuracy level for impervious surfaces was 0.90 while the accuracy level for the tree canopy was 0.93. Although the research of Herold et al. (2003) achieved high accuracy for mapping impervious surfaces and forest canopies, their research is urban impervious surfaces based only. Their research did not include classifications of several detailed urban LULC categories such as residential, commercial, and industrial. From the above literature review, systematic classifications of urban LULC classe s into USGS Level II and the USGS Level III have not been tested so far. Although L u and Weng (2004) utilized an altered V I overall accuracy is not very h ig h. In addition, the research of Herold et al. (2003) did not have a comprehensive classification of urban LULC classe s, altho ugh they utilized the CART method to classify urban impervious areas. This chapter explores a method to classify urban LULC classe s into the USGS Level II and the USGS Level III and simultaneously yields high accuracy. The CART method is explored, and anci llary data are included in the method so as to yield high accuracy. Spectral Characteristics of Urban LULC The difficulty of classifying urban LULC classe s with high accuracy based on satellite imageries relies on three factors. First, an urban environmen t consists of a combination of different land covers that may have similar spectral characteristics (Paul, 2007). For example, a suburban residential development may consist of open space, which may be in the form of golf courses, parks, etc., that have si milar spectral signatures to the natural, non urban grasslands or forest covers. Also, some exposed bare land and agricultural land may present similar spectral values to urban impervious areas (Kim, 2007). Second, urban LULC classe s are rarely

PAGE 44

44 homogenous (Myint et al., 2007) and can be mixed with each other in a satellite image pixel. For example, on a Landsat TM image (30 30 meters), an urban road with a vegetation covered median may have vegetation covers mixed with impervious covers such as concrete o r asphalt in the same pixel. For single family residential units, it may have residential rooftops mixed with hard surfaces, swimming pools, exposed soil, or vegetation (Myint et al., 2007). In addition, urban features consist of various types of materials such as glass, concrete, metals, plastics, asphalt, grass, shingles, shrubs, trees, soil, water, etc., which are very complex and have completely different spectral characteristics (Myint, 2001). Third, urban growth in the suburbs is often characterized a s low density single family development, in which single family housing units may be set apart from each other at certain distances and where their texture information is beyond the detection of satellite remote sensors. Based on the research of Cowen et a l. (1995), for clear identification of an object of interest, the minimum resolution of high quality imagery is one half of the diameter of the smallest object (Cowen et al., 1995; Jensen and Cowen, 1999), which means for a single family house with 10 mete rs width, the minimum spatial resolution to identify that single family house is 5 5 meters. This creates a serious obstacle since satellite imageries are not able to provide some important spectral bands for LULC classification based on this resolution. When high resolution imagery is used, e.g., QuickBird and IKONOS, mid infrared and thermal bands are absent because these high resolution sensors are not equipped with the above specified spectral bands that are critical for urban LULC identifications. Ac cording to Jensen and Cowen (1999), certain electromagnetic spectra such as visible color (0.4 0.7 m), near infrared (0.7 1.1 m), and middle infrared (1.5 2.5 m) are useful to extract USGS Level III land cover in an urban environment, and the ther mal infrared portion of the spectrum (3 12 m) is useful to detect urban temperature (Jensen and Cowen, 1999).

PAGE 45

45 According to Toll (1984), Landsat 4 and 5 TM Band 5 (1.55 1.75 m), Band 6 (10.4 12.5 m), and Band 7 (2.08 2.35 m) can improve the over all discrimination level of urban LULC classes over Landsat MSS because of added spectral bands (Toll, 1984). Research Question The leading research question in this chapter is how to develop a method that uses the Landsat TM and ETM+ imageries to classify urban LULC classe s into USGS Level II and USGS Level III, and simultaneously yields high accuracy. Based on the current research, it is difficult to classify urban LULC classe s into USGS Level II and USGS Level III by reaching an accuracy level of more th an 85 percent when using Landsat sensors because of the limitations of the conventional supervised and unsupervised classification methods. How to explore a way that better addresses the spectral characteristics of urban LULC classe s, while simultaneously delivering high accuracy results, is a serious research question. This chapter compares and explores the CART method, the V I S method, and the conventional supervised methods in order to classify urban LULC classe s to the level similar to USGS Level II an d USGS Level III. Methodologies Classification System To classify urban LULC classe s, a workable classification system should be adopted. Currently, there are two classification systems available to remote sensing fields: one is the USGS Classification Sy Based Classification Standards (LBCS) (Jensen and Cowen, 1999). For urban LULC classes, the USGS Classification System does not specify single family and multi family residential classes at L evel II, which are the two most important land use categories to measure urban growth in this research, and depends upon the user to define them at Level III. In addition, the USGS Classification System does not define golf courses, zoos, and urban parks a t Level II. However,

PAGE 46

46 commercial and industrial uses are both defined at Level II. These make the classification of urban LULC classe s difficult without having a standard level to incorporate important urban provide land cover and vegetation information in an urban area (Jensen and Cowen, 1999). Therefore, this research adopts a revised USGS Classification System, referenced from the National Land Cover Database (NLCD) classification system. The detailed clas ses are as follows (LCI, 2007, p.1; Anderson et al., 1976, p.10 21): S INGLE FAMILY This i ncludes areas with a mixture of constructed materials and vegetation. Constructed materials account for 30 80 percent of the cover. Vegetation may account for 20 to 70 percent of the cover. These areas commonly include single family attached and detached housing units and mobile homes. Population densities will be lower than in the multi family residential areas. M ULTI F AMILY This includes highly developed areas where people reside in high numbers. Examples include apartment complexes, condominiums, and row houses. Vegetation accounts for less than 20 percent of the cover. Constructed materials account for 80 to100 percent of the cover. C OMMERCI AL I NSTITUTIONAL T R ANSPORTATION This i ncludes areas dominated by retails, sales, services, and offices; various educational, religious, health, correctional, and military facilities; and utility and transportation infrastructures, e.g. power plants, roads, railroads, parkin g lots, etc. I NDUSTRIAL W AREHOUSES This i ncludes a wide range of land uses from light manufacturing to heavy manufacturing plants, which also include mineral mines and warehouses. G RASSLANDS These a re areas dominated by upland grasses, forbs, and shrub s. In rare cases, herbaceous cover is less than 25 percent but exceeds the combined cover of the woody species present. These areas are not subject to intensive management, but they are often utilized for grazing. F ORESTS These a re areas characterized by tree cover (natural or semi natural woody vegetation, generally greater than 6 meters tall); tree canopy accounts for 25 100 percent of the cover. A GRICULTURAL This i s land used primarily for production of food and fiber. R ECREATIONAL AND O THERS This i n cludes golf courses, parks, zoos, conservations and reserves, and cemeteries.

PAGE 47

47 W ETLANDS These a re areas where the soil or substrate is periodically saturated with or covered with water. W ATER These are all areas of open water, generally with less than 25 percent cover of vegetation/land cover. B ARREN These are areas basically characterized as landfills or characterized by bare rock, gravel, sand, silt, clay, or other earthen material, with little or no "green" vegetation present, regardless of its inherent ability to support life. Vegetation, if present, is mor e widely spaced and scrubby than that in the "green" vegetated categories; lichen cover may be extensive. The above classification system crosses USGS Classification Level I, Level II, and Level III. According to Jensen and Cowen (1999), the resolution for a remote sensor for USGS Level I is spatial 20 100 meters and spectral V NIR MIR Radar; for Level II it is spatial 5 20 meters and spectral V NIR MIR Radar; and for Level III it is spatial 1 5 meters and Pan V NIR MIR. The Landsat ETM+ imagery is no t able to provide the spatial resolution at Level III as suggested by Jensen and Cowen (1999). However, according to Toll (1984), as spectral resolution is more meaningful than spatial resolution in detecting urban LULC classe s, the 30 30 meters Landsat spatial resolution, for example, can still be used for the urban growth analysis. Data Inventory and Pilot Area The satellite data source involved in this section for a pilot area is the 2003 Landsat ETM+ with imagery data acquired on February 11, 2003, d etermined based on cloud conditions with 0 percent cloud coverage. The training data are acquired from the 2004 digital ortho quarter quads (DOQQs), which is the most recent available for download. A pilot area in Alachua County is chosen to test the CART method, the V I S me thod, and the supervised method The pilot area, coded as Q4619SW in DOQQ images, is located in central Gainesville, covering an area of 20,700 acres (Figure 2 3). A specific method that accurately classifies urban LULC classe s and yiel ds the highest level is selected to classify the urban LULC classe s for the entire county. The selection of the pilot area is based on the factors that urban LULC classe s and natural LULC

PAGE 48

48 classe s are intermingled in an area where the sensitivity of a class ification method can be tested as to whether it is able to yield high accuracy for all eleven LULC classe s. This quadrant was selected because the area is a typical urban area where it covers a variety of important urban and natural landscapes such as the University of Florida campus, adjacent commercial and single family and multi family uses, forests, wetlands, industrial uses, shopping centers, agricultural and nurseries, golf courses, parks, ranchland and grassland, and water. As a result, the pilot are a comprises all the urban uses and the most natural LULC classe s. This is different from the pilot area that is either occupied by all urban LULC classe s or by all natural LULC classe s. As mentioned above, successful classifications of urban and natural LU LC classe s in this quadrant help extend the method to the entire county. Based on the above quadrant, the 2003 ETM+ imagery is masked so that the extent of the IDRI SI Andes, ENVI 4.4, and ArcGIS 9.3. CART Method This chapter tests the CART method based on these two software packages, IDRISI Andes and ENVI 4.4. Huang and Jensen (1997) described the knowledge based building procedure in remote sensing into three steps : training, decision tree generation, and from decision trees to production rules (p.1,186 p.1,187). First, for training, it is important to have the learning process begun based on the induction from the training data. Second, a decision tree is generat ed based on the training data. In this step, a recursive procedure is processed so that no more pixels can be divided in the end. An optimal size of the tree is also determined in this step. Several splitting rules such as the Gini, the Entropy, and the To wing can be applied in order to generate a tree with the least errors in each node (Huang and Jensen, 1997; Breiman et al., 1984; Zambon et al., 2006). Third, a decision tree is transformed to production rules. For example, in a child node,

PAGE 49

49 i.e., a leaf, t method in this research. First, various relevant layers such as the ETM+ layer and ancillary layers are stacked; second, sample regions of interest (ROIs) are collected, which can be collected from the DOQQ training datasets; third, a splitting rule is selected as to whether it is Gini, Entropy, Ratio Gain, or something else; fourth, trees are built based on the sample ROIs as well as the selected splitting rule; fifth, tree is pruned; sixth, classification maps are created; and seventh, accuracies are assessed. Bec ause of the different splitting rules offered by the two different software packages, in which the ENVI 4.4 RuleGen 1.02 offers the splitting rule called Quick, Unbiased, and Efficient Statistical Tree (QUEST ) and IDRISI Andes applies three splitting rules, namely gain ratio, entropy, and Gini. Splitting rules in the different software packages produce different classification maps with different results. This study finds a splitting rule that best represents the ground truth by yielding the highest accuracy level. The ENVI 4.4 RuleGen 1.02 has the QUEST module which allows numeric and categorical variables to be stacked together. This is different from IDRSI Andes, in which categorical layers are treated as n umeric layers. The QUEST module also provides options to input standard errors (SEs) and cross validations (CVs) for tree pruning. For the splitting rules, the QUEST module provides the choices between univariate and linear. Other options are also provided such as the input of Alpha value, minimum node size, variable selection method, output Pstricks tree, and so on. For IDRISI Andes, it p rovides three splitting rules. The gain of a single classification is defined as the entropy af ter classification X an d Gain ( X ) tests the maximization of the

PAGE 50

50 information gain. The entropy algorithm is given to oversplitting since every split can potentially contribute to information gain. The gain ratio algorithm attempts to overcome this potential bias through a normali zation process. The entropy rule aims to identify splits where as many groups as possible are divided as precisely as possible and forms groups by minimizing the within group diversity. The Gini splitting rule is a measure of impurity at a given internode that is at the maximum when all pixels are equally distributed among all classes (IDRISI Andes Help, no date) ENVI 4.4 RuleGen 1.02 For the CART method, based on the above descriptions, a n NDVI layer is first developed from the ETM+ image by using the ERDAS Imagine 9.1 software. A Principal Components (PCs) layer and a Tasseled Cap layer are also developed by using the ERDAS Imagine 9.1 software. In addition, an unsupervised classification layer, namely, the ISODATA layer, is created in ERDAS Imagine 9. (2004) research, in which Yang (2000) and Lo and Choi (2004) found that an iteration of 60 80 is optimal for a successful unsupervised classification. These layers are stacked together with three additional layers, which are mentioned below for decision tree building in ENVI 4.4 RuleGen 1.02. The reason for stacking various layers is to find various variables for decision tree building. It is optimal to find as many variables and/or layers as possible that are related to the research purposes. The above variables and/or layers are used to represent vegetation as well as other variables by simplifying band combinations. For this end, three additional variables and/or layers are found, which a re the parcel layer, the parcel assessed value layer, and the parcel residential unit layer. The vector parcel data are sorted based on the parcel land use code and categorized according to the previously defined eleven LULC classes and then rasterized in

PAGE 51

51 ArcGIS 9.3. After this is complete d the land use codes in the parcel data are reclassified into the eleven classes. The reason to include the parcel data into the ancillary data is to reduce the occurrence of pixel confusion given the mixed or spectral similarity for urban LULC classe s. This is a very important step since without parcel signatures urban pixels can be easily mixed. For example, the University of Florida campus is a single parcel containing various LULC classe s. Without parcel signatures, the LULC classe s on the campus such as nurseries, student apartments, recreational facilities, and so on cannot be discriminated accurately based on the 30 meter ETM+ imageries. Therefore, additional parcel signatures must be sought in addition to the exis ting parcels on the campus. In order to differentiate the spectral similarity of the industrial and commercial uses, a parcel assessed value layer is introduced in addition to the parcel layer. Because industrial and commercial uses share different assesse d values, in which industrial uses are assessed much lower in value than commercial uses, the addition of the assessed values into the ancillary data helps identify correctly the industrial and commercial uses. To further differentiate the single family an d multi family uses, a parcel number of unit layer is included because single family and multi family uses have different numbers of units in a parcel. The inclusion of this layer helps clarify the single family use from the multi family use because of the ir spectral similarities. After this is accomplished, the two layers are stacked in ENVI 4.4 RuleGen 1.02 along with the above mentioned NDVI, PCs, Tasseled Cap, ISODATA, and parcel layers. At the time of layer stacking, sample LULCs are collected from the DOQQ image by using the ROI tool in ENVI 4.4. ROIs are polygons identified by the user to define a sample boundary for each LULC class. The sampling training sites are selected based on the stratified random samples that are generated from the ERDAS Im agine 9.1 software. As a result, seven samples

PAGE 52

52 (257 pixels) are collected for water, five samples (257 pixels) for wetlands, nine samples (1,141 pixels) for urban recreational and others, thirty four samples (810 pixels) for commercial institutional transp ortation, seventeen samples (1,251 pixels) for multi family, twenty nine samples (1,072 pixels) for single family, nine samples (422 pixels) for industrial warehouses, forty nine samples (442 pixels) for forests, and thirty three samples (210 pixels) for g rasslands. These sample polygons are spread throughout the pilot area. Because the barren use is not found on the DOQQ image, there are no ROIs for the barren class. When sample ROIs are collected, they are reconciled to the stacked image because the 1 met er resolution of sample ROIs are different from the 30 meter stacked image generated from the ETM+ images. A decision tree is built in ENVI 4.4 thereafter based on the reconciled ROIs and the stacked image. The computer picks up DN values from the stack ed image and splits these DN values into binary groups based on the rules the computer generates from the training samples. It proceeds as a recursive process until it is not possible to continue further. The algorithm selected in the calculation is the QU EST with other parameters set as default. (Trees are not provided in this research because its size has exceeded the page limits.) No manual pruning is conducted and the pruning process is executed automatically by the computer. After the decision tree is built, it is input into the ENVI 4.4 to map the ten LULC classe s (barren not included). When the map is created, 196 random points are selected by using stratified random sampling on the classified image to do an accuracy assessment. These 196 random point s are the outcome with an expected accuracy of 85 percent, a Z score of 1.96, and an allowable uncertainty of 0.05. The LULC of random points are verified from the DOQQ image and field trips. IDRISI Andes A similar process is undertaken based on the IDRIS I Andes software. Because the stacked layer has already been created in ENVI 4.4, it is imported into IDRISI Andes by containing a

PAGE 53

53 bunch of sub layers. After the sub layers are imported, a vector digitized training sites file is created based on the previo usly generated ROIs in the ENVI 4.4. Then, a splitting rule is selected so that a tree can be built. The tree pruning process is also conducted automatically. After the tree is built, a classification map is produced, and the map is imported into the ERDAS Imagine 9.1 software for accuracy assessment. The accuracy assessment is processed also based on 196 stratified sampling points. This research tests three splitting rules in the IDRISI Andes software package. They are gain ratio, entropy, and Gini. Becaus e the three trees produce the maps that have apparent accuracy results, the map with the most obviously correct result is used to do an accuracy assessment, which can save research time. V I S Method This study uses an ETM+ image to extract five components based on the V I S model and uses the maximum likelihood method to extract these five components (Hung and Ridd, 2002). Procedurally, first, end members are collected from DOQQ images and later reconciled to the ETM+ images for the five classes as mention ed above using the ROI tool in ENVI 4.4. When enough end members are collected, five rule maps are produced with each rule representing each component. Then, the rule maps are transformed to be exponential and are further adjusted according to the five cla ss compositions collected from the training sample points on the DOQQ image. As a result, nine quadrant fishnets in 30 by 30 meters dimension are created randomly in ArcGIS 9.3 on the DOQQ images so that sub pixel spectra can be unmixed and different compo nent percentages within a pixel can be determined. A linear regression model is applied in this regard. After rule maps are adjusted, they are classified into ten classes, namely single family, multi family, commercial institutional transportation, industr ial warehouses, grasslands, forests, agricultural, recreational and others, wetlands, and water (no barren is found in the pilot

PAGE 54

54 area), using the CART method. Finally, 196 stratified random sampling points are found, and accuracy is assessed. Supervised M ethod The supervised method is implemented in the ENVI 4.4 software. The classification is based on the classifier algorithms, which namely are parallelepiped, minimum distance, Mahalanobis distance, and maximum likelihood. Because of the previously create d ROIs, the process of the supervised method is fairly straightforward. The Landsat ETM+ imagery with ROIs is input into the computer, and the results using different classifier algorithms are produced. As introduced in the Literature Review because of th e spectral characteristics of urban LULC classe s, the overall accuracy level using this method is expected to be lower than using the CART method. In fact, after producing maps applying the above four types of classifier algorithms, it is found that the su pervised method is unable to distinguish the urban LULC classe s for their spectral characteristic for mixed pixels and similar pixels. It is also unable to handle these many urban classes as proposed in this research with respect to high accuracy based on the Landsat ETM+ imageries. Therefore, it is meaningless to measure the individual accuracy level of each supervised classifier algorithm. This research does not go into that direction; instead, this research uses the supervised classification maps as a re ference to the maps created by the CART method to see whether the CART method can reveal every detail of urban LULC classe s. More details are discussed further in the next part of the chapter. Pilot Area Results CART Method The result s of using the CART method to classify LULC classe s are quite encouraging. As Figures 2 4, 2 5, 2 6 and 2 7 show, the ENVI 4.4 RuleGen 1.02 QUEST algorithm, the IDRISI Andes gain ratio, entropy, and Gini splitting rules all have good turnouts in classifying urban

PAGE 55

55 LULC classe s. On these maps, single family, multi family, commercial institutional transportation, industrial warehouses, and recreational others have excellent classifications with clear boundar ies that match soundly with the parcel boundaries and ground truths. In addition, wetlands, water, forests, agricultural, and grasslands the land covers that are within the parcels are all nicely presented on these maps. These illustrate that the CART meth od is an effective tool to handle the data at both the parcel level and at the sub parcel level. Comparing the maps produced by the different splitting rules, it is found that there are still slight differences for each rule to map urban LULC classe s. Thi s research finds that the ENVI 4.4 QUEST algorithm performs soundly in classifying urban LULC classe s with high accuracy. This is due to the fact that the QUEST algorithm takes categorical variables into consideration. For the IDRISI Andes software package although some studies suggest that the Gini splitting rule yields higher accuracy than other splitting rules, such as the entropy (Zambon et al., 2006) and gain ratio, this study finds that the gain ratio splitting rule outperforms the Gini or the entrop y rule from direct visual observations. For example, in the Gini rule, the natural curve of Banias Arm Lake is flattened out by nearby wetlands as well as a nearby recreational use while d shape. This may be due to the fact that the gain ratio rule normalized split gains during node splitting processes. The overall accuracy of the RuleGen 1.02 QUEST algorithm reaches 87.24 percent based on 196 stratified random sampling points. The specif ic accuracy breakdowns for each class are presented in Tables 2 1, 2 2, and 2 3. From the tables, it is evident that water and agricultural have the highest accuracy level while forests, single family, commercial institutional transportation, recreational and others, industrial warehouses, grasslands, and wetlands also have a fairly good accuracy level. The multi family use has a relatively lower accuracy level,

PAGE 56

56 however. This is because the multi family use can easily be confused with single family and comm ercial institutional transportation uses. From the accuracy assessment, specifically, single family is slightly mixed with multi family, commercial institutional transportation, and forests. Commercial institutional transportation is slightly mixed with si ngle family, forests, recreational and others, multi family, and grassland. Multi family is slightly mixed with single family, and commercial institutional transportation. Industrial warehouses is slightly mixed with commercial institutional transportation The overall accuracy of the IDRISI Andes gain ratio splitting rule was 85.20 percent based on 196 stratified random sampling points. The specific breakdowns for each class are listed in Tables 2 4, 2 5, and 2 6. From the tables, urban LULC classe s such as industrial warehouses, commercial institutional transportation, single family, multi family, forests, water, and agricultural are nicely classified. Different from the results of the ENVI 4.4 QUEST module, recreational and others, grasslands, and wetla nds in IDRISI Andes have a relatively lower accuracy level. Specifically, grasslands are slightly mixed with commercial institutional transportation; wetlands are slightly mixed with single family; and recreational and others are slightly mixed with grassl ands, wetlands, and commercial institutional transportation. Overall, the CART method is proficient in classifying urban LULC classe s at detailed levels. It is capable of handling both the parcel level and sub parcel level data and can well address sub pixel, spectral similarity, and mixed pixel issues for urban LULC classifications. In particular, in this study, a large parcel, the University of Florida campus, is found together with the surrounding smaller parcels, but urban LULCs at the sub campus le vel are well presented. In many cases, by using the parcel ancillary data, the CART method can mask out the mixed and similar pixels and represent them by a single class. This is especially useful when researchers

PAGE 57

57 want to identify residential areas that ar e covered by forests and grasslands, in which low density single family units can be easily overlooked by satellite sensors. The identification of the residential areas at the parcel level could avoid underestimates of urban built areas due to vegetation c over. The overall accuracy levels for using the two software packages are both above the 85 percent threshold. As what is found in this study, the CART method is able to identify sub parcel land covers. However, the identification of the sub parcel land c overs is mostly suitable for large parcels such as the University of Florida campus. For small parcels, the sub parcel covers are masked out, however. As compared to the maps produced by the supervised classification method, the CART method is more capable of illustrating urban land uses rather than urban land covers. Contrarily, the conventional supervised method is more able to present urban land covers, however. V I S Method The results for the V I S model are based on the rule maps that are generated f rom the ENVI 4.4 software. Because there are five classes, namely, bright impervious surface, dark impervious surface, forests, grasslands, and soil, for the V I S model, there will be five rule maps generated for these five classes. The original five rule maps are illustrated in Figure 2 8 with the brightest presenting the highest DN values. Because the rule maps are in a discriminant function format, they will be converted into the exponential function format and then normalized. Therefore, the above Fig ure 2 8 maps are converted to the maps in Figure 2 9, using the Band Math tool in ENVI 4.4. Based on the nine quadrant fishnets of 30 by 30 meters in dimension that are randomly collected in ArcGIS 9.3 on the DOQQ images, sub pixel spectra on Figure 2 9 ar e further unmixed, and different component percentages within a pixel are determined for the rule maps in Figure 2 9. As a result, a linear

PAGE 58

58 regression model is applied in this case. The adjusted rule maps based on the linear regression model are extracted in ENVI 4.4 and presented in Figure 2 10; the highest DN values in these maps show each of the five land covers, which are presented by the brightest areas in Figure 2 10. These five land covers are further illustrated in Figure 2 11. As mentioned above, t he final V I S classification map is presented in Figure 2 11. From 2 11, it is evident that the single family use is missing while the multi family use, the commercial institutional transportation use, and the industrial warehouses use are represented by bright impervious surfaces and dark impervious surfaces. The final classification map should use the CART method in order to yield the eleven LULC classes described in the research question session. The result is presented in Figure 2 12, in which the abov e four urban uses such as the single family use, the multi family use, the commercial institutional transportation use, and the industrial warehouses use are well represented. The accuracy assessment that uses 196 stratified sampling points is summarized i n Tables 2 7, 2 8, and 2 9. The overall accuracy level for the V I S model using the CART method reaches 74.88 percent. Supervised Method A supervised classification method is also conducted based on the maximum likelihood algorithm. Comparing the outcome s of the different classifier algorithms, it is found that the maximum likelihood algorithm produces classification maps that are closer to the ground truths than other algorithms from visual interpretations whereas it still has a disparity from accurate c lassifications of ground LULC classe s (Figure 2 13). From the map, it is clear that some urban features are mixed with each other. For example, agricultural is mixed with grasslands; multi family is mixed with commercial institutional transportation; multi family is mixed with single family; industrial warehouses is mixed with commercial institutional transportation; single family is mixed with forests; and recreational others is mixed with grasslands, single family, and

PAGE 59

59 agricultural land. However, a forest corridor on the northwestern part of the quadrant can be clearly identified, where creeks are flowing in between. This is beyond the scope of the CART method that uses parcel layers as ancillary data because forests and other LULC features within the resi dential parcels are masked out. Pilot Area Summaries of CART Method, V I S Method, and Supervised Method From the above comparisons, the strengths and weaknesses of each method are summarized in Table 2 10. These descriptions provide guidelines of urban L ULC classifications for the entire county, which is explained in detail in the next part of the research. Based on the results in Table 2 10, this study uses the CART method to map urban LULC classifications for the entire county. County Wide Urban LULC C lassifications Using the Preferred CART Methodology The county wide urban LULC classifications apply the CART method. The ENVI 4.4 RuleGen 1.02 QUEST module is employed in this regard. The classifications are based on 1982, 1994, and 2003, respectively, using the Landsat TM and ETM+ imageries. The training data use the county 1982 aerial photography image, 1995 and 2004 DOQQ images. Differen t from the above pilot area study, the layer stacking for the county wide classifications applies thirteen layers, which is the same number of layers for all three timeframes. These layers include (1) the Landsat TM or ETM+ layer, (2) the Tasseled Cap laye r, (3) the PCs layer, (4) the NDVI layer, (5) the ISODATA layer, (6) the parcel layer, (7) the residential versus commercial institutional transportation parcel layer, (8) the single family versus multi family parcel layer, (9) the single family versus nat ural land parcel layer; (10) the commercial institutional transportation layer versus natural land parcel layer, (11) the forests versus agricultural land parcel layer, (12) the recreational others versus agricultural land parcel layer, and (13) the grassl ands versus agricultural land parcel layer. The residential versus commercial institutional transportation

PAGE 60

60 parcel layer is the one that uses the codes to differentiate between the residential use and the commercial institutional transportation use in the p arcel data. It is the same for the layers of (8) to (13) in order to differentiate these uses by the parcel codes. The above layers are empirically tested in order to compare the accuracy levels for each combination of layer stacking before and after a cer tain parcel layer is added. This effort is made in order to reduce the pixel confusions given by the raw TM and ETM+ data because pixel confusions can happen between residential uses and the commercial institutional transportation use, the single family us e and the multi family use, commercial institutional transportation use and the natural land, forests and the agricultural land, the recreational others land and the agricultural land, and grasslands and the agricultural land. The ISODATA layer is operated based on 60 iterations that are the same as the classification method tested for the pilot area based on the relevant research results. In addition, the parcel layers from (6) to (13) are all categorical datasets. The layers from (1) to (5) above are cont inuous datasets, however. In particular, for the layers from the above (6) to (13), 1 is coded for one category, and 10 is coded for the other category. The coding is properly adjusted before these layers are added into the layer stacking, and they are sel ected as categorical datasets in the ENVI 4.4 RuleGen 1.02 QUEST module. The remaining settings use default ones in the ENVI RuleGen 1.02 QUEST module. Because the initial ROI signatures create a large quantity of data to run in the computer, which overbur den ENVI 4.4, reduced ROI signatures are sought based on the random sampling technique. As a result, for the 2003 classification maps, after the adoption of the random sampling techniques, there are 10,952 pixels for the single family use, 10,247 pixels fo r the multi family use, 11,238 pixels for the commercial institutional transportation use, 5,878 pixels for the industrial warehouses use, 35,013 pixels for the grasslands use, 37,751 pixels for the

PAGE 61

61 forests use, 12,827 pixels for the agricultural use, 12,3 40 pixels for the recreational others use, 40,611 pixels for the wetlands use, 11,346 pixels for water, and 1,067 pixels for the barren use. Similarly, for the 1994 classification maps, there are 21,224 pixels for the single family use, 6,891 pixels for th e multi family use, 11,195 pixels for the commercial institutional transportation use, 4,122 pixels for the industrial warehouses use, 38,043 pixels for the grasslands use, 30,350 pixels for the forests use, 16,072 pixels for the agricultural use, 12,303 p ixels for the recreational others use, 41,953 pixels for the wetlands use, 12,514 pixels for water, and 695 pixels for the barren use. For the 1982 classification maps, there are 12,812 pixels for the single family use, 4,365 pixels for the multi family us e, 8,889 pixels for the commercial institutional transportation use, 2,656 pixels for the industrial warehouses use, 58,960 pixels for the grasslands use, 48,045 pixels for the forests use, 14,922 pixels for the agricultural use, 11,902 pixels for the recr eational others use, 15,690 pixels for the wetlands use, 10,080 pixels for water, and 695 pixels for the barren use. The ROI pixels of LULC classe s in different timeframes are illustrated in Table 2 20. The Central Processing Unit ( CPU ) time for producing each classification map for the three timeframes is 15 to 20 hours to finish a CART tree, which is later converted into classification maps using the trees generated. The final 1982 classification map is illustrated in Figure 2 14. The final 1994 classific ation map is illustrated in Figure 2 15. The final 2003 classification map is illustrated in Figure 2 16. The accuracy assessments for each of the three timeframes are based on 1,100 stratified random sample points, which are 100 points for each class, and they are generated in ERDAS Imagine 9.1. These accuracy assessments are illustrated in Tables 2 11, 2 12, 2 13, 2 14, 2 15, 2 16, 2 17, 2 18, and 2 19, respectively.

PAGE 62

62 Table 2 1. Error matrix (ENVI 4.4 RuleGen 1.02 QUEST algorithm) Reference Data Classified Data Single Family Commercial Institutional Transportation Forests Recreational Others Multi Family Grasslands Water Industrial Warehouses Wetlands Agricultural Row Total Single Family 33 1 1 0 1 0 0 0 0 0 36 Commercial Institutional Transportation 1 32 2 1 1 1 0 1 1 0 40 Forests 1 0 46 0 1 1 0 1 0 0 50 Recreational Others 0 0 1 8 0 0 0 0 0 0 9 Multi Family 2 3 0 0 13 0 0 0 0 0 18 Grasslands 0 1 0 0 0 8 0 0 0 0 9 Water 0 0 0 0 0 1 18 0 1 0 20 Industrial Warehouses 0 1 0 0 0 0 0 10 0 0 11 Wetlands 0 0 0 0 0 0 0 0 2 0 2 Agricultural 0 0 0 0 0 0 0 0 0 1 1 Column Total 37 38 50 9 16 25 18 12 4 1 196

PAGE 63

63 Table 2 2. Accuracy totals (ENVI 4.4 RuleGen 1.02 QUEST algorithm ) Class Name Reference Totals Classified Totals Number Correct Producers Accuracy Users Accuracy Single Family 37 36 33 89.19% 91.67% Commercial Institutional Transportation 38 40 32 84.21% 80.00% Forests 50 50 46 92.00% 92.00% Recreational Others 9 9 8 88.89% 88.89% Multi Family 16 18 13 81.25% 72.22% Grasslands 11 9 8 72.73% 88.89% Water 18 20 18 100.00% 90.00% Industrial Warehouses 12 11 10 83.33% 90.91% Wetlands 4 2 2 50.00% 100.00% Agricultural 1 1 1 100.00% 100.00% Totals 196 196 171 Overall classification a ccuracy = 87.24% Table 2 3. KAPPA (K^) s tatistic s (ENVI 4.4 RuleGen 1.02 QUEST a lgorithm) Class Name Kappa Conditional Kappa for each c ategory Single Family 0.8973 Commercial Institutional Transportation 0.7519 Forests 0.8926 Recreational Others 0.8835 Multi Family 0.6975 Grasslands 0.8823 Water 0.8899 Industrial Warehouses 0.9032 Wetlands 1.0000 Agricultural 1.0000 Overall Kappa Statistics 0.8473

PAGE 64

64 Table 2 4 Error matrix (IDRISI Andes Ratio G ain rule) Reference Data Classified Data Industrial Warehouses Single Family Recreational Others Forests Commercial Institutional Transportation Multi Family Water Grasslands Wetlands Agricultural Row Total Industrial Warehouses 12 0 0 0 0 0 0 1 0 0 13 Single Family 0 33 0 0 0 2 0 0 0 0 35 Recreational Others 0 0 6 0 1 0 0 2 1 0 10 Forests 1 1 0 42 1 0 0 1 2 0 48 Commercial Institutional Transportation 0 0 0 1 30 0 0 0 0 0 31 Multi Family 0 0 0 0 6 13 0 0 0 0 19 Water 0 0 0 0 0 0 19 0 0 0 19 Grasslands 0 0 0 0 2 0 0 7 0 0 9 Wetlands 0 1 1 5 0 0 0 0 4 0 11 Agricultural 0 0 0 0 0 0 0 0 0 1 1 Column Total 13 35 7 48 40 15 19 11 7 1 196

PAGE 65

65 Table 2 5 Accuracy totals (IDRISI Andes Ratio G ain rule) Class Name Reference Totals Classified Totals Number Correct Producers Accuracy Users Accuracy Industrial Warehouses 13 13 12 92.31% 92.31% Single Family 35 35 33 94.29% 94.29% Recreational Others 7 10 6 85.71% 60.00% Forests 48 48 42 87.50% 87.50% Commercial Institutional Transportation 40 31 30 75.00% 96.77% Multi Family 15 19 13 86.67% 68.42% Water 19 19 19 100.00% 100.00% Grasslands 11 9 7 63.64% 77.78% Wetlands 7 11 4 57.14% 36.36% Agricultural 1 1 1 100.00% 100.00% Totals 196 196 167 Overall classification a ccuracy = 85.20% Table 2 6. KAPPA (K^) s tati stics (IDRISI Andes Ratio Gain r ule) Class Name Kappa Conditional Kappa for each c ategory Industrial Warehouses 0.9176 Single Family 0.9304 Recreational Others 0.5852 Forests 0.8345 Commercial Institutional Transportation 0.9595 Multi Family 0.6580 Water 1.0000 Grasslands 0.7646 Wetlands 0.3401 Agricultural 1.0000 Overall Kappa Statistics 0.8256

PAGE 66

66 Table 2 7 Error matrix (CART method V I S model) Reference Data Classified Data Agricultural Water Wetlands Recreational Others Commercial Institutional Transportation Multi Family Single Family Industrial Warehouses Forests Grasslands Row Total Water 0 7 0 0 0 0 0 0 0 0 7 Wetlands 0 0 3 0 0 0 0 0 0 0 3 Recreational Others 0 0 0 5 0 0 0 0 1 1 7 Commercial Institutional Transportation 1 0 0 0 41 1 0 0 5 3 51 Multi Family 0 0 0 0 5 12 3 0 4 0 24 Single Family 0 0 0 0 3 7 63 1 9 0 83 Industrial Warehouses 0 0 0 0 0 0 0 5 0 0 5 Forests 0 0 0 0 3 1 3 0 22 1 30 Grassland 0 0 0 0 0 1 0 0 0 0 1 Agricultural 0 0 0 0 0 0 0 0 0 0 0 Column Total 1 7 3 5 52 22 69 6 41 5 211

PAGE 67

67 Table 2 8 Accuracy totals (CART method V I S model) Class Name Reference Totals Classified Totals Number Correct Producers Accuracy Users Accuracy Water 7 7 7 100.00% 100.00% Wetlands 3 3 3 100.00% 100.00% Recreational Others 5 7 5 100.00% 71.43% Commercial Institutional Transportation 52 51 41 78.85% 80.39% Multi Family 22 24 12 54.55% 50.00% Single Family 69 83 63 91.30% 75.90% Industrial Warehouses 6 5 5 100.00% 71.43% Forests 41 30 22 91.30% 75.90% Grasslands 5 1 0 0.00% 0.00% Agricultural 1 0 0 0.00% 0.00% Totals 211 211 158 Overall classification a ccuracy = 74.88% Table 2 9. KAPPA (K^) statistics (CART method V I S m odel) Class Name Kappa Conditional Kappa for each c ategory Single Family 0.6419 Commercial Institutional Transportation 0.7398 Forests 0.6690 Recreational Others 0.7074 Multi Family 0.4418 Grasslands 0.0243 Industrial Warehouses 1.0000 Agricultural 0.0000 Water 1.0000 Wetlands 1.0000 Overall Kappa Statistics 0.6735

PAGE 68

68 Table 2 10 Comparisons of the strengths and weaknesses of urban LULC classification methods LULC Classification M ethod Purpose Strengths Weaknesses CART Method A knowledge based method. Some researchers apply it to classify urban LULC classe s that may have mixed pixels in Landsat TM and ETM+ images. It is used t o yield high accuracy for land use classifications. For urban research, current research has only applied it to classify urban impervious surfaces as opposed to natural areas. However, it is capable of classifying LULC to USGS Level III concerning Landsat TM and research includes classifying urban LULC using Landsat TM and ETM+ sensors equivalent to USGS Level III. Is able to deal with mixed pixels because it is not a statistically based approach such as minimum distance; Performs well at the sub pixel level because it circumvents the urban sub pixel issue by applying ancillary data; Yields high accuracy; Can incorporate either numerical or categorical data to be included as ancillary data; Is capable of classifying urban L ULC classe s equivalent to USGS Level III; Raw data can be Landsat data and does not necessarily have to have high resolution data; Best in classifying urban LULC classe s at the county level, which is beyond the capability of higher resolution sensors to ha ndle because of excessive amount of data; Relies on ancillary data; Needs to test different ancillary data and finds the most optimal combinations for layer stacking; Still needs conventional methods to help create ancillary data, such as ISODATA, NDVI, PC s Tasseled Cap and so on; Needs to combine with other methods such as GIS for image processing; and Makes it easier to find accurate classifications of land uses rather than land covers if parcel data are used as ancillary data (this is because the parce l layer will mask out land cover details).

PAGE 69

69 Table 2 10 Continued LULC Classification Method Purpose Strengths Weaknesses Is able to utilize maximally the advantages of Landsat data (band 1 through band 7) for urban research in terms of spectral characteristics of urban LULC classe s; Has default settings that are good enough; does not need to know complicated algorithms; can perform nicely based on fairly user friendly interface; and Does not need to trim trees, and automatic tree pruning performs well to yield high accuracy. V I S Method A sub pixel LULC method designed to address urban mixed pixels in terms of Landsat TM and ETM+ sensors. It classifies urban LULC classe s based on three elements: vegetation, i mpervious surface, and soil. An urban LULC is determined based on the percentage of each of three elements in a pixel. As a result, end members of vegetation, impervious surface, and soil need to be extracted from satellite images. Attempts to deal with ur ban mixed pixels; Simplifies multiple LULC categories into three; Is capable of dealing with mixed pixels for urban LULC classe s simply based on impervious surfaces and yield high accuracy; and Has the potential to classify multiple vegetation covers. Needs to combine with other methods, such as Artificial Intelligence (AI) to determine actual classes in light of mixed pixels; Is not an easy method to retrieve end members in an accurate fashion because of the technical difficulty; Has weak performance t o classify different urban uses such as single family, multi family, commercial, and industrial;

PAGE 70

70 Table 2 10 Continued LULC Classification Method Purpose Strengths Weaknesses Has relatively low accuracy to classify multiple urban uses; May still need to use the CART method to classify multiple urban classes; Simplifies use classes, which causes loss of pixel abundance in terms of V I S elements; Cannot deal with water or wetland. Has difficulty in defining the urban uses considering their percentages of three elements (there is technical difficulty in classifying multiple urban LULC classe s based on the referencing diagram); Has difficulty in obtaining soil end members in urban environment (using mineral indexes to sort out soil class does not guarantee to yield high accuracy in urban LULC classification).

PAGE 71

71 Table 2 10 Continued LULC Classification Method Purpose Strengths Weaknesses Supervise d a nd Unsupervised (Or Hybrid) Method The conventional supervised and unsupervised method are applied in multiple fields, including the areas of urban and non urban research. The supervised and unsupervised methods are the bases of all LULC classification methods, which is the most flexible method for all circumstances. The hybrid method utilizes the advantages of supervised and unsupervised methods. For urban research, it is still difficult to handle mixed pixels. As a result, it is a per pixel based method. Research finds that the hybrid method combining with other method such as fuzzy logic works better. Supervised method usually yields higher accuracy than unsupervised method (Short, no date); Is not required to be familiar with spectral characteristics of ground truths (Kramber and Morse, 1994); the unsupervised method can help; A ddresses spectral similarities to some degree; has classes that are which can be identified (Liou and Yang, no date) using hybrid method; Uses ISODATA that is able to differentiate needed classes if based on optimal numbers o f classes and iterations; Is more flexible in including fuzzy (soft) classification method into LULC classe s; and Works better if combined with band ratio or vegetation index (or other index) images ( Banman no date). It is not clear whether the hybrid method yields high accuracy in terms of urban research; It is less capable of dealing with urban mixed pixels and spectral similarity issues; It still needs to be familiar with the study area for supervised signatures; otherwise, supervised signatures may contain other uses and covers; and It needs to be a professional who is an expert on supervised and unsupervised methods.

PAGE 72

72 Table 2 11 Error matrix (1982 classification map) Reference Data Classified Data Single Family Multi Family Commercial Institutional Transportation Industrial Warehouses Grasslands Forests Agricultural Recreational Others Wetlands Water Barren Row Total Single Family 33 0 0 0 2 5 0 0 0 0 0 40 Multi Family 0 4 0 0 0 0 0 0 0 0 0 4 Commercial Institutional Transportation 0 0 30 0 4 2 0 0 0 0 0 36 Industrial Warehouses 0 0 0 3 0 0 0 0 0 0 0 3 Grasslands 3 0 2 0 343 16 0 0 2 0 0 366 Forests 2 0 2 0 11 405 0 0 12 0 0 432 Agricultural 0 0 1 0 0 1 55 0 0 0 0 57 Recreational Others 0 0 0 0 0 1 0 64 3 0 0 68 Wetlands 0 0 0 0 0 1 0 0 40 0 0 41 Water 0 0 0 0 0 0 0 0 0 52 0 52 Barren 0 0 0 0 0 0 0 0 0 0 1 1 Column Total 38 4 35 3 360 431 55 64 57 52 1 1,100

PAGE 73

73 Table 2 12 Accuracy totals (1982 classification map) Class Name Reference Totals Classified Totals Number Correct Producers Accuracy Users Accuracy Single Family 38 40 33 86.84% 82.50% Multi Family 4 4 4 100.00% 100.00% Commercial Institutional Transportation 35 36 30 85.71% 83.33% Industrial Warehouses 3 3 3 100.00% 100.00% Grasslands 360 366 343 95.28% 93.72% Forests 431 432 405 93.97% 93.75% Agricultural 55 57 55 100.00% 96.49% Recreational Others 64 68 64 100.00% 94.12% Wetlands 57 41 40 70.18% 97.56% Water 52 52 52 100.00% 100.00% Barren 1 1 1 100.00% 100.00% Totals 1,100 1,100 1,030 Overall classification a ccuracy = 93.64% Table 2 13. KAPPA (K^) statistics (1982 classification map) Class Name Kappa Conditional Kappa for each c ategory Single Family 0.8187 Multi Family 1.0000 Commercial Institutional Transportation 0.8279 Industrial Warehouses 1.0000 Grasslands 0.9066 Forests 0.8972 Agricultural 0.9631 Recreational Others 0.9375 Wetlands 0.9743 Water 1.0000 Barren 1.0000 Overall Kappa Statistics 0.9122

PAGE 74

74 Table 2 1 4 Error matrix (1994 classification map) Reference Data Classified Data Single Family Multi Family Commercial Institutional Transportation Industrial Warehouses Grasslands Forests Agricultural Recreational Others Wetlands Water Barren Row Total Single Family 69 1 5 0 1 4 0 0 1 0 0 81 Multi Family 0 5 0 0 0 0 0 0 0 0 0 5 Commercial Institutional Transportation 0 0 13 0 1 0 0 0 0 0 0 14 Industrial Warehouses 0 0 0 4 1 0 0 0 0 0 0 5 Grasslands 1 0 4 0 212 15 1 0 1 0 0 234 Forests 5 0 1 0 19 394 2 0 13 0 0 434 Agricultural 0 0 0 0 0 3 50 0 0 0 0 53 Recreational Others 0 0 0 0 0 0 0 53 2 0 0 55 Wetlands 0 0 1 0 8 10 0 0 143 2 0 164 Water 0 0 0 0 0 0 0 0 0 54 0 54 Barren 0 0 0 0 0 0 0 0 0 0 1 1 Column Total 75 6 24 4 242 426 55 53 160 56 1 1,100

PAGE 75

75 Table 2 1 5 Accuracy totals (1994 classification map) Class Name Referenc e Totals Classifie d Totals Numbe r Correct Producer s Accuracy Users Accurac y Single Family 75 81 69 92.00% 85.19% Multi Family 6 5 5 83.33% 100.00% Commercial Institutional Transportation 24 14 13 54.17% 92.86% Industrial Warehouses 4 5 4 100.00% 80.00% Grasslands 242 234 212 87.60% 90.60% Forests 426 434 394 92.49% 90.78% Agricultural 53 53 50 94.34% 94.34% Recreational Others 53 55 53 100.00% 96.36% Wetlands 160 164 143 89.38% 87.20% Water 56 54 54 96.43% 100.00% Barren 1 1 1 100.00% 100.00% Totals 1,100 1,100 998 Overall classification accuracy = 90.73% Table 2 16 KAPPA (K^) statistics (1994 classification map) Class Name Kappa Conditional Kappa for each c ategory Single Family 0.8410 Multi Family 1.0000 Commercial Institutional Transportation 0.9270 Industrial Warehouses 0.7993 Grasslands 0.8795 Forests 0.8496 Agricultural 0.9405 Recreational Others 0.9618 Wetlands 0.8502 Water 1.0000 Barren 1.0000 Overall Kappa Statistics 0.8790

PAGE 76

76 Table 2 17 Error matrix (2003 classification map) Reference Data Classified Data Single Family Multi Family Commercial Institutional Transportation Industrial Warehouses Grasslands Forests Agricultural Recreational Others Wetlands Water Barren Row Total Unclassified 0 0 0 0 0 1 0 0 0 0 0 1 Single Family 80 1 2 0 3 7 0 0 0 0 0 93 Multi Family 1 9 3 0 0 1 0 0 0 0 1 15 Commercial Institutional Transportation 0 0 15 0 0 0 0 1 0 0 0 16 Industrial Warehouses 0 0 0 6 0 0 0 0 0 0 0 6 Grasslands 1 0 0 0 203 18 0 0 8 0 0 230 Forests 5 0 1 0 19 378 1 0 15 0 0 419 Agricultural 0 0 0 0 1 0 56 0 0 0 0 57 Recreational Others 0 0 0 0 0 0 0 51 2 0 0 53 Wetlands 0 1 0 0 2 7 0 2 152 2 0 166 Water 0 0 0 0 0 0 0 0 1 42 0 43 Barren 0 0 0 0 0 0 0 0 0 0 1 1 Column Total 87 11 21 6 228 412 57 54 178 44 2 1,10 0

PAGE 77

77 Table 2 18 Accuracy totals (2003 classification map) Class Name Referenc e Totals Classifie d Totals Numbe r Correct Producer s Accurac y Users Accuracy Single Family 87 93 80 91.95% 86.02% Multi Family 11 15 9 81.82% 60.00% Commercial Institutional Transportation 21 16 15 71.43% 93.75% Industrial Warehouses 6 6 6 100.00% 100.00% Grasslands 228 230 203 89.04% 88.26% Forests 412 419 378 91.75% 90.21% Agricultural 57 57 56 98.25% 98.25% Recreational Others 54 53 51 94.44% 96.23% Wetlands 178 166 152 85.39% 91.57% Water 44 43 42 95.45% 97.67% Barren 2 1 1 50.00% 100.00% Totals 1,100 1,100 993 Overall classification a ccuracy = 90.27% Table 2 19 KAPPA (K^) statistics (2003 classification map) Class Name Kappa Conditional Kappa for each c ategory Single Family 0.8482 Multi Family 0.5960 Commercial Institutional Transportation 0.9363 Industrial Warehouses 1.0000 Grasslands 0.8519 Forests 0.8436 Agricultural 0.9815 Recreational Others 0.9603 Wetlands 0.8994 Water 0.9758 Barren 1.0000 Overall Kappa Statistics 0.8746

PAGE 78

78 Ta ble 2 20 ROI pixels of LULC classe s in 1982, 1994, and 2003 (pixels) LULC 2003 1994 1982 Single Family 10,952 21,224 12,812 Multi Family 10,247 6,891 4,365 Commercial Institutional Transportation 11,238 11,195 8,889 Industrial Warehouses 5,878 4,122 2,656 Grasslands 35,013 38,043 58,960 Forests 37,751 30,350 48,045 Agricultural 12,827 16,072 14,922 Recreational Others 12,340 12,303 11,902 Wetlands 40,611 41,953 15,690 Water 11,346 12,514 10,080 Barren 1,067 695 695

PAGE 79

79 Figure 2 1. I S model (Ridd, 1995. p.2,173) Figure 2 2. V I S model and point T (Hung, 2003, p.13; Hung and Ridd, 2002, p.1,174 )

PAGE 80

80 Figure 2 3. Pilot area location for LULC classification

PAGE 81

81 Figure 2 4. Urban LULC classifications using ENVI 4.4 RuleGen 1.02 QUEST module

PAGE 82

82 Figure 2 5. Urban LULC classifications using IDRISI Andes gain ratio rule

PAGE 83

83 Figure 2 6. Urban LULC classifications using IDRISI Andes entropy rule

PAGE 84

84 Figure 2 7. Urban LULC classifications using IDRISI Andes Gini rule

PAGE 85

85 Figure 2 8. Original maps for five components. A) Bright Impervious Surface, B) Dark Impervious Surface, C) Forests, D) Grasslands, and E) Soil.

PAGE 86

86 Figure 2 9. Exponential transformed maps for five components. A) Bright Impervious Surface, B) Dark Impervious Surface, C) Forests, D) Grasslands, and E) Soil.

PAGE 87

87 Figure 2 10. Adjusted rule maps for five components. A) Bright Impervious Surface, B) Dark Impervious Surface, C) Forests, D) Grasslands, and E) Soil.

PAGE 88

88 Figure 2 11. V I S model final classification map

PAGE 89

89 Figure 2 12. V I S model using CART method to classify LULC

PAGE 90

90 Figure 2 13. Urban LULC classifications using the supervised method

PAGE 91

91 Figure 2 14. 1982 classification map for Alachua County

PAGE 92

92 Figure 2 15. 1994 classification map for Alachua County

PAGE 93

93 Figure 2 16. 2003 classification map for Alachua County

PAGE 94

94 CHAPTER 3 URBAN MODEL BUILDING THE MULTINOMIAL LOGI STIC REGRESSION MODE L Chapter Overview This chapter uses the ML R model to simulate urban growth in Alachua County through modeling the historical urban development in the county for the past 30 years. This chapter simulates the urban growth in the county for 2003, based on eleven LULC classe s identified in Chapter 2 for 1982, 1994, and 2003, respectively, according to the Landsat TM and ETM+ data. These eleven LULC cla sses are single family, multi family, commercial institutional transportation, industrial warehouses, recreational others, agricultural, forests, grasslands, wetlands, water, and barren. The study uses the images that are classified for 1982, 1994, and 200 3, respectively, to simulate the urban LULC classe s for 2003. Logistic Regression Model Logistic Regression is a model to predict membership of a dependent variable based on a series of independent variables, which can be continuous, discrete, dichotomou s, or some combination of several data types. The output of the modeling results is categorical or dichotomous belonging to a specific membership or not or a continuous probability value, between 0 and 1, for the membership. Logistic Regression can be appl ied to an urban study to predict the occurrence of urban development in a probabilistic manner, based on a series of independent variables such as demographic, social economic, ecologic physical, and spatial factors that are identified as significant (Hu a nd Lo, 2007). The Logistic Regression model considers historical factors towards urban development. Unlike the Cellular Automata (CA) model, which simulates urban spatial development without delving into the reasons behind the development, Logistic Regress ion is capable of informing those reasons (Hu and Lo, 2007). The Logistic Regression model is applied to deforestation, agricultural, and urban growth (Hu and Lo, 2007). For urban growth, practically, Hu and Lo (2007) used it to simulate urban

PAGE 95

95 growth in At lanta, Georgia. They found that the Logistic Regression model is able to include a variety of variables into analysis from physical, spatial, demographic, social, and economic aspects as well as land use development policies and environmental protection. T his is advantageous over non statistic models such as the CA model. More importantly, the Logistic Regression model requires much less CPU time for data processing. This is a big advantage over the CA model. However, Logistic Regression is a stationary mod el that treats land use development as if it occurs at the same time. It tells where the development takes place rather than when and hence is less temporally dynamic (Hu and Lo, 2007). In addition, the Logistic Regression model does not consider personal or household preferences for land use development nor personal behavior in land use development, which is unlike the Agent Based Model (ABM) that considers personal behaviors and choices. The logistic regression model will be explained in detail later in t his chapter, which will provide reasoning as to why the logistic regression model is chosen in this study for urban LULC modeling. The Accuracy of Urban Growth Modeling The models to simulate urban land use changes have been developed since the 1950s (Ba tty and Longley, 1994). The development of models for urban land use change attempts to simulate urban land uses from various angles by studying the inter relationship between human behaviors and land use development. The development of urban simulation mo dels is much the accuracy level of an urban development simulation model reflects how well humans standings of cities are still limited, it is still difficult for humans to simulate urban development with great accuracy. The only way humans can do this is to characterize the nature of urban development as much as possible and work as accurately as poss ible to reflect the real world urban development patterns. On the other

PAGE 96

96 hand, almost all models for predicting an event have errors ranging from sample errors to random errors 1 and to systematic errors, or bias. 2 The urban models are no exceptions. As a result, it is always necessary to conduct accuracy assessments for urban simulation models. Moreover, urban development is much more historically related. The prediction of urban land use changes typically starts from analyzing the land use historical deve lopment. Since most cities originate from organic, unplanned states (Batty and Longley, 1994), some may still be developed in an organic way; they make urban simulation difficult and, as a result, may introduce additional errors compared to a planned city. unpredictable and uncertain in nature. For land use development, land use decisions may be beyond the prediction of a scientific model. For example, different owners may deal with their land in a different way. Consider ing the same geophysical and economic conditions of their land, some may sell their land more quickly than others simply because of personal financial reasons. This may be beyond the prediction of an urban land use model. To deal with land use development uncertainties, consequently, it is common to use probability to address the issue. However, the introduction of probability to urban land use change models will unavoidably bring errors because, given a typical sample size, it is not possible to realize th e predicted land uses that 100 percent match the observed land uses: there is always a difference between the observed and the predicted. Lastly, there is a policy uncertainty for urban land use development, in which it is not possible for an urban land us e change model to predict whether there will be more conservative or more liberal policies for urban land use development and whether more strict environmental laws will be enacted in the future given the quantitative nature of the urban land use change mo dels. 1 The difference between observed values and predicted values (Mowrer and Congalton, 2000). 2 and Congalton, 2000).

PAGE 97

97 Uncertainty of Future Land Use Development Because the urban models are rule based, there are orders or rationales behind them. As a result, predictions of future land use world circumstances, however, a city may still be developed in its own way, being presented in a relatively organic fashion. As mentioned previously, because there are uncertainties related to people and their land use decisions, at the micro level, the actual urban land use changes may funds, even though there are land use control tools available such as the future land use plan and zoning. For example, a well predicted land for single family use may actually be developed as multi family, and it is common that the land that is suitable for development, which is predicted Uncertainty of Future Land Use Policies There is a land use policy uncertainty in the development of urban land use models. In reality, the future land us e policies may vary from the current ones at the local, state, and federal levels. For example, a local government may change its attitude towards land use development and may become more liberal or more conservative towards land use development. The same is true for the state and federal government. For example, whether the state or federal government will pass the growth control laws to discourage land use development or whether the Congress will enact stricter environmental laws will be determined by the political dynamics at the executive and legislative levels, which are beyond land use modeling. In addition, natural disasters, such as fires, earthquakes (NCGIA no date),

PAGE 98

98 and hurricanes, which are unpredictable in nature will inevitably influence land use development and thus influence the accuracy of urban models. Moreover, economic downturn and upturn (NCGIA, no date) will affect land use development as well a nd hence influence the accuracy of urban land use simulations. As a result, it is not likely that a model 100 percent matches the actual development; it is expected that the created models reflect the actual land use patterns to the maximum possible extent and best estimate the occurrence of future land use development. Relevant Literature Review for Logistic Regression discussions, which is the model building relate d to the MLR model for urban growth simulation. Poelmans and Rompaey (2010) utilized the Landsat MSS, TM, and ETM+ images to simulate urban growth based on 1976, 1988, and 2000, respectively, employing a hybrid model of the logistic regression and the CA m odels. Learned urban simulation studies that were mostly based on the logistic regression, CA, or combination of them, they proposed a hybrid model to simulate urban growth for the Flanders Brussels region of Belgium. Their land classification included urb an land, arable land, grassland, forest, and water, and their independent variables included distance to cities, slope, employment potentials, distance to roads, and zoning status. These land categories as well as identified variables allowed the authors t o produce thirty one models to simulate urban growth based on probabilities. They found that the CA model contributed less to the accuracy of the model than the logistic regression model. They also found that zoning status was the most important factor det ermining the accuracy of their model by comparing the odd ratios for each variable, and also that the combination of the logistic regression and the CA model dramatically enhanced the accuracy of the model. However, the drawback of their research is that t heir raw data, derived from the Landsat images, is not of high accuracy. The accuracy levels for 1976, 1988, and 2000 in their research were 77.6 percent, 82.8 percent, and

PAGE 99

99 82.3 percent, respectively. The 1988 and 2000 data had just reached the benchmark o f the 80 percent accuracy level, whereas the 1976 data was lower than the necessary level of 80 percent. sought to simulate urban growth. Also, the logistic regression model is adopted to improve the accuracy of the model. These elements are inspirational for this research, in which various spatial independent variables can be useful to simulate urban uses which can lead to the destinations of this research. Almeida et al. (2003) also applied a hybrid CA and logistic regression model to simulate urban uses based on five categories: residential use, industrial use, services, and mixed use, for the Town of Bauru, Brazil. They used the hybrid model to select possible explan atory (independent) variables that were associated with each land use identified above through looking values were selected when they were greater than a benc hmark, which created a total of nineteen explanatory variables. Then, these nineteen variables were input into the software, known as DINAMICA. During the process of calculating the probability values in the logistic regression model, the Wald Chi square t est and the G statistics were assessed in order for them to exclude the variables having the least significance level. As a result, MINITAB was used in this regard so as to obtain the P value for each variable based on the maximum likelihood method. Finall y, DINAMICA was employed again to compute the probabilities for each of the five urban uses mentioned above. Calibration was also conducted at this step so as to fine tune the parameters such as the number of iterations, average size and variance of patche s, and so on, in the CA model. The final map includes a predicted land use map for 1988 based on the five land use categories.

PAGE 100

100 There are three merits in the research of Almeida et al. (2003) First, they simulate urban uses for more than two categories, in wh ich five land use categories are included, not merely one, e.g., urban vs. non urban. Second, they employ the hybrid model to simulate urban growth based on the probabilities for each land use category. As a result, five different urban uses can be allocat ed through comparing the ordinal values of each use. Third, they use the P value to select independent variables that are significant to their research. These three merits of the research of Almeida et al. (2003) are also the direction that this study is going to take. Landis (1994) proposed a California Urban Future (CUF) model to simulate urban growth in the northern bay area of California which comprised fourteen counties. His model has four modules: population growth, spatial database, spatial allocation, a nd annexation, which is essentially a vector based urban growth analysis with the simulation based on the Census Topologically Integrated Geographic Encoding and Referencing ( TIGER ) data. On the wth based on the urban category only, i.e., the residential use (Landis and Zhang, 1997). It does not address detailed urban uses for the CUF model. To remedy the inadequacies of the CUF model, Landis and Zhang (1997) developed a CUF II model to simulate urban growth. The reason that the CUF II model is closely related to this research is that it proposed an MLR technique to model urban development and redevelopment. In addition, the CUF II model added four additional uses by expanding the original single urban use to residential use, commercial use, industrial use, public use, and transportation use. The CUF II model is inspirational for urban growth modeling in this study as it suggests using the MLR model, for the first time, to simulate urban developmen t and redevelopment. It also carries some heuristic elements such as inclusions of various urban uses,

PAGE 101

101 e.g., residential use, commercial use, industrial use, public use, and transportation use, as mentioned above, considering urban development and redevelo pment. The CUF and CUF II models that are inspirational for this study are also represented in additional elements such as the adoption of population forecasts, the proposal of different scenarios, and the bid of competitive urban uses for a typical site, which will be discussed in detail in Chapter 4. Hu and Lo (2007) developed a method to simulate urban growth in the Atlanta region in Georgia, where it incorporated 13 counties, using the Landsat images. The land use categories of their research are based on six LULC categories: high density urban, low density urban, bare land, crop or grassland, forest, and water. For the reason that their study area comprised 13 counties, they modeled urban growth based on urban use only, along with some natural land. Th eir remote sensing data covered the period of 1987 and 1997, respectively, for the study area. They proposed twenty independent variables, which include spatial and social data such as income per capita, poverty rate, distance to nearest urban cluster, dis tance to CBD, distance to active economic center, distance to the nearest major road, and so on. Hu and Lo (2007) used the (2007) research is inspirational for t his study in terms of model building, model interpretation, and sensitivity analysis, for which this research will also follow their research sequence and offer the similar contents as what they provided. They will be described in detail in later of this c hapter. Methodologies Currently, there are three data sources available for county based urban research: parcel data, aerial photographic images, and satellite images. Parcel data is a commonly used data The most important advantage of the parcel data is its coverage, which covers an entire county. Unlike some data

PAGE 102

102 sources, e.g., zoning data and future land use data, which are based on political boundaries, such as incorporated cities and towns, the parcel data is based on the entire county, which facilitates urban studies to be conducted without combining data from piecemeal sources. The parcel data can be used directly based on urban research needs, which usually need to be aggregated into urban uses thr ough combinations of various property use codes. As a result, parcel data is able to provide land use information for important urban use types such as single family, multi family, commercial, industrial, institutional, and recreational. It also provides t he information for natural land uses, such as forests, grasslands, agricultural land, and so on. Moreover, the parcel data provides information on vacant land, some of which are categorized based on their usages. For example, parcel data categorizes vacant residential uses, vacant commercial uses, vacant industrial uses, and vacant institutional uses, which can be used for infill development purposes. Because the parcel data is based on property value assessments, some land use types are absent from it, s uch as transportation corridors, residential right of ways, and wetlands. In addition, parcel data reflects land uses rather than land covers. For example, the University of tailed land cover information such as buildings, lawns, parking lots, forests, agricultural land, and so on within the parcel. Moreover, parcel data with its land categories do not reflect their actual uses. Because of the shortcomings with the parcel dat a, it is necessary to find alternative data sources. Currently, there are remote sensing data available for urban studies based on aerial photographic images and satellite images. The challenge of using satellite images to classify urban LULC classe s rests on three factors as introduced in the previous chapter. First, urban LULC classe s usually have similar spectral characteristics, which preclude them from being

PAGE 103

103 distinguished accurately (Paul, 2007). Second, urban LULC classe s are barely homogeneous, in wh ich different types of LULC classe s mix together in a pixel, which makes the pixel hard to differentiate based on its pixel value. Third, satellite sensors may not be sharp enough to detect low density urban development at the urban fringe. These three fac tors make urban LULC classification difficult to classify that are equivalent to the USGS Level III standard (Anderson et al., 1976), in which residential uses such as single family and multi family are differentiated, which is critical to urban growth res earch. Also, these three factors prevent researchers from obtaining accurate ground truth information for urban LULC classe s. Because of this, county level studies, which require urban uses to be differentiated further into USGS Level III, are seldom condu cted. In addition, based on the recent trend, many urban researchers used AI models to simulate urban growth. Because some AI models, e.g., the CA model and the ABM, consider people and their choices in land use decisions, as a result the utilization of other model types, e.g., the logistic regression model, to mimic detailed urban development has b een neglected. In fact, the logistic regression model is a heuristic platform for urban growth modeling to build upon as long as urban LULC classe s are accurat ely classified and important urban use categories are provided. The logistic regression model offers a number of advantages compared to some AI models such as the CA model. First, the logistic regression model considers multiple factors such as demographi c, social economic, ecologic physical, and spatial factors (Hu and Lo, 2007) that are significant to urban land use changes. It can also include governmental land use policies and environmental protection factors into the model (Landis, 1994; Landis 1995). In addition, the logistic regression model takes historical factors into consideration (Landis and Zhang, 1997). Consequently, longitudinal data can be imported into the model to test whether historical

PAGE 104

104 development of land use patterns influence the curre nt land use patterns and whether historical factors can be considered to predict the future land use patterns. Generally speaking, the logistic regression model is capable of revealing the reasons behind urban land use patterns and changes (Hu and Lo, 2007 ). Most importantly, the logistic regression model requires less CPU time to process the data, which is time saving for simulating urban land use development especially in large areas, with respect to large quantities of data processing needs. For the coun ty based study, the logistic regression works well with research needs, which is capable of predicting urban uses and allocating urban uses based on multi use categories. Logistic Regression Model Construction This study adopts the MLR model to simulate urban growth. The MLR model is a model evolved from the binary logistic regression model (IDRISI Help, no date), in which multi variables instead of two variables are included into the model to simulate urban growth based on the probability of a dependent variable it receives. It is mainly applied to predict categorical or dichotomous variables, or continuous variables, which are in probability between 0 and 1, given one or more predictors or explanatory variables. Logistic regression is similar to the Ord inary Least Squares ( OLS ) regression model (Field, 2000). Like the OLS regression model, the logistic regression model reflects the relationships between a dependent variable and one or more independent variables (Hosmer Jr and Lemeshow 1989). It has coefficients for each independent variable and also a constant intercept (Field, 2000). It can incorporate categorical and continuous data for independent variables into the model. The logistic regression model expresses the probability of whether a dependent variable belongs to a certain category or is a member in a logit format (Menard, 1995). A logit is also called the natural logarithm of the odds, which represents the natural logarithm of the probability ratio of the dependent variable that belongs to a member or the probability that does not belong

PAGE 105

105 to a member (Menard, 1995). For the value of the dependent variable, if a value is closer to 1, it indicates that it is more likely for that variable to belong to a member in a study; convers ely, if the value is closer to 0, it indicates that it is more likely that it does not belong to the member in question (Field, 2000). As a result, 0 often refers to not belonging to a member and 1 refers to belonging to the member. The logit can be writte n as: ( 3 9) 3 Based on the above description, the logit can also be written as the lineal form of independent variables and the dependent variable for an MLR model: ( 3 10) 4 Because it is obvious that ; as a result, in the fo rm of exponentiation, Equation 3 10 can be re written as: ( 3 11) 5 And (3 12) 6 And (3 13) 7 Where (3 14) 3 So urce: Menard, 1995 4 Source: Menard, 1995 5 Source: Menard, 1995 6 Source: Menard, 1995 7 Source: Field, 2000

PAGE 106

106 In order to obtain the coefficients as well as the constant intercept, the logistic regression model adopts a maximum likelihood approach to determine the coefficients and the constant, which is different from the OLS regression model, in which the coefficients and the constant are calculated directly. Based on this maximum likelihood approach, the logistic regression model tries tentat ive s olutions and then slightly revises those tentative numbers until the changes, given by different trials, are the closest so that they can be neglected. This is actually an iterative process, which can only be performed by computers (Menard, 1995). Three te chniques can be adopted for the goodness of fit analysis for an MLR model: the Chi square test, the Taylor Russell Diagram analysis, and the ROC curve. The Chi square test is used to analyze nominal or ordinal data to see whether they are statistically sig nificant. In the logistic regression model, Chi square is used to examine whether the observed and the expected are statistically significant. The Taylor Russell Diagram is used to analyze how good the prediction of a model is, given the observation and pr ediction interactions based on four possible outcomes such as true positive, true negative, false positive, and false negative. The ROC curve is used to measure the degrees of agreements between the simulated use and the actual use occurred (Pontius and Sc hneider, 2001). It answers the question as to how well an LULC can change given high suitability scores provided by the model for an area (Hu and Lo, 2007). Statistically, the ROC value is between 0.5 and 1. If a ROC value is 1, it means that high suitabil ity scores in an area result in a perfect agreement for the area to change its LULC classe s. Conversely, if a ROC value is 0.5, it means that high suitability scores are assigned more randomly at locations across an area (Pontius and Schneider, 2001). This chapter will use the Chi square test as well as the ROC curve to conduct sensitivity analyses. The Taylor Russell Diagram is not used because the Taylor Russell Diagram addresses the sensitivity of modeling

PAGE 107

107 results such as true positive, true negative, fa lse positive, and false negative that are redundant with the Chi square test and the ROC curve test. The ROC curve also addresses these values nevertheless with more heuristic statistics not available in the Taylor Russell Diagram. Logistic Regression Resu lts This study uses both the LOGISTICREG and MULTILOGISTICREG modules in IDRISI so as to conduct urban growth simulations for Alachua County. The simulations are preceded based on four urban uses identified in the LULC classification process in Chapter 2: single family, multi family, commercial institutional transportation, and industrial warehouses. Accordingly, four MLR models are established based on the four urban LULC classes identified, with each LULC class having a corresponding MLR model. First, re levant independent variables for each of the four LULC classe s are found. They are input into the model to test whether they are statistically significant to the dependent variable and also whether they yield high accuracy. When an independent variable is found to be statistically significant, it will be input into the MLR model for refining purposes. As a result, non statistically significant independent variables are not included in the refining models. Benchmarks are also set for increasing the overall a ccuracy level of a dependent variable. For residential uses, including the single family use and the multi family use, at least 80 percent of the overall accuracy should be reached. For commercial institutional transportation and industrial warehouses deve lopment, at least 90 percent accuracy should be achieved. As a result, an incremental process is undertaken for each of the four urban LULC classe s in order to find the relevant independent variables that yield the highest overall accuracy. As a result, pr ospective independent variables are added into the model one by one to test whether they increase the general accuracy level. This is an empirical process, which assesses spatial characteristics of the four LULC classe s, such as the proximity to existing r oads and existing residential development,

PAGE 108

108 and so on, rather than social economic factors such as population, income, and poverty. The reason for evaluating spatial characteristics of each dependent variable is that the social economic factors are in fact reflected and implied in the land use development, which do not need to be assessed specifically For example, the increased single family use in the study area reflects the population growth, and the increased commercial institutional transportation use i n the area reflects the economic development of the study area. Therefore, this study does not simulate these background factors in particular. Generally speaking, this study tests fourteen independent variables to simulate the single family use, eight var iables for the multi family use, fourteen for the commercial institutional transportation use, and eight for the industrial warehouses use. For refined models, this study tests five independent variables for the single family use, six independent variables for the multi family use, eight independent variables for the commercial institutional transportation use, and six independent variables for the industrial warehouses use. Second, when probability maps are produced for four land uses of interest (usually square test, Pseudo R sensitivity tests generate positive results such as good agree ment between the predicted and the observed for goodness of fit, the four simulated uses are mosaiced to create a 2003 LULC simulation map, along with a number of natural land such as recreational others, agricultural, wetlands, forests, grasslands, water, and barren. Finally, the four prediction maps will be used to allocate future land uses based on the eleven LULC classe s identified. This step will be illustrated in detail in Chapter 4. In Chapter 4,

PAGE 109

109 four urban uses together with seven natural land are allocated for 2020 and 2030, respectively. This chapter does not cover the allocation process, however. Model Calibration Single Family To predict the single family use, fourteen independent variables are found. They are (1) existing single family use in 1982 based on remote sensing classifications; (2) existing single family use in 1994 based on remote sensing classifications; (3) existing single family use in 2003 based on the parcel data; (4) proximity to industrial warehouses in 1994; (5) proximity to existing 1982 single family use based on remote sensing classifications; (6) proximity to existing 1994 single family use based on remote sensing classifications; (7) proximity to 1982 road network; (8) proximity to 1994 road network; (9) future land use for the single family use; (10) 1994 single family density; (11) proximity to schools; (12) located within urban cluster areas; (13) zoning for the single family use; and (14) vacant land for the single family use in 2003. They are listed in Table 3 1. The dependent variable is the 2003 single family use based on the classified use from the remote sensing data. It is noted that these independent variables are all in thematic maps where 1 and 0 numbers are assigned to represent a cell within a specific targe t area of interest or outside a target area. In addition, zonal statistics are applied to proximity analyses, in which proximity to industrial, proximity to existing single family uses, proximity to roads, and proximity to schools are partitioned by mean v alues of the zonal statistics. To avoid autocorrelation provided by the independent variables, 0.2 percent of stratified sampling is applied. This is because 0.2 percent of random sampling is tested against the simulated single family use for its autocorr elation characteristics, which results in a random distribution for the single score are 0.04 and 0.62 (Table 3 5 and Table 3 6), respectively, obtained from extracting 0.2 percent stratified random po ints in ENVI 4.4 from the 2003 single family prediction map, and they are tested in

PAGE 110

110 ArcGIS 9.3. The outputs from operations in LOGISTICREG and MULTILOGISTICREG in IDRISI include a probability map (Figure 3 1), a residual map, and a statistical sheet, in wh ich a bunch of statistics are calculated such as coefficient, mean, and standard deviation for each independent variable, 2 Log Likelihood, Pseudo R Square value, Chi square value, probability cutting threshold, 2 2 contingency tables for true positive and false positive percentage values, and ROC values based on 100 intervals. When examining the calculated P values of each independent variable (Table 3 1), it is found that five independent variables are statistically significant to the dependent variabl e of the 2003 predicted single family use, given that P values of these independent variables are less than 0.1 for two tailed P values (0.05 for one tailed P values) (Table 3 1). These five independent variables, named as refined variables, are (1) existi ng single family use in 1982, (2) existing single family use in 2003 based on the parcel data, (3) proximity to 1982 roads, (4) future land use for the single family use, and (5) 2003 vacant land for the single family use. The probability map, after refine ment, creates a raw dataset for urban use allocations, which will essentially be introduced in Chapter 4. The independent variables that are statistically significant will be thrown into the model to test the goodness of fit and the overall accuracy. The c utting threshold before refinement is set at 0.6813, in which probability values that are greater than the cutting threshold are identified as the single family use. The actual predicted 2003 single family use employs the cutting threshold following the re fined models. The 2003 single family use prediction map (before refinement) is presented in Figure 3 2. Model Calibration M ulti family The process to produce a multi family prediction map is essentially the same as the single family use. During the proces s, eight independent variables are found. They are (1) existing multi family land use in 1982 based on remote sensing classifications; (2) existing multi family

PAGE 111

111 land use in 1994 based on remote sensing classifications; (3) existing multi family land use in 2003 based on the parcel data; (4) vacant land for the multi family land use in 1994 with ecosystems, parks, and conservations masked; (5) proximity to the existing 1994 multi family land use based on remote sensing classifications; (6) proximity to major roads; (7) future land use for the multi family use; and (8) zoning for the multi family land use. The dependent variable is the 2003 multi family use derived from the 2003 remote sensing urban LULC classification map. They are input into the LOGISTICREG and the MULTILOGISTICREG modules in IDRISI. The eight independent variables are all thematic data, in which 1 and 0 are categorized for the cells within an area of interest or outside the area, respectively. In addition, 1.5 percent stratified sampling is applied because the random sampling based on this percentage has random distribution of the multi family use, with autocorrelation excluded. The output maps are a probability map (Figure 3 3) and a residual map for the multi family use. Specifically, the c utting threshold is 0.0470 before refinement because of the sampling applied. Based on this number, a 2003 multi family prediction map is produced (Figure 3 4). Moreover, the calculated P values indicate that six of the eight independent variables are stat istically significant in terms of the 0.10 two tailed P values (Table 3 2). They are (1) existing multi family land use in 1982 based on remote sensing classifications; (2) existing multi family land use in 1994 based on remote sensing classifications; (3) existing multi family land use in 2003 based on parcel data; (4) proximity to existing 1994 multi family land use based on remote sensing classifications; (5) future land use for the multi family use; and (6) zoning for the multi family land use. This mea ns these six variables contribute significantly to the dependent variable. Model Calibration Commercial Institutional T ransportation There are sixteen independent variables for the commercial institutional transportation use. They are (1) existing commercial institutional transportation use in 2003 based on the parcel

PAGE 112

112 data; (2) existing commercial institutional transportation use in 198 2 based on remote sensing classifications; (3) existing commercial institutional transportation use in 1994 based on remote sensing classifications; (4) vacant land for the commercial institutional transportation use in 2003 with ecosystem, parks, and cons ervations masked; (5) vacant land for the commercial institutional transportation use in 1994 with ecosystem, parks, and conservations masked; (6) 1994 vacant land proximity to 2003 single family and multi family uses; (7) 1994 vacant land proximity to 199 4 single family and multi family uses; (8) proximity to 2003 commercial institutional transportation use; (9) proximity to 1994 commercial institutional transportation use; (10) proximity to 2003 TIGER roads; (11) proximity to major roads; (12) future lan d use for the commercial institutional transportation use; (13) proximity to road intersection; (14) zoning for the commercial institutional transportation use; (15) 2003 vacant land proximity to 2003 single family and multi family uses; and (16) 2003 vaca nt land proximity to the 2003 multi family use. The dependent variable is the 2003 commercial institutional transportation use derived from the remote sensing data. The same as the single family use and the multi family use, the independent variables for t he commercial institutional transportation use are all thematic, in which 1 and 0 are categorized according to the cell locations within or outside a target area of interest. Like the above two uses, stratified random sampling is tested on the commercial i nstitutional transportation use, in which 8.5 percent sampling is applied, which outputs the highest accuracy rate, while it is still randomly distributed. The output files are a probability map, a residual map, and a statistical sheet, which is the same a s the above single family use and the multi family use. The probability map is presented in Figure 3 5. The cutting threshold is 0.1440 before refinement. Because the program becomes over burdened with the sixteen independent variables operated in the IDRI SI software, two independent variables are not

PAGE 113

113 included into the data running. These two variables include proximity to major roads and 2003 vacant land proximity to 2003 single family and multi family uses. Based on this programming running, i.e., fourtee n variables instead of the sixteen, a commercial institutional transportation prediction map is produced, which is presented in Figure 3 6. In particular, eight independent variables are proved to be statistically significant. These variables include (1) e xisting commercial institutional transportation use in 2003 based on the parcel data; (2) existing commercial institutional transportation use in 1982 based on remote sensing classifications; (3) existing commercial institutional transportation use in 1994 based on remote sensing classifications; (4) vacant land for the commercial institutional transportation use in 2003 with ecosystem, parks, and conservations masked; (5) vacant land for the commercial institutional transportation use in 1994 with ecosyste m, parks, and conservations masked; (6) proximity to 2003 commercial institutional transportation use; (7) proximity to 1994 commercial institutional transportation use; and (8) future land use for the commercial institutional transportation use. Their tw o tailed P values are all less than the 0.10 level (Table 3 3). They will be input into the model to test the goodness of fit for model refining purposes. Model Calibration Industrial W arehouses The industrial warehouses dependent variable has eight inde pendent variables. They are (1) 2003 industrial warehouses land use proximity to major roads; (2) existing industrial warehouses land use in 2003 based on the parcel data; (3) existing industrial warehouses land use in 1982 based on remote sensing classifi cations; (4) proximity to existing residential uses; (5) existing industrial warehouses land use in 1994 based on remote sensing classifications; (6) 1994 vacant land proximity to 1994 industrial warehouses land use; (7) future land use for the industrial warehouses land use; and (8) zoning for the industrial warehouses land use. The dependent variable is the 2003 industrial warehouses use given by the remote sensing data. Like

PAGE 114

114 the single family use, the multi family use, and the commercial institutional tr ansportation use, the independent variables for the industrial warehouses use are all thematic, with 1 categorized as located within the specific area of interest and 0 as out of the specific area of interest. In addition, 15 percent stratified random samp ling is adopted because it creates a random distribution for random points located within the simulated 2003 industrial warehouses areas at this level and also generates high accuracy. The productions include an industrial warehouses use probability map (F igure 3 7), a residual map, and a statistical sheet. The cutting threshold is 0.1281 before refinement. In particular, among the eight independent variables, six independent variables are proved to be statistically significant to the dependent variables be cause their P values are all less than the 0.10 level (Table 3 4). These six independent variables are (1) existing industrial warehouses land use in 2003 based on the parcel data; (2) existing industrial warehouses land use in 1982 based on remote sensing classifications; (3) existing industrial warehouses land use in 1994 based on remote sensing classifications; (4) 1994 vacant land proximity to 1994 industrial warehouses land use; (5) future land use for the industrial warehouses land use; and (6) zoning for the industrial warehouses land use. The 2003 predicted industrial warehouses use is illustrated in Figure 3 well as the z score test for each of the above four dependent variables considering the situation of with and without stratified sampling are listed in Tables 3 5 and 3 6, respectively. Refined LULC Logistic Regression Model A refined model is the one based on refined independent variables. Because the above independent variables for each dependent variable have P values representing the correlations with the dependent variables, a refined model takes the independent variables that are statistically significant to the dependent variable. However, compared to the original models, generally speaking, refined models provide relatively the same results as the original models in

PAGE 115

115 terms of the overall accuracy and the goodness of fit based on the four parameters examined: the From Table 3 7, the refined single family model has almost the same values for the overall ROC value and the McFadden va lue, while having higher values for the overall accuracy value and the family use, in Table 3 7, the overall accuracy level is the same are slightly less than the original model. In terms of the commercial institutional transportation use, it presents the same pattern as the multi family use, in which the overall accuracy value is the same as the original model, while the refined model ha s slightly lower values for the overall warehouses use for the refined model has the same value as the original model for the overall accuracy value while it has a slightly higher value for th ROC value and the McFadden value. Overall, the refined model has advantages for the overall accuracy level as the single family has a slightly higher overall accuracy value for the refined model while the other three uses have the same overall accuracy level. These values are explicable because the non statistical significance variables are excluded leaving the statistically significant variables in the model that raises the overall accuracy leve l. Because the difference between the original models and refined models are minor, this study uses the refined models in model building, which will input into the next step for the sensitivity analyses as well as the simulation of 2003 urban uses. Single family The refined single family model has five independent variables, which are (1) existing single family land use in 1982 based on remote sensing classifications; (2) existing single family land use in 2003 based on the parcel data; (3) proximity to 19 82 road; (4) future land use for the

PAGE 116

116 single family land use ; and (5) vacant land for the single family land use in 2003. The cutting threshold for the refined single family model is 0.6871. The overall accuracy for refined single family is 97.99 percent. The refined 2003 predicted single family probability map is shown in Figure 3 9. The refined 2003 single family prediction map is shown in Figure 3 10. The independent variables for refined single family model are shown in Table 3 8. Multi family The refi ned multi family model has six independent variables. They are (1) existing multi family land use in 1982 based on remote sensing classifications; (2) existing multi family land use in 1994 based on remote sensing classifications; (3) existing multi family land use in 2003 based on the parcel data; (4) proximity to existing 1994 multi family land use based on remote sensing classifications; (5) future land use for the multi family land use; and (6) zoning for the multi family land use. The cutting threshold for the refined multi family model is 0.0334. The overall accuracy for refined multi family is 98.83 percent. The refined 2003 predicted multi family probability map is shown in Figure 3 11. The refined 2003 multi family prediction map is shown in Figure 3 12. The independent variables for the refined multi family model are shown in Table 3 9. Commercial institutional transportation The refined commercial institutional transportation model has eight independent variables. They are (1) existing commercial i nstitutional transportation land use in 2003 based on the parcel data; (2) existing commercial institutional transportation land use in 1982 based on remote sensing classifications; (3) existing commercial institutional transportation land use in 1994 base d on remote sensing classifications; (4) vacant land for the commercial institutional transportation land use in 2003 with ecosystem, parks, and conservations masked; (5) vacant land for the commercial institutional transportation land use in 1994 with eco system, parks, and

PAGE 117

117 conservations masked; (6) proximity to 2003 commercial institutional transportation land use; (7) proximity to 1994 commercial institutional transportation land use; and (8) future land use for the commercial institutional transportation land use. The cutting threshold for the refined commercial institutional transportation model is 0.1408. The overall accuracy for the refined multi family is 99.53 percent. The refined 2003 predicted commercial institutional transportation probability map is shown in Figure 3 13. The refined 2003 commercial institutional transportation prediction map is shown in Figure 3 14. The independent variables for the refined commercial institutional transportation model are shown in Table 3 10. Industrial warehouse s The refined industrial warehouses model has six independent variables, which are (1) existing industrial warehouses use in 2003 based on the parcel data; (2) existing industrial warehouses use in 1982 based on remote sensing classifications; (3) existing industrial warehouses use in 1994 based on remote sensing classifications; (4) 1994 vacant land proximity to 1994 industrial warehouses use; (5) future land use for the industrial warehouses use; and (6) zoning for the industrial warehouses use. The cutti ng threshold for the refined industrial warehouses model is 0.1609.The overall accuracy for refined industrial warehouses is 99.71 percent. The refined 2003 predicted industrial warehouses probability map is shown in Figure 3 15. The refined 2003 industria l warehouses prediction map is shown in Figure 3 16. The independent variables for the refined industrial warehouses model are shown in Table 3 11. Sensitivity Analysis The sensitivity analysis is conducted based on the number of the dependent variables si nce the number of dependent variables determines the number of the sensitivity analysis. Because there are four dependent variables in this study, four sensitivity analyses are conducted based on three groups of parameters: the Chi square value, the Pseudo R

PAGE 118

118 V value, and the ROC curve. The Chi square value is from the Chi square test, which is to calculate the difference between the predicted land use and the existing land use based on observed values and expected values. In t he case of this study, the Chi square test is conducted for the predicted 2003 land use changes and the observed 2003 land use changes ba sed on the following algorithm: (3 15) 8 Where is the observed value and is the expected value. can be written as: (3 16) 9 The goodness of fit Pseudo R Square values includes the McFadden value (McFadd en, 1973), which is written as: (3 17) 10 And: (3 18) 11 Where refers to the observed value at location ; refers to the predicted value at location And: (3 19) 12 Where refers to the number of cases, for which or 8 Source: Sirkin, 1999 9 Source: Sirkin, 1999 10 Source: McFadden, 1973 and Ainsworth, no date 11 Source: Ainsworth, no date 12 Source: Ainsworth, no date

PAGE 119

119 The optimal goodness of fit McFadden value is between 0.2 and 0.4 (Ainsworth, no date and Hu and Lo, 2007); the higher the value, the better the goodness of fit for a simulation model to fit the actual use. squ are based test for nominal based correlation analysis (Garson, 2009). It has a value between 1 and 1, in which 1 and 1 refer to a perfect relationship between the independent variables and the dependent variable and 0 indicates no relations (Seaman, 2001 (3 20) 13 w here is the Chi square value; is the number of sample size; and is the smaller number of (rows 1) or (columns 1) ( Cramr, 1999; Garson, 2009 ). The ROC curve is to test the validation of a model of how actual changes interact with the predicted changes in a quantifiable manner through different scenarios. It is based on a 2 2 contingency table for an observed use (or actual change) versus a predicted use (or predicted change) in light of a depen dent variable (Table 3 12). It calculates the true positive rate, A / (A+C), as well as false positive rate, B / (B+D), for each scenario and adds them up cumulatively to generate an overall ROC value, a single index, for a dependent variable (Pontius and Schneider, 2001). The optimal goodness of fit ROC value is from 0.5 to 1, with 1 indicating the perfect match between the predicted use and the actual use and 0.5 indicating a random distribution of the suitability and/or probability values across the land scape (Pontius and Schneider, 2001). The scenarios are a slice of the overall suitability or probability map into several equal interval groups, in which the highest probability group is usually assigned as group 1 and so on for the rest until the smallest probability is assigned. These scenarios in different 13 Source: Cramr, 1999 and Garson, 2009

PAGE 120

120 groups show a ROC curve to be calculated according to a typical dependent variable in light of represent t wo land categories, observed versus predicted, simulation of one land class will have one contingency table. As a result, the validation of more than one set of land types, e.g., four dependent variables in this study, needs four contingency tables and acc ordingly four ROC curves (Pontius and Schneider, 2001). In addition, the larger the number of groups, the higher the accuracy of the ROC curves (Pontius and Schneider, 2001). The ROC curve is generated in IDRISI. This study slices the probability map for e ach dependent variable into 20 groups. The 2 2 contingency table is presented in Table 3 12 and the overall ROC value is calculated as: (3 21) 14 w here is the rate of false positives for scenario i and is the rate of true positives for scenario i The symbol n is the number of suitability groups. Single family Chi Square Test Based on the predicted 2003 single family use, compared to the existing 1994 single family use, a prediction change map is produced. Likewise, compared to the existing 2003 single family use and the 1994 single family use, an observed change map is created. The changes versus non changes in both the predicted and observed maps are summarized in Table 3 13. Subsequently, the above values are transformed into a contingency table for expected values (Table 3 14) based on Equation 3 7 above. The calculation of expected values for the non change category of single family is: 14 Source: Pontius and Schneider, 2001

PAGE 121

121 As a result, the expected value is presented in the contingency table as Table 3 14 illustrates At this time, a null hypothesis is assumed: : There is no relationship between the Observed and the Predicted. The Chi square value is tested to see if it can reject the null hypothesis. Therefore, the Chi square value is calculated based on the observ ed table (Table 3 1 3 ) and the expected table (Table 3 1 4 ). 2,611,776 2,498,484 113,292 1.2935 E 10 5,137 30,276 143,568 113,292 1.2935E10 89,401 25,700 138,992 113,292 1.2935E10 92,344 121,279 7,987 113,292 1.2935E10 1,606,996 Because Table 3 1 3 and Table 3 1 4 are 2 2 tables, the (degrees of freedom) value is 1. As a result, the critical value for 1 is: 0.05 = 3.84 < 1,793,878 rejec t 0.01 = 6.64 < 1,793,878 0.001 = 10.83 < 1,793,878

PAGE 122

122 0.00000001 = 32.84 < 1,793,878 so P < 0.00000001 = 0.0000 From the above, it is obvious that the calculated Chi square value is greater than the critical value at significance level of 0.05, which is 3.84; therefore, the null hypothesis is rejected. Also, there is a strong correlation between the predicted and ob served values because the P value is close to 0. Thus, the results of the MLR model are reliable and can be used to predict the future single family development for the county. Pseudo R The McFadden Pseudo R Square value as we ll 15. The McFadden is 0.8204, which shows a perfect goodness of fit for the entire single family use model to represent the actual use. The 2 table (Sirkin, 1999), showing a strong correlation between the dependent variable and the independent variables. The overall accuracy value is 97.99 percent, which is presented in Table 3 16. ROC Curve The ROC curve is generated based on 0.2 percent stratified random sampling with 20 equal interval thresholds. The number of cells that are selected for the ROC analysis is illustrated in the contingency table as Table 3 17 illustrates. The ROC table generated in IDRISI for the ROC analysis is further pr ojected onto a chart in EXCEL, which is presented in Figure 3 17. The overall ROC value for 20 thresholds is 0.979, showing a fairly good agreement between the predicted single family use based on five refined independent variables and the actual single fa mily use in the county. Multi family Chi Square Test Similar to the single family use, the Chi square test for the multi family use is based on the predicted 2003 multi family use and the existing 1994 multi family use. As a result, a prediction change m ap is produced. Likewise, compared to the existing 2003 multi

PAGE 123

123 family use and the 1994 multi family use, an observed change map is also generated. The predicted and observed maps for the changes versus non changes are summarized in Table 3 18. Based on Equa tion 3 7 an expected value for the non change category is calculated as follows: The above expected values can be listed in the contingency table as Table 3 19 illustrated. At this stage, a null hypothesis is assumed: : There is no relationship between the Observed and the Predicted. The Chi square value is calculated based on the above observed table as well as the expected table as follows to see if it can reject the null hypothesis: 2,748,811 2,741,374 7,437 55,308,969 20.18 2,030 9,467 7,437 55,308,969 5842.29 30,622 38,059 7,437 55,308,969 1453.24 7,568 131 7,437 55,308,969 422,205.87 Because Table 3 1 8 and Table 3 1 9 are 2 2 tables, the (degrees of freedom) value is 1. As a result, the critical value for 1 is: 0.05 = 3.84 < 429,522 reject

PAGE 124

124 0.01 = 6.64 < 429,522 0.001 = 10.83 < 429,522 0.00000001 = 32.84 < 429,522 so P < 0.00000001 = 0.0000 Based on the above, it is obvious that the calculated Chi square value is greater than the critical value at significance level of 0.05, which is 3.84, a nd the null hypothesis is rejected. In addition, there is a strong correlation between the predicted and observed values because the calculated P value is close to 0. As a result, the logistic regression model is dependable to predict the future multi fami ly use in the county. Pseudo R The IDRISI MULTILOGISTICREG module provides the McFadden Pseudo R McFadden is 0.3260, which is perfect goodness of fit for the multi fam ily model to simulate the shows a relatively strong correlation between the dependent variable and the independent variables. The Pseudo R V values are presented in Table 3 20. The overall accuracy value for the multi family use is 98.83 percent, which is presented in Table 3 21. ROC Curve The ROC analysis is conducted based on 1.5 percent stratified random sampling with 20 equal interval t hresholds. The number of cells that are selected for the ROC analysis is presented in Table 3 22. The ROC table generated in IDRISI for the ROC analysis is projected onto a chart in EXCEL, presented in Figure 3 18. The overall ROC value for 20 thresholds i s 0.588, which shows a not very strong goodness of fit between the predicted values and the actual values for the multi family use. Because the ROC is more than 0.5, the goodness of fit for the multi family use is still acceptable.

PAGE 125

125 Commercial institutional transportation Chi Square Test Similar to the single family use and the multi family use, the Chi square test for the commercial institutional transportation use is based on the predicted 2003 commercial institutional transportation use and the existing 1994 commercial institutional tra nsportation use. Consequently, a prediction change map is produced. Similarly, compared to the existing 2003 commercial institutional transportation use and the 1994 commercial institutional transportation use, an observed change map is also prepared. The predicted and observed maps for the changes versus non changes are summarized in Table 3 23. Similarly, an expected value for the non change category is calculated based on Equation 3 7 as follows: The above expected values can be listed in a conting ency table (Table 3 2 4 ) for expected values as follows: At this point, the null hypothesis is assumed as: : There is no relationship between t he Observed and the P redicted. The Chi square value is calculated based on the above observed matrix as well a s the expected matrix to test if it can reject the null hypothesis: 1,302,828 1,307,892 5,064 25,644,096 19.6072 21,375 16,311 5,064 25,644,096 1,572.1964 1,451,848 1,446,784 5,064 25,644,096 17.7249

PAGE 126

126 12,980 18,044 5,064 25,644,096 1421.1980 The (degrees of freedom) value is 1 because Table 3 2 3 and Table 3 2 4 are 2 2 tables. In this case, the critical value for 1 is: 0.05 = 3.84 < 14,390,821 reject 0.01 = 6.64 < 14,390,821 0.001 = 10.83 < 14,390,821 0.00000001 = 32.84 < 14,390,821 so P < 0.00000001 = 0.0000 Based on the above, it is obvious that the null hypothesis is rejected because the calculated Chi square value is greater than the critical value at the significance level of 0.05. Also, there is a strong correlation between the predicted and observed valu es because the P value is close to 0. Thus, the logistic regression model is dependable to predict the commercial institutional transportation use in the study area. Pseudo R The IDRISI MULTILOGISTICREG module provides the McF adden Pseudo R is calculated as 0.8014, which is perfect goodness of fit for the predicted commercial institutional s hows a strong correlation between the dependent variable and the independent variables. These values are presented in Table 3 25. The overall accuracy value for the commercial institutional transportation use is 99.53 percent, which is presented in Table 3 26.

PAGE 127

127 ROC Curve Similar to the above single family use and the multi family use, the ROC analysis for the commercial institutional transportation use is preceded based on 8.5 percent stratified random sampling with 20 equal interval thresholds. The number of cells that are selected for the ROC analysis is presented in Table 3 27. The ROC table generated in IDRISI for the ROC analysis is projected in Figure 3 19. The overall ROC value for 20 thresholds is 0.853, a fairly good match between the predicted use based on eight refined independent variables and the actual commercial institutional transportation use. Industrial warehouses Chi Square Test The Chi square test for the industrial warehouses use is based on the predicted 2003 industrial use and the existing 1994 industrial use. In this case, a prediction change map is produced. Also, compared to the existing 2003 industrial warehouses and the 1994 industrial warehouses use, an observed change map is created. The predicted and observed maps for the ch anges versus non changes are summarized in Table 3 28. An expected value for the non change category is calculated based on Equation 3 7 as follows: The expected values are input into the contingency table (Table 3 2 9). The null hypothesis is as sumed at this point: : There is no relationship between the observed and the predicted. The Chi square value is calculated to test if it can reject the null hypothesis based on the above observed table as well as the expected table as follows:

PAGE 128

128 2,777,009 2,773,010 3,999 15,992,001 5.7670 1,603 5,602 3,999 15,992,001 2,855 6,399 10,398 3,999 15,992,001 1,538 4,020 21 3,999 15,992,001 761,524 Because Tables 3 2 8 and 3 2 9 are 2 2 tables, the (degrees of freedom) value is 1. The critical value for 1 is: 0.05 = 3.84 < 764,385 reject 0.01 = 6.64 < 764,385 0.001 = 10.83 < 764,385 0.00000001 = 32.84 < 764,385 so P < 0.00000001 = 0.0000 It is evident that the null hypothesis can be rejected because the Chi square value is greater than the critical value at the significance level of 0.05. In addition, there is a st rong correlation between the predicted and observed values because the P is close to 0. Therefore, the logistic regression model is dependable to predict the industrial warehouses use in the study area. Pseudo R The IDRISI MULTILOGISTICREG module provides the McFadden Pseudo R McFadden is 0.6259, which is perfect goodness of fit between the simulated industrial use and 0.6488, which is close to 0.71 for a 2 2 table

PAGE 129

129 (Sirkin, 1999) and shows a strong correlation between the dependent variable and the independent variables. The above values are presented in Table 3 30. The overall accuracy value for the industrial warehou ses use is 99.71 percent, which is presented in Table 3 31. ROC Curve Similar to the above three uses, the ROC analysis for the industrial warehouses use is preceded based on 15 percent stratified random sampling with 20 equal interval thresholds. The nu mber of cells that are selected for ROC analysis is presented in Table 3 32. The ROC table generated in IDRISI for the ROC analysis is illustrated in Figure 3 20. The overall ROC value for 20 thresholds is 0.821, a fairly good match between the predicted i ndustrial warehouses use based on refined six independent variables and the actual industrial warehouses use. 2003 Urban LULC Simulation Because of the validities of the above models to simulate the four dependent variables, a final 2003 urban LULC predi ction map is produced. The final 2003 prediction map essentially mosaics all the above four predicted urban land uses based on the cutting thresholds for each dependent variable provided by the refined models, along with mosaicing the remaining natural lan d. However, because of the data transfer from the IDRISI software, which causes the four dependent maps to be shifted in ArcGIS 9.3, i.e., those four predicted maps create gaps in IDRISI compared to their equivalent maps in ArcGIS 9.3, a geo reference proc ess is employed, which is operated in the ERDAS Imagine 9.1 software in order to shift the IDRISI maps back to the ArcGIS 9.3 geo referencing. Then, these geo referenced maps are mosaiced to create a final 2003 LULC prediction map, which is illustrated in Figure 3 21. This study also conducts overall accuracy assessment for the simulated 2003 classification map. The overall accuracy level reaches 97.30 percent. The statistics regarding the confusion matrix of the 2003 simulation map is presented in Tables 3 33 through 3 37.

PAGE 130

130 Table 3 1. Independent variables for the single family use Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Existing SF in 1982 based on remote sensing classifications Thematic data 1: SF; 0: Non SF 0.869271 2.385172 0.187856 4.627328 <0.0001*** Existing SF in 1994 based on remote sensing classifications Thematic data 1: SF; 0: Non SF 0.13758 1.147494 0.253376 0.542988 0.5871 Existing SF in 2003 based on parcel data Thematic data 1: SF; 0: Non SF 8.463601 4739.091 0.29786 28.41469 <0.0001*** Proximity to industrial warehouses in 1994 Thematic data 1: Out of 1,000 meter radius; 0: Within 1,000 meter radius* 0.091319 0.912727 0.500023 0.18263 0.8551 Proximity to existing 1982 SF based on remote sensing classifications Thematic data 1: Within 150 meter radius; 0: Out of 150 meter radius** 0.089720 1.093868 0.4031 0.222574 0.8239

PAGE 131

131 Table 3 1. Continued Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Proximity to existing 199 4 SF based on remote sensing classifications Thematic Data 1: Within 50 Meter Radius; 0: Out Of 50 Meter Radius* 0.522662 1.686511 0.342529 1.525891 0.127 Proximity to 1982 roads Thematic Data 1: Within 530 Meter Radius; 0: Out Of 530 Meter Radius** 0.859717 2.362492 0.4885 1.759912 0.0784*** Proximity to 1994 road Thematic Data 1: Within 520 Meter Radius; 0: Out Of 520 Meter Radius** 0.743280 0.475552 0.488734 1.52083 0.1283 Future land use for SF Thematic Data 1: Future SF Use; 0: Future Non SF Use 1.638371 5.146778 0.295178 5.550451 <0.0001*** 1994 single family density Thematic Data 1: Areas Surrounded By High Density SF Houses; 0: Other Areas 0.329253 0.719461 0.473763 0.69498 0.4871 Proximity to schools Thematic Data 1: Areas Within 5,000 Meter Radius; 0: Areas Out Of 5,000 Meter Radius** 0.264259 1.302465 0.482561 0.547617 0.584

PAGE 132

132 Table 3 1. Continued Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Located within urban cluster are as Thematic Data 1: Within Urban Cluster Areas; 0: Out Of Urban Cluster Areas 0.595032 0.551545 0.440117 1.35199 0.1764 Zoning for SF Thematic Data 1: Zoned For SF ; 0: Zoned For Non SF 0.02532 0.975002 0.31321 0.08083 0.9356 Vacant land for SF in 200 3 Thematic Data 1: Vacant Land In 2003; 2: Occupied Land In 2003 2.191825 8.951535 0.258796 8.469316 <0.0001*** Note: *: Based on zonal statistical mean value **: Based on arbitrary partition *** Less than 0.10 for two tailed P values, which is statistical significance for the independent variable

PAGE 133

133 Table 3 2. Independent variables for the multi family use Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Existing MF in 1982 based on remote sensing classifications Thematic Data 1: MF ; 0: Non MF 2.280251 9.779132 0.071839 31.741126 <0.0001*** Existing MF in 1994 based on remote sensing classifications Thematic Data 1: MF ; 0: Non MF 1.851125 6.366975 0.071839 25.767683 <0.0001*** Existing MF in 2003 based on parcel data Thematic Data 1: MF ; 0: Non MF 7.202940 1343.374633 0.064695 111.336890 < 0.0001*** Vacant land for MF in 1994 with ecosystems, parks, and conservations masked Thematic Data 1: Vacant For MF ; 0: Occupied 0.360957 1.434702 0.498237 0.724468 0.4688 Proximity to existing 1994 MF based on remote sensing classifications Thematic Data 1: Within 240 Meter Radius; 0: Out Of 240 Meter Radius* 1.676713 5.347947 0.37419 4.480913 0.0101

PAGE 134

134 Table 3 2. Continued Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Proximity to major roads Thematic Data 1: Within 500 Meter Radius; 0: Out Of 500 Meter Radius** 0.671131 1.956448 0.485658 1.381900 0.167 Future land use for MF Thematic Data 1: Future MF Use; 0: Future Non MF Use 0.340167 1.405182 0.074757 4.550300 <0.0001*** Zoning for MF Thematic Data 1: Zoned For MF ; 0: Zoned For Non MF 1.089908 2.973400 0.073797 14.769000 <0.0001*** Note: *: Based on zonal statistical mean value **: Based on arbitrary partition *** Less than 0.10 for two tailed P values, which is statistical significance for the independent variable.

PAGE 135

135 Table 3 3. Independent variables for the commercial institutional transportation use Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Existing COM INST TRANS in 2003 based on the parcel data Thematic Data 1: COM INST TRANS ; 0: Non COM INST TRANS 5.005301 149.202003 0.134753 37.144265 <0.0001**** Existing COM INST TRANS in 1982 based on remote sensing classifications Thematic Data 1: COM INST TRANS ; 0: Non COM INST TRANS 1.403852 4.070849 0.116184 12.083003 <0.0001**** Existing COM INST TRANS in 1994 based on remote sensing classifications Thematic Data 1: COM INST TRANS ; 0: Non COM INST TRANS 3.146571 23.256183 0.130084 24.188763 <0.0001**** Vacant land for COM INST TRANS in 2003 with ecosystem, parks, and conservations masked Thematic Data 1: Vacant For COM INST TRANS ; 0: Occupied 25.076209 1.28689E 11 0.498138 50.339883 <0.0001****

PAGE 136

136 Table 3 3. Continued Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Vacant land for COM INST TRANS in 1994 with ecosystem, parks, and conservations masked Thematic Data 1: Vacant For COM INST TRANS ; 0: Occupied 1.853776 6.383880 0.496737 3.731907 0.0002**** 1994 vacant land proximity to 2003 SF and MF uses Thematic Data 1: Within 250 Meter Radius; 0: Out Of 250 Meter Radius* 1.53E 01 1.164872 4.95E 01 0.308204 0.7579 1994 vacant land proximity to 1994 SF and MF uses Thematic Data 1: Within 300 Meter Radius; 0: Out Of 300 Meter Radius* 0.265254 0.767011 0.496888 0.533831 0.5935 Proximity to 2003 COM INST TRANS Thematic Data 1: Within 630 Meter Radius; 0: Out Of 630 Meter Radius* 20.095309 533680983.9 0.498258 40.331131 <0.0001**** Proximity to 1994 COM INST TRANS Thematic Data 1: Within 550 Meter Radius; 0: Out Of 550 Meter Radius* 1.240891 3.458692 0.499053 2.486491 0.0129**** Proximity to 2003 TIGER roads Thematic Data 1: Within 300 Meter Radius; 0: Out Of 300 Meter Radius** 0.354306 1.425191 0.470957 0.752311 0.4519

PAGE 137

137 Table 3 3. Continued Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Proximity to major roads Thematic Data 1: Within 1,000 Meter Radius; 0: Out Of 1,000 Meter Radius** N/A N/A N/A N/A N/A Future land use for COM INST TRANS Thematic Data 1: Future COM INST TRANS Use; 0: Future Non COM INST TRANS Use 0.429676 1.536760 0.192272 2.234730 0.0254**** Proximity to road intersection Thematic Data 1: Within2 ,000 Meter Radius; 0: Out Of 2,000 Meter Radius*** 0.046365 1.047457 0.499109 0.09290 0.926 Zoning for COM INST TRANS Thematic Data 1: Zoned For COM INST TRANS ; 0: Zoned For Non COM INST TRANS 0.055038 0.946450 0.4603 0.147405 0.7089 2003 vacant land proximity to 2003 SF and MF uses Thematic Data 1: Within 250 Meter Radius; 0: Out Of 250 Meter Radius* N/A N/A N/A N/A N/A 2003 vacant land proximity to 2003 MF uses Thematic Data 1: Within 400 Meter Radius; 0: Out Of 400 Meter Radius* 0.184621 1.202763 0.489345 0.377282 0.706 Note: *: Based on zonal statistical mean value. **: Based on arbitrary partition. ***: Average value. ****: Less than 0.10 for two tailed P values, which is statistical significance for the independent variable.

PAGE 138

138 Table 3 4 Independent variables for the industrial warehouses use Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) 2003 industrial proximity to major roads Thematic Data 1: Within 400 Meter Radius; 0: Out Of 400 Meter Radius* 0.131403 1.140423 0.467009 0.281372 0.7784 Existing industrial in 2003 based on parcel Thematic Data 1: Industrial ; 0: Non Industrial 4.010653 55.182906 0.0194 206.734703 <0.0001**** Existing industrial in 1982 based on remote sensing classification Thematic Data 1: Industrial ; 0: Non Industrial 2.398089 11.002126 0.053974 44.430439 <0.0001**** Proximity to existing residential uses Thematic Data 1: Out Of 160 Meter Radius; 0: Within 160 Meter Radius* 0.481350 0.617948 0.493858 0.974674 0.3297 Existing industrial in 1994 based on remote sensing classification Thematic Data 1: Industrial ; 0: Non Industrial 5.16036964 174.228846 0.069397 74.360126 <0.0001****

PAGE 139

139 Table 3 4. Continued Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) 1994 vacant land proximity to 1994 industrial Thematic Data 1: Within 1,050 Meter Radius; 0: Out Of 1,050 Meter Radius* 2.719848 15.178008 0.454024 5.990537 <0.0001**** Future land use for industrial Thematic Data 1: Industrial ; 0: Non Industrial 1.336143 3.804342 0.123114 10.852894 <0.0001**** Zoning for industrial Thematic Data 1: Industrial ; 0: Non Industrial 2.984637 19.779312 0.111794 26.697645 <0.0001**** Note: *: Based on zonal statistical mean value. ****: Less than 0.10 for two tailed P values, which is statistical significance for the independent variable. Table 3 5. Autocorrelation and four dependent variables for the test Dependent Variables With Stratified Sampling Without Stratified Sampling Percent Of Sampling Value Character Value Character 2003 SF 0.2% 0.04 Random 0.8064 Clustered 2003 MF 1.5% 0.06 Random 0.7937 Clustered 2003 COM INST TRANS 8.5% 0.05 Random 0.6498 Clustered 2003 INDUS WARE 15% 0.04 Random 0.7617 Clustered

PAGE 140

140 Table 3 6. Autocorrelation and four dependent variables for the Z score test Dependent Variables With Stratified Sampling Without Stratified Sampling ( Normality Assumption ) 2003 SF 0.62 2,241.4285 2003 MF 0.3 2,206.2896 2003 COM INST TRANS 0.85 1,806.1605 2003 INDUS WARE 0.69 2,117.3811 Table 3 7. Comparison between the original models and the refined models Parameters SF MF COM INST TRANS INDUS WARE Original Model Refined Model Original Model Refined Model Original Model Refined Model Original Model Refined Model Overall ROC* 0.9922 0.9913 0.8420 0.8140 0.9933 0.9895 0.9628 0.9505 Overall A ccuracy 97.95% 97.99% 98.83% 98.83% 99.53% 99.53% 99.71% 99.71% McFadden 0.8243 0.8204 0.3344 0.3260 0.8018 0.8014 0.6284 0.6259 Cramer's V 0.8761 0.8804 0.5190 0.5188 0.8403 0.8395 0.6465 0.6488 *: Based on 100 intervals.

PAGE 141

141 Table 3 8 Independent variables for refined single family use Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Existing SF in 1982 based on remote sensing classifications Thematic Data 1: SF ; 0: Non SF 1.168234 3.216307 0.187856 6.218773 <0.0001*** Existing SF in 2003 based on the parcel data Thematic Data 1: SF ; 0: Non SF 8.709365 6,059.393491 0.297860 29.239794 <0.0001*** Proximity to 1982 road Thematic Data 1: Within 530 Meter Radius; 0: Out Of 530 Meter Radius** 0.158141 1.171332 0.488500 0.323728 0.7461 Future land use for SF Thematic Data 1: Future SF Use; 0: Future Non SF Use 1.161380 3.194339 0.295178 3.934508 <0.0001*** Vacant land for SF in 2003 Thematic Data 1: Vacant Land In 2003; 2: Occupied Land In 2003 2.293864 9.913165 0.258796 8.863598 <0.0001*** Note: **: Based on zonal statistical mean value *** Less than 0.10 for two tailed P values, which is statistical significanc e for the independent variable.

PAGE 142

142 Table 3 9 Independent variables for refined multi family use Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Existing MF in 1982 based on remote sensing classifications Thematic Data 1: MF ; 0: Non MF 2.369196 10.688797 0.071839 32.979248 <0.0001*** Existing MF in 1994 based on remote sensing c lassifications Thematic Data 1: MF ; 0: Non MF 1.700038 5.474156 0.071839 23.664557 <0.0001*** Existing MF in 2003 based on the parcel data Thematic Data 1: MF ; 0: Non MF 7.414149 1,659.296689 0.064695 114.601578 <0.0001*** Proximity to existing 1994 MF based on remote sensing classifications Thematic Data 1: Within 240 Meter Radius; 0: Out Of 240 Meter Radius* 1.896014 6.659295 0.374190 5.066981 <0.0001*** Future land use for MF Thematic Data 1: Future MF Use; 0: Future Non MF Use 0.382844 1.466449 0.074757 5.121180 <0.0001*** Zoning for MF Thematic Data 1: Zoned For MF ; 0: Zoned For Non MF 1.249038 3.486986 0.073797 16.925321 <0.0001*** Note: *: Based on zonal statistical mean value *** Less than 0.10 for two tailed P values, which is statistical significance for the independent variable

PAGE 143

1 43 Table 3 10 Independent variables for refined commercial institutional transportation use Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Existing COM INST TRANS in 2003 based on the parcel data Thematic Data 1: COM INST TRANS ; 0: Non COM INST TRANS 5.007813 149.577323 0.134753 37.162909 <0.0001**** Existing COM INST TRANS in 1982 based on remote sensing classifications Thematic Data 1: COM INST TRANS ; 0: Non COM INST TRANS 1.429635 4.177176 0.116184 12.304925 <0.0001**** Existing COM INST TRANS in 1994 based on remote sensing classifications Thematic Data 1: COM INST TRANS ; 0: Non COM INST TRANS 3.188094 24.242178 0.130084 24.507964 < 0.0001**** Vacant land for COM INST TRANS in 2003 with ecosystem, parks, and conservations masked Thematic Data 1: Vacant For COM INST TRANS ; 0: Occupied 25.137195 0.000000 0.498138 50.462313 <0.0001****

PAGE 144

144 Table 3 10. Continued Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Vacant land for COM INST TRANS in 1994 with ecosystem, parks, and conservations masked Thematic Data 1: Vacant For COM INST TRANS ; 0: Occupied 1.874564 6.517974 0.496737 3.773755 0.0002**** Proximity to 2003 COM INST TRANS Thematic Data 1: Within 630 Meter Radius; 0: Out Of 630 Meter Radius* 20.159299 568,947,659 0.498258 40.459559 <0.0001**** Proximity to 1994 COM INST TRANS Thematic Data 1: Within 550 Meter Radius; 0: Out Of 550 Meter Radius* 1.308222 3.699589 0.499053 2.621408 0.0088**** Future land use for COM INST TRANS Thematic Data 1: Future COM INST TRANS Use; 0: Future Non COM INST TRANS Use 0.419155 1.520676 0.192272 2.180010 0.0293**** Note: *: Based on zonal statistical mean value. ****: Less than 0.10 for two tailed P values, which is statistical significance for the independent variable

PAGE 145

145 Table 3 1 1 Independent variables for refined industrial warehouses use Independent Variables Data Type Note Coefficient Odds Ratio Standard Deviation Z Score P Value (Two Tailed) Existing industrial warehouses in 2003 based on the parcel data Thematic Data 1: Industrial ; 0: Non Industrial 4.209021 67.290613 0.019400 216.959832 <0.0001**** Existing industrial warehouses in 1982 based on remote sensing classifications Thematic Data 1: Industrial ; 0: Non Industrial 2.408243 11.114410 0.053974 44.618566 <0.0001**** Existing industrial warehouses in 1994 based on remote sensing classifications Thematic Data 1: Industrial ; 0: Non Industrial 5.122834 167.810220 0.069397 73.819239 <0.0001**** 1994 vacant land proximity to 1994 industrial warehouses Thematic Data 1: Within 1,050 Meter Radius; 0: Out Of 1,050 Meter Radius* 2.811813 16.640055 0.454024 6.193093 <0.0001**** Future land use for industrial warehouses Thematic Data 1: Industrial ; 0: Non Industrial 1.304183 3.684679 0.123114 10.593299 <0.0001**** Zoning for industrial warehouses Thematic Data 1: Industrial ; 0: Non Industrial 3.106883 22.351276 0.111794 27.791146 <0.0001**** Note: *: Based on zonal statistical mean value. ****: Less than 0.10 for two tailed P values, which is statistical significance for the independent variable.

PAGE 146

146 Table 3 1 2 2 2 contingency table for the ROC curve Model Reality Urban Land Class (1) Non Urban Land Class (0) Total Urban Land Class (1) A B A+B Non Urban Land Class (0) C D C+D Total A+C B+D A+B+C+D Table 3 13. Contingency table of change versus non change for observed values for the single family use (unit: cells) Observed Prediction Non Change Change Marginal Non Change 2,611,776 30,276 2,642,052 Change 25,700 121,279 146,979 Marginal 2,637,476 151,555 2,789,031 Table 3 1 4 Contingency table of change versus non change for expected values for the single family use (unit: cells) Expected Prediction Non Change Change Marginal Non Change 2,498,484 143,568 2,642,052 Change 138,992 7,987 146,979 Marginal 2,637,476 151,555 2,789,031 Table 3 1 5 Pseudo R for the single family use Goodness of Fit ( Parameters ) Results McFadden 0.8204 P Level 0.0000 Cramer's V 0.8804 Table 3 1 6. Overall accuracy for the single family use Observed Prediction Cat 0 (non SF) Cat 1 (SF) % Correct Cat 0 (non SF) 2,516,623 53,545 97.92% Cat 1 (SF) 2,431 216,432 98.89% Overall % 90.32% 9.68 % 97.99%

PAGE 147

147 Table 3 1 7 ROC analysis and 2 2 contingency table based on 0.2 percent stratified sampling for the single family use Simulated b y Threshold Reality (Reference Image) 1 ( SF ) 0 (Non SF ) 1 (Sf) A (number of cells) B (number of cells) 0 (Non SF ) C D For t he Given Reference Image: A+C= 454 B+D= 7,467 Table 3 1 8 Contingency table of change versus non change for observed values for the multi family use (unit: cells) Observed Prediction Non Change Change Marginal Non Change 2,748,811 2,030 2,750,841 Change 30,622 7,568 38,190 Marginal 2,779,433 9,598 2,789,031 Table 3 1 9 Contingency table of change versus non change for expected values for the multi family use (unit: cells) Expected Prediction Non Change Change Marginal Non Change 2,741,374 9,467 2,750,841 Change 38,059 131 38,190 Marginal 2,779,433 9,598 2,789,031 Table 3 20 Pseudo R for the multi family use Goodness of Fit ( Parameters ) Results McFadden 0.3260 P Level 0.0000 Cramer's V 0.5188 Table 3 2 1 Overall accuracy for the multi family use Observed Prediction Cat 0 (Non SF ) Cat 1 ( SF ) % Correct Cat 0 (Non SF ) 2,743,136 1,107 99.96% Cat 1 ( SF ) 31,545 13,243 29.57% Overall % 99.49% 0.51 % 98.83%

PAGE 148

148 Table 3 2 2 ROC analysis and 2 2 contingency table based on 1.5 percent stratified sampling for the multi family use Simulated b y Threshold Reality (Reference Image) 1 ( MF ) 0 (Non MF ) 1 ( MF ) A (number o f cells) B (number of cells) 0 (Non MF ) C D For t he Given Reference Image: A + C = 425 B + D = 38,187 Table 3 2 3 Contingency table of change versus non change for observed values for the commercial institutional transportation use (unit: cells) Observed Prediction Non Change Change Marginal Non Change 1,302,828 21,375 1,324,203 Change 1,451,848 12,980 1,464,828 Marginal 2,754,676 34,355 2,789,031 Table 3 2 4 Contingency table of change versus non change for expected values for the commercial institutional transportation use (unit: cells) Expected Prediction Non Change Change Marginal Non Change 1,307,892 16,311 1,324,203 Change 1,446,784 18,044 1,464,828 Marginal 2,754,676 34,355 2,789,031 Table 3 2 5 Pseudo R for the commercial institutional transportation use Goodness of Fit (P arameters) Results McFadden 0.8014 P Level 0.0000 Cramer's V 0.8395 Table 3 2 6 Overall accuracy for the commercial institutional transportation use Observed Prediction Cat 0 (Non COM ) Cat 1 ( COM ) % Correct Cat 0 (Non COM ) 2,741,496 2,823 99.90% Cat 1 ( COM ) 10,420 34,292 76.70% Overall % 98.67% 1.33 % 99.53%

PAGE 149

149 Table 3 2 7 ROC analysis and 2 2 contingency table based on 8.5 percent stratified sampling for the commercial institutional transportation use Simulated By Threshold Reality (Reference Image) 1 ( COM ) 0 (Non COM ) 1 ( COM ) A (number of cells) B (number of cells) 0 (Non COM ) C D For th e Given Reference Image: A + C =3,603 B + D =305,688 Table 3 2 8 Contingency table of change versus non change for observed values for the industrial warehouses use (unit: cells) Observed Prediction Non Change Change Marginal Non Change 2,777,009 1,603 2,778,612 Change 6,399 4,020 10,419 Marginal 2,783,408 5,623 2,789,031 Table 3 2 9 Contingency table of change versus non change for expected values for the industrial warehouses use (unit: cells) Expected Prediction Non Change Change Marginal Non Change 2,773,010 5,602 2,778,612 Change 10,398 21 10,419 Marginal 2,783,408 5,623 2,789,031 Table 3 30 Pseudo R for the industrial warehouses use Goodness of Fit (P arameters) Results McFadden 0.6259 P Level 0.0000 Cramer's V 0.6488 Table 3 3 1 Overall accuracy for the industrial warehouses use Observed Prediction Cat 0 (Non COM ) Cat 1 ( COM ) % Correct Cat 0 (Non COM ) 2,774,034 1,738 99.94% Cat 1 ( COM ) 6,264 6,995 52.76% Overall % 99.69% 0.31 % 99.71%

PAGE 150

150 Table 3 3 2 ROC analysis and 2 2 contingency table based on 15 percent stratified sampling for the industrial warehouses use Simulated By Threshold Reality (Reference Image) 1 ( INDUS ) 0 (Non INDUS ) 1 ( INDUS ) A (number of cells) B (number of cells) 0 (Non INDUS ) C D For t he Given Reference Image: A + C =1,999 B + D =577,774

PAGE 151

151 Table 3 33. Accuracy assessment for 2003 urban LULC simulation map based on predicted pixels (1) Class Single Family Multi Family Commercial Institutional Transportation Industrial Warehouses Grasslands Forests Agricultural Recreational Others Wetlands Water Barren Total Single Family 218,674 716 0 0 33,765 9,098 88 82 6,137 827 0 269,387 Multi Family 38 39,533 676 36 161 816 52 72 698 50 1 42,133 Commercial Institutional Transportation 114 3,766 96,836 1,276 4,785 9,259 1,331 2,243 3,670 531 43 123,854 Industrial Warehouses 37 773 1,460 11,362 3,338 10,050 4,748 92 2,472 1,020 204 35,556 Grasslands 0 0 0 0 522,074 0 0 0 0 0 0 522,074 Forests 0 0 0 0 0 1,021,887 0 0 0 0 0 1,021,887 Agricultural 0 0 0 0 0 0 140,794 0 0 0 0 140,794 Recreational Others 0 0 0 0 0 0 0 128,462 0 0 0 128,462 Wetlands 0 0 0 0 0 0 0 0 390,722 0 0 390,722 Water 0 0 0 0 0 0 0 0 0 111,787 0 111,787 Barren 0 0 0 0 0 0 0 0 0 0 2,375 2,375 Total 218,863 44,788 98,972 12,674 564,123 1,051,110 147,013 130,951 403,699 114,215 2,623 2,789,031 Overall Accuracy = 3,760,630/3,865,155 = 97.2957% Kappa Coefficient = 0.9667

PAGE 152

152 Table 3 34. Accuracy assessment for 2003 urban LULC simulation map based on predicted percent (2) Class Single Family Multi Family Commercial Institutional Transportation Industrial Warehouses Grasslands Forests Agricultural Recreational Others Wetlands Water Barren Total Single Family 99.91 1.60 0 0 5.99 0.87 0.06 0.06 1.52 0.72 0 9.66 Multi Family 0.02 88.27 0.68 0.28 0.03 0.08 0.04 0.05 0.17 0.04 0.04 1.51 Commer cial Institutional Transportation 0.05 8.41 97.84 10.07 0.85 0.88 0.91 1.71 0.91 0.46 1.64 4.44 Industrial Warehouses 0.02 1.73 1.48 89.65 0.59 0.96 3.23 0.07 0.61 0.89 7.78 1.27 Grasslands 0 0 0 0 92.55 0 0 0 0 0 0 18.72 Forests 0 0 0 0 0 97.22 0 0 0 0 0 36.64 Agricultural 0 0 0 0 0 0 95.77 0 0 0 0 5.05 Recreational Others 0 0 0 0 0 0 0 98.10 0 0 0 4.61 Wetlands 0 0 0 0 0 0 0 0 96.79 0 0 14.01 Water 0 0 0 0 0 0 0 0 0 97.87 0 4.01 Barren 0 0 0 0 0 0 0 0 0 0 90.55 0.09 Total 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00

PAGE 153

153 Table 3 35. Accuracy assessment for 2003 urban LULC simulation map (3) Class Predicted (Pixels) Predicted (Percent) Observed (Pixels) Observed (Percent) Single Family 218,674 9.66% 218,863 7.85% Multi Family 39,533 1.51% 44,788 1.61% Commercial Institutional Transportation 96,836 4.44% 98,972 3.55% Industrial Warehouses 11,362 1.27% 12,674 0.45% Grasslands 522,074 18.72% 564,123 20.23% Forests 1,021,887 36.64% 1,051,110 37.69% Agricultural 140,794 5.05% 147,013 5.27% Recreational Others 128,462 4.61% 130,951 4.70% Wetlands 390,722 14.01% 403,699 14.47% Water 111,787 4.01% 114,215 4.10% Barren 2,375 0.09% 2,623 0.09% Total 2,684,506 100.00% 2,789,031 100.00%

PAGE 154

154 Table 3 36. Accuracy assessment for 2003 urban LULC simulation map (4) Class Commission (Percent) Omission (Percent) Commission (Pixels) Omission (Pixels) Single Family 18.83% 0.09% 50,713/269,387 189/218,863 Multi Family 6.17% 11.73% 2,600/42,133 5,255/44,788 Commercial Institutional Transportation 21.81% 2.16% 27,018/123,854 2,136/98,972 Industrial Warehouses 68.04% 10.35% 24,194/35,556 1,312/12,674 Grasslands 0.00% 7.45% 0/522,074 42,049/564,123 Forests 0.00% 2.78% 0/1,021,887 29,223/1,051,110 Agricultural 0.00% 4.23% 0/140,794 6,219/147,013 Recreational Others 0.00% 1.90% 0/128,462 2,489/130,951 Wetlands 0.00% 3.21% 0/390,722 12,977/403,699 Water 0.00% 2.13% 0/111,787 2,428/114,215 Barren 0.00% 9.45% 0/2,375 248/2,623 Table 3 37. Accuracy assessment for 2003 urban LULC simulation map (5) Class Prod Acc .(Percent) User Acc .(Percent) Prod Acc .(Pixels) User Acc .(Pixels) Single Family 99.91% 81.17% 218,674/218,863 218,674/269,387 Multi Family 88.27% 93.83% 39,533/44,788 39,533/42,133 Commercial Institutional Transportation 97.84% 78.19% 96,836/98,972 96,836/123,854 Industrial Warehouses 89.65% 31.96% 11,362/12,674 11,362/35,556 Grasslands 92.55% 100.00% 522,074/564,123 522,074/522,074 Forests 97.22% 100.00% 1,021,887/1,051,110 1,021,887/1,021,887 Agricultural 95.77% 100.00% 140,794/147,013 140,794/140,794 Recreational Others 98.10% 100.00% 128,462/130,951 128,462/128,462 Wetlands 96.79% 100.00% 390,722/403,699 390,722/390,722 Water 97.87% 100.00% 111,787/114,215 111,787/111,787 Barren 90.55% 100.00% 2,375/2,623 2,375/2,375

PAGE 155

155 Figure 3 1 Predicted 2003 single family probability map

PAGE 156

156 Figure 3 2 2003 single family prediction map

PAGE 157

157 Figure 3 3 Predicted 2003 multi family probability map

PAGE 158

158 Figure 3 4 2003 multi family prediction map

PAGE 159

159 Figure 3 5 Predicted 2003 commercial institutional transportation probability map

PAGE 160

160 Figure 3 6 2003 commercial institutional transportation prediction map

PAGE 161

161 Figure 3 7 Predicted 2003 industrial warehouses probability map

PAGE 162

162 Figure 3 8 2003 industrial warehouses prediction map

PAGE 163

163 Figure 3 9 Predicted 2003 single family probability map (refined)

PAGE 164

164 Figure 3 10 Refined 2003 single family prediction map

PAGE 165

165 Figure 3 11 Predicted 2003 multi family probability map (refined)

PAGE 166

166 Figure 3 12 Refined predicted 2003 multi family prediction map

PAGE 167

167 Figure 3 13 Predicted 2003 commercial institutional transportation probability map (refined)

PAGE 168

168 Figure 3 14 Refined 2003 commercial institutional transportation prediction map

PAGE 169

169 Figure 3 15 Predicted 2003 industrial warehouses probability map (refined)

PAGE 170

170 Figure 3 16 Refined 2003 industrial warehouses prediction map

PAGE 171

171 Figure 3 17 ROC curve for the single family use

PAGE 172

172 Figure 3 18 ROC curve for the multi family use

PAGE 173

173 Figure 3 19 ROC curve for the commercial institutional transportation use 0 20 40 60 80 100 120 0 4.138 9.175 14.219 19.251 24.302 29.358 34.417 39.471 44.519 49.574 54.629 59.682 64.738 69.796 74.853 79.875 84.904 89.916 94.957 100 true positive % (sensitivity) false positive % (specificity) Commercial Institutional Transportation ROC Curve (Percent) commercial-institutionaltransportation ROC curve

PAGE 174

174 Figure 3 20 ROC curve for the industrial warehouses use

PAGE 175

175 Figure 3 21 2003 Alachua County LULC simulation

PAGE 176

176 CHAPTER 4 URBAN LAND USE ALLOC ATIONS Chapter Overview The urban land use allocation is the next step following the MLR model building in Chapter 3. The allocation is essentially a conversion process that converts the continuous probability maps into categorical discrete maps based on the different urban uses. The urban allocation process provides tangible maps derived from the abstract probability maps and fulfills the urban use allocations through a series of processes such as urban development scenario identification, urban development and land forecasts, co nflict analyses, and final allocations. The fundamental goal of the allocation is to produce urban allocation maps based on different development scenarios to simulate different urban development arrangements. These will be introduced in detail in this cha pter. Urban allocation narration sequence The narrations of overall urban use allocation are organized in three parts: (1) literature review and summaries of research that has been conducted so far; (2) narrations of allocation methods; and (3) narration s of urban use allocation results. The literature review will introduce the research currently underway to learn about methods and results. The urban allocation methods will introduce the technique s that will be applied in this study. The results are the o utcomes of the applied allocation methodology. Urban allocation scale The study area is Alachua County, which comprises agricultural land plus a number of small cities and towns, with the largest city being the City of Gainesville. As predominantly a rura l county, Alachua County has consumed significant amounts of natural land for urban development and suffered serious urban sprawl problems over the years. This rural county is a typical example that can be dissected and examined for its loss of precious na tural land over the years, which can give the reasons of how natural land is consumed and

PAGE 177

177 converted into urban uses and why urban sprawl was triggered in such a way. The examination can inductively lead to understandings of how urban sprawl happens in Alac hua County as exemplified and how the rule can be extracted and used, deductively, for other Florida counties and cities for its heuristic character. Thus, the research that finds rules from this rural county for urban growth is expected to encourage resea rch for other Florida cities and counties. As a result, this study starts not from a big county with a large urban population, such as the City of Jacksonville or the City of Orlando, but from a relatively smaller county in population still rural in charac ter. The above narration is the case that can horizontally induce urban growth rules across the political boundaries of different counties. It can also be induced vertically across different timeframes for urban growth research in the county, in which st rategies dealing with future urban growth, i.e., in 2020 and 2030, can be put forth based on the analysis of the current situation in 2010 as well as the past situations in 1982, 1994, and 2003, respectively. As a result, the quantitative counts of land an d spatial distribution of the current and past urban development are assessed. The future urban growth in the quantitative amount of land and spatial spreading of development will also be assessed as this urban allocation process provides clues to the quan tity and quality of urban development for the next twenty years. The allocation process will also offer statistics of how many acres of urban development will be and where they will be. In fact, this urban allocation establishes foundations for policy form ation of urban growth in the future. Urban allocation method This study incorporates five scenarios in order to simulate urban development based on five conditions. These five scenarios include the BAU scenario, the infill scenario, the increased density scenario, the redevelopment scenario, and the conservation scenario. They will be illustrated in detail in the Methodology part of the chapter. These five

PAGE 178

178 scenarios almost comprise all development possibilities, prospects, and visions for the future urban growth of the county. With these five scenarios, the future urban growth in Alachua County will basically follow along these development paths and fulfill these development manners. Aside from the above five scenarios, the urban allocation method also in cludes three forecasting methods, which correspond to the above five scenarios, because the different development scenarios entail different forecasting methods. These three forecasting methods include the BAU (baseline) forecasting method, the increased density forecasting method, and the redevelopment forecasting method. They will be explained in detail in the Methodology part of this chapter. Urban uses that will be allocated The urban allocation in this study deals w ith four urban uses, which are single family, multi family, commercial institutional transportation, and industrial warehouses. These four urban uses will be allocated based on the five scenarios as mentioned above. In addition, mixed use will be simulated in the redevelopment scenario as well as in the conservation scenario. The natural LULC classe s such as grasslands, forests, agricultural, recreational others, wetlands, water, and barren will not be allocated in this study, however. Urban allocation data This study entails the combination of parcel data, remote sensing data, and probability data that are derived from the remote sensing data to analyze urban growth. On the one hand, the continuous probability data, derived from the MLR model, will be appl ied to create the collapsed maps based on each urban use. These are the maps that will be converted from the continuous data to the discrete data so that they can be later input into the conflict analysis to analyze. These probability data from the MLR mod el essentially come from the

PAGE 179

179 classification of urban and natural LULC classe s, which originated from the remote sensing Landsat TM and ETM+ data. On the other hand, parcel data are utilized based on parcel counts for different urba n uses from the Alachua C ounty counts of different urban uses will be extracted for their acreages and will be used in the infill scenario, increased density scenario, and the redevelopment scenario because these three scenarios call for i nfill acreages for development, and infill acreages come directly from the parcel counts from the parcel data. Furthermore, these parcel data are extracted to create the LULC Current Plan map, depicting the urban uses and natural LULC classe s in 2010, whic h is the base map for the future allocation maps for 2020 and 2030, respectively. Allocation Literature Review The literature review will find equivalent research that has been conducted so far for urban use allocations. Because some of the descriptions in this literature review part will be related to the urban growth model building, i.e., the MLR model, which is the major contents in Chapter 3, this literature review will carry over some of the descriptions of the urban growth model building to this chapt er. However, because this chapter mainly discusses the urban use allocations, the major theme of this chapter will still cover the urban use allocations as the most important part in the chapter. The literature that is reviewed has much to do with the appl ication of the MLR model. Other models, such as the CA model, are not included in this review. Verburg et al. (2007) used a Conversion of Land Use and its Effects at Small Regional Extent ( CLUE S ) model that is derived from the MLR model to simulate agric ultural land abandonment and conversions in Europe based on 1 km resolution satellite images. Their research was conducted based on two timeframes, 2000 and 2030, respectively, in which the agricultural land was analyzed as opposed to other uses such as bu ilt up areas and grasslands. They found that agricultural land conversions varied according to geographic locations, in which

PAGE 180

180 agricultural land was converted in the fastest way in dense urban areas and needed vacant land the most to accommodate urban devel opment and also in marginal areas such as mountainous regions. They also conducted a landscape analysis based on eight land classes identified as built up, arable land, pastures, forests/nature, inland wetlands, irrigated agriculture, abandoned land, and o thers. Since their research was primarily on the agricultural land, their land use classes for urban development was solely based on urban built up, along with several other natural land categories such as pastures, forests, wetlands, agricultural land, an d so on. The research of Verburg et al. (2007) is inspirational as they used the overall non spatial land consumptions to forecast the land demands and then translated them into spatial distributions for urban built up areas and various other natural land. This study uses population forecast to translate it into land consumptions for urban uses as well. Landis (1994) developed a CUF model to simulate urban growth in the northern San Francisco Bay Area of California, which consists of fourteen counties. The CUF model had four modules: population growth, spatial database, spatial allocation, and annexation. The CUF model (p.403), in which population in each individual city and county was estimated using the OLS regression algorithm, rather than projecting the regional population first and then subsequently splitting them into smaller units based on each individual city and county. Second, the CUF model merged various la yers related to urban growth into a single layer in a polygon format. As a result, vector based analysis was conducted at this point, for which data sources were from the Census Bureau TIGER databases such as TIGER roads, TIGER census tracts, TIGER city bo undaries, TIGER hydrology, and TIGER railroads and airports; sphere of influence; slope polygons; highway buffers; earthquake faults; prime agricultural land; marsh and wetlands; and

PAGE 181

181 sewer and water utility service costs. Third, the CUF model allocated the forecasted population based on the undeveloped land available for development derived from the above vector spatial maximum profits, were simulated. As a result, u ndeveloped land was weighted as scores and sorted according to their profit potentials calculated from the model. In this step, non developable land due to environmental constraints, ownership, and public policy issues was excluded from allocation. Finally the forecasted population was allocated to the areas within the sphere of influence based on the scores of each undeveloped land received from the weighting factors as mentioned above. Municipal annexation and urban incorporation factors were also consid ered as the last module in the CUF model. As a result, Landis (1994) used a simple regression model to simulate the expansions of urban boundaries within the area. In addition to the northern Bay Area of California, Landis (1995) also applied the CUF mode l to the San Francisco Bay Area based on three scenarios: business as usual, maximum environmental protection, and compact city. The business as usual scenario followed the current land development policies at the local level, in which the current local la nd use practices would be continued in the future. The maximum environmental protection scenario proscribed development from occurring in steeper slope areas, wetlands, and agricultural land. Compact city allowed development to only occur in the area havin g at least 18 persons per acre, setting aside at least 20 percent of population to infill development, and within 1,000 meters of existing urban boundaries (Landis, 1995). The CUF model is close to the model of this study in terms of the model structure s ince the CUF model incorporates population forecast and population allocation, 1 the two important steps for urban growth modeling. The CUF model is based on population forecasts for each individual 1 Can be translated into land use forecast and land use allocation.

PAGE 182

182 city and county as well as the principles that either the forecasted population is completely allocated within the undeveloped areas or the undeveloped land is totally consumed up, leaving the additional population to be allocated to adjacent communities. In addition, the CUF model allocates population based on a score for potential profit each land unit receives, which is quite similar to the suitability analysis (Carr and Zwick, 2007), in which each land unit obtains a y of based on three scenarios is inspirational as it incorporated three different land use policies to urban development, in which this study will adopt the concept of this approach to simulate the urban growth for Alachua County in the future as well and also adopt the concept of 1,000 meters of existing urban boundaries as urban buffer areas. In addition, the infill development is also inspirational to this study. However, the CUF model addresses urban growth solely based on the urban category, or more precisely, the residential use (Landis and Zhang, 1997). It did not address additional urban land categories such as residential, commercial, and industrial, together with other non urban natural land, which will also influence urban land use development. As a result, it is a basic urban growth model without considering various uses involved for urban growth. At this point, conflict analysis (Carr and Zwick, 2007) such as the competition among various urban uses are absent from the model. To remedy these shortcomings, Landis and Zhang (1997) proposed a CUF II model to simulate urban growth. The major overhaul of the CUF II model is its inclusion of the MLR algorithm, i n which probabilities are used to measure the probability of a site to be developed and/or redeveloped. Compared to the old CUF model, the CUF II model added four additional elements: expansion of the original single urban use to residential use, commercia l use, industrial

PAGE 183

183 use, public use, and transportation use; inclusion of urban development together with urban redevelopment; bid of different urban uses to be located in a typical site; and inclusion of historical data to calibrate the model. Compared to t he old CUF model, the CUF II model provided the results of the spatial locations of the land uses for their correlations with different variables the model specified, which resulted in the current urban growth in different counties in the San Francisco Bay Area, based on the development and redevelopment categories. For example, for vacant land development, cities with larger population size would be more likely developed in terms of residential use than cities with smaller population size in the area. For the redevelopment, for example, however, sites that were closer to transportation corridors would less likely be redeveloped into residential use. In this case, the CUF II model did not allocate land to meet the land demands given various different urban u se categories the model identified, in which the CUF I model had. In addition, the CUF I and CUF II models adopt a regional approach for urban growth, in which, e.g. cities with larger population will be more likely to be developed for residential use. Thi s is quite different from the scope of this study in which only a local county is considered for urban use allocations. Nevertheless, the CUF II is inspirational for urban growth modeling because it, for the first time, proposes to use the MLR model to sim ulate urban development and/or redevelopment. It also carries some heuristic elements such as inclusions of various urban uses, bid of different competitive uses for a site, and inclusion of historic data to calibrate the model, in which these elements can be used in this study (some parts are illustrated in Chapter 3). Hu and Lo (2007) utilized the Landsat images to simulate urban growth in the Atlanta region in Georgia, which included 13 counties, based on six LULC categories: high density urban, low den sity urban, bare land, crop or grassland, forest, and water. Because their study

PAGE 184

184 area covered 13 counties, they modeled urban growth based on urban use only, along with other natural land. Their remote sensing data covered the period of 1987 and 1997, resp ectively, for the study area. During the analyzing process, they combined high density urban and low density urban into one variable, i.e., urban, which was used as a dependent variable for model building. They found twenty independent variables such as po pulation density, income per capita, poverty rate, percentage of white people, employment rate, slope, distance to nearest urban cluster, distance to CBD, distance to active economic center, distance to the nearest major road, number of urban cells within a 7 by 7 window, existing/planned conservation, existing high density urban, existing low density urban, existing bare land, existing cropland/grassland, existing forest, and easting and northing coordinates. Based on these variables, Hu and Lo (2007) prod uced a number of maps to predict urban growth with certain percentages, e.g., 20 percent, 25 percent, or 30 percent of urban areas, within the study area. These percentages can be translated into the specific timeframes that meet the thresholds specified a bove. For example, the urban areas in Atlanta were 16.4 percent in 1997, in which 16.4 percent of urban development referred to 1997; in 2005, the urban areas in Atlanta reached 25 percent of the total area, in which 25 percent of the urban areas referred to 2005, and so on for the rest of urban area percentages mentioned above. Although Hu and Lo (2007) successfully modeled urban growth based on different periods translated from different percentages of urban areas, their research is not very convenient fo r readers to interpret the modeling results, given the percentage of urban development. Research that addresses the urban growth based on specific years will be more meaningful than the percentage development. This is what this study is going to achieve. Carr and Zwick (2007) developed a Land Use Conflict Identification Strategy (LUCIS) model to allocate urban, agricultural, and conservation uses, the three uses upon which their

PAGE 185

185 y an analysis that deals with these three uses based on a conflict nature that one typical use, e.g., urban, is in conflict with the other two uses, i.e., agricultural and conservation, when they are allocated to a specific site. Although Carr and Zwick (2 model essentially calls for the thesis that these three uses are competing with each other to bid for a typical site during the allocation process. The LUCIS model is fundamentally a suitability analysi s model. Carr and Zwick (2007) constructed the LUCIS model based on five steps: (1) goals and objectives identification; (2) preparation of data inventory; (3) suitability analysis; (4) preference identification; and (5) conflict analysis. Goals and object ives are the intended setting for a specific scheme that acts as an input into the suitability analysis. Each of the three uses has its equivalent goals and objectives which are hierarchically dispersed across its compositional structures so that one use, e.g., urban, has an overall goal and is composed of several sub categorical goals and objectives as input into the suitability analysis. It is the same as for the agricultural and conservation uses. This identification of goals and objectives establishes f oundations for the suitability analysis, as mentioned above, as well as the conflict analysis because the suitability analysis and the conflict analysis utilize the goals and objectives to create suitability maps for each land use category. The data inven tory for the LUCIS model is basically parcel data as well as pertinent GIS shapefiles. The three land uses have their existing datasets retrieved from the county parcel data for urban, agricultural, and conservation, respectively, in which detailed subgrou ps of the three uses can be extracted when needed. For example, the urban use includes retail, commercial, industrial, warehouses, and so on, which are retrieved from the county parcel data when they are required. As a result, some non existing GIS shapefi les can be created by employing the existing

PAGE 186

186 GIS vector datasets. Thus, raster datasets can be created for various analyses through manipulating the parcel data and pertinent existing GIS shapefiles. The suitability analysis is the crux of the LUCIS model suitability analysis essentially tells how well a piece of land fits a land use arrangement whether it will be best developed as urban, or preserved as agricultural or conservation. The suitability analysis primarily translates th e goals and objectives in the above Step 1 into tangible maps so that uses that have specific, equivalent goals and objectives are depicted on maps to illustrate where the suitable land is in terms of a specific goal and objective. To achieve this purpose, the goals and objectives are hierarchically established and categorized as overall goals and objectives, general goals and objectives, and sub goals and sub objectives. A typical map is created for each sub objective, from the bottom up, which depicts hig h suitability or low suitability for each sub goal and sub objective. They are then combined to create a suitability map that illustrates a more generalized goal and objective for a typical use (which is at the middle hierarchy in the goals and objectives organizational structure). This newly created map for the suitability of a more generalized objective, as mentioned above, will be the input files for the next step, i.e., the preference identification. Because there are a number of sub objectives involved in the suitability analysis, a multiple utility assignment (MUA) technique, which weighs different sub objectives, is applied by combining all the sub objectives into this more generalized objective (the middle hierarchy of the goals and objectives). It i s noted, however, that the suitability analysis only combines the sub objective to the more generalized objective level (the middle objective as mentioned above) in the hierarchical structure of the goals and objectives. A higher level, which is the overal l goals and objectives, will be combined in the next

PAGE 187

187 step, Step 4, the preference identification. In addition, the highest suitability of a typical use has the score of 9 while the lowest score is 1. The preference identification is basically the step tha t combines the sub categorized goals and objectives into the overall goals and objectives. To do this work, an Analytic Hierarchy Process ( AHP ) is employed. Throughout the AHP process, different goals at this step are combined based on the weight of each goal, which is to form a map for the overall goal. As a result, three uses to be allocated have three maps with each map corresponding with each use. These three maps are further input into the next step Step 5 the conflict analysis with data manipulations. Conflict analysis is the last step in the LUCIS model, which allocates the three uses based on the conflict nature that these three uses are c ompeting with each other for a specific site. In this process, areas that cannot be developed or have already been developed will be excluded from the analysis process. Also, the preference maps will be normalized by making the values from 1 to 9 to betwee n 0 and 1. Then, the normalized maps will be collapsed into three categories, which are low, medium, and high preferences represented by 1, 2, and 3, respectively, with the highest preferences coded as 3. The collapsed methods used in the LUCIS model compr ise four collapsing methods, namely, the natural breaks method, the manual method, the equal interval method, and the standard deviation method (Carr and Zwick, 2007). When all the processes are completed, the collapsed maps will be combined into a conflic t map, and final allocation will be conducted based on the conflict maps that have been created. In the LUCIS model, future populations and future land use acreage demands are forecasted for urban use as a final allocation process. Gross urban density (GU D) is proposed for future land demands given the forecasts of future population. As a result, baseline future urban

PAGE 188

188 acreage demands are calculated using the above urban gross density. Carr and Zwick (2007) provided six steps for the final allocation, which are (1) allocating urban use if urban use is not in conflict with the other two uses; (2) allocating urban use for additional cells when urban use is in is based on forecasts for the future 50 to 60 years; (4) allocating remaining cells to agricultural use when agricultural use is not in conflict with the other two uses; (5) allocating remaining cells to conservation land when conservation does not conflict w ith the other two uses; and (6) allocating remaining cells to agricultural or conservation uses when they are in conflict but with higher normalized preference values (Carr and Zwick, 2007, p. 167). (2007) model is inspirational to this study for its conflict analysis and final allocation elements. As a result, this study adopted some of the techniques that are utilized by Carr and Zwick (2007) in their models. Similar to Carr and Zwick (2007), this study will also apply a population forecast and the GUD to forecast future urban development demands in the next 20 years (2020 2030). In addition, a concept of conflict analysis is implemented, and four urban uses in this study, i.e., single family, multi family, commercial institutional transportation, and industrial warehouses, are put forward into the analysis so that they compete with each other for a specific site. However, because this study adopts four uses, instead of three, the conflict analysis i s somewhat more complex than the LUCIS model. As a result, more conflict scores (conflict combinations) are received from this study than the one in the LUCIS model (Table 4 23). In addition, this study does not deal with urban development from the angle o f suitability but from the angle of probability as it extracts datasets from the MLR model. Therefore, this study does not need normalization for each use because the raw data from the MLR model, which has a value between 0 and 1, has already been normaliz ed. The remaining conflict analysis procedures

PAGE 189

189 in this study are essentially the same as the LUCIS model, in which the collapsing process using the natural breaks method and the creation of conflict maps are recommended because they are essentially in the same process as the LUCIS model. For the population forecasts, this study does not employ a regional population forecast method; rather it only utilizes the population for Alachua County because this study is not a region based study. In addition, this st udy does not have goals and objectives that will be input into the conflict analysis because the model building suggests probability for a specific use rather than suitability. As a result, the model building approach from the MLR model in this study is fu ndamentally different from the LUCIS model, and suitability analysis is not conducted in this study, either. Therefore, this study does not have goals and objectives identification, suitability analysis, and preference identification: no MUA and AHP are in volved in the allocation process. In addition, this study does not have a population forecast for the future 50 60 years either, as the LUCIS model proposes; but rather, the population in the future 20 years (to 2030) is projected in this research. In ter ms of the preparation of data inventory, this study utilizes remote sensing data as well as the county parcel data for allocation. Because this study calls for several scenarios other than the baseline forecasting, which are introduced in the later part of this chapter, infill development parcel data. As a result, unlike the LUCIS model in which GIS vector datasets are used and converted to raster datasets, this study uses both the parcel data and remote sensing data, both the vector data and the raster data for study. In addition, because the county parcel data does not provide wetlands information, which this study does require and include, remote sensing data are utilized in this regard. Also, because the remote sensing data do not provide infill information,

PAGE 190

190 parcel data are employed at this point. Therefore, the advantages of both the parcel data and the remote sensing data are utilized in this study. The lit erature review describes the most recent research that has been done so far across the academic field. In addition, from the above literature, many of the current urban studies are conducted based on the urban use that Hu and Lo (2007) suggest. Detailed ur ban categories such as single family, multi family, commercial, institutional, transportation, industrial, and warehouses are often not included. In addition, there is one study that attempts to use the MLR model to simulate urban development and redevelop ment (Landis and Zhang, 1997). Although some studies address urban uses in detailed categories (Landis and Zhang, 1997), they did not predict future land uses through allocating land uses (Landis and Zhang, 1997). Some studies, although allocating future l and uses, were neither based on detailed urban use categories nor in actual years (Hu and Lo, 2007). This research will address urban growth issues based on three perspectives: simulations involving various urban uses, conflict analysis adoptions, and vari ous scenario predictions with clear timelines, borrowing the ideas from the LUCIS model and the CUF I and the CUF II models. Also, mixed use will be simulated in this research, which further increases the broadness and breath of the research. Methodology F ive Scenarios The methodology of urban land use allocation is based on five scenarios: namely, the BAU scenario, the infill scenario, the increased density development scenario, the redevelopment scenario, and the conservation scenario. These five scenari os allocate land use and land cover based on the probability maps generated by the MLR model that is illustrated in Chapter 3. The probability maps are the maps that contain the probability numbers for each of the four urban land uses, namely, single famil y, multi family, commercial institutional transportation, and

PAGE 191

191 industrial warehouses. These probability maps are extracted from the MLR model using two different masks based on the different scenario types: one is for the BAU scenario and the other is for t he other four scenarios because the other four scenarios share the same geographic boundaries for development. Specifically, masks are the geographic boundaries that exclude existing areas and occupied urban uses, which are already developed, not suitable for development, or are forbidden for development. This study allocates four urban uses as introduced above based on these two mask types which are introduced in detail later in this chapter. According to the above five scenarios, the natural LULC classe s such as grasslands, forests, agricultural, recreational others, wetlands, water, and barren are emphatically not included in the allocation process. They are to remain the same as they were in 2003. BAU scenario The BAU development refers to the scenario that growth occurs according to the current development arrangement, in which the BAU scenario imitates generalized development but without many development restrictions. Development restrictions are development patterns that are imposed by the state or lo cal authorities, in which the outward development is curbed by imposing environmental regulations on the land so that the outward development can move inward, suggested by cases such as in the patterns of infill, increased density, redevelopment, and conse rvation. In the BAU development scenario, however, outward development is allowed, and development on wetlands (Landis, 1995) as well as prime farmlands is also allowed. As a result, k setting, and the urban uses, existing conservation land, water and creek buffers, utilities, high voltage power lines, stormwater ponds, and parks and cemet eries are masked out, however. Besides, the BAU scenario does not restrict development within urban buffer areas, i.e., within 1,000 meters of the

PAGE 192

192 municipal and the Urban Cluster areas (urban area) in Alachua County, that is delineated by the Alachua Count y Growth Management Department (Figures 4 1 and 4 2) or outside the urban buffer areas. It can be developed, as mentioned above, anywhere in the county as long as there is vacant la nd available for development. Figure 4 3 illustrates the mask of the BAU scenario. In this figure, the shaded areas are the areas that are masked out, and the remaining areas the white areas are the land that is available for development. From the figure, it is clear that the central city, the City of Gainesville, is almost all masked out. The lower right portions of the map are also masked out Infill scenario cal Office. This vacant land is set aside not for present day use but for future use. The vacant land that is set aside for future use can be any of the four urban uses : single family, multi family, commercial institutional transportation, or industrial warehouses. For example, vacant residential uses, coded as 0000, are set aside as residential use mostly single family use in cant commercial, vacant institutional, and vacant respectively. In this case, when urban land uses are allocated, it will be the first priority to satisfy th e vacant infill urban land needs that correspond to a typical urban use before they are allocated to other vacant, non infill land. The infill development mask differs from the one in the BAU scenario in that the infill development excludes future conserva tion for development (Figure 4 4). From Figure 4 4, most of the areas in the eastern and southeastern Alachua County in infill development are masked

PAGE 193

193 out, which means development in these parts of the areas are restricted or forbidden. These restricted and forbidden areas include wetlands, prime farmlands, and future conservation areas such as the Florida Forever land and the Alachua County Forever land in addition to the masked out areas in the BAU scenario. However, because of the priority of the vacant p arcel data, it is imperative to satisfy the vacant, infill parcels before they go to the vacant non infill land, as they will in the BAU scenario. Consequently, there will be overlap between the vacant infill land and the vacant non infill land because vac ant infill land can be located in vacant non infill land, i.e., the conflict areas, which will be introduced in depth in the later part of the chapter. Because there is no difference between the vacant infill land within the 1,000 meters urban buffer areas (urban area) and those outside 1,000 meters (greenfield area) (Figure 4 1), development can occur either inside or outside the 1,000 meters of urban buffer areas (Figure 4 5). Vacant land that is inside or outside the 1,000 meters of urban buffer areas ha s essentially the same opportunities for development when the infill land is actually allocated. Increased density development scenario The increased density development scenario presents the same mask as the infill development does. The increased densit y development takes place inside the 1,000 meters of urban buffer areas (Figure 4 5). Increased density occurs when the current density the BAU scenario or the infill scenario results in too much outward urban development. As a result, increased density is called for in order to reduce the development outside urban buffer areas. Specifically, the increased density development is the increase of GUD and correspondently increases all other densities such as the gross single family density, the gross multi fam ily density, the gross residential density (GRD), the gross commercial institutional transportation density, and the gross industrial warehouses density. The increased density development aims to reduce the occurrence of urban sprawl by significantly reduc ing the development acreages that

PAGE 194

194 happened outside 1,000 meters of the urban buffer area. This study adopts a reasonably increased GUD value compared to the ones in other cities and/or other counties in the State of Florida. The increased density developme nt scenario is a continuous scenario evolved from the infill development. As a result, the increased density development incorporates infill development in the allocation process. When a certain land use is to be allocated, the first step will be the adopt ion of the infill process to see if a development can be accommodated in vacant infill lots. Because the infill process can be either located inside 1,000 meters of the urban buffer areas or outside the urban buffer areas (Figure 4 5), when the infill proc ess is complete, it will then be allocated inside the increased density development areas where they are typically located within 1,000 meters of urban buffer areas because this study imitates increased density that is regulated to occur within 1,000 meter s of the urban buffer area. As a result, the increased density development can also be written as the infill/increased density scenario. Redevelopment scenario The redevelopment scenario adopts a method to further reduce the occurrence of urban sprawl by allocating certain amounts of urban uses by a process of redevelopment in the central urban areas. When a typical urban use is allocated based on this scenario, a certain portion of an urban use will be set aside within the existing urban areas where the e xisting urban structures will be demolished and new structures will hence be established there on the existing urban land. After this process is complete, the urban use will be allocated to see if it can be developed inside the vacant urban infill areas. W hen this test is completed, the urban uses will be allocated upon the increased density areas within the 1,000 meters of the urban buffer areas. When the increased density allocation is accomplished in the increased density development areas, finally, mixe d use will be allocated within the 1,000 meters of urban buffer areas. The mixed use is based on the locations of multi family and commercial institutional transportation uses combined because

PAGE 195

195 mixed use is the combination of multi family and commercial ret ail, service, and office. As a result, areas containing multi family and commercial institutional transportation will be scrutinized to see if they are suitable for mixed use development. This process is undertaken by examining whether the areas are along the 200 meters of major roads because these road front areas are typically the areas apt to mixed use development. Generally speaking, based on the above, the redevelopment scenario is a combination of multiple development scenarios such as infill, incre ased density development, and mixed use as introduced above, which can also be written as redevelopment/infill/increased density/mixed use. As a result, the mask of the redevelopment scenario is essentially the same as the infill scenario as well as the in creased density development scenario as introduced previously. C onservation scenario The conservation scenario allocates urban land based on the mask of all environmental conservation land, existing restricted urban areas, existing occupied land, as well a s existing urban forbidden areas. This scenario is used to test if vacant land within urban conflict areas is adequate to accommodate the development by considering the exclusion of the environmental land in a maximum manner. As a result, it precludes the development not only in the existing conservation areas but also in the future conservation areas such as the Florida Forever Conservation land and the Alachua County Forever Conservation land. For urban restricted and forbidden areas, contrary to the BAU scenario, the conservation scenario does not allow development to occur on wetlands and prime farmlands. Consequently, the mask of the conservation scenario shares the same mask as the infill development, the increased density development, and the redevelo pment. The conservation scenario follows a similar allocation sequence for a typical land use as introduced above. First, the land use will be allocated within the existing urban areas for certain

PAGE 196

196 portions of redevelopment. Second, the land use will be al located to see if it is within the increased density development areas. Third, the land use will be allocated to see if it accommodates mixed use. The conservation scenario does not include the infill process. This is the major difference between the conse rvation scenario and the other three development scenarios such as the infill scenario, the increased density development scenario, and the redevelopment scenario. Based on the above scenario descriptions, the conservation scenario can also be written as c onservation/ redevelopment/increased density/mixed use. Because of the infill development, the increased density development, and the redevelopment utilize the same mask as the conservation scenario, the mask of the conservation is the same as the above th ree scenarios, which is illustrated in Figure 4 4. Different from the above three scenarios (i.e., infill, increased density development, and redevelopment), however, the conservation scenario takes land directly from urban conflict areas. The urban confli ct areas refer to the places that urban uses compete with each other to be allocated, which can take place either within 1,000 meters of urban buffer areas or outside. They will be introduced later in this chapter. Forecasting Development Acreages for Fiv e Scenarios Baseline forecasting method (BAU forecasting and infill forecasting methods) The development forecasts are preceded based on the timeframes of 2020, and 2030, respectively. The baseline forecasting method is also known as the BAU forecasting me thod as well as the infill forecasting method because the BAU scenario and the infill scenario use this technique to forecast future urban growth. In particular, the baseline forecasting method and the follow up other three development scenarios employ a l inear regression model to forecast urban growth acreages. This is because the population growth in Alachua County presents a linear growth pattern in 1982, 1994, 2003, and 2010, as well as in 2020 and 2030 in the future (Figure 4 6). During the forecasting procedure, GUD is adopted (Carr and Zwick, 2007).

PAGE 197

197 The GUD refers to the total population occupying total urban land (Carr and Zwick, 2007) in the study area, where the total urban land comprises single family, multi family, commercial institutional trans portation, and industrial warehouses combined. When doing the forecasting, the first stage is to calculate the GUD for 2010. Because this study proposes that the GUD in 2010 remain constant (Carr and Zwick, 2007) as compared to those in 2020 and 2030, resp ectively, it will be easy to calculate the total urban acreages in 2020 and 2030, given that the value of the GUD in 2010 and the populations that have been forecasted are based on the Census (Table 4 6). When the total urban acreages in 2020 and 2030 are attained, the next goal is to calculate the acreages of single family use, multi family use, commercial institutional transportation use, and industrial warehouses us e in 2020 and 2030, respectively; the acreage of each urban use in 2010 is directly counted from the 2010 parcel maps because these data are acreages based on diff erent urban uses in 2020 and 2030, respectively, it is necessary to estimate the single family and the multi family population. This study proposes that the ratio of the single family population to the multi family population in different years, i.e., the SF/MF value, presents a linear decline format. As Figure 4 7 shows, the ratio of the single family population to the multi family population in 1982 was 1.53, while this ratio was 1.48 in 1994, 1.45 in 2003, 1.30 in 2010, and will be 1.26 in 2020, and 1.19 in 2030, which demonstrates a continual linear trend of decreasing throughout the years (Figure 4 7). The above SF/MF values are calculated using the SPSS V.19 software. Based on these numbers, it is easy to calculate the single family population and the multi family population in 2020 and 2030, respectively, which are illustrated in Tables 4 1 and 4 2. After this process is complete and because the gross single family density

PAGE 198

198 and the gross multi family density (the ratio of single family population to the single family acreage and the ratio of multi family population to the multi family acreage) remain constant throughout the years in 2010, 2020, and 2030, respectively, for the BAU scenario, the single family and the multi family acreages are designated si mply by single family or multi family population being divided by the gross single family density or the gross multi family density in the years 2020 and 2030, respectively. Then, urban acreage changes based on different urban use categories are calculated current year values for each of the four urban use types. Thus, it will be the same case for the values. The results are presented in Tables 4 1 and 4 2 respectively. The baseline forecasting method for the commercial institutional transportation use and the industrial warehouses use advocates a concept of employment population. The employment population is the count of the employment population based on different indus trial and/or business types extracted from the Florida Statistical Abstract 2003 (BEBR, 2003). Once the employment populations for commercial institutional transportation and industrial warehouses are determined for a typical year, i.e., 2003, the gross co mmercial institutional transportation density and the gross industrial warehouses density can also be determined given their acreages in 2003. As soon as their densities are established for 2003, their future densities, i.e., 2010, 2020 and 2030, can also be established following the same density as it was in the 2003 densities because the densities will remain constant for the BAU scenario, similar to the gross single family density as well as the gross multi family density. When the densities are resolved the next stage is to calculate the commercial institutional transportation and the industrial warehouses acreages. Because the ratio of the acreages of the commercial institutional

PAGE 199

199 transportation use to the acreages of the industrial warehouses use in 20 03 is a constant, i.e., 7.8, in this study, given that their total acreages derived from the total urban acreages minus total residential acreages, the final acreages for the commercial institutional transportation use and the industrial warehouses use are computed. The results are presented in Tables 4 4 and 4 5, respectively. In addition to the concept of the GUD, the GRD is also applied (Table 4 3). The GRD is the ratio of the total residential population to the total acreage of residential use, includin g both single family use and multi family use. The total residential population is the sum of single family population and multi family population. From Table 4 3, the GRD values increase from 2003 through 2030, which means that more people will reside on the same acreages of land compared to the values in the previous years. Finally, urban acreage changes based on the four urban land uses are computed and illustrated in Table 4 7. It represents the acreages and cells for each of the four urban uses that wi ll be developed in 2010, 2020, and 2030, respectively. In the table, cell numbers are utilized for the allocation of each of the four urban uses and meanwhile for projecting them on maps, which is explained in detail in the later part of the chapter. The v alues that are critical to the baseline forecasting method are illustrated in the Tables 4 1 through 4 8. Increased density forecasting method The increased density forecasting method applies an increased GUD compared to the baseline forecasting method. Th e increased density forecasting method is used by the increased curb urban sprawl by accommodating more people in the same acreage of land compared to the non s hrinkage baseline forecasting method. This study adopts an increased density that is 15 percent more than the baseline GUD. As a result, the original baseline GUD, which is 2.61

PAGE 200

200 people/acre, is elevated to 3.0 people/acre in the increased density forecasti ng method (Table 4 14). This new GUD is in reference with the one in Orange County, Fla., where the City of Orlando is located. The new GUD is a reasonably increased value not too high because Alachua County is a rural county, compared to the values in ot her Florida counties and cities. Once the increased GUD is determined, similar to the baseline forecasting method, the total urban acreages based on 2020 and 2030, respectively, are computed simply by the total population being divided by the increased ne w GUD. The next step is to determine the ratio of the total residential acreage to the total urban acreage. These numbers are from the baseline forecasting method, which are illustrated in Table 4 8. From Table 4 8, the percentage of the residential acreag e out of the total urban acreage in 2020 is 67.92 percent while it is 54.78 percent in 2030. Correspondingly, the percentage of the residential acreage change out of the total urban change in the increase density forecasting method is also 67.92 percent in 2020 and 54.78 percent in 2030, and they are the same as the percentages for the baseline forecasting method. Based on the calculated percentages of the residential acreage changes in 2020 and 2030 as well as the total urban acreages in 2020 and 2030, res pectively, in the increased density forecasting method, it is not difficult to calculate the residential acreage changes in 2020 and 2030. The results are illustrated in Table 4 11. The calculation of the commercial institutional transportation acreage as well as the industrial warehouses acreage also comes from the percentages of commercial institutional transportation use as well as industrial warehouses occupying the total urban acreages for the baseline forecast in Table 4 8. Finally, given that the pe rcentages of the single family makes up out of the total residential acreages in 2020 and 2030, respectively, in the baseline forecasting method (Table 4 8), the single family and the multi family acreages in 2020 and 2030 in the increased density

PAGE 201

201 forecast ing method can be obtained simply from the total residential acreages in 2020 and 2030, respectively, multiplying the occupancies of the single family and multi family uses by the total urban acreages (Table 4 8). The results are illustrated in Tables 4 9 and 4 10. As mentioned previously, given the percentages of the commercial institutional transportation use as well as the ratio of the industrial warehouses use to the total urban acreages, it is not difficult to estimate the acreages of the commercial in stitutional transportation use as well as the industrial warehouses use (Tables 4 12 and 4 13). The total change for each of the four urban land uses is presented in Table 4 15. From the above tables, with the increase of the GUD in 2020 and 2030, the gros s single family density, the gross of multi family density, the gross residential density, the gross commercial institutional transportation density, and the gross industrial warehouses density are all increased compared to the ones in the baseline forecas t. Redevelopment forecasting method The last forecasting method is the redevelopment forecasting method, in which the redevelopment scenario and the conservation scenario accept this forecasting method to predict future urban growth. Similar to the increa sed density method, the redevelopment forecasting method also attempts to reduce the development that goes outside the urban buffer areas by allocating certain acres inwards for redevelopment; because of that, urban sprawl in the greenfield areas will be c ontrolled. This study suggests a 15 percent redevelopment ratio in the urban buffer areas. Based on this ratio, significant amounts of new urban development can be curtailed and reduced. The calculation of the redevelopment forecasting method is essential ly similar as the one in the increased density forecasting method, in which the changes of the total urban acreage are calculated first before the redevelopment process takes place. During this forecasting process, the same values in the total urban acreag e changes in the increased density forecasting method

PAGE 202

202 (Table 4 14) are adopted, which are shown in Table 4 21; so are the values for other acreages such as the single family acreage, the multi family acreage, the total residential acreage, the commercial i nstitutional transportation acreage, and the industrial warehouses acreage, which are the same as the increased density forecasting method proposed and are adopted by the redevelopment forecasting method (Tables 4 9 through 4 14). Then, redevelopment value s are calculated based on the ratio of 15 percent as mentioned above, which leaves 85 percent of urban development for generalized development. Finally, the GUD as well as the associated densities for different urban land uses are estimated based on the 15 percent redevelopment ratio as well as the population forecasts given by the BEBR data and the population estimates for single family and multi family for 2020 and 2030, respectively, as illustrated in the baseline forecasting method. From Tables 4 16 thr ough 4 22, it is obvious that the densities increase, considering the 15 percent redevelopment ratio, which means that the redevelopment scenario increases densities inside the urban buffer areas. The values of the redevelopment forecasting method are pres ented in Tables 4 16 through 4 22. The Conflict Analysis The conflict analysis is the one type of analysis that offers ways to allocate land for a typical urban use in order to find out whether it belongs to the single family use, the multi family use, the commercial institutional transportation use, or the industrial warehouses use. The conflict analysis is based on the thesis that the four urban uses compete with each other for a typical site (Landis, 1995). As a result, the four urban uses are put in a p arallel manner (Carr and Zwick, 2007) so that each urban use can be compared with others side by side so as to determine which use the site is best suited for. For example, for a specific site, the conflict analysis will result in a score in the format of XXXX, e.g., 2341 (or any other combinations of the four digits), in which the thousands digit, 2, refers to the single family use, the hundreds digit, 3, refers to the

PAGE 203

20 3 multi family use, the tens digit, 4, refers to the commercial institutional transportati on use, and the units digit, 1, refers to the industrial warehouses use. Continued from the above example, because 4 is the largest number in the combination in which it is the highest score, the commercial institutional transportation use is the preferred use for this site; thus, this site belongs to the commercial institutional transportation use. When the same two or three highest scores occur in the combination, e.g., 4423 or 3331, it means that these two or three of urban uses are in minor conflict (Ca rr and Zwick, 2007) in that each of the two or three uses have the same opportunities to be allocated to a specific use. In the above 4423 example, the single family use and the multi family use are in minor conflict, and these two uses have the same chanc es to be allocated to their specific uses. Similarly, as the case of 3331, the single family use, the multi family use, and the commercial institutional transportation use all have the same chances to be allocated to their specific uses. When all the four scores are the highest, e.g., 4444 or 3333, it means that the four urban uses are all in major conflict (Carr and Zwick, 2007) with regard to a high preference status, in that all the four urban uses have the same chances to be developed according to a cer tain use on a given site. However, when all the four scores are the lowest, e.g., 1111, it means all the four urban uses are all in major conflict with regard to a low preference position, which refers to the situation that none of the uses are suitable fo r their certain uses to be allocated on a given site. Carr and Zwick (2007) proposed a four step method for the conflict analysis to allocate land uses. They include: (1) creation of a mask that excludes the areas that are developed, not suitable for devel opment, or restricted for development; (2) normalization of the suitability maps; (3) collapse of the normalized preference maps; and (4) combination of the collapsed lues

PAGE 204

204 are between 1 and 9 (Carr and Zwick, 2007). Therefore, their method needs normalization so as to adjust the values between 0 and 1, as mentioned previously. Because, in this study, the probability maps extracted from the logistic regression model have continuous values between 0 and 1, they are already normalized, and no normalization is needed in this regard. As a result, this study adopts a three step method for the conflict analysis: (1) creation of a mask to exclude the areas that are already devel oped, not suitable for development, or restricted or forbidden for development; (2) collapse of the preference maps; and (3) creation of the conflict maps. Table 4 23 illustrates the conflict scores and their equivalent descriptions in the conflict analysi s in that all the combinations of the conflict scores are put forward. The conflict analysis is the selection of the conflict uses based on different scores from Table 4 23, in which masks are applied that conform to different development scenarios. In th is study, two conflict masks are proposed: one is the BAU mask; the other is the infill development mask. The infill development mask is suitable for the infill scenario, the increased density development scenario, the redevelopment scenario, and the conse rvation scenario, respectively, because all these scenarios utilize this same mask for the conflict analysis. Creation of the masks As mentioned above, the mask creation incorporates two mask types: one is for the BAU scenario; the other is for the infill scenario, which is also the mask for the increased density development scenario, the redevelopment scenario, and the conservation scenario, because the last four scenarios use the same mask. These two masks are illustrated in Figures 4 3 and 4 4, respecti vely. The function of a mask is to exclude the areas that are already developed, not suitable for development, or forbidden for development; the leftover areas, after masking, are those areas that can be developed.

PAGE 205

205 From the BAU mask, it is evident that ma ny areas in the eastern and western parts of the county are vacant and available for development because the future conservation land is not included for masking (Figure 4 3). Because of the large areas that are not excluded for development, development ch oices in the county are ample in the BAU scenario. As a result, sprawled urban development that is scattered around the county outside 1,000 meters of urban buffer areas is evident based on the recent urban development practices from 2003 to 2010. Paradoxi cally, the opportunity for urban development in the infill development is much less than such opportunity in the BAU scenario. Because large areas of future conservation land are included in the masking process for the infill development they mostly are l ocated in the eastern and southeastern parts of the county; many development opportunities in the eastern part of the county are rigorously restricted (Figure 4 4). The developable land is mostly concentrated in the western part of the county, however. The purpose of the creation of the two types of masks is to extract the preference maps that include developable land only because the raw preference maps that are derived from the logistic regression model contain both the developable and non developable/ developed land. This is an important step because the conflict maps cannot be properly produced without it, and this masking technique successfully leads to the creation of the conflict maps by leaving out undevelopable areas so that the remaining parts ca n be chosen for development. When mask creation is completed, the next job is to collapse the preference maps that have been masked into four categories in accordance with the four urban uses so that the conflict maps can be created thereupon. Collapse of the preference maps When the preference maps originating from the logistic regression model are masked, the next phase is to collapse the preference maps. The collapsed maps use the natural breaks scheme

PAGE 206

206 to divide the preference numbers, which are given th e probability values between 0 and 1, into four categories. This study has tested different collapsing methods: the natural breaks method, the manual method, the equal interval method, and the standard deviation method; the natural breaks method handles a bimodal or multimodal distribution (Carr and Zwick, 2007) and creates four collapsed groups. The other three collapsing methods yield only three or even two collapsed groups. Consequently, the natural breaks method is adopted, which brings forth the four g roups symbolized namely by 1, 2, 3, and 4, in which 4 is the highest score representing the highest probability for development and 1 is the lowest score suggesting the lowest probability. Thus, the continuous numbers in the preference maps become discrete in the collapsing maps. This collapsing process establishes a basis for the creation of the conflict maps which will be introduced in the next section. Figure 4 8 illustrates the example of a preference map, and Figure 4 9 illustrates a sample of a collap sed map for the BAU scenario. Creation of the conflict maps Once the collapsed maps are prepared for the five scenarios, the follow up effort is to create conflict maps. Because this study has proposed two masks for different development scenarios, the con flict maps will have the same two equivalent mask types also: one is for the BAU scenario; the other is for the infill scenario as well as for the other three scenarios. When making a conflict map, collapsed maps for each urban use are combined by taking t he score of the single family use multiplied by 1,000 plus the score of the multi family use multiplied by 100 plus the score of the commercial institutional transportation use multiplied by 10 and plus the score of the industrial warehouses use multiplied by 1. For example, for the BAU scenario, this study has created four collapsed maps for each urban use. The single family use has its own collapsed map while the other three urban uses all have their own collapsed maps. When they are

PAGE 207

207 combined to create a conflict map, they are input into the Single Output Map Algebra function in ArcGIS 9.3 to fulfill this task by making: single family 1,000 + multi family 100 + commercial institutional transportation 10 + industrial warehouses 1 The final scores for the BAU scenario are presented in Table 4 24, and the final scores for the infill scenario as well as for the other three scenarios are presented in Table 4 25 because these three scenarios have a same conflict score. From the two tabl es, it is clear that the BAU scenario has a lot more vacant land for development compared to the infill scenario as well as the other three development scenarios because the BAU scenario has more land that comes to participate in the allocation process. Sp ecifically, for the BAU scenario, the allocation directly goes to the conflict areas in order to seek vacant land. For the infill scenario, however, the infill development will first fill out the infill lots that are already set aside as vacant land for a typical Finally, when the infill lots are filled up, the development will go to the conflict areas so as to allocate land, according to the conflict scores a development will receive. Therefore, different from the BAU scenario, for the infill development, the sequence begins with infill allocation and end with the conflict allocation. For the increased density development, the development will allocate land i n the infill lots that are either located within 1,000 meters of the urban buffer areas or outside. When this land is filled up, then it will go to the conflict areas to seek vacant land, which are typically located inside the urban buffer areas. This is t he same as for the redevelopment scenario in which the land within the infill lots as well as the 1,000 meters of urban buffer areas is filled up first, before going to the conflict areas to find available land. For the conservation scenario, however, it d oes not go to the infill vacant lots as a first step but goes

PAGE 208

208 to the conflict areas directly in order to seek available land. These are the major differences when these five scenarios are compared and allocate urban land. When the conflict areas are alloca ted for a specific scenario, it will first go to the specific urban use that dominates in the conflict map. For example, when a score of 4211 is suggested (Table 4 24), it is a single family dominated area. As a result, the development pattern will be the single family development. When all the areas dominated by single family are filled up, then the allocation process will go and seek the single family conflict areas with other urban uses. When these amounts of land are consumed also, the allocation proces s will finally utilize all of the conflict areas for all four urban uses. The commercial institutional transportation development in this study is the case in which the development finally finds the scores of 1111 in both the two conflict types because the commercial institutional transportation dominant areas and the areas of the commercial institutional transportation use in minor conflict with other uses cannot satisfy the development needs. As a result, it goes to the all low preference all in conflict areas to look for needed cells. Generally speaking, for a proper allocation of urban uses, it is imperative to follow the development sequence that single family dominates over multi family, multi family dominates over commercial institutional transportati on, and commercial institutional transportation dominates over industrial warehouses. These allocations all put single family as the first priority and the remaining urban uses in secondary status. For example, based on this rule, if a single family use is in conflict with other urban uses, the cells belong to the single family use. Once single family acreage demands are satisfied, then, the cells in conflict can be allocated to other uses.

PAGE 209

209 From Tables 4 24 and 4 25, it is obvious that, generally speaking the BAU scenario has adequate land for development while the infill scenario and the other three scenarios have reduced approximately 50 percent of the total acres of land. In particular, the commercial institutional transportation use dominating areas ( i.e., 1231, 1141, 1132, 1131, and 1121) are not sufficiently large enough to meet the development needs. As a result, they will have to be allocated from 1111, the all low preference all in conflict areas. Figures 4 10 and 4 11 are the conflict maps for th e BAU scenario as well as the infill development scenario. Final Allocation Timeframes of the final allocation The timeframes of the final allocation follow the timeframes of the foresting methods, in which 2020 and 2030 are proposed. Because of the two ti meframes, the final allocation will offer two maps for each scenario: namely, 2020 allocation, and 2030 allocation. These allocations are basically the results of the conflict maps as well as the infill parcels that are derived from the above conflict anal the urban uses being allocated based on the five scenarios in accordance with the forecasts of the urban development acreages for each urban use and for each scenario. Because t he 2010 allocation map depicts the LULC classe s in 2010 that have already been passed, the 2010 directly. Therefore, a total of twenty two maps are created based on th e five scenarios, with each scenario having two maps for a year, i.e., 2020 or 2030, and two maps for the year 2010, which is the same for all five scenarios in that year (see Figures 4 12 through 4 33). Urban uses to be allocated in final allocation The urban uses that will be allocated are the four uses mentioned previously: namely, the single family use, the multi family use, the commercial institutional transportation use, and the

PAGE 210

210 industrial warehouses use. Specifically, mixed use is allocated in the r edevelopment scenario as well as in the conservation scenario. Consequently, these two scenarios contain the mixed use in their final allocation maps in addition to the four urban uses mentioned above. In addition to the urban uses, natural LULC classe s su ch as grasslands, forests, agricultural, recreational others, wetlands, water, and barren will be mapped also. Specifically for the barren land, landfill is suggested by this natural LULC. In addition to the mixed use, existing and future conservation land is marked out as a result in the conservation scenario maps. Final allocation sequence The allocation sequence follows the scenario sequence described in the forecasting section. The first scenario that will be allocated is the BAU scenario; the second t o be allocated is the infill development scenario; the third is the increased density development scenario; the fourth is the redevelopment scenario; and the fifth is the conservation scenario. Based on the allocation forecasting method for each scenario, the overall land consumptions that are forecasted for the five scenarios are actually decreasing, and they are mostly located within the urban buffer areas. This development pattern presents an optimistic form for curbing urban sprawl in Alachua County wit h respect to urban land acreage consumptions as well as urban development locations. Final allocation cell size and spatial reference The raster maps adopt a 30 30 meters resolution, in accordance with the resolution of the remote sensing datasets. As a result, the computation of cell counts and cell acreages are based on this resolution. One cell, i.e., 30 meter 30 meter, is 900 square meters, which is equivalent to 0.22 acres. Therefore, the acreage computation is based on this conversion rule to cal culate to and fro from the metric square meters to the English acres and vice versa. In particular, the spatial reference for the allocation maps is: NAD_1983_UTM_Zone_17N. The datum is: D_North_American_1983.

PAGE 211

211 Final allocation preference map slicing, polic y allocation, and cell statistics Once the preference maps are masked, they need to be transformed from the continuous data to the discrete data so that cells with certain values can be extracted, calculated, and allocated. The continuous data are the data that contain floating points that are continuously stretched between 0 and 1. There are no unique values that can be derived from the value of a certain cell in the continuous data. As a result, the continuous data do not have an attribute table in ArcGIS 9.3. The discrete data, however, incorporate integers, which have unique values for cells. They have attribute tables for the values of all cells. This study adopts a technique called ing a large area into a number of smaller equal zones (Carr and Zwick, 2007) for allocations based on their cell values. As a result, the cells are extracted and grouped according to the magnitude of their values. The biggest preference values are assigned to the biggest sliced values (Carr and Zwick, 2007) when they are being allocated after slicing. As a result, the highest sliced values (i.e., the highest preference values) are the first to be allocated, and then follows the cells with smaller numbers so that the smaller numbers are allocated later in a secondary status, and the smallest are allocated in the last. The common slicing method has two alternatives: one is to use the Slice tool in ArcGIS 9.3; the other is to use Single Output Map Algebra by m aking a preference map and multiplying a certain value, e.g., 100,000. Because the Slice tool in ArcGIS 9.3 sometimes does not create sufficiently small units for allocation, which the allocation process prefers, this study uses the second method to slice certain areas in a preference map because this second method can potentially create sufficiently small units of land by multiplying a sufficiently large constant, e.g., 100,000, in this study, with the preference maps. Overall, the slice technique can be u sed to extract discrete values from a continuous dataset, e.g., a preference map, which need conflict

PAGE 212

212 areas or infill lots as masks in order to seize those masked cells for allocation. In this case, the vector infill parcel map must be converted to a raste r dataset first. This study applies the slice technique (the second method) to slice preference maps for certain areas covered by cells in conflict or infill lots. When a cell is still large after slicing, this study adopts a policy allocation method to f urther extract necessary cells from it. For example, an area contains 12,070 cells after slicing and only 7,000 or so cells are needed based on the development forecasts. This study utilizes a policy allocation method to allocate these 7,000 cells, in whic h cells that are located within a jurisdiction with clear geographic boundaries are chosen as high priority compared to cells that urban development. In the above example, within 1,000 meters of urban buffer area is taken for development, which yields 7,132 cells, and these 7,132 cells are what should be allocated. For the areas that are overlapped, e.g., industrial warehouses infill development overlaps wit h the commercial institutional transportation conflict areas, this study adopts a Cell Statistics tool to extract overlapped cells so that the formal counts of industrial warehouses infill cells as well as the commercial institutional transportation acreag es can exclude the overlapping cells. The Single Output Map Algebra module can also be used in this regard. This study uses cell statistics to accomplish this goal, however. Final Allocation Results BAU S cenario As mentioned previously, because the year 2010 has passed, the allocation for 2010 has the result, the allocation map for 2010 is a direct count from the parcel d ata for the four urban uses that reflect the parcel acreages and their changes throughout the years. This 2010 map, titled

PAGE 213

213 reality, the development for 2020 an d 2030 for all five scenarios starts from this map. Figures 4 12 through 4 17 illustrate the allocation based on t he parcel maps from the county the BAU scenario for 2020 and 2030, respectively. From the map, it is evident that the urban sprawl has occurred from 2010 because cells have been scattered throughout the county starting from 2010 although the level of the sprawl is not very severe in that year. The 2020 and 2030 BAU maps show a continued sprawled development pattern, which is quite serious because the leapfrog development continues outside urban buffer areas in the greenfield areas where the greenfield land is gradually being filled out to become urbaniz ed areas in the outskirts of Alachua County, outside the urban buffer areas. In 2010, the leapfrog single family development outside urban buffer areas is largely located in the areas of the College West Estates, Scattered Oaks Estates, Sunny Meadows, and Windy Acres. It is also located in the area of Windy Hills; Pinesville; the area of the University Country Estates; Wacahoota; the area of the Grassy Lake Estates; and the Monteocha vicinity. The leapfrog urban development in the greenfield starts to grow from these areas in 2020 and in 2030; as a result, the single family development continues in the above areas for the above years. The growth becomes larger in area and more concentrated, which spreads to nearby parcels and fills the gaps between the urban buffer areas in the greenfield areas in 2020 and 2030. Even the areas in east Alachua County, where the single family development is minor in 2010, have incredible growth in the areas such as the Wacahoota vicinity, the Micanopy vicinity, the Winsor vicin ity, the Melrose vicinity, the Monteocha vicinity, the Wood Meadows Neighborhood, the Traxler vicinity, the Bellamy Road Estates, and the Bellamy Forest Neighborhood in 2020. Besides, additional single family growth

PAGE 214

214 is scattered around in the outskirt of t he county other than the development in the above areas in 2020. In 2030, the single family growth pattern continues to fill out the gaps between the urban buffer areas, and the vacant land in the gap areas are largely filled out. However, there are still significant areas inside the urban buffer areas that are vacant, which are suitable for development but are not developed. Overall, it is obvious that for the BAU urban development, the growth pattern is a sprawled type, which is characterized as severe le apfrog development in the greenfield areas outside the urban buffer areas. Another noticeable development is the commercial institutional transportation development, which is developed along State Route 441 on the north side of Gainesville as well as south of the Gainesville Regional Airport next to the Alachua County Agricultural Center and Fairgrounds. The area that is in the Jonesville vicinity south of NW 39th Avenue is also a big commercial institutional transportation development. Nevertheless, this d evelopment is all within the urban buffer areas and there is no sprawled pattern outside the urban buffer areas for the development of this land use. Tables 4 26 through 4 29 illustrate the developed acreage of land in 2010, 2020, and 2030, respectively, w ith comparisons between the allocated acreages and the demanded acreages in those years. Infill S cenario Figures 4 18 through 4 21 describe allocations for the infill development scenario. The infill scenario causes the development to occur in the infill parcels before the acreages are allocated to the conflict areas. The infill acreages can either be located within 1,000 meters of urban buffer areas or outside the areas. The infill scenario begins in 2020 because the infill scenario allocation in 2010 sh ares the same map with the BAU scenario. The single family development in the infill scenario for 2020 is all located within the infill areas, which have 57,938 cells, equivalent to 12,885 acres (Table 4 31). The single family development mostly

PAGE 215

215 occurs wit hin the 1,000 meters of urban buffer areas in 2020. They are basically located in the nodal areas where major transportation corridors intersect. These nodal areas include the City of Alachua, the City of High Springs, the City of Newberry, the City of Arc her, the City of Waldo, and the City of Hawthorne as Figure 4 2 shows. The areas in east Alachua County and west of family development. Because most of the single family development occurs within the urban buffe r areas, not much leapfrog development is evident in the infill scenario in 2020. There are a few, minor acreages of the commercial institutional transportation use that are scattered outside the urban buffer areas, however. Multi family and industrial war ehouses uses are all located within 1,000 meters of urban buffer areas. Leapfrog development is quite evident for the single family development in 2030. The leapfrog single family development occurs fairly intensively in the greenfield areas outside urban buffer areas in 2030. Examples include the area of County Route 232 intersecting with County Route 235; the area of State Route 121 intersecting with NW 156 Avenue; the area of NE 156 Avenue intersecting with NE 39th Street; the area of County Route 1469 i ntersecting with NE US 301; the area of County Route 234 intersecting with County Route 1474; and the area of SW 137 Avenue intersecting with SW 91 Street. Figure 4 roads. The single family infill scenario for 2030 allocates a to tal of 25,990 cells, equivalent to 5,780 acres (Table 4 30). The multi family development allows some acreages to occur in the infill areas and some in the conflict areas in 2020, but when the infill acreages are used up in 2020, the development, therefore, goes to the conflict areas (Table 4 32). For the commercial institutional transportation use, there is some development (about 45 percent) in the infill areas, and when it is used up in 2020, the development goes to the conflict areas in 2030. The commercial institutional

PAGE 216

216 transportation use is all located within urban buffer areas and all within the conflict are as in 2030, however. All the industrial warehouses development happens in the infill areas both for 2020 and 2030 (Table 4 30). The four urban uses with their development in the infill areas as well as in the conflict areas are illustrated in Table 4 30. T he four urban uses with their acreage demand and allocation are illustrated in Tables 4 31 through 4 34. Increased Density Development Scenario urban development acreage s for the four urban development patterns. The single family development is characterized as the infill development in 2020 and 2030, but with inward development patterns inside the 1,000 meters of urban buffer areas. Most of the multi family development i s located within conflict areas for 2020, and the multi family development all occurs in the conflict areas in 2030 after the development consumes all the infill acreages in 2020. For the commercial institutional transportation development, the development pattern is infill oriented with a few being located in the conflict areas in 2020. The commercial institutional transportation development in 2030 is mostly located in the conflict areas when the infill acreages are filled out. The industrial warehouses d evelopment is all located in the infill areas in both 2020 and 2030. The proposed development acreages in the infill areas as well as in the conflict areas in 2020 and 2030 are shown in Table 4 35. The allocated land versus demanded land for the four urban uses is shown in Tables 4 36 through 4 39. Increased single family use in 2020 and 2030 all occurs within the urban buffer areas, which means no leapfrog sprawl takes place during the above period. The same is true for the multi family use, the commercial institutional transportation use, and the industrial warehouses use: all of the development happens within urban buffer areas. Apart from some leapfrog development in 2010, the urban development in the increased development scenario in

PAGE 217

217 2020 and 2030 is al l located within the urban buffer areas, with no leapfrog development in those years. Figures 4 22 through 4 25 show that growth in 2020 happens in the central city the City of Gainesville as well as in the nodal areas where major road corridors intersec t. These nodal areas include the City of Alachua, the City of High Springs, and the City of Newberry (Figure 4 2). The growth inside the Urban Cluster areas in 2020 is largely located in east Gainesville close 22 through 4 2 5). Growth in 2030 continues the pattern in 2020, but more intensively in the City of Gainesville as well as in the nodal areas. The growth fills in the gap areas in east Gainesville, outside 1,000 meters of urban buffer areas, in north Gainesville, in wes t Gainesville, in south Gainesville, as well as around the airport areas. Significant amounts of development are evident in the above areas; they are fairly intensive although they are all located within the urban buffer areas. Redevelopment Scenario The redevelopment scenario further shrinks the development acreages for the four urban uses similar to the increased density development scenario. The redevelopment allocation combines the redevelopment scenario, the infill scenario, the increased density deve lopment scenario, and the mixed use together. Because there is 15 percent of urban development occurring within the central city the City of Gainesville as well as the Urban Cluster areas for redevelopment, the new urban development generally occurs within the central city, in the Urban Cluster areas, as well as in the nodal areas where major transportation corridors intersect such as the City of Alachua (Figure 4 2). Most of the redevelopment growth is located within the City of Gainesville city limits as well as in the county Urban Cluster areas in 2020 and 2030, respectively. In 2020, the single family development is largely located within both the Urban Cluster areas in east Gainesville and west Gainesville. The multi family development, in 2020, is

PAGE 218

218 loca ted in the Urban Cluster areas as well in both east Gainesville and west Gainesville, although the development is not much. The commercial institutional transportation use, in 2020, is largely located within west Gainesville, in the nodal areas around the City of Alachua, and along State Route 441. The industrial warehouses use is located at the intersection of State Route 441 and NW 34th Street in 2020. The above urban development is all located within the urban buffer areas, however, with no sprawl in 202 0, although there is some sprawled development in 2010. In 2030, the single family development continues to grow in the Urban Cluster areas, most growth occurs within the nodal areas of the City of High Springs instead of the City of Alachua. In 2030, multi family development grows within west Gainesville (on the west side of the Urban Cluster areas). The amount of development is still not very much compared to t he amount of development in 2020. In 2030, the major commercial institutional transportation development is located in the area where Waldo Road intersects with NE 39th Avenue. Also, some development is located north of NE 39th Avenue and east of the North Main Street area as well as south of State Route 441 and west of NW 43rd Street. In 2030, the industrial warehouses use is largely located at the intersection of NW County Route 235 and NW 173 Street as well as in the area north of the Gainesville Regiona l Airport (Figure 4 2). In 2020, all single family development occurs in infill parcels within 1,000 meters of the urban buffer areas. Most of the multi family development occurs within the conflict areas with a small amount of multi family development (54 acres) in the infill areas. Most of the commercial institutional transportation use is located within the infill areas while a small fraction of the commercial institutional transportation use (78 acres) is located in the conflict area. The industrial war ehouses use is all situated within the infill areas, however.

PAGE 219

219 In 2030, single family development still occupies the infill areas because the infill parcels for single family development, when the infill m ulti family cells are filled up in 2020, the development in 2030 receives land in the conflict areas. The commercial institutional transportation occupies both the infill areas and the conflict areas in 2030 because there is still some land available for i nfilling (583 acres). Same as 2020, the industrial warehouses use takes all infill land based on the development sequence that infill development is prioritized over the development in the conflict areas. The mixed use development usually occurs inside the commercial institutional transportation use, the multi family development, and along major transportation corridors. This study imitates land that is 200 meters away from the major transportation corridors to be seized for mixed use development, in which they are also located inside a typical commercial institutional transportation use as mixed use land. As a result, they are basically located within the central city the City of Gainesville in the Urban Cluster areas, or along State Route 441 for 2020 and 2030. In the City of Gainesville, the mixed use land is located along these transportation corridors: NW 53rd Avenue, NE 39th Avenue, North Main Street, NE 31st Avenue, West Newberry Road, NW 23rd Avenue, SW 24th Avenue, SW 75th Street, and SW Archer Road (Figure 4 2). The mixed use land occupies a total of 687 acres in 2020 and it develops to a total of 2,029 acres in 2030. Table 4 40 presents the acreages that are located within infill areas as well as in the conflict areas for the four urban uses in 2020 and 2030, respectively. Tables 4 41 through 4 44 illustrate the urban acreages demand verses urban acreages allocation for the four urban uses in 2020 and 2030, respectively. Figures 4 26 through 4 29 show the urban redevelopment in 2020 and 2030, respect ively.

PAGE 220

220 Conservation Scenario so that land to be allocated is smaller in acreage compared to the BAU scenario, the infill scenario, and the increased density develop ment scenario so as to curb urban sprawl. This scenario adopts a 15 percent redevelopment benchmark, same as the value for the redevelopment scenario. However, different from the infill scenario, the increased density scenario, and the redevelopment scenar io, in which infill land is taken priority over the land in the conflict areas, the conservation uses land only from the conflict areas. As a result, the total development acreages for the conservation scenario for the four urban uses can be located either within the 1,000 meters of urban buffer areas or outside. Because all land is located within conflict areas, the conservation scenario does not have overlaps because it is already separated in the conflict analysis where land is divided into categories ba sed on different conflict scores each urban cell receives (Table 4 25). Figures 4 30 through 4 33 illustrate the urban development in 2020 and 2030, respectively, based on the conservation scenario. Taking an in depth look into the allocation map for the conservation scenario, most of the single family development (80 percent) will be located inside the 1,000 meters of urban buffer areas in 2020 although there is some single family development that is scattered outside the urban buffer areas. Furthermore, not much single family development occurs within the City of Gainesville city limits, most of which, however, is located in the Urban Cluster areas as well as within the city limits of other incorporated cities in the county such as the City of Alachua, t he City of High Springs, the City of Newberry, and the City of Archer (Figure 4 2). In 2020, about 83 percent of the multi family development is situated within 1,000 meters of urban buffer areas although the multi family development is not much, which has only a total of 247 acres in 2020. For the commercial institutional transportation use,

PAGE 221

221 however, 99 percent of the development occurs within 1,000 meters of the urban buffer areas, most of which is located within the City of Gainesville city limits. Most of the industrial warehouses development takes place along State Route 441 and NW County Route 235 inside 1,000 meters of the urban buffer areas, which occupies about 63 percent of the total industrial warehouses acreages. Overall, the urban sprawl with ou tward leapfrog development outside 1,000 meters of urban buffer areas is not severe in the conservation scenario compared to the BAU scenario, in that both scenarios take land from the conflict areas without infill cells, because most of the development in the conservation scenario allocates land inside 1,000 meters of urban buffer areas. For 2030, about 57 percent of the single family development happens within 1,000 meters of urban buffer areas. Scattered single family development is evident outside 1,0 00 meters of the urban buffer areas where they are concentrated in the gap areas surrounded by the urban buffer areas; however, their locations are outside 1,000 meters of urban buffer areas. For the multi family development, about 86 percent of the develo pment is located within 1,000 meters of the urban buffer areas and they are mostly situated near SW 20th Avenue and NW 39th Avenue in 2030 (Figure 4 2). The commercial institutional transportation development is all located inside the 1,000 meters of urban buffer areas in 2030, where their sites are on the north side of the Gainesville Regional Airport, south of NE 39th Avenue, intersection of SW 24th Avenue and SW 75th Street, intersection of State Route 121 and SW 63rd Boulevard, north of SW 63rd Avenue, intersection of NW 43rd Street and State Route 441, and intersection of State Route 121 and County Route 231 (Figure 4 2). The industrial warehouses development has 46 percent of land developed within 1,000 meters of the urban buffer areas in 2030. The maj or development outside 1,000 meters of the urban buffer areas for the industrial warehouses use is in the Cadillac

PAGE 222

222 vicinity. From the above comparisons, it is apparent that the single family development has the most prominent leapfrog development which con stitutes urban sprawl because of the single family use in terms of scattered and leapfrog developments, however. Like the redevelopment scenario, the conservation scenario has also mixed use development. The mixed use in the conservation scenario is essentially the same as the redevelopment scenario, most of which are located inside a commercial institutional transportation use as well as along major transportation corridors. These major transportation corridors accountable for mixed uses are NW 53rd Avenue, NE 39th Avenue, North Main Street, NE 31st Avenue, West Newberry Road, NW 23rd Avenue, SW 24th Avenue, SW 75th Street, and SW Archer Road (Figure 4 2). Same as the redevelopment scenario, the mixed use land in the conservation scenario has a total of 687 acres in 2020 and a total of 2,029 acres in 2030. Table 4 45 illustrates the acreages of development that are located inside the 1,000 meters of urban buffer are as and outside based on 2020 and 2030, respectively. Tables 4 46 through 4 49 illustrate the acreages demanded and allocated for the four urban uses in 2010, 2020, and 2030, respectively, in terms of the conservation scenario.

PAGE 223

223 Table 4 1. The baseline forecasting method for urban development acreages (single family) Baseline Year Single Family Acres Single Family Population (People) Gross Single Family Density (People/Acre) Changes (Acres) Single Family 2003 48,673 121,415 2.49 -2010 60,899 131,732 2.16 12,226 2020 73,739 159,507 2.16 12,840 2030 79,455 171,871 2.16 5,716 Table 4 2. The baseline forecasting method for urban development acreages (multi family) Baseline Year Multi Family Acres Multi Family Population (People) Gross Multi Family Density (People/Acre) Changes (Acres) Multi Family 2003 3,469 83,699 24.13 -2010 3,559 101,684 28.57 90 2020 4,431 126,593 28.57 872 2030 5,055 144,429 28.57 624 Table 4 3. The baseline forecasting method for urban development acreages (total residential) Baseline Year Total Acres Total Population (People) Total Gross Residential Density (People/Acre) Changes (Acres) Total Residential 2003 52,142 205,114* 3.93 -2010 64,458 233,416* 3.62 12,317 2020 78,170 286,100** 3.66 13,712 2030 84,510 316,300** 3.74 6,340 (Source *: U.S. Census Bureau (1980, 1990, 2000, 2010); **: Medium n umber (BEBR, March 200 9))

PAGE 224

224 Table 4 4. The baseline forecasting method for urban development acreages (commercial institutional transportation) Baseline Year Tota l Acres Commercia l Institutional Transportation Employment (People) Gross Commercial Institutional Transportation Density (People/Acre) Changes (Acres) Commercial Institutional Transportation 2003 22,010 68,042*** 3.09 -2010 22,167 68,497 3.09 157 2020 27,891 86,183 3.09 5,724 2030 32,530 100,518 3.09 4,639 (Source: ***: BEBR (2003)) Table 4 5. The baseline forecasting method for urban development acreages (industrial warehouses) Baseline Year Total Acres Industrial Warehouses Employment (People) Gross Industrial Warehouses Density (People/Acre) Changes (Acres) Industrial Warehouses 2003 2,819 9,130*** 3.24 -2010 2,823 9,145 3.24 4 2020 3,576 11,586 3.24 753 2030 4,170 13,511 3.24 594 (Source: ***: BEBR (2003)) Table 4 6. The baseline forecasting method for urban development acreages (total urban) Baseline Year Total Acres Total Population (People) Total Gross Urban Density (People/Acre) Changes (Acres) Total Urban 2003 76,971 205,114* 2.66 -2010 89,448 233,416* 2.61 12,477 2020 109,637 286,100** 2.61 20,189 2030 121,210 316,300** 2.61 11,573 (Source *: U.S. Census Bureau (1980, 1990, 2000, 2010); **: Medium n umber (BEBR, March 2009))

PAGE 225

225 Table 4 7. The baseline forecasting method for urban acreage changes Year Sing le Family Multi Family Commercial Institutional Transportation Industrial Warehouses Acres Cells Acres Cells Acres Cells Acres Cells 2010 12,226 54,976 90 404 157 706 4 18 2020 12,840 57,738 872 3,920 5,724 25,737 753 3,388 2030 5,716 25,702 624 2,807 4,639 20,860 594 2,671 Total 30,782 138,415 1,586 7,131 10,520 47,303 1,351 6,077 Table 4 8. Baseline urban acreage changes with percentages Urban Acreage Changes (Acres) Percentage Of Urban Acreage Change (%) Year 2010 2020 2030 2010 2020 2030 Single Family 12,226 12,840 5,716 97.99 % 63.60 % 49.39 % Multi Family 90 872 624 0.7 2 % 4.32 % 5.39 % Residential 12,317 13,712 6,340 98.71% 67.92% 54.78% Commercial Institutional Transportation 157 5,724 4,639 1.26% 28.35% 40.08% Industrial Warehouses 4 753 594 0.03% 3.73% 5.13% Total 12,478 20,189 11,573 100.00% 100.00% 100.00% Table 4 9. The increased density forecasting method for urban development acreages (single family) Increased Density Year Single Family Acres Single Family Population (People) Gross Single Family Density (People/Acre) Changes (Acres) Single Family 2003 48,673 121,415 2.49 -2010 60,899 131,732 2.16 12,226 2020 64,663 159,507 2.47 3,764 2030 69,635 171,871 2.47 4,972 (Source *: U.S. Census B ureau (1980, 1990, 2000, 2010))

PAGE 226

226 Table 4 10. The increased density forecasting method for urban development acreages (multi family) Increased Density Year Multi Family Acres Multi Family Population (People) Gross Multi Family Density (People/Acre) Changes (Acres) Multi Family 2003 3,469 83,699 24.13 -2010 3,559 101,684 28.57 90 2020 3,815 126,593 33.18 256 2030 4,358 144,429 33.14 543 (Source *: U.S. Census Bureau (1980, 1990, 2000, 2010)) Table 4 11. The increased density forecasting method for urban development acreages (total residential) Increased Density Year Total Acres Total Population (People) Total Gross Residential Density (People/Acre) Changes (Acres) Total Residential 2003 52,142 205,114 3.93 -2010 64,458 233,416 3.62 12,317 2020 68,478 286,100 ** 4.18 4,020 2030 73,993 316,300 ** 4.27 5,515 (Source *: U.S. Census Bureau (1980, 1990, 2000, 2010); **: Medium n umber (BEBR, March 2009)) Table 4 12. The increased density forecasting method for urban development acreages (commercial institutional transportation) Increased Density Year Total Acres Commercial Institutional Transportation Employment (People) Gross Commercial Institutional Transportation Density (People/Acre) Changes (Acres) Commercial Institutional Transportation 2003 22,010 68,042 *** 3.09 -2010 22,167 68,497 3.09 157 2020 23,845 86,183 3.61 1,678 2030 27,881 100,518 3.61 4,035 (Source: ***: BEBR (2003))

PAGE 227

227 Table 4 13. The increased density forecasting method for urban development acreages (industrial warehouses) Increased Density Year Total Acres Industrial Warehouses Employment (People) Gross Industrial Warehouses Density (People/Acre) Changes (Acres) Industrial Warehouses 2003 2,819 9,130 *** 3.24 -2010 2,823 9,145 3.24 4 2020 3,043 11,586 3.81 221 2030 3,560 13,511 3.80 517 (Source: ***: BEBR (2003)) Table 4 14. The increased density forecasting method for urban development acreages (total urban) Increased Density Year Total Acres Total Population (People) Total Gross Urban Density (People/Acre) Changes (Acres) Total Urban 2003 76,971 205,114 2.66 -2010 89,448 233,416 2.61 12,477 2020 95,367 286,100 ** 3.00 5,919 2030 105,433 316,300 ** 3.00 10,067 (Source *: U.S. Census Bureau (1980, 1990, 2000, 2010); **: Medium n umber (BEBR, March 2009)) Table 4 15. The increased density forecasting method for urban acreage changes Year Single Family Multi Family Commercial Institutional Transportation Industrial Warehouses Acres Cells Acres Cells Acres Cells Acres Cells 2010 12,226 54,976 90 404 157 706 4 18 2020 3,764 16,924 256 1,151 1,678 7,545 221 993 2030 4,972 22,356 543 2,442 4,035 18,145 517 2,323 Total 20,962 94,257 889 3,997 5,870 26,396 742 3,335

PAGE 228

228 Table 4 16. The redevelopment forecasting method for urban development acreages (single family) Redevelop ment Year Single Family Acres Single Family Population (People) Gross Single Family Density (People/Acre) Changes (Acres) After 15% R edevelop ment Single Family 2003 48,673 121,415 2.49 7,448 -2010 60,899 131,732 2.16 12,226 12,226 2020 64,098 159,507 2.49 3,764 3,199 2030 68,324 171,871 2.52 4,972 4,226 (Source *: U.S. Census Bureau (1980, 1990, 2000, 2010)) Table 4 17. The redevelopment forecasting method for urban development acreages (multi family) Redevelop m ent Year Multi Family Acres Multi Family Population Gross Multi Family Density (People/Acre) Changes (Acres) After 15% Redevelopm ent Multi Family 2003 3,469 83,699 24.13 --2010 3,559 101,684 28.57 90 90 2020 3,776 126,593 33.52 256 218 2030 4,238 144,429 34.08 543 462 (Source *: U.S. Census Bureau (1980, 1990, 2000, 2010)) Table 4 18. The redevelopment forecasting method for urban development acreages (total residential) Redevelop m ent Year Total Acres Total Population (People) Total Gross Residential Density (People/Acre) Changes (Acres) After 15% Redevelopm ent Total Residential 2003 52,142 205,114 3.93 --2010 64,458 233,416 3.62 12,317 12,317 2020 67,875 286,100 ** 4.22 4,020 3,417 2030 72,562 316,300 ** 4.36 5,515 4,688 (Source *: U.S. Census Bureau (1980, 1990, 2000, 2010); **: Medium n umber (BEBR, March 2009))

PAGE 229

229 Table 4 19. The redevelopment forecasting method for urban development acreages (commercial institutional transportation) Redevelopment Year Total Acres Commercial Institutional Transportation Employment (People) Gross Commercial Institutional Transportation Density (People/Acre) Changes (Acres) After 15% Redevelopment Commercial Institutional Transportation 2003 22,010 68,042 *** 3.09 --2010 22,167 68,497 3.09 157 157 2020 23,594 86,183 3.65 1,678 1,426 2030 27,024 100,518 3.72 4,035 3,430 (Source ***: BEBR (2003)) Table 4 20. The redevelopment forecasting method for urban development acreages (industrial warehouses) Redevelopment Year Total Acres Industrial Warehouses Employment (People) Gross Industrial Warehouses Density (People/Acre) Changes (Acres) After 15% Redevelopment Industrial Warehouses 2003 2,819 9,130 *** 3.24 --2010 2,823 9,145 3.24 4 4 2020 3,010 11,586 3.85 221 188 2030 3,449 13,511 3.92 517 439 (Source ***: BEBR (2003))

PAGE 230

230 Tab le 4 21. The redevelopment forecasting method for urban development acreages (total urban) Redevelop ment Year Total Acres Total Population (People) Total Gross Urban Density (People/Acre) Changes (Acres) After 15% Redevelopment Total Urban 2003 76,971 205,114* 2.66 8,650 8,650 2010 89,448 233,416 2.61 12,477 12,477 2020 94,479 286,100 ** 3.03 5,919 5,031 2030 103,036 316,300 ** 3.07 10,067 8,557 (Source *: U.S. Census Bureau (1980, 1990, 2000, 2010); **: Medium n umber (BEBR, March 2009)) Table 4 22. The redevelopment forecasting method for urban acreage changes Year Single Family Multi Family Commercial Institutional Transportation Indu strial Warehouses Acres Cells Acres Cells Acres Cells Acres Cells 2010 12,226 54,976 90 404 157 706 4 18 2020 3,199 14,386 218 978 1,426 6,413 188 844 2030 4,226 19,003 462 2,075 3,430 15,423 439 1,975 Total 19,651 88,365 769 3,458 5,013 22,542 631 2,837

PAGE 231

231 Table 4 23. Conflict scores and equivalent descriptions Conflict Score Conflict Score Descriptions 4444 all high preference all in conflict 4443 high preference single family conflicts with high preference multi family and high preference commercial institutional transportation 4442 high preference single family conflicts with high preference multi family and high preference commercial institutional transportation 4441 high preference single family conflicts with high preference multi family and high preference commercial institutional transportation 4434 high preference single family conflicts with high preference multi fam ily and high preference industrial warehouses 4424 high preference single family conflicts with high preference multi family and high preference industrial warehouses 4414 high preference single family conflicts with high preference multi family and high preference industrial warehouses 4144 high preference single family conflicts with high preference commercial institutional transportation and high preference industrial warehouses 4244 high preference single family conflicts with high preference commercial institutional transportation and high preference industrial warehouses 4344 high preference single family conflicts with high preference commercial institutional transportation and high preference industrial warehouses 4433 high preference single family conflicts with high preference multi family 4432 high preference single family conflicts with high preference multi family 4431 high preference single family conflicts with high preference multi family 4423 high preferen ce single family conflicts with high preference multi family 4422 high preference single family conflicts with high preference multi family 4421 high preference single family conflicts with high preference multi family 4413 high preference single family conflicts with high preference multi family 4412 high preference single family conflicts with high preference multi family 4411 high preference single family conflicts with high preference multi family 4341 high preferen ce single family conflicts with high preference commercial institutional transportation 4241 high preference single family conflicts with high preference commercial institutional transportation 4141 high preference single family conflicts with high preference commercial institutional transportation 4341 high preference single family conflicts with high preference commercial institutional transportation 4342 high preference single family conflicts with high preference commercial institutional transp ortation 4343 high preference single family conflicts with high preference commercial institutional transportation 4334 high preference single family conflicts with high preference commercial industrial warehouses

PAGE 232

232 Table 4 23. Continued Conflict Conflict Score Descriptions Score 4324 high preference single family conflicts with high preference commercial industrial warehouses 4314 high preference single family conflicts with high preference commercial industrial warehouses 4234 high preference single family conflicts with high preference commercial industrial warehouses 4224 high preference single family conflicts with high preference commercial industrial warehouses 4214 high preference single family conflicts with high preference commercial industrial warehouses 4134 high preference single family conflicts with high preference commercial industrial warehouses 4124 high preference single family conflicts with high preference commercial industrial warehouses 4114 high preference s ingle family conflicts with high preference commercial industrial warehouses 4333 single family preference dominates 4332 single family preference dominates 4331 single family preference dominates 4323 single family preference dominates 4322 single family preference dominates 4321 single family preference dominates 4313 single family preference dominates 4312 single family preference dominates 4311 single family preference dominates 4211 single family preference dominates 4212 single family preference dominates 4213 single family preference dominates 4221 single family preference dominates 4222 single family preference dominates 4223 single family preference dominates 4231 single family preference dominates 4232 single family preference dominates 4233 single family preference dominates 4111 single family preference dominates 4112 single family preference dominates 4113 single family preference dominates 4121 single family preference dominates 4122 single family preference dominates 4123 single family preference dominates 4131 single family preference dominates

PAGE 233

233 Table 4 23. Continued Conflict Conflict Score Descriptions Score 4132 single family preference dominates 4133 single family preference dominates 3333 all moderate preference all in conflict 3444 high preference multi family conflicts with high preference commercial institutional transportation and high preference industrial warehouses 3441 high preference multi family conflicts with high preference commercial institutional transportation 3442 high preference multi family conflicts with high preference commercial institutional transportation 3443 high preference multi family conflicts with high preference commercial institutional transportation 3414 high preference multi family conflicts with high preference industrial warehouses 3424 high preference multi family conflicts with high preference industrial warehouses 3434 high preference multi family conflicts with high preference industrial warehouses 3144 high preference commercial institutional transportation conflicts with high preference industrial warehouses 3244 high preference commercial institutional transportation conflicts with high pref erence industrial warehouses 3344 high preference commercial institutional transportation conflicts with high preference industrial warehouses 3411 multi family preference dominates 3412 multi family preference dominates 3413 multi family preference dominates 3421 multi family preference dominates 3422 multi family preference dominates 3423 multi family preference dominates 3431 multi family preference dominates 3432 multi family preference dominates 3433 multi family preference dominates 3141 commercial institutional transportation preference dominates 3241 commercial institutional transportation preference dominates 3341 commercial institutional transportation preference dominates 3142 commercial institutional transportation preference dominates 3242 commercial institutional transportation preference dominates 3342 commercial institutional transportation preference dominates 3143 commercial institutional transportation preference dominates 3243 commercial institutional transportation preference dominates 3343 commercial institutional transportation preference dominates 3114 industrial warehouses preference dominates 3214 industrial warehouses preference dominates 3314 industrial warehouses preference dominates

PAGE 234

234 Table 4 23. Continued Conflict Conflict Score Descriptions Score 3124 industrial warehouses preference dominates 3134 industrial warehouses preference dominates 3224 industrial warehouses preference dominates 3234 industrial warehouses preference dominates 3324 industrial warehouses preference dominates 3334 industrial warehouses preference dominates 3332 moderate preference single family conflicts with moderate preference multi family and moderate preference commercial institutional transportation 3331 moderate preference single family conflicts with moderate preference multi family and moderate preference commercial institutional transportation 3313 moderate preference single family conflicts with moderate preference multi family and moderate preference industrial warehouses 3323 moderate preference single family conflicts with moderate preference multi family and moderate preference industrial warehouses 3343 commercial institutional transportation preference dominates 3133 moderate prefere nce single family conflicts with moderate preference commercial institutional transportation and moderate preference industrial warehouses 3233 moderate preference single family conflicts with moderate preference commercial institutional transportation an d moderate preference industrial warehouses 3433 multi family preference dominates 2222 all moderate preference all in conflict 2444 high preference multi family conflicts with high preference commercial institutional transportation and high preference industrial warehouses 2441 high preference multi family conflicts with high preference commercial institutional transportation 2442 high preference multi family conflicts with high preference commercial institutional transportation 2443 high preference multi family conflicts with high preference commercial institutional transportation 2414 high preference multi family conflicts with high preference industrial warehouses 2424 high preference multi family conflicts with high preference industrial warehouses 2434 high preference multi family conflicts with high preference industrial warehouses 2144 high preference commercial institutional transportation conflicts with high preference industrial warehouses 2244 high preference commerc ial institutional transportation conflicts with high preference industrial warehouses 2344 high preference commercial institutional transportation conflicts with high preference industrial warehouses 2411 multi family preference dominates 2412 multi family preference dominates 2413 multi family preference dominates 2421 multi family preference dominates

PAGE 235

235 Table 4 23. Continued Conflict Conflict Score Descriptions Score 2422 multi family preference dominates 2423 multi family preference dominates 2431 multi family preference dominates 2432 multi family preference dominates 2433 multi family preference dominates 2141 commercial institutional transportation preference dominates 2241 commercial institutional transportation preference dominates 2341 commercial institutional transportation preference dominates 2142 commercial institutional transportation preference dominates 2242 commercial institutional transportation preference dominates 2342 commercial institutional transportation preference dominates 2143 commercial institutional transportation preference dominates 2243 commercial institutional transportation preference dominates 2343 commercial institutional transportation preference dominates 2114 industrial warehouses prefe rence dominates 2214 industrial warehouses preference dominates 2314 industrial warehouses preference dominates 2124 industrial warehouses preference dominates 2134 industrial warehouses preference dominates 2224 industrial warehouses preference dominates 2234 industrial warehouses preference dominates 2324 industrial warehouses preference dominates 2334 industrial warehouses preference dominates 2223 industrial warehouses preference dominates 2221 moderate preference single family conflicts with moderate preference multi family and moderate preference commercial institutional transportation 2212 moderate preference single family conflicts with moderate preference multi family and moderate preferen ce industrial warehouses 2232 commercial institutional transportation preference dominates 2242 commercial institutional transportation preference dominates 2122 moderate preference single family conflicts with moderate preference commercial institution al transportation and moderate preference industrial warehouses 2322 multi family preference dominates 2422 multi family preference dominates 1111 all low preference all in conflict 1444 high preference multi family conflicts with high preference commercial institutional transportation and high preference industrial warehouses 1441 high preference multi family conflicts with high preference commercial institutional transportation 1442 high preference multi family conflicts with high preference co mmercial institutional transportation

PAGE 236

236 Table 4 23. Continued Conflict Conflict Score Descriptions Score 1443 high preference multi family conflicts with high preference commercial institutional transportation 1414 high preference multi family conflicts with high preference industrial warehouses 1424 high preference multi family conflicts with high preference industrial warehouses 1434 high preference multi family conflicts with high preference industrial warehouses 1144 high preference commercial institutional transportation conflicts with high preference industrial warehouses 1244 high preference commercial institutional transportation conflicts with high preference industrial warehouses 1344 high preference commercial institutional transportation conflicts with high preference industrial warehouses 1411 multi family preference dominates 1412 multi family preference dominates 1413 multi family preference dominates 1421 multi family preference dominates 1422 multi family preference dominates 1423 multi family preference dominates 1431 multi family preference dominates 1432 multi family preference dominates 1433 multi family preference dominates 1141 commercial institutional transportation preference dominates 1241 commercial institutional transportation preference dominates 1341 commercial institutional transportation preference dominates 1142 commercial institutional transportation preference dominates 1242 commercial institutional transportation preference domi nates 1342 commercial institutional transportation preference dominates 1143 commercial institutional transportation preference dominates 1243 commercial institutional transportation preference dominates 1343 commercial institutional transportation preference dominates 1114 industrial warehouses preference dominates 1214 industrial warehouses preference dominates 1314 industrial warehouses preference dominates 1124 industrial warehouses preference dominates 1134 industrial warehouses preference dominates 1224 industrial warehouses preference dominates 1234 industrial warehouses preference dominates 1324 industrial warehouses preference dominates 1334 industrial warehouses preference dominates 1112 industrial warehouses preference dominates 1113 industrial warehouses preference dominates

PAGE 237

237 Table 4 23. Continued Conflict Conflict Score Description Score 1121 commercial institutional transportation preference dominates 1131 commercial institutional transportation preference dominates 1141 commercial institutional transportation preference dominates 1211 multi family preference dominates 1311 multi family preference dominates 1411 multi family preference dominates Table 4 24. Conflict scores for the BAU scenario Conflict Scores Cells Acres Single Family Dominates 2111 104,976 23,346 3111 39,501 8,785 3112 115 26 3121 2 0 3211 272 60 4111 12,398 2,757 4112 67 15 4211 45 10 Subtotal 157,376 34,999 Multi Family Dominates 2311 12 3 1411 96 21 1311 495 110 1211 5,986 1,331 Subtotal 6,594 1,466 Commercial Institutional Transportation Dominates 1231 3 1 1141 11 2 1132 9 2 1131 474 105 1121 220 49 Subtotal 717 159 Industrial Warehouses Dominates 2114 1 0 1214 2 0 1213 2 0 1114 410 91 1113 662 147 1112 17,660 3,927 Subtotal 18,737 4,167

PAGE 238

238 Table 4 2 4 Continued Conflict Scores Cells Acres Minor Conflict 2212 1 0 2211 1,762 392 2121 4 1 2112 804 179 1221 11 2 1212 76 17 1122 3 1 Subtotal 2,661 592 Major Conflict 1111 1,563,159 347,631 Subtotal 1,563,159 347,631 Total 1,749,244 389,014 Table 4 25. Conflict scores for the infill scenario Conflict Scores Cells Acres Single Family Dominates 4311 10 2 4211 22 5 4112 52 12 4111 9,956 2,214 3211 154 34 3112 96 21 3111 34,088 7,581 2111 75,787 16,854 Subtotal 120,165 26,723 Multi Family Dominates 2411 7 2 2311 205 46 1412 5 1 1411 305 68 1312 36 8 1311 1,892 421 1211 1,679 373 Subtotal 4,129 918 Commercial Institutional Transportation Dominates 1121 325 72 Subtotal 325 72 Industrial Warehouses Dominates 2114 1 0 1213 2 0 1114 359 80 1113 502 112 1112 13,997 3,113 Subtotal 14,861 3,305

PAGE 239

239 Table 4 25. Continued Conflict Scores Cells Acres Minor Conflict 3311 45 10 2211 785 175 2112 574 128 1314 1 0 1221 2 0 1212 10 2 1122 25 6 Subtotal 1,442 321 Major Conflict 1111 844,868 187,890 Subtotal 844,868 187,890 Total 985,790 219,230 Table 4 26. Single family acreage demand and allocation in 2010, 2020, and 2030 for the BAU scenario BAU Single Family 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 12,226 54,976 12,840 57,738 5,716 25,702 Allocated 12,226 54,976 12,853 57,796 5,740 25,812 Table 4 27. Multi family acreage demand and allocation in 2010, 2020, and 2030 for the BAU scenario BAU Multi Family 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 90 404 872 3,920 624 2,807 Allocated 90 404 991 4,455 666 2,996 Table 4 28. Commercial institution transportation acreage demand and allocation in 2010, 2020, and 2030 for the BAU scenario BAU Commercial Institutional Transportation 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 157 706 5,724 25,737 4,639 20,860 Allocated 157 706 5,710 25,675 4,841 21,767 Table 4 29. Industrial warehouses acreage demand and allocation in 2010, 2020, and 2030 for the BAU scenario BAU Industrial Warehouses 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 4 18 753 3,388 594 2,671 Allocated 4 18 825 3,710 797 3,584

PAGE 240

240 Table 4 30. Infill scenario infill development acreages and conflict development acreages in 2020 and 2030 Infill Scenario 2020 2030 Within Infill Areas* Within Conflict Areas ** Total Within Infill Areas* Within Conflict Areas ** Total Acres Cells Acres Cells Acres Cells Acres Cells Acres Cells Acres Cells Single Family Allocated 12,885 57,938 0 0 12,885 57,938 5,780 25,990 0 0 5,780 25,990 Multi Family Allocated 55 247 828 3,723 883 3,970 0 0 835 3,755 835 3,755 Commercial Institutional Transportation Allocated 2,663 11,974 3,214 14,453 5,877 26,427 0 0 4,664 20,974 4,664 20,974 Industrial Warehouses Allocated 759 3,414 0 0 759 3,414 707 3,181 0 0 707 3,181 Total 16,362 73,573 4,042 18,176 20,404 91,749 6,487 29,171 5,499 24,729 11,987 53,900 *: located either within 1,000 meters of urban buffer areas or outside **: without overlapping with the infill areas Table 4 31. Single family acreage demand and allocation in 2010, 2020, and 2030 for the infill scenario Infill Single Family 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 12,226 54,976 12,840 57,738 5,716 25,702 Allocated 12,226 54,976 12,885 57,938 5,780 25,990

PAGE 241

241 Table 4 32. Multi family acreage demand and allocation in 2010, 2020, and 2030 for the infill scenario Infill Multi Family 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 90 404 872 3,920 624 2,807 Allocated 90 404 883 3,970 835 3,755 Table 4 33. Commercial institution transportation acreage demand and allocation in 2010, 2020, and 2030 for the infill scenario Infill Commercial Institutional Transportation 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 157 706 5,724 25,737 4,639 20,860 Allocated 157 706 5,877 26,427 4,664 20,974 Table 4 34. Industrial warehouses acreage demand and allocation in 2010, 2020, and 2030 for the infill scenario Infill Industrial Warehouses 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 4 18 753 3,388 594 2,671 Allocated 4 18 759 3,414 707 3,181

PAGE 242

242 Table 4 35. Increased density development scenario development acreages and conflict development acreages in 2020 and 2030 Increased Density Development Scenario 2020 2030 Within infill a reas* Within conflict a reas ** Total Within i nfil l a reas* Within conflict a reas ** Total Acres Cells Acres Cells Acres Cells Acres Cells Acres Cells Acres Cells Single Family Allocated 3,859 17,353 0 0 3,859 17,353 5,322 23,929 0 0 5,322 23,929 Multi Family Allocated 54 244 245 1,102 299 1,346 0 0 594 2,670 594 2,670 Commercial Institutional Transportation Allocated 1,678 7,546 78 352 1,756 7,898 318 1,428 3,787 17,029 4,105 18,457 Industrial Warehouses Allocated 331 1,489 0 0 331 1,489 552 2,480 0 0 552 2,480 Total 5,922 26,632 323 1,454 6,245 28,086 6,192 27,837 4,381 19,699 10,573 47,536 *: located within 1,000 meters of urban buffer areas **: without overlapping with the infill areas Table 4 36. Single family acreage demand and allocation in 2010, 2020, and 2030 for the increased density development scenario Increased Density Single Family 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 12,226 54,976 3,764 16,924 4,972 22,356 Allocated 12,226 54,976 3,859 17,353 5,322 23,929

PAGE 243

243 Table 4 37. Multi family acreage demand and allocation in 2010, 2020, and 2030 for the increased density development scenario Increased Density Multi Family 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 90 404 256 1,151 543 2,442 Allocated 90 404 299 1,346 594 2,670 Table 4 38. Commercial institution transportation acreage demand and allocation in 2010, 2020, and 2030 for the increased density development scenario Increased Density Commercial Institutional Transportation 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 157 706 1,678 7,545 4,035 18,145 Allocated 157 706 1,756 7,898 4,105 18,457 Table 4 39. Industrial warehouses acreage demand and allocation in 2010, 2020, and 2030 for the increased density development scenario Increased Density Industrial Warehouses 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 4 18 221 993 517 2,323 Allocated 4 18 331 1,489 552 2,480

PAGE 244

244 Table 4 40. Redevelopment scenario development acreages and conflict development acreages in 2020 and 2030 Redevelopment Scenario 2020 2030 Within Infill Areas* Within Conflict Areas ** Total Within Infill Areas* Within Conflict Areas ** Total Acres Cells Acres Cells Acres Cells Acres Cells Acres Cells Acres Cells Single Family Allocated 3,389 15,239 0 0 3,389 15,239 4,234 19,037 0 0 4,234 19,037 Multi Family Allocated 54 244 245 1,101 299 1,345 0 0 455 2,047 455 2,047 Commercial Institutional Transportation Allocated 1,403 6,307 78 352 1,481 6,659 583 2,620 2,998 13,479 3,580 16,099 Industrial Warehouses Allocated 232 1,042 0 0 232 1,042 435 1,955 0 0 435 1,955 Total 5,078 22,832 323 1,453 5,401 24,285 5,252 23,612 3,453 15,526 8,704 39,138 *: located within 1,000 meters of urban buffer areas **: without overlapping with the infill areas Table 4 41. Single family acreage demand and allocation in 2010, 2020, and 2030 for the redevelopment scenario Redevelopment Single Family 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 12,226 54,976 3,199 14,386 4,226 19,003 Allocated 12,226 54,976 3,389 15,239 4,234 19,037 Table 4 42. Multi family acreage demand and allocation in 2010, 2020, and 2030 for the redevelopment scenario Redevelopment Multi Family 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 90 404 218 978 462 2,075 Allocated 90 404 245 1,101 455 2,047

PAGE 245

245 Table 4 43. Commercial institution transportation acreage demand and allocation in 2010, 2020, and 2030 for the redevelopment scenario Redevelopment Commercial Institutional Transportation 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 157 706 1,426 6,413 3,430 15,423 Allocated 157 706 1,481 6,659 3,580 16,099 Table 4 44. Industrial warehouses acreage demand and allocation in 2010, 2020, and 2030 for the redevelopment scenario Redevelopment Industrial Warehouses 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 4 18 188 844 439 1,975 Allocated 4 18 232 1,042 435 1,955 Table 4 45. Conservation scenario development acreages and conflict development acreages in 2020 and 2030 Conservation Scenario 2020 2030 Within Urban Buffer Areas Outside Urban Buffer Areas Total Within Urban Buffer Areas Outside Urban Buffer Areas Total Acres Cells Acres Cells Acres Cells Acres Cells Acres Cells Acres Cells Single Family Allocated 2,670 12,006 685 3,078 3,355 15,084 2,423 10,897 1,847 8,305 4,270 19,202 Multi Family Allocated 205 920 42 189 247 1,109 451 2,029 72 326 524 2,355 Commercial Institutional Transportation Allocated 1,540 6,926 8 37 1,549 6,963 3,544 15,936 0 0 3,544 15,936 Industrial Warehouses Allocated 120 541 72 322 192 863 198 891 237 1,064 435 1,955 Total 4,535 20,393 807 3,626 5,343 24,019 6,616 29,753 2,156 9,695 8,773 39,448

PAGE 246

246 Table 4 46. Single family acreage demand and allocation in 2010, 2020, and 2030 for the conservation scenario Conservation Single Family 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 12,226 54,976 3,199 14,386 4,226 19,003 Allocated 12,226 54,976 3,355 15,084 4,270 19,202 Table 4 47. Multi family acreage demand and allocation in 2010, 2020, and 2030 for the conservation scenario Conservation Multi Family 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 90 404 218 978 462 2,075 Allocated 90 404 247 1,109 524 2,355 Table 4 48. Commercial institution transportation acreage demand and allocation in 2010, 2020, and 2030 for the conservation scenario Conservation Commercial Institutional Transportation 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 157 706 1,426 6,413 3,430 15,423 Allocated 157 706 1,549 6,963 3,544 15,936 Table 4 49. Industrial warehouses acreage demand and allocation in 2010, 2020, and 2030 for the conservation scenario Conservation Industrial Warehouses 2010 2020 2030 Acres Cells Acres Cells Acres Cells Demanded 4 18 188 844 439 1,975 Allocated 4 18 192 863 435 1,955

PAGE 247

247 Figure 4 1. Areas within urban buffer areas and outside urban buffer areas ( Raw data s ource: Alachua County Growth Management Department )

PAGE 248

248 Figure 4 2. Major roads in Alachua County, incorporated towns and cities, and Urban Cluster areas ( Raw data s ource: Alachua County Growth Management Department )

PAGE 249

249 Figure 4 3. The BAU mask ( Raw d ata s ource: Alachua County Growth Management Department )

PAGE 250

250 Figure 4 4. The infill development mask ( Raw data s ource: Alachua County Growth Management Departm ent )

PAGE 251

251 Figure 4 5. Five scenarios with their relationship with the urban area and the greenfield areas, where the urban area is within 1,000 meters of the urban buffer areas and the greenfield area is outside 1,000 meters of the urban buffer areas.

PAGE 252

252 Figure 4 6. Population growth in Alachua County from 1982 2030. Considering the natural ( Source: U.S. Census Bureau (1980, 1990, 2000, 2010) and BEBR (2009, March) ) 141,716 170,567 205,114 233,416 286,100 316,300 0 50,000 100,000 150,000 200,000 250,000 300,000 350,000 1982 1994 2003 2010 2020 2030 Population Alachua County Population Growth 1982 2030 Total Population Population Linear Trend

PAGE 253

253 Figure 4 7. Ratio of single family population /multi family population from 1982 2030

PAGE 254

254 Figure 4 8. BAU preference map for single family development

PAGE 255

255 Figure 4 9. Collapsed map for single family use for the BAU scenario

PAGE 256

256 Figure 4 10. Conflict map for the BAU scenario

PAGE 257

257 Figure 4 11. Conflict map for the infill scenario

PAGE 258

258 Figure 4 12. 2010 Alachua County LULC Current Plan 1

PAGE 259

259 Figure 4 13. 2010 Al achua County LULC Current Plan 2

PAGE 260

260 Figure 4 14. 202 0 Alachua County LULC Alternative Business As Usual Scenario 1

PAGE 261

261 Figure 4 15. 2020 Alachua County LULC Alternat ive Business As Usual Scenario 2

PAGE 262

262 Figure 4 16. 203 0 Alachua County LULC Alternative Business As Usual Scenario 1

PAGE 263

263 Figure 4 17. 203 0 Alachua County LULC Alternat ive Business As Usual Scenario 2

PAGE 264

264 Figure 4 18. 2020 Alachua County LULC Alternative Infill Development Scenario 1

PAGE 265

265 Figure 4 19. 2020 Alachua County LULC Alternati ve Infill Development Scenario 2

PAGE 266

266 Figure 4 20. 203 0 Alachua County LULC Alternative Infi ll Development Scenario 1

PAGE 267

267 Figure 4 21. 203 0 Alachua County LULC Alternative Infill Development Sce nario 2

PAGE 268

268 Figure 4 22. 2020 Alachua County LULC Alternative Increased Density Development Scenario 1

PAGE 269

269 Figure 4 23. 2020 Alachua County LULC Alternative Increase d Density Development Scenario 2

PAGE 270

270 Figure 4 24. 203 0 Alachua County LULC Alternative Increased Density Development Scenario 1

PAGE 271

271 Figure 4 25. 203 0 Alachua County LULC Alternative Increase d Density Development Scenario 2

PAGE 272

272 Figure 4 26. 2020 Alachua County LULC Alternative Redevelopment Scenario 1

PAGE 273

273 Figure 4 27. 2020 Alachua County LULC Alte rnative Redevelopment Scenario 2

PAGE 274

274 Figure 4 28. 203 0 Alachua County LULC Alternative Redevelopment Scenario 1

PAGE 275

275 Figure 4 29. 203 0 Alachua County LULC Alte rnative Redevelopment Scenario 2

PAGE 276

276 Figure 4 30 2020 Alachua County LULC Alternative Conservation Scenario 1

PAGE 277

277 Figure 4 31. 2020 Alachua County LULC Alt ernative Conservation Scenario 2

PAGE 278

278 Figure 4 32. 203 0 Alachua County LULC Alternative Conservation Scenario 1

PAGE 279

279 Figure 4 33. 203 0 Alachua County LULC Alt ernative Conservation Scenario 2

PAGE 280

280 CHAPTER 5 CONCLUSIONS AND DISCUSSION This research has successfully classified satellite imagery for eleven urban and natural LULC classe s for 1982, 1994, and 2003, and also successfully projected four urban LULC classes for both 2020 and 2030 in Alachua County, Florida, based on five sets of articulated assumptions From the study, past, present, and future spatial patt erns of urban development are well presented in order to compare changes over time resulting from the variation in adopted assumptions. U sing remote sensing data alone or parcel data alone did not result in a satisfactory method of LULC classification so remote sensing data and GIS parcel data are used concomitantly. The comparison of classification strategies, included in the study, proves that this combined approach results in the highest degree of accuracy. T he high accuracy of classification maps in Alachua County in 1982, 1994, and 2003 cannot have been obtained w ithout applying ancillary data such as parcel information This is the major advantage of the CART method as it includes ancillary data into considerations. In addition, t he CART method best represent s lan d uses rather than land covers, especially for small parcels. However, the land covers in large p arcels can still be distinguished by the CART method. As a result, sub parcel land cover details such as wetlands can be identified well in this research. Generally speaking, the inclusion of parcel data as well as the information derived from the parcel data is the most important step for successful county wide LULC classifications. Because the computing power that processes the ROIs in ENVI 4.4 RuleGen 1.02 (Chapter 2) to classify urban LULCs is overburdened, reduced ROI signatures are sought based on the random sampling tool. If it is known in advance that the ROIs will overburden the computer,

PAGE 281

281 reduced ROI signatures will be sought in the beginning. This effort can save signi ficant CPU computing time: at least half of the pixel amounts can be cut at the outset. The lo gistic regression model built in Chapter 3 was used to simulate urban growth in Alachua County using four urban LULC c lasse s. This class level is equivalent to USGS Level III. Because the urban LULC is accurately classified in Chapter 2, t he logistic regression model is able to yield high accuracy for each of the four urban LULC classes for urban growth simulation This s tudy simulated urban growth in the county for 2003 LULC classes based on the urban LULC classifications in 1982, 1994, and 2003, which are based on model construction, model calibration, model interpretation, model refinement, and sensitivity analysis Whe n compared with actual 2003 urban land use, the model accurately predicted the spatial distribution of the four LULC classes. This logistic regression model has advantages compared to other AI models such as the CA model, the ABM model, the Neural Network (NN) model, and t he Genetic Algorithm (GA) model. For example, the logistic regression model includes consideration of historical factors with longitudinal data being imported into the model and it is also capable of including spatial variables into the model. As a result, the research goals and objectives of this dissertation are effectively achieved ; t he overall accuracy level for the 2003 simulated LULC map reaches 97.30%. Because of the empirical trials of various spatial independent variables into the model that are tested incrementally on a one by one basis, the accuracy level for each of the four u rban dependent variables can reach more than 97%. In Chapter 4, t his study allocates urban LULC for Alachua County for 2020 and 2030 The all ocation of urban LULC is based on the articulation of five development scenarios, which

PAGE 282

282 cover almost all the possi bilities for possible future development patterns for each scenario can be assessed and mapped Recapping the entire dissertation, it is found that although this dissertation has accomplished the aut the classification methods other than the CART method, the V I S method, and supervised classification methods (Chapter 2). The new and intuitiv e classification methods including t he NN method, the thermal method, and the fuzzy supervised method are worth trying Although adding these classification methods to the cou nty based LULC classifications may produce the same results, the county based LULC classification research will be mo re interesting and exciting. Similarly, the satellite SPOT data can also be tried for LULC classifications to further elevate accuracy; QuickBird and IKONOS imageries may also be used in this regard. In addition, the author for redevelopment using the MLR model. According to Landis and Zhang (1997), the MLR model is capable of simulating redevelopment. More research shall be conducted in this regard to simulate redevelopment in Alachua County; for example, a hybrid model tha t combines the MLR model and the CA model can be considered in this manner Correspondingly, because of time constrain t s the author has conducted research for a rural county Alachua County only. The author hopes tha t similar urban LULC simulation can be c onducted for an urban county such as Duval County (the City of Jacksonville), Orange County (the City of Orlando), or Miami Dade County (the City of Miami) so as to compare the rural county with an urban county. Furthermore, because the MLR model has the d isadvantage that it is less temporally dynamic (Chapter 3), this model needs a complicated allocation process to overcome T he author hopes to try the CA model and the Markov Chain Model (MCM) to simulate urban growth for Florida counties. As a result, mid

PAGE 283

283 term simulation results can be obtained which may be interesting and exciting Three modules can be applied: the CELLATOM module, the CA_MARKOV module, and the MARKOV module T hey are available in the IDRISI software. Suitability analyses will be needed i n this regard for model building rather than probabilities in the MLR model. consider LULC changes for natural land, the future study shall take this factor into consideration. To do this work, a Conversion of Land Use and Its Ef fects ( CLUE ) model can be applied. The CLUE model has a LULC conversion matrix, in which codes can be used to represent the changes and non changes between different pairs of LULC classe s (Verburg, no date); for example, 0 represents that LULC classe s can be changed from LULC A to LULC B and 1 represents that LULC classe s cannot be changed. As a result, more research should be undertaken to see if the CLUE model is capable of simulating redevelopment because the LULC conversion matrix in the CLUE model ma y potentially be used to simulate redevelopment. This study adopted a twenty year timeframe (to 20 30) for urban growth simulation T he examination of land use change would be more meaningful if the time frame could be extended beyond 20 years. A research time frame that extend ed 50 years or more is optimal. In sum, this study classifies eleven LULC classes up to the level of USGS Level III and simultaneously results in high accuracy. Because of the highly accurate LULC that have been classified the MLR m odel is built with high accuracy to predict urban LULC classes. In addition, because the MLR model is built with high accura cy, urban LULC allocation is conducted with h igh accuracy as well which can allow researchers to i future urba n growth. These are the noteworthy elements of this study. In fact, the satellite

PAGE 284

284 imagery classification s the MLR model building, and urban LULC allocation in this study have potential to be widely applied in ways that benefit both the public and the priv ate sectors

PAGE 285

285 LIST OF REFERENCES Ainsworth. (no date). Logistic Regression: Continued. Psy 524. Course Presentation Retrieved December 30, 2009, from http://www.csun.edu/~ata20315/psy524/docs/Psy524%20lecture%2019%20logistic_cont.p pt Almeida, C.M., Monteiro, A.M.V., and Cmara, G. (2003). Modeling the Urban Evolution of Land Use Transitions Using Cellular Automata and Logistic Regression. In Geoscienc e and Remote Sensing Symposium, IGARSS '03, Proceedings, 2003 IEEEInternational. 21 25 July 2003 (pp. 1564 1566). Anderson, J.R., Hardy, E.E., Roach, J.T., and Witmer, R.E. (1976). A Land Use and Land Cover Classification System for Use with Remote Sensor Data Geological Survey Professional Paper 964. United States Government Printing Office, Washington Retrieved October 26, 2008, from http://landcover.usgs.gov/pdf/anderson.pdf Banman, C. (no date). Supervised and Unsupervised Land Use Classification Retrieved October 31, 2009, from http://www.emporia.edu/earthsci/student/banman5/perry3.html Batty, M. and Longley, P. (1994), Fractal Cities: A Geometry of Form and Function. London: Academic Press. BEBR (2003). Florida Statistical Abstract 2003. Bureau of Economic and Business Research: Warrington College of Business Administration. Gainesville: The University of Florida Press BEBR (2009, March). Florida Population Studies: Projections of Florida Population by County, 2008 2035 Bureau of Economi c and Business Research: Warrington College of Business Administration. 42(153): 1 8. Gainesville: The University of Florida Press Benfield F.K., Raimi M.D., and Chen D.D.T. (1999 ) Once There Were Greenfields: How Urban ironment, Economy, and Social Fabric Natural Resources Defense Council. The Boston Indicators Project (no date). Car ownership and Vehicle Miles Traveled, Boston and Metro Boston. Retrieved October 24, 2008, from http://www.bostonindicators.org/indicatorsproject/transportation/indicator.aspx?id=2056 Breiman, L., Friedman, J., Olshen, R., and Stone, C. (1984). Classification and Regression Trees Belmont, Calif: Wadsworth International Group. Broos, M.J. and Day, R. (no date). Measuring Urban Sprawl and Predicting Land Use Change using Geospatial Technologies. Retrieved October 25, 2008, from http://lal.cas.psu.edu/Research/sprawl.asp

PAGE 286

286 Bureau of Transportation Statistics. (2002). National Transportation Statistics 2001 (July, 2002). U.S. Department of Transportation. Retrieved December 21, 2008, from http://answers.google.com/answers/threadview?id=146894 Carr, M.H. and Zwick, P.D. (2007). Smart Land Use Analysis: The LUCIS Model Redlands, CA: ESRI Press. Census Bureau (1980, 1990, 2000, 2010 ) Profile of General Population and Housing Characteristics Retrieved September 18, 2011 from http://www.census.gov Cieslewicz, D.J. (2002). The Environmental Impacts of Sprawl. In Squires, G. D. (Ed.) Urban Sprawl: Causes, Consequences & Policy Responses (p.23 38). Washington, D.C.: The Urban Institute Press. Cowen, D.J., Jensen, J.R., Bresnahan, G., Ehler, D., Traves, D., Huang, X., Weisner, C., and Mackey, H.E. (1995). The Design and Implementation of an Integrated GIS for Environmental Application. Photogrammetric Engineering & Remote Sensing 61(1): 1393 1404 Cramr, H. (1999). Mathematical Methods of Statistics. Princeton: Princeton Press. Dewey, J.F. and Denslow, D.A., (2001). Grow th and Infrastructure in Alachua County: Does Conventional Development Pay Its Share of Public Costs ? Bureau of Economic and Business Research, University of Florida. Downs, A. (1994). New Visions for Metropolitan America Washington DC: The Brookings Inst itution. Duany, A., Plater Zyberk, E., and Speck, J. (2000). Suburban Nation New York: North Point Press. Field, A. (2000). Discovering Statistics using SPSS for Windows. London, Thousand Oaks, Dew Delhi: Sage Publications. Garson, G. D. (2009). Nominal Association: Phi, Contingency Coefficient, Tschuprow's T, Cramer's V, Lambda, Uncertainty Coefficient From Statnotes, North Carolina State University, Public Administration Program Retrieved December 30, 2009, from http://faculty.chass.ncsu.edu/garson/PA765/assocnominal.htm Herold, N.D., Koeln, G. and Cunnigham, D. (2003). Mapping Impervious Surfaces and Forest Canopy Using Classification and Regression Tree (CART) Analys is. ASPRS 2003 Annual Conference Proceedings. May 2003. Anchorage, Alaska. Retrieved April 19, 2008, from http:// www.nemo.uconn.edu/tools/impervious_surfaces/pdfs/Herold_etal_2003.pdf Hosmer, D.W., Jr. and Lemeshow, S. (1989). Applied Logistic Regression. Wiley Series in Probability and Mathematical Statistics. New York, Chichester, Brisbane, Toronto, Singapore: A Wi ley Interscience Publication.

PAGE 287

287 Hu, Z. and Lo, C.P. (2007); Modeling Urban Growth in Atlanta Using Logistic Regression. Computers, Environment and Urban Systems, 31(6): 667 688. Huang, X. and Jensen, J.R. (1997). A Machine Learning Approach to Automated Know ledge Base Building for Remote Sensing Image Analysis with GIS Data. Photogrammetric Engineering & Remote Sensing 63(10): 1185 1194. Hung, M.C. (2003). Remote Sensing and GIS for Urban Environmental Modeling, Monitoring and Visualization. (Doctoral Disse rtation. University of Utah. 2003). Hung M C and Ridd, M.K. (2002). A Subpixel Classifier for Urban Land Cover Mapping Based on a Maximum Likelihood Approach and Expert System Rules. Photogrammetric Engineering & Remote Sensing 68(11): 1173 1180. IDRISI H elp. (no date). IDRISI Andes Clark Labs. Jackson, K.T. (1985). Crabgrass Frontier : The Suburbanization of the United States. New York, Oxford: Oxford University Press. Jargowsky P.A. (2002). Sprawl, Concentration of Poverty, and Urban Inequality. In Squires, G.D. (Ed.) Urban Sprawl: Causes, Consequences and Policy Responses (p.39 117). Washington, DC: Urban Institute. Jensen, J.R. and Toll, D. (1982). Detecting Residential Lan d Use Development at the Urban Fringe. Photogrammetric Engineering & Remote Sensing 48(4): 629 643. Jensen, J.R. and Cowen, D.C. (1999). Remote Sensing of Urban/Suburban Infrastructure and Social Economic Attributes. Photogrammetric Engineering & Remote S ensing 65(5): 611 622. Jensen, J.R. (2005). Introductory Digital Imaging Processing: A Remote Sensing Perspective (Third Edition). Upper Saddle River, New Jersey: Pearson Prentice Hall. Kahn, M.E. (2006). The Benefits of Sprawl Environmental and Urban Economics. Thoughts on Environmental and Urban Issues from an Economics Perspective. Retrieved December 6, 2009, from http://greeneconomics.blogspot.com/2006/03/benefits of sprawl.html Kim, H.J. (2007). Spatiotemporal Analysis of Urban Growth in Local Communities (Doctoral Dissertation. Kent State University. 2007). Kramber, W.J. and Morse, A. (1994). Integrating Image Interpretation and Unsupervised Classification. Retri eved October 31, 2009, from http://libraries.maine.edu/Spatial/gisweb/spatdb/acsm/ac94037.html Landis, J.D. (1994). The California Urban Futures Model: A New Generation of Metropolitan Simulation Models. Environment and Planning B: Planning and Design. 21(4): 399 420.

PAGE 288

288 Landis, J.D. (1995). Imagining Land Use Futures: Applying the California Urban Futures Model. Journal of the American Planning Association. 61(4): 438 457. La ndis, J.D. and Zhang, M. (1997). Modeling Urban Land Use Change: the Next Generation of the California Urban Futures Model. Paper Presented at the National Center for Geographic Information and Analysis (NCGIA) the 1997 Land Use Modeling Workshop. Retrieve d December 30, 2009, from http://www.ncgia.ucsb.edu/conf/landuse97/papers/landis_john/paper.html Lawrence, R. and Wright, A. (2001). Rule Based Classification Systems U sing Classification and Regression Tree (CART) Analysis. Photogrammetric Engineering & Remote Sensing 67(10): 1137 1142. Levy, J. (2003). Contemporary Urban Planning (Six Edition). Upper Saddle River, New Jersey: Prentice Hall. Lewis, R.J. (2000). An Int roduction to Classification and Regression Tree (CART) Analysis. Presented at the 2000 Annual Meeting of the Society for Academic Emergency Medicine in San Francisco, California. Retrieved April 19, 2008, from http://www.saem.org/download/lewis1.pdf LCI (Land Cover Institute). (2007). NLCD Land Cover Class Definitions. Retrieved October 26, 2008, from http://landcover.usgs.gov/classes.php#herb Liou, C Y and Yang, K D.O. (no date). Unsupervised Classification of Remote Sensing Imagery with Non negative Matrix Factorization. Retrieved October 31, 2009, from http://red.csie.ntu.edu.tw/publications/ICONIP05_Unsupervised%20Classification%20of% 20Remote%20Sensing%20Imagery.pdf Lo, C.P. and Choi, J. (2004). A Hybrid Approach to Ur ban Land Use/Cover Mapping Using Landsat 7 Enhanced Thematic Mapper Plus (ETM+) Images. International Journal of Remote Sensing. 25:14, 2687 2700. Lu, D. and Weng, Q. (2004). Spectral Mixture Analysis of the Urban Landscape in Indianapolis with Landsat ET M+ Imagery. Photommetric Engineering & Remote Sensing 70(9): 1053 1062. Lucy, W.H. and Phillips, D.L. (2006). Chicago, Illinois; Washington, DC: American Planning Association. Menard, S. (1995). Applied Logistic Regr ession Analysis Thousand Oaks, London, New Delhi: Sage Publications. McFadden, D. S. (1973). Conditional Logit Analysis of Qualitative Choice Behavior In Zarembka, P. (Ed.), Frontiers in Econometrics (p.105 142). New York: Academic Press.

PAGE 289

289 Mowrer H.T. and Congalton, R.G. (2000). Introduction: the Past, Present, and Future of Spatial Uncertainty Analysis In Mowrer, H.T. and Congalton, R.G.(Ed.) Quantifying Spatial Uncertainty in Natural Resources. (p.xv xxiv). Chelsea, Michigan: Ann Arbor Press. M umford, L. (1961). The City in History: Its Origins, Its Transformations, and Its Prospects San Diego, New York, London: Harcourt Brace & Company. Myint, S.W. (2001). A Robust Texture Analysis and Classification Approach for Urban Land use and Land Cover Feature Discrimination. Geocarto International 16(4): 29 40. Myint, S.W., Wentz, E.A., and Purkis, S.J. (2007). Employing Spatial Metrics in Urban Land Use/Land Cover Mapping: Comparing the Getis and Geary Indices Photogrammetric Engineering & Remote Se nsing 73(12): 1403 1415. NCGIA. (no date). Meeting Summary: the State of the Art of Land use Modeling. Retrieved October 28, 2009, from http://www.ncgia.ucsb.edu/conf/landuse97/summary. html Paul, O.V.D. (2007). Remote Sensing: New Applications for Urban Areas Proceedings of the IEEE. Volume 95, Issue 12, Dec. 2007, 2267 2268. Phinn, S., Stanford, M., Scarth, P., Murray, A.T., and Shyy, P.T. (2002). Monitoring the Composition of Urban Environments based on the Vegetation Impervious Surface Soil(VIS) Model by Subpixel Analysis Techniques. International Journal of Remote Sensing 23(20): 4131 4153. Poelmans, L. and Rompaey, A.V. (2010). Complexity and Performance of Urban Expansion Models. Computers, Environment and Urban Systems. 34(1): 17 27. Pontius, R.G., Jr. and Schneider, L.C. (2001). Land Cover Change Model Validation by an ROC Method for the Ipswich Watershed, Massachusetts, USA. Agriculture, Ecosystems & Environment 85(1 3): 239 248. Ridd, M.K. (1995). Exploring a V I S (vegetation impervious surface soil) Model for Urban Ecosystem Analysis through Remote Sensing: Comparative Anatomy for Cities. International Journal Remote Sensing. 16(2): 2165 2185. Rogan, J., Miller, J., Stow, D., Franklin, J., Levien, L., and Fischer, C. (2003). Land Cover Change Monitoring with Classification Trees Using Landsat TM and Ancillary Data. Phot ogrammetric Engineering & Remote Sensing 69(7): 793 804. Seaman, M. (2001). Categorical Data Retrieved December 30, 2009, from http://edpsych.ed.sc.edu/seaman/edrm711/quest ions/categorical.htm Sirkin, R.M. (1999). Statistics for the Social Sciences (2 Edition) Thousand Oaks, London, New Delhi: Sage Publications. Short, N.M., Sr. (no date). Supervised Classification. End to End Remote Sensing Tutorial Page 1 17. RST. Retrieved October 31, 2009, from http://rst.gsfc.nasa.gov/Sect1/Sect1_17.html

PAGE 290

290 Squires, G.D. (2002). Urban Sprawl and the Uneven Development of Metropolitan American In Squires, G.D. (Ed.) Urban Sprawl: Causes, Consequences & Policy Responses (p.1 22). Washington, D.C.: The Urban Institute Press. Toll, D.L. (1984). An Evaluation of Simulated Thematic Mapper Data and Landsat MSS Data for Discriminating Suburban and Regional Land Use and Land Cov er. Photogrammetric Engineering & Remote Sensing 50(12): 1713 1724. Verburg, P. (no date). The CLUE S model: Tutorial CLUE s (version 2.4) and DYNACLUE (version 2). Wageningen University, Netherlands. Retrieved August 21, 2009, from http://www.cluemodel.nl/ Verburg, P.H., Eickhout, B., and Meijl, H.V. (2007). A Multi Scale, Multi Model Approach for Analyzing the Future Dynamics of European Land Use. The Annals of Regional Science 42(1): 57 77. Ward, D., Phinn, S.R. and Murray. A.T. (2000). Monitoring Growth in Rapidly Urbanizing Areas Using Remotely Sensed Data. Professional Geographer 52(3): 371 386. Wassmer, R.W. (2001). The Connection between Local Government Finance and the Generation of U rban Sprawl in California Retrieved December 6, 2009, from http://www.csus.edu/calst/government_affairs/reports/ffp44.pdf Yang, X. (2000). Integrating Image Analysis and Dynam ic Spatial Modeling with GIS in a Rapidly Suburbanizing Environment (Doctoral Dissertation. University of Georgia. 2000). Zambon, M., Lawrence, R., Bunn, A., and Powell, S. (2006). Effect of Alternative Splitting Rules on Image Processing Using Classifica tion Tree Analysis. Photogrammetric Engineering & Remote Sensing 72(1): 25 30.

PAGE 291

291 BIOGRAPHICAL SKETCH Yong Hong Guo was born in Shanghai, China, in 1968 He graduated from Tongji University in Shang hai, China, in 1991, holding a b he graduated from the Virginia Polytechnic Institute and State University ( Virginia Tech) having a m he Urban and Regional Planning Department at the University of Florida to pursue the degree of Doctor of Philosophy in urban planning. He receiv ed his Ph D f rom the University of Florida in the spring of 2012. He has been a member of the American Institute and Certified Planners (AICP) since 2007.