Citation
Comparison of Landsat MSS and TM Imagery for Long Term Forest Land Cover Change Assessment

Material Information

Title:
Comparison of Landsat MSS and TM Imagery for Long Term Forest Land Cover Change Assessment
Creator:
GENC, LEVENT ( Author, Primary )
Copyright Date:
2008

Subjects

Subjects / Keywords:
Datasets ( jstor )
Forests ( jstor )
Image classification ( jstor )
Information classification ( jstor )
Land cover ( jstor )
Landsat ( jstor )
Lowland forests ( jstor )
Pixels ( jstor )
Plantations ( jstor )
Remote sensing ( jstor )

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
Copyright Levent Genc. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Embargo Date:
8/1/2005
Resource Identifier:
53208037 ( OCLC )

Downloads

This item is only available as the following downloads:


Full Text

PAGE 1

COMPARISON OF LANDSAT MSS AND TM IMAGERY FOR LONG TERM Forest LAND COVER CHANGE ASSESSMENT By LEVENT GENC 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 2003

PAGE 2

Copyright 2003 by Levent Genc

PAGE 3

To my wife, Hanife and My lovely daughter, Destina Ekingen

PAGE 4

ACKNOWLEDGMENTS First, I wish to thank the members of my supervisory committee for their help throughout this effort. Dr. Scot Smith, the committee chair, was invaluable in his assistance, knowledge, and enthusiasm for the work. I thank Dr. Smith for providing significant financial support throughout my graduate education. His efforts went above and beyond the normal expectations, ensuring the completion of this research. I also thank him for being a very good friend, without whom I may not have been at this stage. Dr. Bon Dewitt added his knowledge and expertise to this effort, and was always willing to help answer my questions at a moment’s notice. Dr. Dewitt also provided extensive assistance in editing this work, given me an experience I will never forget. Dr. Michael Binford was incredibly supportive and helpful, and his advice and understanding throughout my “journey” were greatly appreciated. Dr. Binford also provided significant financial support throughout my graduate education, in addition to his insight and knowledge, for which I am grateful. Dr. Grenville Barnes showed great interest and enthusiasm for this research, and his knowledge was very helpful. I also thank Dr. David Gibson for his kindness and great interest and enthusiasm for this research, and his knowledge was very helpful. I also thank Dr. Portier for his help with the statistical analysis. I thank Charisse Griffith-Charles for her support and helps. I also thank Dr. Mark Lee, Dr. Joong Yong Park, and Dr. Balaji Ramachandran, three friends; who, “having iii

PAGE 5

been there” before provided much support and advice. Many thanks also go to Francis McCorry. I thank the Turkish Government for its generous financial support during my graduate studies at the University of Florida. Special thanks go to Dr. Osman Tekinel, Dr. Abdurrahim Korukcu, Dr. Ismet Arici, and Dr. Nafi Baytorun for their understanding and encouragement during my graduate study. I also thank NASA's Land Cover Land Use Change Program for financial support via project NRA 99-OES-06 during this research. My deepest thanks and appreciation go to my parents, Sultan and Celal, for the constant support and unconditional love they have provided throughout my life. I appreciate the endless love and support I receive from my sisters, Gunay and Senay, my brothers, Bulent and Savas, and their respective families. I would like to give my most special thanks and love to my wonderful wife Hanife Genc and my beautiful daughter Destina Ekingen Genc for their constant love, help, support, understanding, and encouragement since the day we started this journey. iv

PAGE 6

TABLE OF CONTENTS page ACKNOWLEDGMENTS.................................................................................................iii LIST OF TABLES...........................................................................................................viii LIST OF FIGURES.............................................................................................................x ABSTRACT.....................................................................................................................xiv CHAPTER 1 INTRODUCTION........................................................................................................1 Monitoring Land Use-Land Cover Change..................................................................3 Remote Sensing and Forest Ages.................................................................................6 Description of Problems Between MSS and TM.........................................................8 Image Enhancement for Land Cover Mapping..........................................................14 Low Pass Filtering......................................................................................................14 Principal Component Analysis (PCA).......................................................................14 Tasseled Cap Transformation (TCT).........................................................................15 Calibration of Remote Sensing Data..........................................................................15 Outline........................................................................................................................18 2 HYPOTHESIS............................................................................................................20 3 DETERMINATION OF DATASETS FOR CLASSIFCATION ACCURACY IMPROVEMENT.......................................................................................................21 Study Area..................................................................................................................24 Image Processing to Create the Datasets....................................................................25 Geometric Correction.................................................................................................25 Preparation of New Datasets for the Test Site...........................................................28 Pre-determination.........................................................................................28 Decisions......................................................................................................34 Preparing New Datasets for the Test Site...................................................................35 Determination of calibration coefficients for created new datasets.............37 Evaluation of calibration coefficients..........................................................40 Image Enhancements for New Datasets.....................................................................47 3x3 low pass filtering...................................................................................47 Principal component transformation and tasseled cap transformations.......51 Comparison of TCT of Landsat MSS and Landsat TM Data.....................................73 v

PAGE 7

Image Classification...................................................................................................74 Land Cover Determinations.......................................................................................74 Clear-cut_1-3 yr pine plantation..................................................................81 4-8 yr pine plantation...................................................................................81 > 8 yr pine plantation...................................................................................83 Wetland forest..............................................................................................86 Unsupervised Classification:......................................................................................86 Supervised Classification...........................................................................................89 Training samples..........................................................................................91 Maximum likelihood classification (MLC)..................................................93 Determination of Classification Accuracy for Test Site:............................................95 Error Matrix................................................................................................................97 Test Sample Design....................................................................................................97 Sample unit...................................................................................................98 Sample size and sampling selection.............................................................98 Evaluation of Error Matrices....................................................................................101 Kappa Analysis.........................................................................................................102 Results of Classification Accuracy for Test Site......................................................104 Summary of Image Classification............................................................................126 Determination of the Most Accurate Dataset...........................................................126 Evaluation of Overall Accuracy Means....................................................................128 Land Cover Recognitions.........................................................................................134 Summary...................................................................................................................138 4 LAND COVER CLASSIFICATION FOR APPLICATION SITE.........................139 Application Site........................................................................................................140 Image characteristics................................................................................................142 Determination of Land Cover Types and Image Classification...............................144 Results.......................................................................................................................148 Evaluation of Area Coverage...................................................................................148 Evaluation of Classification Accuracy.....................................................................139 5 CONCLUSIONS AND RECOMENDATIONS......................................................156 APPENDIX A NORMALIZED DIFFERENCE VEGETATION INDEX (NDVI).........................158 B TEXTURE ANALYSIS...........................................................................................160 C PRINCIPAL COMPONENT ANALYSIS...............................................................163 Determination of Principal Components..................................................................163 Principal Components...............................................................................................167 vi

PAGE 8

LIST OF REFERENCES.................................................................................................170 BIOGRAPHICAL SKETCH...........................................................................................177 vii

PAGE 9

LIST OF TABLES Table page 1-1 Characteristics of Landsat-MSS and Landsat TM multispectral scanners.........11 3-1 Determination of overall classification accuracy for various datasets................31 3-2 Description of datasets used for test site.............................................................38 3-3 Degree of Correlation between each band and each Principal Component for MSS_ 30_O_1986......................................................59 3-4 Degree of Correlation between each band and each Principal Component for MSS_30_ S_1986.......................................................59 3-5 Degree of Correlation between each band and each Principal Component for TM_ 30_O_1986........................................................59 3-6 Degree of Correlation between each band and each Principal Component for TM_ 30_S_1986.........................................................59 3-7 Degree of Correlation between each band and each Principal Component for MSS_ 60_O_1986......................................................64 3-8 Degree of Correlation between each band and each Principal Component for MSS_ O_S_1986........................................................64 3-9 Degree of Correlation between each band and each Principal Component for TM_ 60_O_1986........................................................64 3-10 Degree of Correlation between each band and each Principal Component for TM_60_ S_1986.........................................................64 3-11 Numbers of pixels selected for training sets.......................................................91 3-12 Individual land cover area recorded after classification.....................................94 3-13 Error matrix used to generate accuracy assessment.......................................103 3-14 Error Matrix of MSS_30_O_1986 for test site.................................................107 3-15 Error Matrix of TM_30_O_1986 for test site...................................................107 viii

PAGE 10

3-16 Error Matrix of MSS_30_S_1986 for test site..................................................109 3-17 Error Matrix of TM_30_S_1986 for test site....................................................109 3-18 Error Matrix of MSS_60_O_1986 for test site.................................................113 3-19 Error Matrix of TM_60_O_1986 for test site...................................................113 3-20 Error Matrix of MSS_60_S_1986 for test site..................................................115 3-21 Error Matrix of TM_60_S_1986 for test site....................................................115 3-22 Error Matrix of MSS_PC_TC_30_O_1986 for test site...................................118 3-23 Error Matrix of TM_PC_TC_30_O_1986 for test site.....................................118 3-24 Error Matrix of MSS_PC_TC_30_S_1986 for test site....................................120 3-25 Error Matrix of TM_PC_TC_30_S_1986 for test site......................................120 3-26 Error Matrix of MSS_PC_TC_60_O_1986 for test site...................................123 3-27 Error Matrix of TM_PC_TC_60_O_1986 for test site.....................................123 3-28 Error Matrix of MSS_PC_TC_60_S_1986 for test site....................................125 3-29 Error Matrix of TM_PC_TC_60_S_1986 for test site......................................125 3-30 Overall accuracies with 4 replicates used to validate the most accurate dataset.................................................................................................127 3-31 Calculated means for overall accuracies...........................................................129 3-32 Ranking means of overall accuracies................................................................131 3-33 LSD means determination y 9 to y 2 ....................................................................131 3-34 LSD means determination y 10 to y 13 .................................................................132 3-35 LSD means determination y 12 to y 16 .................................................................132 3-36 Significant and non-significant level using Fisher LSD method for overall accuracy’s means..............................................................................................133 3-37 Land cover conversion from-to tables for 30 m and 60 m classified images...137 4-1 Look up table for classification........................................................................151 4-2 Classification accuracy for application site from 1975 to 2000.......................154 ix

PAGE 11

LIST OF FIGURES Figure page 1-1 Landsat MSS sensor records reflectance intensities in four wavelength bands with 57x57 ground cell size: (a)0.5 -0.6 m (green), (b) 0.6-0.7 m (red), (c) 0.7-0.8 m (near-IR), (d) 0.8-1.1 m (IR), and (e). RGB: 421.........................12 1-2 Landsat TM sensor records 6 wavelength bands with approximately 30x30 ground cell size: (a) 0.45-0.52 m (blue), (b) 0.52-0.60 m (green), (c) 0.63-0.69 m (red), (d) 0.76-0.90 m (Near infrared), (e) 1.55-1.75 m (mid infrared), (f) 2.08-2.35 m (mid infrared), (g) RGB..........................................13 3-1 Procedure of land cover classification for test site.............................................23 3-2 Composite of bands 4, 3, and 2 (RGB) image of WRS 17-39 scene for the test site in Hamilton County....................................................................24 3-3 Geometric correction for study area images.......................................................27 3-4 Data preparation for 30 and 60 m datasets. Section A ) MSS dataset, Section B ) TM dataset..........................................................39 3-5 Coefficients of determination (R 2 ) and Mean Square Error (S-value) for the original MSS 30m vs. the original TM 30m and the smoothed MSS 30m vs. the smoothed 30m TM datasets A) Original Red Band, B) Original Green Band, C) Original MSS NIR3 and Original TM NIR, and D) Original MSS NIR4 and original TM NIR. E) Smoothed Red Band, F) Smoothed Green Band, G) Smoothed MSS NIR3 and Smoothed TM NIR, and H) Smoothed MSS NIR4 and Smoothed TM NIR.......................................................................................41 3-6 Coefficients of determination (R 2 ) and Mean Square Error (S-value) for the MSS 30m principal components vs. the TM 30m principal components and the MSS 30m tasseled cap indexes and the TM 30m tasseled cap indexes (original and smoothed): A) O_PC1, B) O_PC2, C) S_PC1, D) S_PC2, E) O_BI, F) GI_O, G) S_BI, and D) S_GI..............................................................................42 3-7 Coefficients of determination (R 2 ) and Mean Square Error (S-value) for the original MSS 60m vs. the original TM 60m and the smoothed MSS 60m vs. the smoothed 60m TM datasets A) Original Red Band, B) Original Green Band, C) Original MSS NIR3 and Original TM NIR, and D) Original MSS NIR4 and original TM NIR. E) Smoothed Red Band, F) Smoothed Green Band, G) x

PAGE 12

Smoothed MSS NIR3 and Smoothed TM NIR, and H) Smoothed MSS NIR4 and Smoothed TM NIR.......................................................................................43 3-8 Coefficients of determination (R 2 ) and Mean Square Error (S-value) for the MSS 60m principal components vs. the TM 60m principal components and the MSS 60m tasseled cap indexes and the TM 60m tasseled cap indexes (original and smoothed): A) O_PC1, B) O_PC2, C) S_PC1, D) S_PC2, E) O_BI, F) GI_O, G) S_BI, and D) S_GI..............................................................44 3-6 Principal component transformation for MSS_30_O_1986 dataset...................55 3-7 Principal component transformation for MSS_30_S_1986 dataset....................56 3-8 Principal component transformation for TM_30_O_1986 dataset.....................57 3-9 Principal component transformation for TM_30_S_1986 dataset......................58 3-10 Principal component transformation for MSS_60_O_1986 dataset...................60 3-11 Principal component transformation for MSS_60_S_1986 dataset....................61 3-12 Principal component transformation for TM_60_O_1986 dataset.....................62 3-13 Principal component transformation for TM_60_S_1986 dataset......................63 3-14 Tasseled cap transformation for 30m and 60m MSS_O_1986...........................69 3-15 Tasseled cap transformation for 30m and 60m MSS_O_1986...........................70 3-16 Tasseled cap transformation for 30m and 60m TM_O_1986.............................71 3-17 Tasseled cap transformation for 30m and 60m TM_O_1986............................72 3-18 Common spectral reflectance characteristics for healthy green grass, dead grass, and bare dry soil for the wavelength interval from 0.4 to 1.1 ........73 3-23 Classification schema..........................................................................................76 3-25 Forest age determination area and training area determinations for test site (Image Jan 7, 2000).............................................................................................78 3-26 Forest age at the selected points outlined on Figure 3-25...................................79 3-27 Wetland appears on DOQ images inset from Landsat TM, March 24, 1986.....80 3-28 Clear cut and 1-3 yr pine plantation A) New Clear-Cut area B) 1 yr old pine plantation, and C) 3 yr old pine plantation.........................................................82 3-29 4-8 yr pine plantation A) 3-5 yr pine B) 5-8 yr old pine plantation....................84 xi

PAGE 13

3-30. > 8 years pine plantation A) 15-20 year old pines B) >30 year old pine plantation.............................................................................................................85 3-31 Wetland forest A) Wetland forest (riparian forest) B) mixed pine trees with ottomland hardwood...........................................................................................87 3-32 20 clusters, from which 4 classes were determined by using unsupervised classification on test site.....................................................................................90 3-33 Training data sample selection............................................................................92 3-34 Equiprobability contours defined by a maximum likelihood classifier .............95 3-35 Examples of reference data locations for various classes................................100 3-36 Classified images A) MSS_30_O classification B) TM_30_O Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest........................106 3-37 Classified images A) MSS_30_S classification B) TM_30_S Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest........................108 3-38 Classified images A) MSS_60_O classification B) TM_60_O Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest........................112 3-39 Classified images A) MSS_60_S classification B) TM_60_S Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest........................114 3-40 Classified images A) MSS_PC_TC_30_O Classification B) TM_PC_TC_30_O Classification Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest...................................................................................117 3-41 Classified images A) MSS_PC_TC_30_S classification B) TM_PC_TC_30_S Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest.................................................................................................................119 3-42 Classified images A) MSS_PC_TC_60_O classification B) TM_PC_TC_60_O Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest......................................................122 3-43 Classified images A) MSS_PC_TC_60_S classification B) TM_PC_TC_60_S Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light xii

PAGE 14

green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest.................................................................................................................124 3-44 Image difference between the 30 m classified TM and the 30 m classified MSS images.....................................................................................136 3-45 Image difference between the 60 m classified TM and the 60 m classified MSS images......................................................................................136 4-1 Composite (Bands 3,3, and 2 RGB) image of WRS 17-39 scene, application site Hamilton County and application site Alachua County in southeast costal plain........................................................................................141 4-2 Datasets used for classification.........................................................................143 4-3 Classified images used to produced look up tables..........................................145 4-4 Land cover determination for application site..................................................146 4-5 Agriculture, urban, and roads were masked from the application site (forest area is shown in green)..........................................................................147 4-6 Image classification from 1975 to 2000...........................................................150 4-7 Area of coverage change based on the years....................................................152 C-1 Two dimensional multispectral space showing the individual pixel vectors and their mean position, as defined by M, mean vector (adopted from Richards and Jia, 1999).............................................................163 C-2 Modified coordinate system in which the pixel vectors have no correlated components.......................................................................................................166 C-3 Principal component axes for the data of figure C-1........................................169 xiii

PAGE 15

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 COMPARISON OF LANDSAT MSS AND TM IMAGERY FOR LONG TERM FOREST LAND COVER CHANGE ASSESSMENT By Levent Genc August 2003 Chair: Scot E. Smith Cochair: Bon A. Dewitt Major Department: Civil and Coastal Engineering The main objective of this research is to determine forest cover change from 1975 to 2000 for a region in north Florida. In order to monitor long-term forest cover change for this project, Landsat MSS must be used with Landsat TM because Landsat MSS is the only datasets that is available for civilian use prior to the year of 1982. However, using these two different datasets in a project had been problematic and needed to be studied to obtain higher overall classification accuracy. Landsat MSS and TM classifications were achieved through a common approach. By increasing overall accuracy for both sensor types individually, images from these sensors will be more consistent. In order to achieve the main objective of this study, sixteen derived datasets were constructed and tested from March 24, 1986 Landsat MSS and Landsat TM imagery from the same area using low pass filtering, principal component Analysis (PCA), and tasseled xiv

PAGE 16

cap transformation (TCT) to demonstrate the relationship between these two imageries. Similarities were 88.5% in best case. Results suggest that smoothing operations performed prior to classification improved the classification accuracy because they created a selection of homogeneous training sets. It was found that the smoothing operation performed prior to classification improved the classification accuracy by 7% compared to the original dataset classification. Performing the PCA and TCT to smoothed datasets also improved the classification accuracy by 4% compared to the smoothed datasets only. The first two principal components, PC1 and PC2, added to the first two indexes from TCT, BI and GI, were used to create new 4-band datasets for MSS and TM. These combinations of images were used to determine the forest land cover for the application site from the years 1975 to 2000. It was found that the determination of five land cover classes using these techniques produced moderate overall land cover classification accuracy ranging from 62.7% to 88.5%. However, this smoothing operation before classification created a selection of homogeneous training sets, which can result in the loss of some detail in the land cover classification. Therefore smoothing is not recommended for other studies unless it is required to improve the visualization and specific classes. xv

PAGE 17

CHAPTER 1 INTRODUCTION A primary objective of this dissertation is to develop a methodology for relating Landsat MSS data to Landsat Thematic mapper (TM) data. This was necessary in order to determine forest land cover changes from 19752000. Using historical Landsat Multispectral Scanner (MSS) imagery is often critical to successful long-term land use-land cover change (LULCC) studies. Malila et al. (1984) compared Landsat 4MSS and TM images based on their spectral and spatial characteristics and found that over the agricultural areas both MSS and TM data from different dates were highly correlated between the MSS band 1 and TM Band 2, MSS band 2 and TM band 3, MSS band 3 and TM band 4, and MSS band 4 and TM band 4 (90%, 97%, 97%, and 98% respectively). They also computed the tasseled cap transformation indexes, Brightness Index (BI) and Greenness Index (GI) and their relationship for MSS and TM data. Malila et al. (1984) stated that the correlation between GI from MSS and TM is 99% and that BI from MSS and TM are 75 % correlated. However they did not classify the images. Benson et al. (1985) analyzed same-time images from Landsat MSS and TM data over the forest area using different combinations of Landsat MSS and TM bands to determine the best band combinations for forest cover classification. They found that image composites of TM 5, 4, and 3, MSS 4, 2, and 1, and TM 4, 3, 2 were statistically more interpretable than the image composite of TM 5, 3, 2 for forest cover area. However, Benson et al. (1985) and Malila et al. (1984) did not use the transformations of principal components and tasseled cap transformation indexes to 1

PAGE 18

2 classify and compare Landsat MSS and Landsat TM one to the other. Unlike the studies by Bensen et al (1985) and Malila et al. (1984), This research used composite of selected principal components and BI and GI from a tasseled cap transformation and classified these images using statistical and hybrid classification methods. Tokola et al. (1999) attempted to calibrate Landsat MSS and Landsat TM data using images of different dates and different statistical methods on smoothed and non-smoothed datasets. They found that there is no evidence to suggest that smoothing or calibration of MSS or TM images improved classification accuracy. In this research, raw digital numbers (DN) of Landsat MSS and Landsat TM were used to create new datasets using spatial and spectral transformation techniques to classify the resulting images. In remote sensing literature, there are no studies that show principal component, tasseled cap, and smoothing operations being performed and used to produce forest land cover classifications and the statistical relationships between MSS and TM datasets. In this research, it is hypothesized that creation of new datasets-band combinations using Principal Component Analysis (PCA) and the Tasseled Cap Transformation (TCT)-improves the accuracy of monitoring LULCC. This study has created the potential for more efficient use of MSS and TM data in a related study within a LULCC project in north Florida titled, “Determining Changes in Land-Cover and Land Use (LCLU) Patterns in the Lower Coastal Plain Region from 1975-2000.” (Binford et al., 2002) The National Aeronautics and Space Administration (NASA) initiated the first civilian program specializing in the acquisition of remotely sensed digital satellite data in 1972. The first system was called ERTS (Earth Resources Technology Satellite), and

PAGE 19

3 later renamed Landsat. There have been several Landsat satellites launched since 1972 (Jensen, 2000; Lillesand and Kiefer, 2000). Landsat 1, 2, and 3 are no longer operating, but Landsat 4 and 5 are still in orbit gathering data as well as Landsat 7 Enhanced Thematic Mapper plus (ETM+) (Gibson and Power 2000). The Landsat satellite series 1, offers the longest continuous satellite dataset available to civilian sector for land cover change studies Monitoring Land Use-Land Cover Change Satellites have monitored global LULCC over the last three decades. Forest degradation assessment is a prevalent application of land cover change studies. Forest degradation information is desired for political decision-making and the planning of forest management regimes for rehabilitation (Apan, 1997) as well as for monitoring global carbon production. Since human society utilizes the environment for the fulfillment of its essential needs such as food and recreation, humans should have some concern about maintaining a healthy and productive environment (Rindfuss and Stern, 1998). Questions that concern global change include what an environment should look like, how to keep the environment productive, what forces drive its degradation, and how to manage societal activities on a global scale such that we maintain and/or repair the earth's capacity to sustain the lives and livelihoods of all of its inhabitants. Essentially, scientists attempt to understand the causes, consequences, and areas of interference of land use and land cover for the management of global change. Human interactions with the environment are analyzed to determine whether humans cause LULCC. In order to understand global change, we must observe whether the ways in which humans use the earth's resources in their socio-cultural, technological, economic,

PAGE 20

4 political, and organizational contexts provide the entry point to gaining a better understanding of global change (Rindfuss and Stern, 1998). During the early 1970’s this issue became one of the top concerns for scientists, when the environmental movement began to flourish in the world (Turner, 1991). In the last twenty years deforestation, particularly in the tropics, has become an important global-scale phenomenon. It has been accelerating as a result of population pressure and economic development. The subject has become an important component of both science and policy. In terms of science it has been consistently singled out as a key element of many areas of global change research. It is also a common link between several important international policy issues. During the early 1990s, people recognized that main global changes are largely human-induced and that these changes do and will affect human societies, and thus that they could only be understood by analyzing the input from human life. Since then it has become common to speak of the human dimensions of environmental change and responses to global change (Rindfuss and Stern, 1998; Turner, 1991). Turner (1989) states that the human factor is fundamental within societal forces that in a causal sense link humans to nature and bring about global environmental changes. In this sense, the study of global changes through the lens of nature-society relationships is a profoundly geographical theme. Human driving forces comprise the sum of individual and group actions, but they are more manageably described as collective categories of these actions. Land use is the observed immediate reason and/or manifestation of environmental change. For example, agricultural and forestry practices have changed entire landscapes;

PAGE 21

5 land-management practices more generally alter plant and animal communities both at the species and habitat level, or they affect nutrient cycling and distribution in the soil; creation or changes of transportation routes dissect habitats and alter water and energy flows; industrial emissions affect environmental and human health. The question here is how do humans adjust to a variable and changing environment, which factors facilitate or impede such adjustments and adaptations, and which factors augment or diminish societal vulnerability to, say, climatic variability, and thus what might be the most effective avenues to take in response to global environmental changes (Turner, 1991). Ojima et al. (1994) asked the question whether land use and land cover changes affect global environmental changes. This question has no perfect answer for current LULCC projects. Different land cover types get changed through land use activities resulting in proximate sources of change, which are driven by the larger forces at work in any given social context (Meyer, 1995). The physical environment, of which land cover is but one aspect, is very much influenced by, and in turn influences, the changing global climate. In order to understand global change, it is necessary to understand the relationship between land use and land cover change. This means that the dynamic forces of land cover change must be understood, emphasizing how changes in land use control cause changes in land cover. The differences between land cover and land use change are important to define. Land use characterizes the human use of a land cover type. For instance, foresters and timber companies could use forests and farmers could use grazing land (Meyer, 1995). Land cover is the actual distribution of vegetation, water, desert,

PAGE 22

6 ice, and other physical features of the land, including those created by human activities (Meyer, 1995). However, as Meyer (1995) described, land cover change defines changes from one cover type to another. For instance, the change of forests to pasture is an important land cover change. Land cover change is not always as simple as it seems. Change is the significant modification within a single cover type or the gradual, long-term modification of one cover type to another. Transformation involves alterations of structure or function without a wholesale change from one type to another. Such changes could involve changes in productivity of land. In the end, long term and constant change in local levels would result in global change. For example, the change of forests would be the reason for climate change. Meyer (1995) said that human impact on LULCC at a local level must be the reason for global LULCC and this must be monitored using historical data. There are many remote sensing sensors available for evaluating LULCC. However, only Landsat MSS sensor has been around for almost three decades. In order to monitor the LULCC through the last three decades, Landsat MSS data must be used and must be compatible with the new face of remote sensing sensors. Remote Sensing and Forest Ages Deforestation and degradation of forests are significantly increasing atmospheric carbon dioxide in the world. The magnitude of this problem was assessed by Brown (1993), Fearnside (1996), and Houghton et al. (2000). According to Brown (1993), Kummer and Turner (1994), and Skole (1994), the most effective way to reduce a large amount of the uncertainty in carbon flux is to improve the mapping of area-change of forest age cover using remote sensing technology.

PAGE 23

7 According to Brown and Lugo (1990), regenerated forests represented approximately 30% of the world’s total closed tropical forest area and the rate of their formation exceeded 0.09 million-km 2 per year. These regenerating forests may, due to their rapid growth rate, represent a large “sink” for atmospheric carbon and thereby play an important role in the global carbon budget (Houghton et al., 2000). However, the magnitude of this carbon “sink” is uncertain as information on the extent of regenerating forests and their associated carbon flux is lacking (Curran et al., 1997 ). In mapping forested lands for carbon estimates, it is important to differentiate the different age classes of the trees. Questions exist concerning the ability of satellite sensor data to differentiate age classes within forests. Work by Fearnside (1982) using Landsat Multi-Spectral Scanner (MSS) data suggested that spectral differentiability between mature and young tropical forests was lost after two years. Using SPOT HRV (High Resolution Visible) data and ground measurements, Skole et al. (1994) found that mature forests and newly regenerated young forest growth could be discriminated accurately. Steininger (1996) indicated that mature and young forest spectral signatures are the same at around 14 year old trees for Landsat Thematic Mapper (TM) images. Steininger (1996) relied largely on field surveys to identify 16 stands of secondary forest near Manaus, Brazil. Steininger (1996) found that the age determination of these forests, which ranged from 2 to 19 years, resulted from interviewing people familiar with the ground. Both Steininger (1996) and Boyd et al. (1996) agree that the use of middle and TM band 6 (thermal infrared) wavelengths greatly improve mature and young forest age discrimination.

PAGE 24

8 Landsat images have been found to be particularly well suited for mapping of age of tree due to its relatively low cost and large area coverage (Steininger, 1996; Boyd et al., 1996). A variety of image processing techniques including the Normalized Difference Vegetation Index (NDVI) ( Jensen, 1996; Lillesand and Kiefer, 2000; Genc et al., 2002), and various unsupervised and supervised classification approaches have been used for forest mapping applications (Jensen1996; Kimes et al., 1998; Richard and Jia, 1999). Description of Problems Between MSS and TM Landsat satellites 4 and 5 represented the second phase of the Landsat program. The Landsat-4 and Landsat-5 satellites carry both the TM and MSS remote sensing instruments (Table 1-1). The MSS instrument is similar to MSS instruments carried by Landsat-1, Landast-2, and Landsat-3. However, the addition of the TM instruments to Landsat-4 and landsat-5 required several design changes in the MSS instruments carried on these satellites because of the lower orbit and new satellite platform dictated by the TM instrument. These design changes were made in part to make the Landsat-4 and Landsat-5 MSS data radiometrically and geometrically compatible with MSS data from the earlier Landsat satellites (Jensen, 1996; Lillesand and Kiefer, 2000). MSS (multispectral scanner) images from Landsat series have ground coverage of roughly 185 km x 170 km from a height of approximately 900 km for Landsat 1, 2, and 3 and 705 km for Landsat 4 and 5 (Lillesand and Kiefer, 2000; Jensen, 2000; Gibson and Power, 2000). MSS data are widely used for general geologic studies as well as vegetation inventories. The spatial resolution of MSS data is 56 x 79 m, with 79x79-meter pixels IFOV (Table1-1). A typical scene contains approximately 2340 rows and 3240 columns. The radiometric resolution, brightness value DN’s, is converted to 6-bit

PAGE 25

9 (0-63), but it is stored as 8-bit (Markham and Barker, 1985; Jensen, 1996; Richards and Jia, 1999; Lillesand and Kiefer, 2000; Jensen, 2000). Landsat MSS sensors record reflectance intensities in four wavelength bands: 0.5 -0.6 m (green), 0.6-0.7 m (red), 0.7-0.8 m (near-IR) and 0.8-1.1 m (IR) (Figure 1-1). The TM (thematic mapper) is a multispectral scanning system much like the MSS, except that the TM sensor records reflected and emitted electromagnetic energy from the visible, reflective-infrared, middle-infrared, and thermal-infrared regions of the spectrum (Richards and Jia, 1999; Lillesand and Kiefer, 2000; Jensen, 2000; Gibson and Power, 2000). As is demonstrated in Table 1-1, Landsat TM has higher spatial, spectral, and radiometric resolution than Landsat MSS (Jensen, 1996; Lillesand and Kiefer, 2000). Landsat TM sensors record reflectance intensities in 7 wavelength bands: 0.45-0.52 m (blue), 0.52-0.60 m (green), 0.63-0.69 m (red), 0.76-0.90 m (Near infrared), 1.55-1.75 m (mid infrared), 10.4-12.5 m (thermal infrared), and 2.08-2.35 m (mid infrared) (Table 1-1 and Figure 1-2). Landsat TM has a swath width of approximately 185 km from a height of approximately 705 km. It is useful for vegetation type and health determination, soil moisture, snow and cloud differentiation, and rock type discrimination (Jensen, 2000). The spatial resolution of TM is 28.5 m x 28.5 m for all bands except the thermal (band 6), which has a spatial resolution of 120 m x 120 m (Richards and Jia, 1999). The larger pixel size of this band is necessary for adequate signal strength. However, the thermal band is resized to 30 m x 30 m to match the other bands (Gibson and Power 2000). The radiometric resolution is 8-bit, meaning that each pixel has a possible range of data values from 0 to 255. Hence, the properties and quality of

PAGE 26

10 Landsat data have varied with time, which introduces difficulties into the comparison of images recorded by different platforms.

PAGE 27

Table 1-1. Characteristics of Landsat-MSS and Landsat TM multispectral scanners Landsat Sensors Spatial Spectral Bands Radiometric Band Range Ground Coverage MSS 57m with 79m IFOV 4 6-bit (Stored 8-bit) 0.50-0.60 m (green) 0.60-0.70 m (red) 0.70-0.80 m (Near infrared) 0.80-1.10 m (Infrared) 185 x 185 km TM 28.5m 120x120m (Band 6) 7 8-bit 0.45-0.52 m (blue) 0.52-0.60 m green 0.63-0.69 m red 0.76-0.90 m (Near infrared) 1.55-1.75 m (mid infrared) 10.4-12.5 m (thermal infrared) 2.08-2.35 m (mid infrared) 185 x 185 km 11

PAGE 28

12 Figure 1-1. Landsat MSS sensor records reflectance intensities in four wavelength bands with 57x57 ground cell size: (a)0.5 -0.6 m (green), (b) 0.6-0.7 m (red), (c) 0.7-0.8 m (near-IR), (d) 0.8-1.1 m (IR), and (e). RGB: 421.

PAGE 29

13 Figure 1-2. Landsat TM sensor records 6 wavelength bands with approximately 30x30 ground cell size: (a) 0.45-0.52 m (blue), (b) 0.52-0.60 m (green), (c) 0.63-0.69 m (red), (d) 0.76-0.90 m (Near infrared), (e) 1.55-1.75 m (mid infrared), (f) 2.08-2.35 m (mid infrared), (g) RGB.

PAGE 30

14 Image Enhancement for Land Cover Mapping Low Pass Filtering A low–pass filter is one of the spatial filters that smoothes the image by attenuating the periodic “salt and paper” noise recorded by remote sensing sensors (Jensen 1996, Richards and Jia, 1999). Image enhancement that de-emphasizes or blocks the high spatial frequency details is a low –pass filter, which uses the moving window function (Wolf and Dewitt, 2000). Wolf and Dewitt stated that the moving window operation has two inputs: original data and a localized response function called a kernel. Principal Component Analysis (PCA) PCA has proven to be of great value in analysis of multispectral remotely-sensed data (Gonzalez and Woods, 1993; Jensen, 1996; Richard and Jia, 1999). The transformation of the raw remote sensor data using PCA can result in new principle component images that may be more interpretable than the original data (Singh and Harrison, 1985). PCA may also be used to compress the information content of n number of bands of imagery into fewer than n number of bands of transformed principle component images (Richards and Jia, 1999). PCA has been used in remote sensing for different purposes. Mathematical derivation of PCA and its application have been demonstrated by many researchers including Gonzalez and Woods (1993), Jensen (1996), Richard and Jia (1999) and Lilesand and Kiefer (2000). PCA has been used to correlate Landsat (TM) imagery for prediction of land cover change (Mather, 1999). Vani et al. (2001) applied PCA to Indian Remote Sensing satellite (IRS) imagery to describe data fusion. Hunter and Power (2002) applied PCA to Compact Airborne Spectrographic Imager (CASI) data and found

PAGE 31

15 that combination of PC2, PC3, and PC4 produced the best vegetation and sediment classification results. Tasseled Cap Transformation (TCT) The tasseled cap transformation was primarily developed for and tested in agricultural applications of remote sensing data. Kauth and Thomas (1976) developed the tasseled cap transformation originally for application to MSS band space. However, Crist and Cicone (1984) developed a similar transformation for TM data occupying the same spectral regions that Kauth and Thomas (1976) examined (Jensen, 1996). The tasseled cap transformation can be described as a vegetation index, but mathematically it is a factor analysis. MSS and TM data are highly correlated permitting band ratio transformations, but because of this, the effective dimensionality of these data may be less than the total number of bands recorded (Crist and Cicone, 1984). Knowing that high correlations exist within the data indicates that factor transformations may be particularly effective at reducing dimensionality while maintaining variability. Tasseled cap is derived from a rotation of principal components. However, the axes are rotated according to a set of coefficients. Standard un-calibrated coefficients are applied in this use of tasseled cap (Kauth and Thomas, 1976; Crist and Cicone, 1984; Jensen, 1996). Calibration of Remote Sensing Data The process of monitoring LULCC can usually be divided into calibration and actual analysis. There is evidence indicating the usefulness of calibration of raw sensor data prior to any multitemporal analysis (Duggin and Robinove, 1990; Coppin and Bauer, 1994). A radiometric refinement technique was developed by Hall et al. (1991) and it has been widely used in land cover change studies. The algorithm calibrates pixel values using the Kauth–Thomas tasseled cap transformation, based on radiometric control sets

PAGE 32

16 and landscape elements whose reflectance is nearly constant over time. The method rectifies the images using a linear transformation and assumes that the differences between images contain only linear differences compared to radiometric control sets. The images under different atmospheric conditions were calibrated within 1% accuracy of absolute reflectance in the visible and near-infrared bands Coppin and Bauer (1994) and Knick et al. (1997). Singh (1986) indicated that change detection methods using regression functions were the best method. However, there are some results that show that the calibration of multitemporal imagery slightly reduces the interpretation accuracy (Collins and Woodcock, 1996). Stein et al. (1999) said that there is some scope for using spatial prediction methods in remote sensing and this helps us to understand band-to-band relations. Forest cover change studies applied absolute calibration to the forest area, but the difficulties in sensor calibration and the lack of atmospheric parameters for different sets of conditions make relative calibration an attractive and simple solution (Singh, 1986). Horan et al. (1974) pointed out that because of “degradation,” optical data from the solar calibrators on Landsat 1 were problematic. Thome et al. (1997) found that problems of altitude control in Landsat 2 made the data difficult to interpret. The first generation Landsat data also suffered from “within-scan-line drop” and “scan-correlated level shift” (Metzler and Malila, 1985). Investigations into the data from the Landsat 5 TM showed the uncertainties in radiometry within-scene calibration to be about 5.0% for bands 1 to 4 (Thome et al., 1997). Technical problems and variations in data quality result in the fact that it is very difficult to compare different data sources directly even within the same satellite

PAGE 33

17 program. Thome et al. (1997) indicated that calibration of the multiple Landsat systems must allow the inter-comparison of datasets from each system. Absolute calibration could be achieved with atmospheric (Richter, 1997) or sensor calibration (Olsson, 1995). Anuta et al. (1983) found that visual assessment of MSS imagery over Lake Michigan revealed major noise patterns in all bands. They suspected that this noise pattern occurred as a result of data processing. However, they found that there was no visual noise effect apparent in the thematic data sets. Anuta et al. (1983), Jensen (1996), Richards and Jia (1999), and Campbell (2002) stated that the spectral dimensionality of the TM sensors is of great interest since it potentially represents nearly double the information over the MSS. Anuta et al. (1983) applied principal component transformation to investigate the dimensionality of MSS and TM over the Chicago O’Hare test site in Illinois, USA. They stated that the first two principal components of MSS and the first four components of TM data contain a maximum amount of information of original datasets. Spectral class analyses in MSS and TM data were applied to determine the spectral information effect on land cover classification. Anuta et al. (1983) applied classification on MSS and TM dataset near Des Moines, Iowa, from Sept. 3, 1982, to determine the classes of agricultural, forest suburban, urban, and water. They used the unsupervised classification algorithm to produce the results. While TM data initially defined 94 classes and obtained 42 final target spectrally separable classes, spectrally separable classes in MSS were 21, which is half of the result obtained with TM. Because the detection and identification of different land cover conditions are critical steps in an effective stratification procedure, the high quality TM image products

PAGE 34

18 from Landsat-4 and Landsat-5 systems should play an even more valuable role in this regard. Benson and Degloria, (1985) stated that the spectral quality of TM data could be evaluated for survey purposes by determining the extent to which natural targets in vegetated environments can be discriminated using specially generated and commercially available image products. Using extracted spectral statistics for specific land cover types, and comparing this with selected photographic reproductions of the same area representing various spectral band combinations, one could qualitatively determine which color composite provides the most information for stratification purposes (Benson and Degloria, 1985). Composites of TM and MSS images were used to identify the major forest cover types using interpretation techniques (Benson and Degloria, 1985). In the last three decades, LULCC has been monitored in nearly all regions of the world. In order to measure long-term land cover change on earth, researchers must use historic Landsat MSS and Landsat TM data. Williams et al. (1984) assumed that using TM band 2, 3, 4 to create new images could act like MSS datasets. Outline Chapter 2 describes the objectives of this study. Chapter 3, Method and Results, explains the generation of new datasets for classification accuracy improvement. Since ancillary data were not available, land covers were determined using Landsat MSS, Landsat TM, and Landsat ETM+ images from 1975 to 2000. Once land covers were determined, the maximum likelihood classifier, which is the most common technique, was applied to several different band combinations consisting of two different data sources: red, green, near infrared, Principal Components, and Tasseled cap indexes. Determination of the best data combinations for land cover

PAGE 35

19 classification was investigated using classification accuracies after the best dataset was determined. In Chapter 4, the dataset determined the best in terms of classification accuracy was used to classify all available images from 1975 to 2000 for the application site in Alachua County, Florida, to determine the attribute table for carbon related analysis. Chapter 5 contains the conclusions. It compares land cover change and overall accuracy for individual years and their coverage areas. It demonstrates how the results from this work can be used for additional research. In order to make this document self-contained, it summarizes several important concepts and procedures in the appendices. Appendix A describes the Normalized Difference Vegetation Index (NDVI) that was used for pre determination of datasets for initial classification. Appendix B also describes the texture analysis that was used as one of the components to determine pre classification before the actual data creation took place. Appendix C explains the procedure of principal component transformation.

PAGE 36

CHAPTER 2 HYPOTHESIS The main objective of this research is to develop a method that permits an increase in land cover classification accuracy (LCCA) using Landsat MSS and Landsat TM data particularly for a project on “LAND USE AND LAND COVER CHANGE: Decadal-Scale Dynamics of Land Ownership, Land Management and Carbon Storage in the Southeastern Lower Coastal Plain of the US (1975-2000)”. In order to achieve the main objective of this study, it is important to demonstrate and understand the exact relationship between spatially and spectrally different Landsat MSS and Landsat TM imagery. For this purpose, March 24 Landsat MSS 1986 and March 24 Landsat TM 1986 imagery from the same area, acquired at the same time, were studied to determine a relationship between these two images. The atmospheric affect difference is zero, since images used in this research were taken at same time. This is assumed to be the perfect condition for a study to understand and demonstrate relationships using certain extracted information, calculated PCA and TCT along with low pass filtering from Landsat MSS and Landsat TM data, and create new band combinations that may improve the accuracy of classification. Since these combinations of images have not been done before, results of this research may make it possible to use MSS and TM data more efficiently in a similar study. . 20

PAGE 37

CHAPTER 3 DETERMINATION OF DATASETS FOR CLASSIFCATION ACCURACY IMPROVEMENT The determination of datasets for classification accuracy improvement is divided into four major stages (Figure 3-1). The first stage in image processing is geometric correction. Once the raw data have been properly geo-referenced to real world coordinates, additional products created through processing could inherit the spatial context of the unprocessed data (Jensen, 1996; Richards and Jia, 1999; Campbell, 2002). In the second stage, after the data have been properly corrected, it is possible to begin manipulating the images to produce information to obtain more interpretable images from the raw images. In half of the tests, prior to transformation, low-pass filtering was performed using a mean value kernel-moving window to attempt to improve the result of classification (Richards and Jia, 1999; King, 2002). After spatial filtering, PCA and TCT were the operations used to obtain data including more readily interpretable datasets, components and indexes, for the test site (Jensen, 1996; Patterson and Yool, 1998; Richards and Jia, 1999; Lillesand and Kiefer, 2000; Campbell, 2002). After transformations, both principal components (PCs) and tasseled cap indexes (TCIs) were used to create new datasets. All datasets for the test site including original raw images were used to determine the relationships among the datasets using linear regression analysis (Tokola et al., 1999) and all datasets were classified with and without low pass filtering. 21

PAGE 38

22 The third stage is the determination of land cover classes for the testing site using Landsat MSS, Landsat TM, and Landsat ETM+ images as well as U.S. Geological Survey (USGS) digital orthophoto quadrangle (DOQ) and field trips. For this purpose also, unsupervised classification was performed to determine the initial land cover pattern (Seyler et al., 2002). The fourth stage is the process of actual classification using unsupervised and supervised classification (Jensen, 1996; Richards and Jia, 1999; ERDAS, 1999; Lillesand and Kiefer, 2000; and Campbell, 2002). Training data were selected and accuracy assessment was performed according to Congalton and Green (1998).

PAGE 39

23 Figure 3-1. Procedure of land cover classification for test site

PAGE 40

24 Study Area The study area was divided into two regions: (a) test site and (b) application site. The detail of the application site is described in the following chapter (4). The test site in Hamilton County, Florida was selected to run the model. A 7 km x 13 km area was selected in which land cover types are clearly determined (no clouds or defects). The selection of the test site used the criterion of the availability of both Landsat MSS and Landsat TM data for March 24, 1986 (Figure 3-2). Once the model was developed for the test site, the results were applied to the application site to produce the land cover classification. Figure 3-2. Composite of bands 4, 3, and 2 (RGB) image of WRS 17-39 scene for the test site in Hamilton County

PAGE 41

25 Image Processing to Create the Datasets The following sections provide a description of manipulations made to the images for the purpose of identification for land cover classifications in order to achieve the requirements of the long-term forest land cover change project. Since Landsat MSS data is the only earth looking imagery that is available for civilian use prior to the year of 1982 (Jensen, 1996), Landsat MSS and recent Landsat TM (or other satellite imagery) must be used in the same project to determine long term forest cover change such as the project on ‘LAND USE AND LAND COVER CHANGE: Decadal-Scale Dynamics of Land Ownership, Land Management and Carbon Storage in the Southeastern Lower Coastal Plain of the US (1975-2000)’. Geometric Correction Geometric registration and correction of images had been completed in previous research. Coordinates of sixty ground-control points (GCP) were collected, evenly distributed over the study area at the intersections of roads, using single receiver code phase position GPS. Initially the September 30, 1997 Landsat TM scene was chosen as the base image for the georectification to the UTM coordinate system (zone 17N), WGS84 datum and field-checked for accuracy because it exhibited good contrast between vegetation and roads as well as no cloud cover. The rest of the 67 Landsat MSS and Landsat TM images from the study area were registered using image-to-image registration techniques using the September 30, 1997 image as reference. The RMS error for the geometric correction (first order polynomial with nearest neighbor resampling) was 9.7 m. After image registration had been done, for accuracy purposes, a corrected base image from September 30, 1997 was checked with the 1:100,000 tiger road map and a DOQ (USGS, 2002) resampled to a 3 m pixel size. DOQ’s are processed, digital aerial

PAGE 42

26 photographs in which image displacement caused by terrain relief and camera tilt has been removed. Subsequently multiple Landsat TM images from 1982 to 1999 and ETM+ 2000 images were registered to the base scene using image-to-image registration. MSS images and TM images acquired at the same moment, December18, 1982, March 24, 1986, and March 16, 1989, were registered to each other using image to image registration with second order polynomial transformation and the nearest neighbor image resampling method (Figure 3-3). MSS images from 1975 to 1981 were registered using the TM December 1982 image as reference using the same techniques as the previous image registration (Figure 3-3). The average RMS error was 8.0 m for all the images with a maximum value of 8.9 m and a minimum of 6.8 m. In order to obtain spatial consistency, all Landsat MSS, Landsat, and Landsat ETM+ images for the application site were resampled to a 30 m pixel size using the affine transformation with the nearest neighbor resampling technique (Wolf and Dewitt, 2000). In addition, the Landsat MSS image from March 24,1986 of the test site area was resampled to 60m in order to analyze the spatial effect on classification.

PAGE 43

27 Figure 3-3. Geometric correction for study area images

PAGE 44

28 Preparation of New Datasets for the Test Site Data preparation is divided into two sections: (1) Pre-determination and (2) Preparation of the new datasets for test site. Pre-determination was done in order to determine the best datasets for the test site and to develop a model. The new datasets were constructed from the dataset that had the highest overall accuracies in the pre-determination phase. Pre-determination Before the final data preparation process started, a total of sixty-eight different datasets were classified using supervised classification as described by Jensen (1996) to determine the best dataset that could be used for land cover classification for this research. The decision was made based on overall accuracy. Band components of the datasets that gave the highest overall accuracies were chosen to construct new datasets. Tested datasets are described and listed in Table 3-1. The following section describes the decision making process for the datasets used. Details of filtering (smoothing), PCA that produced principal component 1 (PC1) and principal component 2 (PC2), and TCT that produce brightness index (BI) and greenness index (GI), and yellowish index (YI), can be seen later on in this chapter (Richards and Jia, 1999; Jensen, 1996; Jensen, 2000; and Campbell, 2002). However some terms will be explained in the appendices because they are not directly related to the focus of this research. Details of these terms, Normalized Difference Vegetation Index (NDVI) can be found in Richards and Jia (1999), Jensen (1996), Jensen (2000), and Campbell (2002) (Appendix A) and texture analysis can be found in Richards and Jia (1999) (Appendix B). Since similarities are high between NDVI and TCT second index, GI, it was preferable to use GI for creating datasets because TCT also provide GI which was used as

PAGE 45

29 another band for creation of dataset. The original MSS 4 band dataset and the original TM 6 band dataset were each resampled to 30 m and 60 m. After resampling, various image enhancement techniques were used to create new datasets and these datasets were tested to determine the best dataset for classification. MSS dataset. After all MSS datasets were classified, it was found that the overall classification accuracies, the sum of the correct classified pixel divided by the sum of the total number of pixel classified (Congalton and Green, 1998), for the original datasets (1_2_3_4_O), using the random sampling method, were 67.1% for 30m and 65.0% for the 60 m dataset (Table 3-1). The original MSS dataset was smoothed and called 1_2_3_4_S (Smoothed MSS band1, band2, band3, band4) datasets, which produced better overall accuracy for both 30m and 60 m (71.1% and 67.5% respectively) than the original datasets. As is demonstrated in Table 3-1 the NDVI_4-band_O (the NDVI band was added into the original 4-band dataset as a 5th band) dataset produced 68.0% and 66.1% overall accuracies for both 30m and 60m dataset respectively. Smoothing the NDVI_4-band_S dataset gave a better overall accuracy than the NDVI_4-band_O dataset for both 30m and 60m (71.8% and 72.3% respectively). Using composite images of PC1, PC2, and PC3 (PC123_O) from the MSS original dataset (1_2_3_4_O MSS) improved the overall accuracies from 67.1% to 74.1% for the 30 m dataset and from 65.0% to 72.4% for the 60 m dataset. Likewise, PC123_S, the smoothed principal component datasets, also improved overall accuracies from 71.1% to 78.2% for the 30 m dataset and from 67.5% to 68.1 % for the 60 m dataset (Table 3-1). Furthermore, overall accuracies, using a composite of PC1, PC2, and PC3 (PC123_O) of original MSS 30 m datasets and PC1, PC2, and PC3 (PC123_S) of the

PAGE 46

30 smoothed MSS 30m, were found to be 74.1% and 78.2% respectively. Overall accuracies, using a composite of PC1, PC2, and PC3 (PC123_O) of original MSS 60 m datasets and PC1, PC2, PC3 (PC123_S) of the smoothed MSS 60m, were found to be 72.4 % and 68.1 % respectively. Overall accuracies, using a composite of PC1 and PC2 (PC12_O) of original MSS 30 m datasets and PC1 and PC2 (PC12_S) of the smoothed MSS 30m, were found 76.6 % and 79.2% respectively. Overall accuracies, using a composite of PC1 and PC2 (PC12_O) of original MSS 60 m datasets and PC1 and PC2 (PC12_S) of the smoothed MSS 60m, were found to be 74.5 % and 73.8 % respectively. Tasseled cap indexes were also used to classify the test site. Overall accuracies using the composite of the original first two indexes, BI and GI (BI_GI_O) MSS dataset for 30 m and the composite of the smoothed first two indexes BI and GI (BI_GI_S) MSS dataset for 30 m were found to be 76.6 % and 78.9% respectively. Overall accuracies for 60 m dataset using the composite of the original first two indexes, BI and GI (BI_GI_O) MSS dataset and the composite of the smoothed first two indexes BI and GI (BI_GI_S) MSS dataset were found to be 77.1 % and 70.5% respectively. It was also found that using PC12_BI_GI_O datasets derived from Landsat MSS for both 30 m and 60 m provide 74.2% and 77.0% overall accuracies respectively. Likewise, PC12_BI_GI_S datasets derived from Landsat MSS provided 79.0% overall accuracy for 30 m and 76 % for 60 m datasets. It was assumed that using first order 3x3 texture of original PC1 and PC2 datasets might improve the overall classification accuracy for both MSS and TM datasets. However, in general, inclusion of the texture band to created a new dataset caused lower accuracy than without it.

PAGE 47

Table 3-1. Determination of overall classification accuracy for various datasets 31 Dataset name Description of Dataset A B C D 1_2_3_4_O Original MSS 4 band image 67.1 65.0 1_2_3_4_S MSS dataset after smoothing 71.1 67.5 1_2_3_4_5_7_O Original TM 6 band image 79.6 79.8 1_2_3_4_5_7_S TM dataset after smoothing 4 band 81.5 82.1 2_3_4_O Original three band (2, 3, 4) TM image 79.6 81.8 2_3_4_S TM three band (1,2,3) after smoothing image 80.1 77.1 NDVI_6-band_O NDVI plus original 6-band TM 80.1 79.2 NDVI_6-band_S NDVI plus smoothed 6-band TM 82.3 78.9 NDVI_4-band_O NDVI plus original 4-band MSS 68.0 66.1 NDVI_4-band_S NDVI plus smoothed 6-band MSS 71.8 72.3 PC123_O PC1, PC2, and PC3 of original image 74.1 83.3 72.4 80.8 PC123_S PC1, PC2, and PC3 of smoothed image 78.2 86.4 68.1 80.5 PC12_O PC1, PC2 of original image 76.6 82.1 74.5 79.0 PC12_S PC1, PC2 of smoothed image 79.2 87.0 73.8 79.5 BI_GI_WI_O BI, GI, and WI of original 6-band TM image 83.6 80.1 BI_GI_WI_S BI, GI, and WI of smoothed 6-band TM image 84.2 79.4 BI_GI_O BI, GI of original image 76.6 84.7 77.1 77.9 BI_GI_S BI, GI of smoothed image 78.9 82.1 70.5 80.3 BI_GI_TXofPC12_O BI, GI of original plus texture of PC12 of original image 71.6 80.1 66.6 59.6 BI_GI_TXofPC12_S BI, GI of smoothed plus texture of PC12 of smoothed image 74.9 82.7 69.1 64.5 NDVI_PC12_BI_GI_TXofPC12_O NDVI, PC1, PC2, BI, GI plus texture of PC12 of original image 71.0 79.5 70.2 78.1 NDVI_PC12_BI_GI_TXofPC12_S NDVI, PC1, PC2, BI, GI plus texture of PC12 of smoothed image 75.6 78.3 71.6 77.9 PC12_BI_GI_TXofPC12_O PC1, PC2, BI, GI plus texture of PC12 of original image 74.1 77.1 66.0 60.3 PC12_BI_GI_TXofPC12_S PC1, PC2, BI, GI plus texture of PC12 of smoothed image 66.2 69.5 61.8 62.8 PC12_BI_GI_O PC1, PC2, BI, GI original image 74.2 83.5 77.0 83.7 PC12_BI_GI_S PC1, PC2, BI, GI smoothed image 79.0 88.5 76.2 84.0 A) MSS 30 Overall Accuracy (%). B) TM 30 Overall Accuracy (%). C) MSS 60 Overall Accuracy (%). D) TM 60 Overall Accuracy (%)

PAGE 48

32 TM dataset. Using the 6-band original TM dataset (1_2_3_4_5_7_O) achieved overall classification accuracies higher than the original 4 band MSS dataset (1_2_3_4_O) for both the 30 m and the 60 m datasets. Overall accuracies were 79.6% for the 30 m dataset and 79.8% for the 60 m dataset (Table 3-1). After the original 6 band TM dataset was smoothed, the resulting 1_2_3_4_5_7_S dataset improved the overall accuracies for both the 30m and the 60 m (81.5% and 82.1% respectively). Spectrally TM band 2 corresponds to MSS band 1 (red) and TM band 3 corresponds to MSS band 2 (green) (Malila, et al. 1985, Jensen, 1996). TM band 4 spectrally falls in approximately the same spectral range that corresponds to MSS band 3 and band 4 (Table 1-1). TM bands 2, 3, and 4 were selected to create a new dataset called 2_3_4_O (original 3 band Landsat TM) and 2_3_4_S (Smoothed 3 band Landsat TM). Using 2_3_4_O and 2_3_4_S improved the overall accuracy. Overall accuracies were found to be 79.6% for the 30 m dataset and 81.8% for the 60 m dataset. Smoothing the 3-band Landsat TM dataset improved the accuracy of the original 30 m dataset (80.1%) and decreased the accuracy for the 60 m dataset (77.1%) (Table 3-1). The NDVI and original 6-band Landsat TM were also used to create 7 band new datasets using original TM called NDVI_6-band_O and NDVI_6-band_S. These dataset were used to produce overall accuracies for the test site’s datasets, which were found to be 80.1% for original and 82.3 % for smoothed datasets. Results show that adding NDVI into the original Landsat TM dataset did improve the overall accuracies. Overall accuracies for NDVI_6-band_O datasets were found to be 80.1% for the 30m dataset and 79.2% for the 60m dataset. Overall accuracy using smoothing NDVI_6-band_O,

PAGE 49

33 NDVI_6-band_S datasets, was found to be 82.3% for the 30m dataset and 78.9% for the 60m dataset. NDVI_PC12_BI_GI_TXofPC12_O and NDVI_PC12_BI_GI_TXofPC12_S Using the first three principal components (PC123_O) of the original 6-band Landsat TM dataset increased the overall classification accuracies to 83.3 from 79.6% for the 30 m dataset and 80.8 from 79.8% for the 60 m dataset when compared with the original 6-band Landsat TM dataset (1_2_3_4_5_7_O). Smoothing the PC123_O (PC123_S) improved the overall classification accuracy to 86.4% from 85.1% for the 30 m dataset and decreased the overall classification accuracy to 80.5 from 82.1% for the 60m dataset when compared with the original 3 bands Landsat dataset. When the first two principal components (PC1 and PC2) were used for classification, overall accuracy decreased for the PC12_O (82.1%) and increased for the PC12_S (87.0%) compared with the PC123_ O and the PC123_S datasets respectively for the 30 m dataset. However for the 60 m dataset both overall classification on PC12_O and PC12_S decreased to 79.0% and 79.5% respectively. For the Landsat TM, tasseled cap first tree indexes brightness index (BI), greenness index (GI), and water index (WI) were also tested for overall accuracy. This was done to check how the first three tasseled cap indexes (BI, GI, and WI) differ from the first two-tasseled cap index (BI and GI) in terms of classification accuracy. The overall accuracies were found for the BI_GI_WI_O and the BI_GI_WI_S for 30 m datasets as 83.6%and 84.2 % respectively. Overall accuracy for 60 m TM datasets using tasseled cap indexes to create the BI_GI_WI_O and the BI_GI_WI_S were found to be 80.1 % and 79.4 % respectively. The overall accuracies, for 30 m dataset using the first two tasseled cap indexes to create the dataset called the BI_GI_O and the BI_GI_S, were found to be 84.7

PAGE 50

34 % and 82.1 % respectively (Table 3-1). For the 60 m TM datasets, the overall accuracies were 77.9 % for the BI_GI_O and 80.3 % for the BI_GI_S. The highest overall accuracy was produced by PC12_BI_GI_S dataset derived from Landsat 30 m and 60 m dataset (88.5% and 84.0 % respectively). Decisions After sixty-four different datasets were examined for overall classification accuracy (Table 3-1), it was determined that PCs and TCT indexes increased the overall accuracies. NDVI and original dataset combination also produced high overall classification accuracy for TM 30m dataset. PC123_S, PC12_S, and BI_GI_WI_S datasets (86.4%, 87.0%, and 84.7% respectively) for the Landsat TM and PC123_S, PC12_S, and BI_GI_S datasets (78.2%, 79.2%, and 78.9% respectively) for the Landsat MSS provided the best overall classification accuracies for the test site. PC1 and PC2 explain virtually all of the variance in a dataset (Jensen, 1996; Lillesand and Kiefer 2000), so it was decided to use the first two PCs for both Landsat MSS and Landsat TM datasets on 30 m and 60 m. A Table 3-1 indicates that the PC12 original dataset for MSS has a higher overall classification accuracy than the PC123 of MSS does. Since the MSS dataset does not provide any WI index, only BI and GI index were used for both Landsat TM and Landsat MSS datasets for consistency. It was assumed that adding BI and GI into PC12 and creating a new dataset would improve the classification accuracy. Detailed investigations as described in the following section were performed to understand the classification accuracy improvement using only PC1, PC2, BI, and GI for both 30 m and 60 m on MSS and TM datasets. Jensen (1996) stated that GI and NDVI of remotely sensed data are highly correlated (99%). Since TCT produced the GI as well as the BI in same image, it was preferable to use GI to create new datasets instead of NDVI.

PAGE 51

35 As is seen in Table 3-1, Composite of PC1, PC2, BI, and GI of original and smoothed datasets, called PC12_BI_GI_O and PC12_BI_GI_S, gave the highest overall classification accuracies for both 30 m and 60 m spatial resolutions. Therefore, these datasets must be investigated in advance for further decisions. Preparing New Datasets for the Test Site After deciding to use PC1, PC2, BI, and GI for image classification on both Landsat MSS and Landsat TM, it was determined that these datasets should be tested step by step to monitor the classification accuracy improvements. Five steps were taken for the new dataset preparation: First, geo-corrected images from the selected test site were resampled to 30 and 60 m spatial resolutions as described previously. Then 4-band Landsat TM dataset, band2, band3, two times band4,t were spectrally extracted from the original 6-band Landsat TM (1, 2, 3, 4, 5, and 7) for 30 m and 60 m datasets that corresponded to 4-band Landsat MSS 30 and 60 m datasets (band1, band2, band3, and band4) and called the MSS_30_O_1986, the TM_30_O_1986, the MSS_60_O_1986, and the TM_60_O_1986 datasets. After creating the original band combinations, the 4-band MSS and 4-band TM datasets were filtered using 3x3 low-pass filtering techniques to create the MSS_30_S_1986, the TM_30_S_1986, the MSS_60_S_1986, and the TM_60_S_1986 datasets. PCA was performed on the MSS_30_O_1986, the TM_30_O_1986, the MSS_30_S_1986, the TM_30_S_1986, the MSS_60_O_1986, the TM_60_O_1986, the MSS_60_S_1986, and the TM_60_S_1986 datasets and four band PC images were created called the MSS_PC_30_O_1986, the TM_PC_30_O_1986, the

PAGE 52

36 MSS_PC_30_S_1986, and the TM_PC_30_S_1986, the MSS_PC_60_O_1986, the TM_PC_60_O_1986, the MSS_PC_60_S_1986, and the TM_PC_60_S_1986 (Figure 3-4). TCT was performed on the MSS_30_O_1986, the TM_30_O_1986, the MSS_30_S_1986, the TM_30_S_1986, the MSS_60_O_1986, the TM_60_O_1986, the MSS_60_S_1986, and the TM_60_S_1986 datasets and three band TC images were created and called the MSS_TC_30_O_1986, the TM_TC_30_O_1986, the MSS_TC_30_S_1986, and the TM_TC_30_S_1986, the MSS_TC_60_O_1986, the TM_TC_60_O_1986, the MSS_TC_60_S_1986, and the TM_TC_60_S_1986 (Figure 3-4 and Table 3-2). New data layer composites were created of PCs and vegetation indexes. The first two PCs from the MSS_PC_30_O_1986, the TM_PC_30_O_1986, the MSS_PC_30_S_1986, the TM_PC_30_S_1986, the MSS_PC_60_O_1986, the TM_PC_60_O_1986, the MSS_PC_60_S_1986, and TM_PC_60_S_1986 datasets and BIs and GIs from the MSS_TC_30_O_1986, the TM_TC_30_O_1986, the MSS_TC_30_S_1986, the TM_TC_30_S_1986, the MSS_TC_60_O_1986, the TM_TC_60_O_1986, the MSS_TC_60_S_1986, and the TM_TC_60_S_1986 datasets were used to create the 4-band data sets called the MSS_PC_TC_30_O_1986, the TM_PC_TC_30_O_1986, the MSS_PC_TC_30_S_1986, the TM_PC_TC_30_S_1986, the MSS_PC_TC_60_O_1986, the TM_PC_TC_60_O_1986, the MSS_PC_TC_60_S_1986, the TM_PC_TC_60_S_1986 datasets (Figure 3-4 and Table 3-2). A total of sixteen datasets were produced from the original 4-band Landsat MSS 30m and 60m and original 6-band Landsat TM (1, 2, 3, 4, 5, and 7) 30m and 60m images

PAGE 53

37 from March 24, 1986 on the application site. As is seen in Figure 3-5, before new data sets were produced, 3x3 low-pass filtering, PCA, and TCT were performed to produce the specific indexes and components to create new datasets. Determination of calibration coefficients for created new datasets A previous study relates MSS and TM dataset acquired at different dates (Tokola et al., 1999). In this present study, since both MSS and TM data were acquired at the same moment, these datasets will be easier to compare than for the case where MSS and TM imagery were acquired on different dates. At the same moment MSS acquired data and TM sensor are assumed to have the same atmospheric conditions and the same land cover (no land cover changes) and should provide for better classification consistency. Determination of coefficients may help to understand the relationships between MSS and TM datasets enabling them to be used in the same study with more effectiveness terms of land cover classification. Jensen (1996) stated that using low pass smoothing is useful for reducing periodic “ salt and paper” noise recorded by electronic remote sensing systems. After all datasets were created, statistical analysis, linear regression, were employed to understand existing relationships between MSS and TM datasets (Figure 3-5 through Figure 3-8). Determination of calibration coefficients could help to understand statistical relationships between MSS and TM datasets. These calibration coefficients between MSS and TM may also help to understand the improvement of the land classification accuracy depending on their R 2 values. The individual bands from MSS datasets were compared with those of the TM datasets to determine calibration coefficients and R 2 . A linear regression model was

PAGE 54

38 employed in the band -by-band regression (Equation 3-1) (Tokola et al., 1999) and plotted using MINITAB 13 statistical software. iixy10 (Eq.3-1) Where; iy denotes the value of pixel i for a band in a MSS dataset and x i the value of pixel i for a TM dataset. Coefficients 0 and 1 are the y-intercept and slope, respectively. Table 3-2. Description of datasets used for test site Name of Datasets Description of Datasets MSS_30_O_1986 MSS original 4 band (band 1,2,3,and 4) resampled to 30m TM_30_O_1986 TM original 4 band (band 2,3,4, and4) resampled to 30m MSS_30_S_1986 MSS smoothed 4 band (band 1,2,3,and 4) resampled to 30m TM_30_S_1986 TM smoothed 4 band (band 2,3,4, and4) resampled to 30m MSS_PC_TC_30_O_1986 Composite of PC1, PC2, BI, and GI of original MSS 30m dataset TM_PC_TC_30_O_1986 Composite of PC1, PC2, BI, and GI of original TM 30m dataset MSS_PC_TC_30_S_1986 Composite of PC1, PC2, BI, and GI of smoothed MSS 30m dataset TM_PC_TC_30_S_1986 Composite of PC1, PC2, BI, and GI of smoothed TM 30m dataset MSS_60_O_1986 MSS original 4 band (band 1,2,3,and 4) resampled to 60m TM_60_O_1986 TM original 4 band (band 2,3,4, and4) resampled to 60m MSS_60_S_1986 MSS smoothed 4 band (band 1,2,3,and 4) resampled to 60m TM_60_S_1986 TM smoothed 4 band (band 2,3,4, and4) resampled to 60m MSS_PC_TC_60_O_1986 Composite of PC1, PC2, BI, and GI of original MSS 60m dataset TM_PC_TC_60_O_1986 Composite of PC1, PC2, BI, and GI of original TM 60m dataset MSS_PC_TC_60_S_1986 Composite of PC1, PC2, BI, and GI of smoothed MSS 60m dataset TM_PC_TC_60_S_1986 Composite of PC1, PC2, BI, and GI of smoothed TM 60m dataset

PAGE 55

39 Figure 3-4. Data preparation for 30 and 60 m datasets: Section A) MSS dataset, Section B) TM dataset

PAGE 56

40 A total of 58,748 DNs from each datasets were converted to ASCII format and used to determine relationships between the Landsat MSS 30m and Landsat TM 30m datasets using linear regression. In addition, a total of 14,720 DNs from Landsat MSS 60m and Landsat TM 60m datasets were also converted to ASCII format and used to determine coefficients for better understanding of the datasets. Evaluation of calibration coefficients All datasets derived from the original Landsat MSS dataset and the original TM dataset were regressed using band-to-band regressions for 30m and 60m datasets as are seen in Figure 3-5 and Figure 3-8 (all coefficients shown are statistically significant at 0.05% risk). The R 2 of the MSS_30_O_R vs. the TM_30_O_R was the same as the MSS_60_O_R vs. the TM_60_O_R and the S-values, mean square errors, of the MSS_30_O_R vs. the TM_30_O_R was 0.2% lower than the MSS_60_O_R vs. the TM_60_O_R. Similarly, the R 2 of the MSS_30_O_G vs. the TM_30_O_G was found to be 46.9 % and the S-value was found to be 1.46 (Figure 5B). The R 2 of the MSS_60_O_G vs. the TM_60_O_G was found as 57.7 % with 1.46 S-value (Figure 7B). When the Landsat MSS NIR3 and the Landsat TM NIR (band 4) were regressed, for both the 30m and the 60m datasets, the R 2 values from both datasets were close (46.3% and 46.5% respectively) and the S values were 2.12 and 2.14 respectively (Figure 3-5C and Figure 3-7C).

PAGE 57

41 Figure 3-5. Coefficients of determination (R 2 ) and Mean Square Error (S-value) for the original MSS 30m vs. the original TM 30m and the smoothed MSS 30m vs. the smoothed 30m TM datasets A) Original Red Band, B) Original Green Band, C) Original MSS NIR3 and Original TM NIR, and D) Original MSS NIR4 and original TM NIR. E) Smoothed Red Band, F) Smoothed Green Band, G) Smoothed MSS NIR3 and Smoothed TM NIR, and H) Smoothed MSS NIR4 and Smoothed TM NIR

PAGE 58

42 Figure 3-6. Coefficients of determination (R 2 ) and Mean Square Error (S-value) for the MSS 30m principal components vs. the TM 30m principal components and the MSS 30m tasseled cap indexes and the TM 30m tasseled cap indexes (original and smoothed): A) O_PC1, B) O_PC2, C) S_PC1, D) S_PC2, E) O_BI, F) GI_O, G) S_BI, and D) S_GI

PAGE 59

43 Figure 3-7. Coefficients of determination (R 2 ) and Mean Square Error (S-value) for the original MSS 60m vs. the original TM 60m and the smoothed MSS 60m vs. the smoothed 60m TM datasets A) Original Red Band, B) Original Green Band, C) Original MSS NIR3 and Original TM NIR, and D) Original MSS NIR4 and original TM NIR. E) Smoothed Red Band, F) Smoothed Green Band, G) Smoothed MSS NIR3 and Smoothed TM NIR, and H) Smoothed MSS NIR4 and Smoothed TM NIR

PAGE 60

44 Figure 3-8. Coefficients of determination (R 2 ) and Mean Square Error (S-value) for the MSS 60m principal components vs. the TM 60m principal components and the MSS 60m tasseled cap indexes and the TM 60m tasseled cap indexes (original and smoothed): A) O_PC1, B) O_PC2, C) S_PC1, D) S_PC2, E) O_BI, F) GI_O, G) S_BI, and D) S_GI

PAGE 61

45 The Landsat MSS NIR4 and the Landsat TM NIR results also were similar to the MSS NIR3 and the TM NIR for both 30m and 60m datasets (Figure 3-5D and Figure 3-7D). After 3x3 filtering was performed on both the 30m and the 60m datasets, the R 2 and the S-values dramatically increased. These smoothed datasets were able to predict the band–to-band regression very well, and the R 2 of the MSS_30_S_R vs. the TM_30_S_R (67.8% for 30 m and 77.8% for 60 m datasets respectively) (Figure 3-5E and Figure 3-7E) were determined to be 20 to 30 % higher than the original datasets (the MSS_30_O_R vs. the TM_30_O_R and the MSS_60_O_R vs. the TM_60_O_R) (Figure 3-5A and Figure 3-7A) and the S-values decrease from 1.2 to 0.76 (Figure 5E and Figure 7E). The value of R 2 for the MSS_30_S_G vs. the TM_30_S_G (78.2%) was 21.3% higher than the MSS_30_O_G vs. the TM_30_O_G and the value of R 2 for the MSS_60_S_G vs. TM_60_S_G (84.2%) dataset was 26.5% higher than MSS_60_O_G vs. the TM_60_O_G. S-value of MSS_30_S_G vs. the TM_30_S_G and MSS_60_S_G vs. the TM_60_S_G were determined as 0.96 and 0.78 respectively (Figure 3-5F and Figure 3-7F). However comparing the R 2 values, the 60m datasets showed higher values and lower S-values. Likewise, the values of R 2 for the MSS_30_S_NIR3 vs. the TM_30_S_NIR and the MSS_60_S_NIR3 vs. TM_60_S_NIR were also higher than original datasets (Figure 3-5G and Figure 3-7G). The value of R 2 for the 30m dataset was 6% lower than the 60m dataset, when we compared the MSS_30_S_NIR3 vs. the TM_30_S_NIR and the MSS_60_S_NIR3 vs. the TM_60_S_NIR. S-values for the MSS_30_S_NIR3 vs. the

PAGE 62

46 TM_30_S_NIR and the MSS_60_S_NIR3 vs. the TM_60_S_NIR were also determined to be 1.47 and 1.15, which is lower than the original 30 and 60 m datasets. As is seen in Figure 3-5H and Figure 3-7H, the values of R 2 for the MSS_30_S_NIR4 vs. the TM_30_S_NIR and the MSS_60_S_NIR3 vs. the TM_60_S_NIR were determined to be 71.3% and 80.1% respectively. These values are 15%and 23% higher than the original 30m and 60m datasets respectively (Figure 3-5H and Figure 3-7H). The S-value for the MSS_30_S_NIR3 vs. the TM_30_S_NIR was determined to be 1.47, which was higher than the original 30m dataset (1.13). On the contrary, the S-value of the MSS_60_S_NIR3 vs. the TM_60_S_NIR was 1.18 , which was lower than the original 60m dataset. When the MSS_O_PC1 vs. the TM_O_PC1 was regressed for 30m and 60m datasets, the R 2 values were found to be 52.2 % with 6.88 S-value and 53.2 % with 6.7 S-value respectively (Figure 3-6A and Figure 3-8A). The Regression analysis of the MSS_30_O_PC2 vs. the TM_30_O_PC2 showed that R 2 was 57.2 % with 9.71 S-value, which was very close to the results of the MSS_60_O_PC2 vs. the TM_60_O_PC2 results (56.8 % and S-value is10.8). The S-value for MSS_30_O_PC2 vs. the TM_30_O_PC2 was very low (0.95), when compared to the MSS_60_O_PC2 vs. the TM_60_O_PC2 (10.8) (Figure 3-6B and Figure 3-8B) When the PC1 was smoothed using low pass filtering, the R 2 value increased dramatically from 52.2% to 77.0% for the 30m PC1 dataset (Figure 3-6C) and increased from 53.2% to 85.9% for the 60m PC1 dataset (Figure 3-8C). Likewise the R 2 of the smoothed PC2 is higher than the original PC2 dataset for both the 30m and the 60m

PAGE 63

47 datasets (Figure 3-6D and Figure 3-8D). The S-values of the 30m and the 60m PC2 were determined to be 6.09 and 6.7 respectively. When the MSS_30_O_BI vs. the TM_30_O_BI was regressed, R 2 was lower but was very close to the value of the 60m’s result (Figure 3-6E and Figure 3-8E). When a smoothing operation was performed on the BI band for the 30m and the 60m dataset, the R 2 increased 15.1% and 19.8% respectively (Figure 3-6G and Figure 3-8G). The R 2 value of the MSS_30_O_GI vs. the TM_30_O_GI was determined to be 49.7% while the MSS_60_O_GI vs. the TM_60_O_GI was 51.2%. However, after the smoothing operation, the R 2 became 70.7% for 30m and 77.7% for 60m dataset (Figure 3-6H and Figure 3-8H). The S-values also decreased for both the 30m and the 60m dataset (2.7 and 3.2 respectively) after the smoothing operation was performed. It may be useful to use higher R 2 and lower S-values, which is found in the 60 m datasets for most of the cases, for classification. Using 60m datasets for classification however did not produce good overall accuracy (Table 3-1). These coefficients could be a reference for calibrating the MSS images and used for classification. Using the smoothing operation (low pass) before classification improved R 2 values for every step, which must be tested in advance through classification section to see whether high R 2 dataset improved the classification accuracy or not. Image Enhancements for New Datasets There were three enhancement techniques performed to create the new datasets. They were: low pass filtering (LPF), PCA, and TCT. 3x3 low pass filtering A 3X3 filtering process was performed for all the datasets (30 m and 60 m) from the test site explained as follows. A spatial filter is a filter for which the Digital Number

PAGE 64

48 (DN i, j ) at location i, j in the output image is a function of some weighted average of DNs located in a particular spatial pattern around the i, j location in the input images. The simplest low-pass filter computes the mean of a particular input pixel DN in value and its eight surrounding pixels, and outputs the mean as a new DN out value (Jensen, 1996). A more complicated filter uses a kernel matrix as weighting factors in a convolution. A kernel generally consists of a set of numbers in a small array (generally odd dimensions) 3x3, 5x5, 7x7 or 9x9.....NxN (Wolf and Dewitt, 2000). Simply put, a 3x3 kernel has nine coefficients, c i , defined at the following locations in the low pass filter kernel (Equation 3-2): (Eq.3-2) 987654321cccccccccKernel The coefficients, c i in the kernel are multiplied by the following individual DN values (DN i ) (Equation 3-3). 998877665544332211_xDNcxDNcxDNcxDNcxDNcxDNcxDNcxDNcDNcpixelConvolved (Eq.3-3) Where : DN 1 = DN (i-1, j-1) DN 2 = DN (i-1, j) DN3= DN(i-1, j+1) DN 4 = DN (i, j-1) DN5= DN(i, j) DN 6 = DN (i, j+1) DN7= DN(i+1, j-1) DN 8 = DN (i+1, j) DN 9 = DN (i, j+1)

PAGE 65

49 The primary input pixel under investigation at any one time is DN 5 = DN (i, j) the convolution of the Kernel (with coefficients equal to 1) and the original data results in a low–pass filtered image as described by Equation 3-4: iniiicxDNcpixelflitered91int_ (Eq.3-4) where int is integer value and ic is sum of kernel values. The moving average window then shifts to the next pixel where the average of all nine DN values is once again computed (Figure 3-5).

PAGE 66

50 A B Figure 3-5. Filtering processes. A) MSS original 30m dataset from March 24,1986 was smoothed and B) TM original 30 m dataset from March 24, 1986 was smoothed

PAGE 67

51 Principal component transformation and tasseled cap transformations Analysis of remotely-sensed data requires synthesis of the information contained in several discrete spectral bands into information, which can be associated with physical characteristics of scene classes. Crist and Cicone (1984) indicated that the determination of physical characteristics on the ground could be separated into three parts: Compressing the n spectral bands of information into a manageable number of features, Understanding the relationships among the spectral bands for the scene classes of interest and Extracting physical scene characteristics from the spectral features. PCA provides data volume reduction, but presents substantial obstacles with regard to physical interoperation of the derived features, particularly between dates or scenes (Jensen, 1996; Richards and Jia, 1999; Campbell, 2002). TCT of MSS data, first developed by Kauth and Thomas (1976), accomplished three functions. Analysis of Landsat data from agricultural regions has shown that, on any given date, the four -band MSS data primarily occupy a single plane, with the various band pairs providing skewed views of that plane. This planar distribution of the data results from correlations between the two visible bands and the two infrared bands, which arise as a result of plant and soil reflectance proprieties. TCT rotate MSS data planes such that the vast majority of data variability is concentrated in two features. For a single plane, this view and one that presents the edge of the plane (thus establishing its two-dimensionality) can be considered “ fundamental views” since they present the most basic structure of the data in the most direct manner (Kauth and Thomas, 1976; Crist and Cicone,1984).

PAGE 68

52 Principal component analysis for datasets. The transformations were computed from the original datasets according to Richard and Jia (1999) (Appendix C): Eigenvalues (variance matrix) and eigenvectors (correlation matrix) were computed for all datasets and are shown in Figure 3-6 through Figure 3-9 and Table 3-3 through Table 3-6 for 30 m datasets and Figure 3-10 through Figure 3-13 and Table 3-7 through Table 3-10 for 60 m datasets. The software, ERDAS Imagine Ver. 8.5, is capable of calculating eigenvalues based on all datasets. The software then calculated covariance of the pixel data in X space and Y space. In addition, the degree of correlation was calculated for both X and Y space. Plots in the center of Figure 3-6 through Figure 3-9 show the 4 dimensions of the principal component transformations. The first two PC’s of the MSS_30_O_1986 described 99 % of the original MSS dataset. It was found that the first two PC’s of the MSS_60_S_1986 also represented 99 % of the original dataset. The first two PC’s of the TM_30_O_1986 represented 99% of the original 3-bands TM dataset, and the first two PC’s of TM_30_S_1986 represented 98 % of the smoothed TM (Figure 3-6 through Figure 3-9). PC3 and PC4 are assumed to be essentially noise of the original datasets for all cases. This assumption is supported by the results of a study in Charleston South Carolina using TM datasets (Jensen, 1996). However, using TM band 2, 3, 4, and 4 images after double tests, TM PC4 shows noticeable textural information, which needs to be studied for future research (Figure3-8). Table 3-2 through Table 3-5 show relationships between the original dataset and PCs. The highest correlations for the PC1 were the NIR3 and the NIR4 (0.60 and 0.73, respectively) for the MSS_30_O_1986. This suggests that the PC1 for the

PAGE 69

53 MSS_30_O_1986 is loaded mostly by the near infrared reflectance band. This make sense because the test area is mostly forest cover and visually bright. The PC2 for the MSS_30_O_1986 is loaded mostly by red and green band (0.52 and 0.78 respectively) (Table 3-3). This suggested that this component is mostly visible and vegetation is noticeably dark in this image (Figure 3-6). The highest correlations for the PC1 were the NIR3 and the NIR4 (0.60 and 0.73, respectively) for the MSS_30_S_1986. This implies that the smoothed Landsat MSS dataset correlated with the near infrared bands by the same amount as the MSS_30_O_1986 dataset. It was noticed that the PC2 for the MSS_30_S_1986 was also loaded by red and green band (0.52 and 0.78 respectively) that is the same as with MSS_30_O_1986 dataset (Table 3-4). For the MSS_30_S_1986 dataset, PC3 and PC4 were again, assumed to be noise. The PC1 for the TM_30_O_1986 dataset was loaded mostly with the NIR (0.69). Because band 3 and band 4 had the same spectral information, their load to the PC1were the same (Table 3-5). However it is different for the PC2. The PC2 contained mostly visible spectral information from red and green bands (0.45 and 0.88 respectively). The PC3 and the PC4 were supposed to be loaded by noise but, as mentioned previously, examination of the Figure 3-8 shows that PC4 had noticeable textural information. This has to be investigated for future study to determine whether this was caused by using NIR band-4 twice or not. Both PC3 and PC4 were not considered for use to create a new dataset (Figure 3-8). The PC1 for the TM_30_S_1986 dataset was loaded mostly again by the NIR (0.70). Because band 3 and band 4 have the same spectral information, their loading to

PAGE 70

54 PC1 was the same. However it is different for the PC2 (Table 3-6). The PC2 contains mostly visible spectral information from red and green bands. As mentioned above, PC3 and PC 4 were not used in the dataset (Figure 3-9). Because of the same source of the data and the same acquired time, PCA results of the 60 m datasets were similar to the 30 m datasets. The first two PC’s of the MSS_60_O_1986 described 99 % of the original MSS dataset, the first two PC’s of the MSS_60_S_1986 represented 99 % of the smoothed MSS dataset, the first two PC’s of the TM_60_O_1986 represented 99% of the original 3-bands TM dataset, and the first two PC’s of theTM_60_S_1986 represented 98 % of the smoothed TM (Figure 3-10 through Figure 3-13). PC3 and PC4 again were not used. Table 3-6 through Table 3-9 shows the relationship between the original dataset and PCs for 60 m datasets. The highest correlations for the PC1 were the NIR3 and the NIR4 (0.60 and 0.76, respectively) for the MSS_60_O_1986. This suggests that the NIR 3 and the NIR4 bands from MSS_60_O_1986 loaded the PC1. The PC2 from MSS_30_O_1986 is loaded mostly by the red and green band (0.51 and 0.81 respectively) (Table 3-7) (Figure 4.10). The PC1 was mostly loaded by the NIR3 and the NIR4 (0.60 and 0.77, respectively) from the MSS_60_S_1986. This suggests that the MSS_60_S_1986 dataset correlated with near infrared bands by almost the same amount of the MSS_60_O_1986 dataset. It is also noticed that the PC2 for the MSS_60_S_1986 is loaded by red and green band (0.49 and 0.84 respectively), which is very close to the MSS_60_O_1986 dataset (Table 3-8.

PAGE 71

55 Figure 3-6. Principal component transformation for MSS_30_O_1986 dataset

PAGE 72

56 Figure 3-7. Principal component transformation for MSS_30_S_1986 dataset

PAGE 73

57 Figure 3-8. Principal component transformation for TM_30_O_1986 dataset

PAGE 74

58 Figure 3-9. Principal component transformation for TM_30_S_1986 dataset

PAGE 75

59 Table 3-3. Degree of Correlation between each band and each Principal Component for MSS_ 30_O_1986 Bands PC1 PC2 PC3 PC4 Red 0.23 0.52 -0.33 0.75 Green 0.23 0.78 0.38 -0.45 NIR3 0.60 -0.08 -0.68 -0.42 NIR4 0.73 -0.34 0.54 0.24 Table 3-4. Degree of Correlation between each band and each Principal Component for MSS_30_ S_1986 Bands PC1 PC2 PC3 PC4 Red 0.23 0.52 -0.59 0.58 Green 0.23 0.78 0.51 -0.27 NIR3 0.60 -0.09 -0.47 -0.64 NIR4 0.73 -0.33 0.41 0.43 Table 3-5. Degree of Correlation between each band and each Principal Component for TM_ 30_O_1986 Bands PC1 PC2 PC3 PC4 Red 0.15 0.48 -0.87 0.21 Green 0.16 0.85 0.50 -0.60 NIR 0.69 -0.15 0.04 -0.12 NIR 0.69 -0.15 0.04 -0.10 Table 3-6. Degree of Correlation between each band and each Principal Component for TM_ 30_S_1986 Bands PC1 PC2 PC3 PC4 Red 0.09 0.45 -0.89 -0.26 Green 0.10 0.88 0.46 0.06 NIR 0.70 -0.09 0.02 -0.28 NIR 0.70 -0.09 0.02 0.35

PAGE 76

60 Figure 3-10. Principal component transformation for MSS_60_O_1986 dataset

PAGE 77

61 Figure 3-11. Principal component transformation for MSS_60_S_1986 dataset

PAGE 78

62 Figure 3-12. Principal component transformation for TM_60_O_1986 dataset

PAGE 79

63 Figure 3-13. Principal component transformation for TM_60_S_1986 dataset

PAGE 80

64 Table 3-7. Degree of Correlation between each band and each Principal Component for MSS_ 60_O_1986 Bands PC1 PC2 PC3 PC4 Red 0.17 0.51 -0.18 0.82 Green 0.19 0.81 0.27 -0.49 NIR3 0.60 -0.04 -0.75 -0.26 NIR4 0.76 -0.28 0.57 0.14 Table 3-8. Degree of Correlation between each band and each Principal Component for MSS_ O_S_1986 Bands PC1 PC2 PC3 PC4 Red 0.12 0.49 -0.28 0.81 Green 0.15 0.84 0.30 -0.43 NIR3 0.60 -0.02 -0.73 -0.33 NIR4 0.77 -0.22 0.55 0.21 Table 3-9. Degree of Correlation between each band and each Principal Component for TM_ 60_O_1986 Bands PC1 PC2 PC3 PC4 Red 0.11 0.47 -0.87 0.21 Green 0.11 0.88 0.50 -0.60 NIR 0.68 -0.14 0.01 -0.12 NIR 0.68 -0.14 0.01 -0.10 Table 3-10. Degree of Correlation between each band and each Principal Component for TM_60_ S_1986 Bands PC1 PC2 PC3 PC4 Red 0.13 0.41 -0.89 -0.26 Green 0.10 0.86 0.46 0.61 NIR 0.61 -0.08 0.03 -0.21 NIR 0.61 -0.08 0.03 0.31

PAGE 81

65 The PC1 for the TM_60_O_1986 dataset was loaded mostly by the NIR (0.68). Because band 3 and band 4 from TM60_O_1986 dataset have the same spectral information their load to PC1 were the same. On the other hand, PC2 was loaded mostly by visible spectral information from red and green bands (0.47 and 0.88 respectively) (Table 3-8). The PC1 from the TM_60_S_1986 dataset was loaded mostly again by the NIR (0.61). The PC2 was loaded mostly by visible spectral information from red and green bands (0.41 and 0.86 respectively). Tasseled cap transformation. Since it was decided to use BI and GI to create new datasets, it is necessary to explain how and why TCT is effective for land cover classification for certain type of land covers. Kauth and Thomas (1976) developed TCT coefficients to identify the vegetation features. Their coefficients were used to transform the MSS_30_O_1986, the TM_30_O_1986, the MSS_30_S_1986, and the TM_30_S_1986 into the new coordinates that are more meaningful in terms of specific information such as Brightness (BI), Greenness (GI), and Yellowness (YI) (Figure 3-14 through Figure 3-17). Kauth and Thomas (1976) defined that these three TCT indexes, BI, GI, and YI of the MSS original image were more meaningful than the original images and this has been applied to many land cover classification studies (Jensen, 1996; Irish, 2000; Campbell, 2002; Lillesand and Kiefer, 2000; King, 2002). This process was repeated for the 60 m datasets. Even though Crist and Cicone (1984) developed TCT coefficients for TM data, in order to create dataset that are assumed to be same effect in this research, Kauth and Thomas (1976) coefficients for 4-band Landsat MSS were performed on the 4-band Landsat MSS and 4-band Landsat TM images (Figures 3-14

PAGE 82

66 through Figure 3-17). Degradation of TM has been used for many studies because spectrally TM band 2 band 3, and band 4 correspond to MSS band 1, band 2, band 3, and band 4 (Jensen, 1996). BI, the first feature of the MSS data, is a weighted sum of all the bands, defined in the direction of principal variation in soil reflectance. Red and NIR bands in Figure 4.16-4.19 are weighted heavily since they are responsible for differences in soil brightness. The second feature, GI, is a function of the near-infrared bands (0.76-0.90m) and visible bands (0.45-0.69m) of chlorophyll (Jensen, 1996). The substantial scattering of infrared radiation resulting from the cellular structure of green vegetation, and the absorption of visible radiation by plant pigment (chlorophyll), combine to produce high GI values for targets with high densities of green vegetation, while the flatter reflectance curves of soil are expressed in low greenness values. A third feature, termed YI, was originally defined in the spectral direction expected to correspond to plane senescence. However this third component was later defined as haze components of MSS data. (Kauth et al., 1979) Crist and Cicone (1984) determined tasseled cap features for TM data as four components: BI, the first feature, is a weighted sum of all six reflective TM bands. Differences in soil characteristics such as particle size distribution will be clearly expressed in brightness, while increases in vegetation density, which would tend to increase near-infrared responses while decreasing visible responses, will cause less substantial changes in brightness. The second feature TCT from TM data set is Greenness, which is similar to

PAGE 83

67 MSS greenness values. Both MSS and TM greenness are contrasted between the sum of the visible bands and the near infrared bands(s). Wetness is the third component of TCT from TM dataset. This feature of TCT contrasts the sum of the visible and near-infrared bands with the sum of the longerinfrared bands. And finally the fourth component of TCT from the TM data set is called haze. Since TCT is also helpful for identifying vegetation health on the ground (Jensen 2000), which is important for understanding the role of healthy vegetation for this project, it was assumed that TCT individual index could be used to improve classification accuracy. Healthy green vegetation generally reflects 40% to 50 % of the incident near-infrared energy (0.7 to 1.1m), with the chlorophyll in the plants absorbing approximately 80% to 90% of incident energy in the visible (0.45 to 0.69m) part of the spectrum (Jensen 2000). However, dead vegetation reflects a greater amount of energy than healthy green vegetation throughout the visible spectrum (Figure 3-18). On the contrary, unhealthy vegetation reflects less than green vegetation in the infrared region (0.7 to 1.1 m). Dry soil has higher reflectance than green vegetation and lower than dead vegetation in the visible region. In addition, dry soil has lower reflectance than green vegetation in the infrared region (Figure 3-18). It was assumed that TCT will help to determine more readable information about forest ages and wetland forest characteristics from original MSS and TM datasets and could improve classification accuracy. One would expect recently cleared forestland to exhibit high values in BI. However, the relative brightness in this plane may be influenced by soil moisture. On the other hand, recently planted pine forest should fall

PAGE 84

68 along the transition plane, until the forest canopy begins to close. Since the study area is covered by different ages of forest plantation, accurate determination of newly planted forest was very important for overall classification success. In order to distinguish newly planted forest from developed pasture and natural high grassy area, field trips were made to the application site for more detailed information about the land cover. Once the forest canopy has closed blocking the soil sub-strata, pixels covering these kinds of spaces should exhibit very high GI and BI values. On the other hand, mature forest should be characterized by pixels with low values in the plane of soil with high GI values but the self-shading of mature forest should produce relatively low BI values compared to recent clear-cut land. If data are dispersed into two perpendicular planes, then PCA may not be effective for defining the actual planes of variation along which data reside (Christ and Cicone, 1984). If this is the case and factors are rotated, they may intersect on more meaningful axes of variation for describing the data. Tasseled cap factors may not necessarily be perfectly orthogonal like principal components (Crist and Cicone, 1984).

PAGE 85

69 Figure 3-14. Tasseled cap transformation for 30m and 60m MSS_O_1986

PAGE 86

70 Figure 3-15. Tasseled cap transformation for 30m and 60m MSS_O_1986

PAGE 87

71 Figure 3-16. Tasseled cap transformation for 30m and 60m TM_O_1986

PAGE 88

72 Figure 3-17. Tasseled cap transformation for 30m and 60m TM_O_1986

PAGE 89

73 0.4 0.5 0.6 0.7 0.8 0.11 Wavelength (m) Green MSS band 1 Red MSS band 2 N IR MSS band 3 N IR MSS band 4 Dead Grass Dry Bare soil Green Grass Percent Reflectance Figure 3-18. Common spectral reflectance characteristics for healthy green grass, dead grass, and bare dry soil for the wavelength interval from 0.4 to 1.1 (adapted from Jensen 2000). Comparison of TCT of Landsat MSS and Landsat TM Data The brightness feature of TM reflects its close similarity to MSS brightness that is also a weighted sum of all bands. However because of the influence of the mindinfrared bands, TM Brightness is not identical to MSS Brightness. Crist and Cicone (1984) reported that MSS Brightness and TM Brightness had a correlation of 0.77, which is higher than raw band-to-band correlation between MSS and TM (Tokola et al., 1999). The blue band from the TM data is highly correlated with other TM visible bands (red and green) of TM data. Even if MSS data does not have a blue band, Greenness of MSS and Greenness of TM have high similarities. According to Crist and Cicone (1984) this correlation could be better than 0.99 allowing scientists to derive the conclusion that these two are identical. The MSS third component was later on described as Yellowness or later Haze by Crist and Cicone (1984). The TM third band is totally new information called wetness because of the involvement of longerinfrared TM bands.

PAGE 90

74 Crist and Cicone (1984) stated that PCA and TCT are not equivalent. PCA can fail to capture the complex structure of TM data (Crist and Cicone, 1984), but keep most of the raw data information in a few bands (Richards and Jia, 1999). While Brightness MSS data is recommended to be comparable to Brightness of TM data, Greenness of MSS and Greenness of TM are almost equivalent (Crist and Cicone, 1984; Jensen, 1996). Image Classification In this study, hybrid classification was performed to produce appropriate classification for land cover mapping (Figure 3-23). The logic of hybrid classification is using unsupervised and supervised classifications in a classification procedure. Both unsupervised and supervised classifications were used to determine land features and produce maps for the test site. Using unsupervised classification prior to supervised classification has many advantages such as economy and time saving (Jensen ,1996; Richards and Jia , 1999; Campbell ,2002). Since there are no historic auxiliary data available for the test site and the application site, ground data was also determined based on the existing Landsat datasets and DOQ images as well as field trips. Land Cover Determinations Before classification was performed, sixty-three Landsat images on the application site (path 17 and row 37) from 1975 to 2000 and DOQ images from 2000 were visually inspected to determine forest ages and characteristics of land features (Figure 3-24 through Figure 3-27). In addition, two field trips were made to the test site (April, 2002 and September, 2002) to confirm land cover characteristics. During the field trips, unsupervised classification maps were used to determine the location of land features and pictures were taken of these locations and GPS points were acquired to produce an archive for future study.

PAGE 91

75 Figure 3-24, Figure 3-25, and Figure 3-26 show the procedure for determining age of classes based on the clear-cut area on the Landsat images. For this reason, 12 points were randomly selected from regions in a 2000 Landsat TM image and Landsat images from 1975 to 2000 were visually inspected to determine how land cover changed over time (Figure 3-25). Wetland forests for the test site were determined on DOQ images (Figure 3-27) and it was assumed that wetland forest had remained the same except for some occasions (fire or cleared at land owners’ preference) The criteria used for the classes are appropriate attribute tables that should be interoperable with carbon data that have been recorded and evaluated by Gholz et al. (1997). After investigation of all available data and field trips, land cover features in the test site were determined to be in four nominal classes. They are: Clear-cut_1-3 yr pine plantation, 4-8 yr pine plantation, >8 yr pine plantation, and Wetland forest

PAGE 92

76 Figure 3-23. Classification schema.

PAGE 93

77 Figure 3-24. Logic of forest age and wetland forest determination

PAGE 94

78 Figure 3-25. Forest age determination area and training area determinations for test site (Image Jan 7, 2000)

PAGE 95

79 Figure 3-26. Forest age at the selected points outlined on Figure 3-25

PAGE 96

80 Figure. 3-27. Wetland appears on DOQ images inset from Landsat TM, March 24, 1986

PAGE 97

81 Clear-cut_1-3 yr pine plantation Forests are cleared when they are mature enough to fetch a good market price or because of some other landowner preference. In the study area sometime fires force the owner to clear-cut. Because fire residues and ashes are removed from these regions as well as damaged trees, bright areas appear on the images after fire. However, determining clear-cut needs careful visual inspections of the sixty-three Landsat images. A clear-cut area is expected to regenerate as soon as possible and usually appear as 1-year-old pine plantation. Because private companies or individual families own forest areas, different regeneration times could occur and this created confusion for the determination of whether clear-cut areas are planted or not on the Landsat images. Because there is no regulation by local or federal governments for regeneration of forest areas, clear-cut areas sometimes are not regenerated immediately or during the following years. Sometime it takes 2-5 years to be replanted. In most cases, however, clear-cut areas appear to be replanted within 3-4 years after being cleared. This also fits carbon data from 1-3 year old pine trees, which have approximately the same values. When clear-cut areas appear on Landsat images, they are visually inspected and a logic is developed to discriminate the classes (Figure 3-23 through Figure 3-26). Figure 3-28 shows a new clear-cut area (Figure 3-28A), first year plantation pine (Figure 3-28B), and 4-year pine plantation (Figure 3-28C). 4-8 yr pine plantation In order to determine the 4-8 year old pine plantation, sixty-three Landsat images were visually inspected. After regeneration of a clear-cut area (it is assumed that when a tree is planted it is already one yr old) in three years, it is assumed that it becomes a 4 yr old pine forest (Figure 3-24 through Figure 3-26).

PAGE 98

82 A B C Figure 3-28. Clear cut and 1-3 yr pine plantation A) New Clear-Cut area B) 1 yr old pine plantation, and C) 3 yr old pine plantation

PAGE 99

83 This logic works, in general, unless unusual circumstances such as fire has occurred. This happened a sample point 1(Figure 3-25 and Figure 3-26). Examination of the sample point 1 in Figure 3-26 (located in Figure 3-25) shows the logic that helps to determine what happened during the 25 year time period. The clear-cut area first appeared on the Landsat images in 1981 and fire occurred in 1984 in this particular area. Because of fire this particular area appeared as clear-cut from 1981 to 1989 (Figure 3-2 and Figure 3-26). The determination of the 4-8 yr old pine plantation is relatively easy on Landsat images, but needs some additional information such as when the fire happened and for this field trips must be done. Figure 3-29A and Figure 3-29B shows the determination of existing 4-8 yr old pine plantations in year 2002. The problem is for 4-8 year old trees in north Florida’s forest are very tall wetland vegetation growing rapidly in 1-2 years, and dominant the land cover after the wetland forest is cleared. This area was spectrally found to be very close to 4-8 yr pine plantations and classified as 4-8 yr pine plantation. > 8 yr pine plantation Determination of tree age, >8 yr pine plantation, was also relatively easy compared with clear-cut_1-3 yr pine and 4-8 yr pine. After a 4 year time period, 4 yr old pine plantation became >8 yr pine plantation class (Figure 3-30A and Figure 3-30B). In the test site some regions such as sample area 12, shown in Figure 3-25, had never been cleared. During the entire 25 yr time span, the land cover type had been a >8 yr pine plantation, and had never changed based on visual inspection on Landsat Images from 1975 to 2000 and personal communication (Figure 3-26). In general, ages of these trees are around 35-50, so 25 years ago these trees were already bigger than 8 yr old trees and classified as > 8 yr pine plantation.

PAGE 100

84 A B Figure 3-29. 4-8 yr pine plantation A) 3-5 yr pine B) 5-8 yr old pine plantation

PAGE 101

85 A B Figure 3-30. > 8 years pine plantation A) 15-20 year old pines B) >30 year old pine plantation

PAGE 102

86 Wetland forest Determination of wetland forest required two visual inspections from both DOQ and Landsat images. It is assumed that as long as there are no natural phenomena such as tornados, severe flooding or fire, wetland and wetland forest cover does not change. Wetland related forest areas were defined on DOQs and compared visually with all Landsat images available from 1970 to 2000 (Figure 3-24 through Figure 3-27). The wetland forest is primarily cypress tree and riparian forest (Figure 3-31A, B). Unsupervised Classification: Unsupervised classification can be defined as the identification of natural groups or features within remotely sensed multispectral data. Only a minimal amount of initial input from the researcher is required (Jensen, 1996). The notion of the existence of inherent groupings of DNs within a scene may not be naturally obvious, but it can be demonstrated that remotely sensed images are usually composed of spectral classes that are reasonably uniform within respect to DNs in several spectral bands. After classification, researchers have to assign these natural or spectral classes to the information classes of interest (Jensen, 1996; Richards and Jia 1999; Lillesand and Kiefer, 2000). It has been reported that understanding the spectral characteristics of the terrain well enough to label certain clusters as specific classes is the key to unsupervised classification (Jensen, 1996). Algorithms for unsupervised classification are varied, including the iterative Self Organizing Data Analysis Technique (ISODATA) (Jensen, 1996; Rees, 2001; Campbell 2002). In this research, after determination of the new datasets, mentioned earlier for the test site, the original images from 1975 to 2000 including all created datasets from Landsat MSS on March 24, 1986 and original Landsat TM on March 24, the 1986 images

PAGE 103

87 A B Figure 3-31. Wetland forest A) Wetland forest (riparian forest) B) mixed pine trees with bottomland hardwood

PAGE 104

88 were used to produce an unsupervised classification of images with 20 clusters, of which 4 classes were chosen out of the 20 clusters for the test site (Figure 3-32). Unsupervised classification or clustering based on the sixty-three images (Landsat MSS from different seasons in the 1975, 1976, 1981, 1982, 1986, 1989, and 1992 images and Landsat TM from different seasons from the 1982 to 2000 images) was performed for the test site to determine initial spectral classes. In addition, a total of sixteen created datasets from the original 4-band Landsat MSS (band 1, 2, 3, and 4) 30m and 60m and the original 6-band Landsat TM (band 1, 2, 3, 4, 5, and 7) 30 m and 60m images from March 24,1986 of the test site were classified using unsupervised classification. The resulting images from the unsupervised classification were examined along with field trips for both 30m and 60m datasets before supervised classification was performed. Unsupervised classification gave a rough idea of what was on the ground at the test site and this was used to determine the forest ages as well as the location of the wetland forest. Many researchers have benefited from unsupervised classification in the determination of land cover features as well as in the production of actual classification maps. Pre classification is a human’s image interpretation using field trips, personal experience, and auxiliary data to determine land cover types, which is assumed to be correct, before remotely-sensed data are classified to determine land covers. Watson and Wilcock (2001) stated that using Landsat TM band 1, 2, 3, 4, 5, and 7 to classify various images from 1987 to 1989 determined how pre-classification, unsupervised classification, could be used to affect accuracy assessment.

PAGE 105

89 Jensen (1996), Lillesand and Kiefer (2000), Rees (2001), and Campbell (2002) stated that image classification is the process of making quantitative decisions from image data, grouping pixels or regions of the image into land cover classes intended to represent different physical objects or types. Usually each pixel is treated as an individual unit composed of values on several spectral bands. According to Campbell (2002), by comparing pixels one to the other, and to pixels of known identity, it is possible to assemble groups of similar pixels into the same land cover types that are associated with the information categories of interest to users of remotely sensed data. In order to use unsupervised classification results to determine the land cover features as a mosaic of uniform parcels, each land cover is identified with the same value or color. Supervised Classification This classification is well known as controlled classification. Researchers supervise the pixel categorization process by using a combination of fieldwork, analysis of aerial photography and maps, and personal experience (Jensen, 1996). For this purpose, Lillesand and Kiefer (2000) and Campbell (2002) stated that representative sample sites of known land features, called training areas, are used to compile a numerical interpretation key that describes the spectral attributes for each land cover class. Campbell (2002) mentioned that training sites must not include unusual regions and must not be chosen between the boundaries of land cover types. Size, shape, and position must favor convenient identification both on the image and on the ground. Pixels located within these areas form the training areas used to guide the classification algorithm to allocate specific spectral values to suitable informational classes (Jensen,

PAGE 106

90 1996; Rees, 2001; Campbell, 2002). Campbell (2002) also stated that selection of training areas is the key step in supervised classification. Figure 3-32. 20 clusters, from which 4 classes were determined by using unsupervised classification on test site

PAGE 107

91 Training samples Training sets were created with the help of site visits, maps produced by using unsupervised classification, DOQ’s, and image interpretation of color composites from the satellite imagery. Training sets were digitized onto color composite datasets from all the March 24, 1986 Landsat MSS and the March 24, 1986 Landsat TM for both 30m and 60m datasets (Figure. 3-33). There were two criteria used for the training set selection Arai (1992) recommended that homogeneous areas are desirable in order to create quality spectral signature generation. Areas of interest (AOI’s) were selected based upon homogeneity within each of the land cover categories, which were determined using unsupervised classification of Landsat 1975 through 2000 images, DOQs, and field trips (USGS 2002). A minimum of “n +1” training pixels per category is required for supervised classification (n = number of classes)(Swain and Davis, 1978). All the training sets selected met this requirement (Table 3-11). Note that there were averages of roughly 9 pixels per sample. After training pixels were selected, supervised classification was performed using maximum likelihood classification (MLC). Table 3-11. Numbers of pixels selected for training sets Class Number of Samples Total Number of Pixels Clear-cut_1-3 yr pine plantation 10 90 3-8 yr pine plantation 6 54 > 8 yr pine plantation 10 90 Wetland Forest 11 99

PAGE 108

92 Figure 3-33. Training data sample selection

PAGE 109

93 Maximum likelihood classification (MLC) In this research supervised classification was performing using the MLC that quantitatively evaluates both the variance and covariance of the category’s spectral response patterns when classifying an unknown pixel, and calculates the probability that a given pixel belongs to a specific class (Richards and Jia, 1999). To do this, it is assumed that the DNs for each class in each band are normally distributed (Lillesand and Kiefer, 2000). Each pixel is assigned to the class that has the highest probability (Richard and Jia, 1999). For the technical details, the reader is directed to the references provided by Jensen (1996), Richards and Jia (1999), Tso and Mather (2001), and Campbell (2002). Image classification is an important part of the field of remote sensing, image analysis, and pattern recognition. According to Jensen (1996), Richard and Jia (1999), Lillesand and Kiefer (2000), and Campbell (2002) there are three main pattern recognitions that can be used to obtain better results from classification: spatial, temporal and spectral recognition approaches. Spectral pattern refers to the family of classification procedures that utilize this pixel-by-pixel spectral information as the basis for automated land cover classification. Spatial pattern recognition involves the characterization of image pixels on the basis of their spatial relationship with pixels surrounding them. Spatial pattern might consider such aspects as image texture, feature size, shape, directionality, and context. Temporal pattern recognitions utilize time to assist in feature identification (Lillesand and Kiefer, 2000; Campbell, 2002,). Jensen (1996) and Lillesand and Kiefer (2000) gave an example of agricultural fields to describe the temporal aspect of classification. If data from two different dates can be analyzed

PAGE 110

94 spatially and spectrally land cover change can be detected that might recognize agricultural fields easily. After classification of the sixteen datasets in Table 3-2, land cover maps with 4 classes for the test site were produced and areas of land cover recorded for individual classes (see example in Table 3-12). An accuracy assessment was performed on all classified images to determine the best dataset and to create statistical data for further analysis about how classified datasets differ from each other. After determination of the best dataset that produced the highest overall classification accuracy this dataset was used to classify application site datasets (see Chapter 4) Figure 3-34 shows equiprobability contours in a scatter diagram. The shape of equiprobability contours expresses the sensitivity of a likelihood classifier to covariance (Lillesand and Kiefer 2000). Table 3-12. Individual land cover area recorded after classification Class Area (Hectares) Area (%) Clear-cut_1-3 yr pine plantation 84.51 9 3-8 yr pine plantation 891.36 13 > 8 yr pine plantation 1295.10 43 Wetland Forest 4255.74 34

PAGE 111

95 Figure 3-34. Equiprobability contours defined by a maximum likelihood classifier (adapted from Lillesand and Kiefer, 2000). Determination of Classification Accuracy for Test Site: In this research, classification accuracy was assessed according to Congalton and Green (1998) for all sixteen datasets and the results are discussed in the following section. When researchers use remotely sensed data for classification, an accuracy assessment must take place to determine the quality of the classification (Congalton and Green, 1998). Jensen (1996), Congalton and Green (1998), Richards and Jia (1999), and Campbell (2002) stated that since remotely sensed data is generally used to make very important decisions about human life and their environment, one has to understand the

PAGE 112

96 accuracy of the map, not only qualitatively but also quantitatively. In addition, quantitative accuracy assessment is the identification and measurement of any maps produced from remotely sensed data (Jensen, 1996; Congalton and Green, 1998; Richards and Jia, 1999; and Campbell, 2002). Quantitative accuracy assessment explains the relationship between classified areas (based on the training area) and reference information, assumed to be correct, for the same site. Given the obvious limitation of non-site specification accuracy assessment, there is a need to know how the map generated from the remotely sensed data compares to the reference data on a location basis (Congalton and Green, 1998). In other words, accuracy relates to bias and precision and the difference between the two is occasionally important as one may be traded for the other (Campbell, 2000). In thematic mapping from remotely sensed data, the term accuracy is used typically to express the degree of correctness of classification. A thematic map derived with a classification may be considered accurate if it provides an unbiased representation of the land cover of the region it portrays (Foody, 2002). In essence therefore, classification accuracy is typically taken to mean the degree to which the derived image classification agrees with reality or conforms to the truth (Congalton and Green, 1998). Thus classification error is a discrepancy between the situation depicted on the thematic map and reality. It is the goal to represent the entire classification with a single value called overall accuracy. First Anderson et al. (1976) stated that an overall accuracy level of 85% was an acceptable corrected accuracy level that later on was universally accepted and used (Jensen, 1996; Congalton and Green, 1998). Individual categories within the classified maps became important when remotely sensed data was used in specific studies. In order

PAGE 113

97 to determine individual class accuracy, error matrices have been used to represent map accuracy (Congalton and Green, 1998). Error Matrix Many methods of accuracy assessment have been discussed in the remote sensing literature (Aronoff, 1982; Rosenfield and Fitzpatrick Lins, 1986; Foody 2002). The most widely promoted and used, however, may be derived from a confusion or error matrix. Congalton and Green (1998) described the error matrix as a square array of numbers that express the number of sample units assigned to a particular category in one classification relative to the number of sample units assigned to a particular category in another classification. One of the data sets is assumed to be correct and this may be generated from a known map, aerial photography, etc. It has been reported by many researchers that the error matrix is a successful and efficient method to represent map accuracy in individual categories (Jensen, 1996; Congalton and Green, 1998; Campbell, 2002). Test Sample Design To obtain unbiased ground reference information to compare with the remote sensing classification map and fill the error matrix with values, the selection of a proper and efficient sample design is one of the most important components of an accuracy assessment. Basic sampling designs, such as simple random sampling, can be appropriate if the sample size is large enough to ensure that all classes are adequately represented. The adoption of a simple sampling design is also valuable in helping to meet the requirements of a broad range of users (Stehman and Czaplewski, 1998). Several critical issues have to be addressed for an accuracy assessment sample that is truly

PAGE 114

98 representative of the map: (a) the appropriate sample unit, (b) the total number of samples to be collected by category and (c) the selection of the samples. Sample unit Sample units have three choices: (a) a single pixel, (b) a cluster of pixels, and (c) polygons.A single pixel, historically, has been a poor choice because it is an arbitrary rectangular delineation of the landscape that may have little relation to the actual delineation of land cover type. A 3 3 cluster of pixels has been the most common choice for the sample unit for classification, but it may still be an arbitrary delineation of the landscape, resulting in the sample unit encompassing more than one map category. It is important to remember that the sample units are portions of the map that will be sampled for accuracy assessment. If the assessment is performed on a cluster of pixels, nothing can be said either about a single pixel or about polygons. More mapping projects from remotely sensed data are generating polygon products. If the objective of the mapping is to produce a polygon map, it is important that the assessment is conducted on the polygon basis. Therefore, the polygon is replacing the cluster of pixels as the sample unit of choice, and it was also applied for this research. Sample size and sampling selection The adequate number of samples for accuracy assessment of individual categories is often difficult to determine. A number of researchers have used an equation based on the binomial distribution or the normal approximation to the binomial distribution to compute the required sample size. Jensen (1996) suggested the equations to calculate sample size (Equation 3-5) and gave an example.

PAGE 115

99 22))(( E qpZN (Eq.3-5) Where N is the sample size to be used to assess the accuracy of land use classifications map, p is the expected percent accuracy, q is 100-p, E is the allowable percent error, and Z is approximately 2 from the standard normal deviate for the 95% two-sided confidence level. For 85% desired accuracy and 5 % allowable error 225)15)(85(2N 204N Though this method is acceptable for selecting the total number of pixels to be sampled, it was not designed to select a sample size for filling an error matrix. For an image having a resolution approximately the same as Landsat TM, if the study area is not larger than 1 million ha, a minimum of 50 samples for each land cover class is a good sample size to demonstrate in the error matrix according to Congalton and Green (1998). If the area is especially large (that is, more than 1 million acres) or the classification has a large number of land use categories (that is, more than 12 classes), the minimum number of samples should be increased to 75 or 100 samples per class. The choice of samples is critical to generating an error matrix that is representative of the entire map. The common sampling scheme, simple random sampling, was applied for collecting reference data for this study. In a simple random sampling, the good statistical properties are the main advantage that result from the random selection of samples. The number of samples can also be adjusted based on the relative importance of that category within the objectives of the mapping. Since the test site is not larger than 1

PAGE 116

100 million acres (13 km by 7 km), 400 random reference sites, 4 classes at 100 pixels each, were generated to assess the accuracy of the classification map in this research (Figure 3-35). Figure 3-35. Examples of reference data locations for various classes

PAGE 117

101 Evaluation of Error Matrices After the reference information was generated, the error matrices were developed and evaluated for every dataset generated from the original Landsat MSS and TM images from March 24, 1986 and Landsat TM from March 24, 1986 (explained in the “preparing new datasets for test site” section in this chapter). Sample units are assigned to a particular category in one classification relative to the number of sample units assigned to a particular category in another classification. For the example error matrix in Table 3-13, the columns represent the reference data, and the rows indicate the classification generated from the Landsat TM dataset. An error matrix is a very effective way to represent map accuracy because the individual accuracies of each category are plainly described, along with both the errors of inclusion (commission errors) and errors of exclusion (omission errors). A commission error is defined as including an area in an incorrect category. An omission error is excluding an area from the category in which it belongs. Every error is either an omission or a commission error. In addition to showing errors of omission and commission, the error matrix can be used to compute overall accuracy, the producer’s accuracy, and the user’s accuracy. Overall accuracy is computed by dividing the sum of the major diagonal (that is, the correctly classified pixels) by the total number of pixels in the entire error matrix. This value is the most commonly reported statistic in accuracy assessment. However, it is also important to represent the accuracy of individual categories, and the producer’s accuracy and the user’s accuracy are ways of computing individual category accuracies. The producer’s accuracy is performed by dividing the total number of correct pixels in a category (class) by the total number of pixels of that category, as derived from the reference data. This accuracy shows how well the producer (the analyst) classified a

PAGE 118

102 certain area. The user’s accuracy is determined by dividing the total number of correct pixels in a category by the total number of pixels that were actually classified in that category. The user’s accuracy is the probability that a pixel classified on the map actually represents that category (Jensen, 1996). Here, more examples are given to help understand the accuracies. An overall accuracy in the error matrix shown in Table 3-13 is 71.00%. However, suppose the requirement is to determine in the ability to classify clear-cut_1-3 yr pine plantation, then a “producer’s accuracy” would be calculated for this category, that is, dividing the total number of correct pixels in the clear-cut_1-3 yr pine plantation (that is, 92) by the total number of clear-cut_1-3 yr pine plantation, as indicated by the reference data (that is, 130, the column total). This division results in a producer’s accuracy of 70.77%. However, the calculation of the “user’s accuracy” for this category, that is, dividing the total number of correct pixels in the tree class (that is, 92) by the total number of pixels classified as tree (that is, 100 or the row total) reveals a value of 92.00%, which is quite good. Although 70.77% of these trees have been correctly identified as clear_cut_1-3 yr pine plantation in classification, 92.00% of the clear_cut_1-3 yr pine plantation on the classification map is actually clear_cut_1-3 yr pine plantation on the ground. Kappa Analysis The Kappa analysis is a discrete multivariate technique of use in accuracy assessment for statistically determining if one error matrix is significantly different from another (Congalton and Green,1998). The result of performing a Kappa analysis is a statistic (an estimate of the Kappa) that is a measure of agreement of accuracy (Congalton and Green, 1998). K

PAGE 119

103 Table 3-13. Error matrix used to generate accuracy assessment Class Name 1 2 3 4 CT NC PA % UA % OA % 1 92 2 4 2 100 92 70.77 92.00 2 24 52 22 2 100 52 68.42 52.00 3 4 12 72 12 100 72 65.45 72.00 4 10 10 12 68 100 68 80.95 68.00 Ref. Totals 130 76 110 84 400 71.00 1Clear_cut_1-3 yr pine plantation 24-8 yr pine plantation 3> 8 yr pine plantation 4 Wetland Forest CTClassified Totals NCNumber Correct PAProducers Accuracy UAUsers Accuracy OAOverall Accuracy Overall Accuracy = (92+52+72+68)/400 = 71.00 % This measure of agreement is based on the difference between the actual agreement in the error matrix (the agreement between the remotely sensed classification and the reference data as indicated by the major diagonal) and the chance agreement, which is indicated by the row and column totals (Lillesand and Kiefer, 2000) The statistic is computed by Equation 3-6 K kiiikiiikiiinnnnnnnK1211** (Eq.3-6) where k is the number of rows in the matrix, n ii is the number of samples in row i, and column i, n i+ and n +i are the marginal totals for row i and column i, respectively, and n is the total number of samples. Here, an example is given to help understand the kappa analysis 284)68725292(1kiiin 000,40)100*84()100*110()110*76()100*130(*1kiiinn

PAGE 120

104 61.0000,40)400(000,40)284(4002 K This value (0.61) indicates that probability that the classification is better than random chance. Results of Classification Accuracy for Test Site All sixteen datasets from Landsat MSS and Landsat TM were classified using supervised classification (Figure 3-36 through Figure 3-45) and used to assess the accuracy for the test site (Table 3-14 through Table 3-29). The idea here is to determine the dataset that produced the best overall and user accuracies for land cover classification and use this dataset to perform the classification to produce long term land cover maps for the application site. After accuracy assessment was performed for every dataset, it was found that when considering the original 4band MSS and the original 4 band TM bands, the TM_30_O_1986 dataset was the best for discriminating among the four forest classes based on the overall accuracy (67.75% and 71.25% from MSS and TM, respectively). When the probability that a pixel was classified into a given class was compared with what it actually represents on the ground (user accuracy), the TM_30_O_1986 dataset produced better UA than the MSS_30_O_1986 does for all classes (Table 3-14 and Table 3-15). The UA for Clear-cut_-1-3 yr pine plantation was found to be 84% for the MSS_30_O_1986 and 88% for TM_30_O_1986 dataset. It was also found that UA for 4-8 yr pine plantation was 1% higher than the MSS_30_O_1986 dataset (48% vs. 49%). In addition, it was found that UA was 1% better for >8 yr pine plantation class when we compared the MSS_30_O_1986 and the TM_30_O_1986. The UA was 8% higher for

PAGE 121

105 Wetland forest class for the TM_30_O_1986 data set when it was compared with the MSS_30_O-1986 dataset. In addition, overall was calculated based on Equation 3-6 by help of ERDAS imagine version 8.6, and the result showed that for the TM_30_O-1986 dataset was 0.04 higher than that for the MSS_30_O-1986 dataset (0.61 vs. 0.57). K K When the MSS_30_O-1986 and the TM_30_O-1986 datasets were smoothed as is demonstrated in Figure 3-5, higher overall accuracies were found for both the MSS_30_S-1986 and the TM_30_S-1986 datasets (69.00% and 73.50, respectively) than for original datasets. The smoothed datasets also produced higher UA for all classes, when they were compared with original datasets (Table 3-16 and Table 4-17). K values were found to be 0.58 for the MSS_30_S-1986 dataset and 0.64 for the TM_30_S-1986 dataset, which were higher than those of the original datasets.

PAGE 122

106 Figure 3-36. Classified images A) MSS_30_O classification B) TM_30_O Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest.

PAGE 123

107 Table 3-14. Error Matrix of MSS_30_O_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 84 13 1 2 100 84 63.16 84.00 2 42 48 8 2 100 48 69.57 48.00 3 2 6 71 21 100 71 67.62 71.00 4 5 2 25 68 100 68 73.12 68.00 Ref. Totals 133 69 105 93 400 271 Overall Classification Accuracy = 67.50% Overall Kappa= 0.57 Table 3-15. Error Matrix of TM_30_O_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 88 9 0 3 100 88 70.40 88.00 2 33 49 16 2 100 49 68.06 49.00 3 3 10 72 15 100 72 67.29 72.00 4 1 4 19 76 100 76 79.17 76.00 Ref. Totals 125 72 107 96 400 285 Overall Classification Accuracy 71.25% Overall Kappa= 0.61

PAGE 124

108 Figure 3-37. Classified images A) MSS_30_S classification B) TM_30_S Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest.

PAGE 125

109 Table 3-16. Error Matrix of MSS_30_S_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 87 5 2 6 100 87 64.93 87.00 2 38 44 16 2 100 44 77.19 44.00 3 3 4 68 25 100 68 68.69 68.00 4 6 4 13 77 100 77 70.00 77.00 Ref. Totals 134 57 99 110 400 276 Overall Classification Accuracy 69.00% Overall Kappa= 0.58 Table 3-17. Error Matrix of TM_30_S_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 89 7 3 1 100 89 68.99 89.00 2 32 53 11 4 100 53 74.73 53.00 3 1 7 76 16 100 76 73.79 76.00 4 7 4 13 76 100 76 78.35 76.00 Ref. Totals 129 71 103 97 400 294 Overall Classification Accuracy 73.50% Overall Kappa= 0.64

PAGE 126

110 The overall accuracy for the TM_60_O_1986 (72.75%) dataset was even higher than the TM_30_O_1986 dataset (Table 3-15) and the overall accuracy of the MSS_60_O_1986 (Table 3-18) dataset (65.00%) was found to be lower than the MSS_30_O_1986 dataset (67.75) (Table 3-14). When UAs of the MSS_30_O_1986 and the MSS_60_O_1986 were compared, UA for Clear-cut_13 yr pine plantation decreased 22% and 4-8 yr pine plantation increased 17%. This could be owing to the resampling method, nearest neighborhood. The UA of the MSS_60_O_1986 for > 8 yr pine plantation was found to be 9% lower than the UA of the MSS_30_O_1986 dataset and 3% higher than the UA that was found for the class of wetland forest (Table 3-14 and Table 3-18). K for the original 60 m MSS dataset was found to be lower than the original 30m MSS dataset (0.53 and 0.57 respectively) however, the original 60m TM dataset produced higher (0.63) than the original 30m dataset (0.61). K Table 3-19 showed that the original 60m TM dataset produced 72.75% overall accuracy and better UAs for all classes, when they were compared with the original 60m MSS dataset (Table 3-18). When the MSS_60_O_1986 and the TM_60_O_1986 datasets were smoothed and compared with 30m datasets, it was found that there was a higher overall accuracy for the MSS_30_S_1986 (69.00%) (Table 3-16) compared with MSS_60_S_1986 (67.00%) (Table 3-20) and lower overall accuracy for the TM_30_S_1986 datasets (73.50%) (Table 3-17) compared with TM_60_S_1986 (78.50%) (Table 3-21). However, MSS_60_S-1986 produced higher UAs for Clear-cut_1-3 yr pine plantation, 4-8 yr pine plantation, and wetland forest (64%, 69%, and 73% respectively)

PAGE 127

111 and the UA for > 8 yr pine plantation was the same with the original 60m dataset (62%), when they were compared with MSS_60_O_1986 datasets (Table 3-18 and Table 3-20). When the 60m TM dataset was smoothed, TM_60_S_1986 produced higher UAs for all classes when compared with the TM_60_O_1986 dataset (Table 3-21). K values were found to be 0.56 for the MSS_60_S-1986 dataset and 0.71 for the TM_60_S-1986 dataset, which were higher than the original 60m MSS datasets. In general the smoothed dataset produced higher overall accuracies and UAs for both the 30m and the 60m datasets when compared with the original 30m and 60m datasets for both MSS and TM.

PAGE 128

112 Figure 3-38. Classified images A) MSS_60_O classification B) TM_60_O Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest.

PAGE 129

113 Table 3-18. Error Matrix of MSS_60_O_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 62 37 0 1 100 62 69.66 62.00 2 26 65 58 4 100 65 57.52 65.00 3 0 8 62 30 100 62 67.39 62.00 4 1 3 25 71 100 71 66.98 71.00 Ref. Totals 89 113 92 106 400 260 Overall Classification Accuracy = 65.00% Overall Kappa= 0.53 Table 3-19 Error Matrix of TM_60_O_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 68 31 0 1 100 68 80.00 68.00 2 17 75 5 3 100 75 65.79 75.00 3 0 5 73 22 100 73 73.00 73.00 4 0 3 22 75 100 75 74.26 76.00 Ref. Totals 85 114 100 101 400 291 Overall Classification Accuracy 72.75% Overall Kappa= 0.63

PAGE 130

114 Figure 3-39. Classified images A) MSS_60_S classification B) TM_60_S Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest.

PAGE 131

115 Table 3-20. Error Matrix of MSS_60_S_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 64 35 0 1 100 64 71.11 64.00 2 25 69 3 3 100 69 60.00 69.00 3 0 8 62 30 100 62 70.45 62.00 4 1 3 23 73 100 73 68.22 73.00 Ref. Totals 90 115 88 107 400 Overall Classification Accuracy 67.00% Overall Kappa= 0.56 Table 3-21. Error Matrix of TM_60_S_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 70 29 0 1 100 70 68.99 70.00 2 15 77 5 3 100 77 74.73 77.00 3 0 3 80 17 100 80 73.79 80.00 4 0 1 12 87 100 87 78.35 87.00 Ref. Totals 85 110 97 108 400 Overall Classification Accuracy 78.50% Overall Kappa= 0.71

PAGE 132

116 In comparing the error matrix of the MSS_PC_TC_30_O_1986 and the TM_PC_TC_30_O_1986 datasets (Figure 3-40), it was found that the 30m dataset of TM and the 30m dataset of MSS produced 83.50% and 74.00% overall accuracies, respectively (Table 3-22 and Table 3-23), which are higher than the original and the smoothed datasets for 30m and 60m spatial resolutions. When the MSS_PC_TC_30_O_1986 and the TM_PC_TC_30_O_1986 datasets (Figure 3-41) were smoothed, overall accuracies of the MSS_PC_TC_30_S_1986 and the TM_PC_TC_30_S_1986 datasets were determined to be 79.00 % and 88.50%, respectively (Table 3-24 and Table 3-25). The of the TM_PC_TC_30_S_1986 (0.85) was highest compared with all other 30m created datasets. K

PAGE 133

117 Figure 3-40. Classified images A) MSS_PC_TC_30_O Classification B) TM_PC_TC_30_O Classification Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest.

PAGE 134

118 Table 3-22. Error Matrix of MSS_PC_TC_30_O_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 85 11 0 4 100 85 80.19 85.00 2 21 64 7 8 100 64 82.05 64.00 3 0 2 82 16 100 82 85.42 82.00 4 0 1 7 92 100 92 76.67 92.00 Ref. Totals 106 78 96 120 400 323 Overall Classification Accuracy 74.25% Overall Kappa= 0.69 Table 3-23. Error Matrix of TM_PC_TC_30_O_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 85 11 0 4 100 85 81.73 85.00 2 19 67 7 7 100 67 82.72 67.00 3 0 2 87 11 100 87 88.78 87.00 4 0 1 4 95 100 95 81.20 95.00 Ref. Totals 104 81 98 117 400 334 Overall Classification Accuracy 83.50% Overall Kappa= 0.78

PAGE 135

119 Figure 3-41. Classified images A) MSS_PC_TC_30_S classification B) TM_PC_TC_30_S Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest.

PAGE 136

120 Table 3-24. Error Matrix of MSS_PC_TC_30_S_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 90 9 0 1 100 90 82.57 90.00 2 19 65 7 9 100 65 83.33 65.00 3 0 2 75 23 100 75 79.79 75.00 4 0 2 12 86 100 86 72.27 86.00 Ref. Totals 109 78 94 119 400 316 Overall Classification Accuracy 79.00% Overall Kappa= 0.72 Table 3-25. Error Matrix of TM_PC_TC_30_S_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 91 9 0 0 100 91 87.50 91.00 2 13 75 6 6 100 75 84.27 75.00 3 0 4 93 3 100 93 90.29 93.00 4 0 1 4 95 100 95 91.35 95.00 Ref. Totals 104 89 103 104 400 354 Overall Classification Accuracy 88.50% Overall Kappa= 0.85

PAGE 137

121 In comparing the error matrix of the MSS_PC_TC_60_O_1986 and the TM_PC_TC_60_O_1986 datasets (Figure 3-42), it was found that the 60m dataset of MSS and the 60m dataset of TM produced 83.75% and 77.00% overall accuracies, respectively (Table 3-26 and Table 3-27). When the MSS_PC_TC_60_O_1986 and the TM_PC_TC_60_O_1986 datasets (Figure 3-43) were smoothed, overall accuracies of the MSS_PC_TC_60_S_1986 and the TM_PC_TC_60_S_1986 datasets were determined to be 84.00 % and 76.25%, respectively (Table 3-28 and Table 3-29).

PAGE 138

122 Figure 3-42. Classified images A) MSS_PC_TC_60_O classification B) TM_PC_TC_60_O Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest.

PAGE 139

123 Table 3-26. Error Matrix of MSS_PC_TC_60_O_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 78 21 0 1 100 78 82.98 78.00 2 16 74 8 2 100 74 74.75 74.00 3 0 3 73 24 100 73 75.26 73.00 4 0 1 16 83 100 83 75.45 83.00 Ref. Totals 94 97 110 400 308 99 Overall Classification Accuracy 77.00% Overall Kappa= 0.69 Table 3-27. Error Matrix of TM_PC_TC_60_O_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 81 18 0 1 100 81 81.73 85.00 2 16 73 8 3 100 73 82.72 67.00 3 0 3 91 6 100 91 88.78 87.00 4 0 2 8 90 100 90 81.20 95.00 Ref. Totals 97 96 107 100 400 335 Overall Classification Accuracy 83.75% Overall Kappa= 0.78

PAGE 140

124 Figure 3-43. Classified images A) MSS_PC_TC_60_S classification B) TM_PC_TC_60_S Classification. Color indicates: Yellow: Clear Cut-1-3 yr pine three, Light green: 3-8 yr pine three, Dark green: > 8 yr pine three, and Red: wetland forest.

PAGE 141

125 Table 3-28. Error Matrix of MSS_PC_TC_60_S_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 75 24 0 1 100 75 79.79 75.00 2 19 72 6 3 100 72 70.59 72.00 3 0 5 71 24 100 71 79.78 71.00 4 0 1 12 87 100 87 75.65 87.00 Ref. Totals 94 102 89 115 400 305 Overall Classification Accuracy 76.25% Overall Kappa= 0.68 Table 3-29 Error Matrix of TM_PC_TC_60_S_1986 for test site Class Name 1 2 3 4 CT NC PA % UA % 1 82 17 0 1 100 82 83.67 82.00 2 16 71 8 5 100 71 74.74 71.00 3 0 5 90 5 100 90 87.38 90.00 4 0 2 5 93 100 93 89.42 93.00 336 98 95 103 104 400 336 Overall Classification Accuracy 84.00% Overall Kappa= 0.78

PAGE 142

126 Summary of Image Classification Traditionally, the error matrix has been used mainly to provide a basic description of thematic map accuracy and for the comparison of accuracies. it is also possible to use the information contained in the matrix to derive considerably more useful information. As noted previously, the error matrix may be useful in refining estimates of the area extent of classes in the region. After determining the best overall classification accuracies and values for the 30m and 60 m datasets from the test site, it was decided to use the MSS_PC_TC_30_S_1986 and the TM_PC_TC_30_S_1986 for producing the land cover maps to prepare look up tables for the application sites. K After determining the best overall classification accuracies for 30m and 60 m datasets from test site, it was found that the TM_PC_TC_30_S_1986 produced the highest overall accuracy. However it was necessary to determine how much better this overall classification accuracy is than the other datasets. The significance level of the overall accuracy was investigated as detailed in the following section before it was decided to use TM_PC_TC_30_S_1986 and MSS_PC_TC_30_S_1986 for the application site classification Determination of the Most Accurate Dataset Overall classification accuracies were computed four times for each of the sixteen datasets from four different randomly selected points by ERDAS imagine Version 8.6 (Table 3-30) and the results are summarized in Table 3-31. The means of the overall accuracies were tested to determine the mean differences using Fisher’s Least Significant Difference (LSD) method. It was initially assumed that the datasets having the highest overall accuracies must be statistically different from those that had lower overall

PAGE 143

127 Table 3-30. Overall accuracies with 4 replicates used to validate the most accurate dataset Dataset 30m R Overall A (%) Dataset 60m R Overall A(%) MSS_30_O_1986 1 67.50 MSS_60_O_1986 9 65.00 MSS_30_O_1986 1 68.10 MSS_60_O_1986 9 61.20 MSS_30_O_1986 1 62.70 MSS_60_O_1986 9 62.10 MSS_30_O_1986 1 66.90 MSS_60_O_1986 9 67.30 TM_30_O_1986 2 71.25 TM_60_O_1986 10 72.75 TM_30_O_1986 2 70.70 TM_60_O_1986 10 75.10 TM_30_O_1986 2 76.20 TM_60_O_1986 10 76.00 TM_30_O_1986 2 73.20 TM_60_O_1986 10 70.10 MSS_30_S_1986 3 69.00 MSS_60_S_1986 11 67.00 MSS_30_S_1986 3 68.90 MSS_60_S_1986 11 66.90 MSS_30_S_1986 3 71.10 MSS_60_S_1986 11 70.10 MSS_30_S_1986 3 69.80 MSS_60_S_1986 11 60.70 TM_30_S_1986 4 73.20 TM_60_S_1986 12 78.90 TM_30_S_1986 4 77.10 TM_60_S_1986 12 77.10 TM_30_S_1986 4 73.00 TM_60_S_1986 12 71.20 TM_30_S_1986 4 74.10 TM_60_S_1986 12 77.80 MSS_PC_TC_30_O_1986 5 74.20 MSS_PC_TC_60_O_1986 13 77.00 MSS_PC_TC_30_O_1986 5 75.40 MSS_PC_TC_60_O_1986 13 71.20 MSS_PC_TC_30_O_1986 5 73.60 MSS_PC_TC_60_O_1986 13 78.30 MSS_PC_TC_30_O_1986 5 77.00 MSS_PC_TC_60_O_1986 13 78.00 TM_PC_TC_30_O_1986 6 83.50 TM_PC_TC_60_O_1986 14 83.70 TM_PC_TC_30_O_1986 6 84.10 TM_PC_TC_60_O_1986 14 80.10 TM_PC_TC_30_O_1986 6 80.90 TM_PC_TC_60_O_1986 14 84.60 TM_PC_TC_30_O_1986 6 86.20 TM_PC_TC_60_O_1986 14 85.10 MSS_PC_TC_30_S_1986 7 79.00 MSS_PC_TC_60_S_1986 15 76.25 MSS_PC_TC_30_S_1986 7 81.10 MSS_PC_TC_60_S_1986 15 76.10 MSS_PC_TC_30_S_1986 7 81.50 MSS_PC_TC_60_S_1986 15 74.00 MSS_PC_TC_30_S_1986 7 81.00 MSS_PC_TC_60_S_1986 15 78.40 TM_PC_TC_30_S_1986 8 88.50 TM_PC_TC_60_S_1986 16 84.00 TM_PC_TC_30_S_1986 8 89.10 TM_PC_TC_60_S_1986 16 87.30 TM_PC_TC_30_S_1986 8 83.00 TM_PC_TC_60_S_1986 16 81.20 TM_PC_TC_30_S_1986 8 87.30 TM_PC_TC_60_S_1986 16 84.60 R= replication

PAGE 144

128 accuracies and this assumption was tested as follows. In order to compare the means of the overall accuracies, an analysis of means was performed to test the hypothesis (Ho= 1 = 2 = 3 =..........= 14 = 15 = 16 ) against the alternative hypothesis, (Ha= 1 2 3 .......... 14 15 16 ) that at least one of the means differs from the rest. If Ho is rejected, the LSD is defined to be the observed difference between two overall accuracy means necessary to declare the corresponding overall accuracy means different (Ott and Longnecker, 2001). For the specified value of , the least significant difference for comparing i to j jiErrorijnnMStLSD112/ (Eq.3-7) where n i and n j are the respective datasets size i and j and t is the critical student t value for the appropriate and number of degrees of freedom (df) and MS Error is the mean square error. For this research n =4 for all datasets nMStLSDLSDErrorij22/ (Eq.3-8) After determining the LSD value, all pairs of overall accuracy means were compared using Fisher’s LSD comparison method, ijjiLSDyy where y is the mean of the selected dataset. Evaluation of Overall Accuracy Means First the Analysis of Variance (AOV) was performed to determine whether there was a statistically significant difference between any of the means. The detatils of this test are not presented here but further information on statistical testing by analysis of variance the reader can consult Ott and Longnecker, (2001). The threshold value for

PAGE 145

129 determining significant difference between two means was calculated to be 2.90 using Equation 3-8. Table 3-31 gives the datasets and equivalent dataset numbers in the R column. These dataset numbers are used in Table 3-32 where the means are ranked for overall accuracies from the lowest to the highest. smallestestlyyarg If this difference was greater than the LSD, it was concluded that the corresponding pairwise overall accuracy means are significantly different from each other. The 2 nd largest and the smallest means were then compared and the difference compared with the LSD smallestestlndyyarg.2 Table 3-31. Calculated means for overall accuracies Datasets R N Mean StDev MSS_30_O_1986 1 4 66.30 2.44 TM_30_O_1986 2 4 72.83 2.48 MSS_30_S_1986 3 4 69.70 1.01 TM_30_S_1986 4 4 74.35 1.89 MSS_PC_TC_30_O_1986 5 4 75.05 1.50 TM_PC_TC_30_O_1986 6 4 83.67 2.18 MSS_PC_TC_30_S_1986 7 4 80.65 1.12 TM_PC_TC_30_S_1986 8 4 86.97 2.75 MSS_60_O_1986 9 4 63.90 2.78 TM_60_O_1986 10 4 73.48 2.64 MSS_60_S_1986 11 4 66.17 3.94 TM_60_S_1986 12 4 76.25 3.44 MSS_PC_TC_60_O_1986 13 4 76.12 3.33 TM_PC_TC_60_O_1986 14 4 83.37 2.25 MSS_PC_TC_60_S_1986 15 4 76.18 1.79 TM_PC_TC_60_S_1986 16 4 84.27 2.50 R= dataset number

PAGE 146

130 Table 3-33, Table 3-34 and Table 3-35 illustrate how the corresponding differences between the means of accuracies were calculated. These were compared with the threshold value and designated as greater than or less than this threshold value. Table 3-36 shows that there are three main groups, which are not significantly different from each other at the significant level of 0.05. As illustrated in Table 3-36, the means of the dataset 1 (MSS_30_O_1986), 3 (MSS_30_S_1986), 9 (MSS_60_O_1986), and 11 (MSS_60_S_1986) are not significantly different at the significance level of 0.05 (Table 3-36) (group A). Similarly the means of dataset 2 (MSS_30_O_1986), 4 (TM_30_S_1986), 5 (MSS_PC_TC_30_O_1986), 10 (TM_60_O_1986), 12 (TM_60_S_1986), 13 (MSS_PC_TC_60_O_1986), and 15 (MSS_PC_TC_60_S_1986) are the same at significance level of 0.05 (group B). It was found that, there are no significant differences among the means of datasets 6 (TM_PC_TC_30_O_1986), 7 (MSS_PC_TC_30_S_1986), 8 (TM_PC_TC_30_S_1986), 14 (TM_PC_TC_60_O_1986), and 16 (TM_PC_TC_60_S_1986) at the significance level of 0.05 (group C) (Table 3-36).

PAGE 147

Table 3-32. Ranking means of overall accuracies 131 OA 9 11 1 3 2 10 4 5 15 13 12 7 14 6 16 8 i y 63.9 66.1 66.3 69.7 72.8 73.4 74.3 75.0 76.1 76.1 76.2 80.6 83.3 83.6 84.2 86.9 Table 3-33. LSD means determination 9y to 2y 9y 11y 1y 3y 2y 9.20.2398yy 9.28.20118 yy 9.26.2018 yy 9.22.1738 yy 9.21.1428 yy 9.23.20916yy 9.21.181116 yy 9.29.17116 yy 9.259.14316 yy 9.24.11216 yy 9.27.1996yy 9.25.17116 yy 9.23.1716 yy 9.29.1336 yy 9.28.1026 yy 9.24.19914yy 9.22.171114 yy 9.20.17114 yy 9.26.13314 yy 9.25.10214 yy 9.27.1697yy 9.25.14117 yy 9.23.1417 yy 9.29.1037 yy 9.28.727 yy 9.23.12912yy 9.21.101112 yy 9.29.9112 yy 9.25.6312 yy 9.24.3212 yy 9.21.12913yy 9.20.101113 yy 9.28.9113 yy 9.24.6313 yy 9.23.3213 yy 9.21.12915yy 9.20.101115 yy 9.28.9115 yy 9.24.6315 yy 9.23.3215 yy 9.21.1195yy 9.29.8115 yy 9.27.815 yy 9.23.535 yy 9.22.225 yy * 9.24.1094yy 9.22.8114 yy 9.20.814 yy 9.26.434 yy 9.25.9910yy 9.23.71110 yy 9.21.7110 yy 9.27.3310 yy 9.29.892yy 9.27.6112 yy 9.25.612 yy 9.21.332 yy 9.28.593yy 9.26.3113 yy 9.24.313 yy * 9.24.291yy * 9.22.0111 yy * 9.20.0111 yy *

PAGE 148

Table 3-34. LSD means determination 10y to 13y 10y 4y 5y 15y 13y 9.25.13108yy 9.26.1248 yy 9.29.1158 yy 9.28.10158 yy 9.28.10138 yy 9.28.101016yy 9.29.9416 yy 9.28.9516 yy 9.21.81516 yy 9.21.81316 yy 9.22.10106yy 9.23.946 yy 9.26.856 yy 9.25.7156 yy 9.25.7136 yy 9.29.91014yy 9.20.9414 yy 9.23.8514 yy 9.22.71514 yy 9.22.71314 yy 9.22.7107yy 9.23.647 yy 9.26.557 yy 9.25.4157 yy 9.25.4137 yy 9.28.21012yy * 9.28.1412 yy * 9.22.1512 yy * 9.21.01512 yy * 9.21.01312 yy * Table 3-35. LSD means determination 12y to 16y 12y 7y 14y 6y 16y 9.27.10128yy 9.23.678 yy 9.26.3148 yy 9.26.368 yy 9.27.2168 yy * 9.20.81216yy 9.26.3716 yy 9.21.21416 yy * 9.26.0616 yy * 9.24.7126yy 9.2376 yy 9.21.71214yy 9.27.2714 yy * 9.24.4127yy 132

PAGE 149

Table 3-36. Significant and non-significant level using Fisher LSD method for overall accuracy’s means er Dataset numb 9 11 1 3 2 10 4 5 15 13 12 7 14 6 16 8 9y 11y 1y 3y 2y 10y 4y 5y 15y 13y 12y 7y 14y 6y 16y 8y 133 A B C

PAGE 150

134 Land Cover Recognitions After the decision as made to use MSS_PC_TC_30_S_1986 and TM_PC_TC_30_S_1986 datasets, image differences were investigated for all datasets (Table 3-2) to determine how the MSS_PC_TC_30_S_1986 and the TM_PC_TC_30_S_1986 datasets differ from the other datasets. The 30 m classified MSS images were subtracted from 30 m classified TM images. Difference images from the 30 m classified TM and the 30 m classified MSS datasets were investigated using change detection to determine the effects of smoothing, PCA, and TCT on image classification (Figure 3-44 A). It was found that changed land cover area was 2865 ha and unchanged area was 7011 ha when the classified MSS_30_O_1986 image was subtracted from the classified TM_30_O_1986 image. Likewise, changed area was 2216 ha and unchanged was 7661 ha when the smoothed 30 m MSS was subtracted from 30 m TM datasets (MSS_30_S_1986 from the classified TM_30_S_1986 images) (Figure 3-44 B). The composite of original PC 12, BI, GI images from MSS were subtracted from PC12_BI_GI_30_O_MSS and PC12_BI_GI_30_O_TM and found that 6770 ha area was unchanged and 3022 ha area was changed when he classified (Figure 3-44 C) MSS_PC_TC_30_S_1986 image subtracted from the classified the TM_PC_TC_30_S_1986 image. As seen Figure 3-44, unchanged area was increased compared with all other image subtraction results. In this subtraction, 7794 ha unchanged and 1998 ha changed area was found on the test area (Figure 3-44 D). The 60 m classified MSS, MSS_60_O_1986, were subtracted from the 60 m classified TM image, MSS_60_O_MSS (Figure 3-45 A) and it was found that changed land cover area was 3514 ha and unchanged area was 6424 ha.. When the

PAGE 151

135 MSS_30_S_1986 image was subtracted from the TM_30_S_1986 image, it was found that unchanged area was 7992 ha and changed area was 1946 ha (Figure 3-45B). When the classified MSS_PC_TC_30_O_1986 image was subtracted from TM_PC_TC_30_O_1986 image, unchanged area was found as 6839 ha and changed area was determined as 3053 ha (Figure3-45 C). Likewise, the classified MSS_PC_TC_30_S_1986 was subtracted from the classified TM_PC_TC_30_S_1986 image. In this subtraction, 7929 ha unchanged and 11963 ha changed area was found in the test area (Figure 3-45 D). After finding changed and unchanged area in the test site, land cover recognition was investigated between the images using from to tables before and after smoothing, PCA, and TCT were performed (Table 3-37). A “ from-to” table is essentially the same as an error matrix used for accuracy analysis. Numbers on columns and rows in Table 3-36 indicated the individual classes as follows: (1) Clear_cut_1-3 yr pine plantation, (2) 4-8 yr pine plantation, (3) > 8 yr pine plantation, and (4) Wetland Forest. As an example, clear_cut_1-3 yr pine plantation class on TM_30_O_1986 image was recognized 69% by MSS_30_O_1986 images as same class. Clear_cut_1-3 yr pine plantation class on the classified TM_30_O_1986 image was recognized 15% as 4-8 yr pine plantation class and 14 % as wetland forest on the classified MSS_30_O_1986. All other classes’ recognition percentages were found as 51% for 4-8 yr pine plantation, 74% for >8 yr pine plantation, and 74% for wetland forest for the classified MSS_30_O_1986 and the classified TM_30_O_1986 image (Table 3-37). The other section of Table 3-37 give similar values for the other compared datasets

PAGE 152

136 A B C D Figure 3-44. Image difference between the 30 m classified TM and the 30 m classified MSS images B A D C Figure 3-45. Image difference between the 60 m classified TM and the 60 m classified MSS images

PAGE 153

137 Table 3-37. Land cover conversion from-to tables for 30 m and 60 m classified images

PAGE 154

138 Summary No significant differences were found between TM_PC_TC_30_O_1986 and TM_PC_TC_30_S_1986, but it was found that there is a significant difference between MSS_PC_TC_30_O_1986 and MSS_PC_TC_30_S_1986 at the significance level of 0.05. In this research, the dataset providing the highest overall classification accuracies and the highest values for MSS was MSS_PC_TC_30_S_1986 and for TM was TM_PC_TC_30_S_1986. K Analysis of land cover recognition indicated that differences between 30 m and 60 m datasets were varied from class to class. However, it was found that smoothing of images provided better land cover recognitions for original datasets and composite of PC1, PC2, BI, and GI images for both 30 m and 60 m datasets. As presented in Chapter 4, the land cover maps for 1975 through 2000 were produced using the datasets that demonstrated the highest overall accuracy.

PAGE 155

CHAPTER 4 LAND COVER CLASSIFICATION FOR APPLICATION SITE To understand LULCC, we must examine how the landscape has changed over an extended period of time, then study why those patterns of change occurred. Our instruments must see large areas synoptically, and have a long time period of archived data with fine temporal resolution. Only the Landsat program has such data. The U.S. National Aeronautics and Space Administration (NASA) initiated the first civilian program specializing in the acquisition of remotely sensed digital satellite data in 1972. There have been 7 Landsat satellites launched (Landsat 6 failed to attain orbit) since 1972, and three (Landsat 4, 5 and 7) are currently in orbit gathering data although new data from Landsat 4 is not available to civilian sector. The Landsat satellite series offers the longest continuous satellite data set available for land cover change studies. Nonetheless, it is very difficult to compare different data sources directly even within the same satellite program because of technical problems and variations in data quality of the Landsat series. Chapter 3 presented the results of the best image processing and classification techniques for different datasets (Landsat MSS and Landsat TM) for this research. The current chapter presents a look up table for carbon data from land cover classification, which was investigated in advance in Chapter 3. It was found that, for classification purposes, the 30m Landsat TM dataset that was a composite of the smoothed PC1, PC2, BI, and GI indexes and the smoothed 30m Landsat MSS dataset that was a composite of the smoothed PC1, PC2, BI, and GI indexes were the best for this research. 139

PAGE 156

140 Application Site The application site is situated in northern Florida in the lower coastal plain in Alachua County (Figure 4-1). A 15 km x 15 km area was selected with good overlap of available cloud-free Landsat MSS and Landsat TM images (Figure 4-2). The area is predominantly flat, less than 100 m above sea level. The area supports pure and mixed pine plantation as well as wetland forest. As with most forestland in north Florida, this area is intensively managed for timber production. The majority of landowners in the area are forest companies and the University of Florida. Stands may be cleaned of bushes during early development and are typically thinned several times during the approximately 15 to 25 year rotation period

PAGE 157

141 Figure 4-1. Composite (Bands 3,3, and 2 RGB) image of WRS 17-39 scene, application site Hamilton County and application site Alachua County in southeast costal plain. .

PAGE 158

142 Image characteristics Landsat MSS, Landsat TM, and Landsat ETM+ images from 1975 to 2000, all acquired during the winter and early spring, were used to generate land cover maps for the application site using the method generated from the test site. The application sites are subsets of the path 17 and row 39 scenes. The nominal selection criteria were to acquire cloud-free images and same-season images (winter). Data used for this project were four band MSS images comprised of MSS-Red, MSS-Green, MSS-NIR (band3), and MSS-NIR (band 4) bands and three bands TM images comprised of TM-Red, TM-Green, and TM-NIR (band4). Even if Landsat data were available from 1972 to the present some images have problems that are obvious and therefore they cannot be used for any classification purpose in this study. For the application site, the MSS 1977, MSS 1978, MSS 1979, and MSS 1980 images were discarded due to errors in the delivered data, which consisted of extensive line dropouts, and random line offsetting. In addition, any available TM 1983, TM 1984 and TM 1993 winter or early spring images have cloud problems (Figure 4-2). A total of 19 images were used to create land cover maps for the application area and create look up tables.

PAGE 159

143 Figure 4-2. Datasets used for classification

PAGE 160

144 Determination of Land Cover Types and Image Classification Image processing for the application site was performed using the dataset determined at the test site and used to produce classified maps and look up table for application site (Figure 4-3). Figure 4-4 shows land cover types that were determined on a DOQ image. The yellow circle is clear-cut_1-3 yr pine plantation, light green square represents 4-8 yr pine plantation, dark green regular pentagon represents > 8 yr pine plantation, red square represents wetland forest, and gray circle with red dot represents other land cover types. Open water was not present in this particular inset. For long-term forest cover change study, any agricultural, urban, road and other land cover types were masked out from forest areas (Figure 4-5). Roads were masked out using the TIGER road map from FGDL website. It was buffered by 20 m on each side to create a reasonable area size. Agriculture and urban sites were masked out using a heads up digitizing method (Figure 4-5). All images from 1975 to 2000 were smoothed using a low pass filter and PCA and TCT were performed. The smoothed PC1, the smoothed PC2, the smoothed BI, and the smoothed GI respectively were added on top of each other to create a new dataset for each year. An additional five field trips were made to the application site during summer 2000, spring 2001, summer 2001, winter 2002 and spring 2002 to confirm the land cover characteristics before classification was performed. It was decided to have two more land cover types than at the test site because of the percentage of open water and masked out area (others) that consists of agriculture, urban, and roads.

PAGE 161

145 Figure 4-3. Classified images used to produced look up tables

PAGE 162

146 5000500250Meter s Figure 4-4. Land cover determination for application site

PAGE 163

147 Figure 4-5. Agriculture, urban, and roads were masked from the application site (forest area is shown in green)

PAGE 164

148 Results Evaluation of Area Coverage A total of nineteen datasets were used to create land cover maps for the application site. The results were presented in Figure 4-6 and Figure 4-7 and their areas of coverage are presented in Table 4-1. The area of coverage of open water for all the years amounted to less than 1%. The area of Clear_cut_1-3 year pine plantation changed year to year. In 1975 this class comprised 3% of the total area while in 1976 this rose to 4%. In 1982, this class increased a further 2% over what it had been in 1981 (6%). In this region, the area of forest cover is very dynamic, increasing and decreasing over the period. However, starting in 1991 and continuing to 1993, there was a large amount of clear cutting causing the land cover to change dramatically. A fire in summer 1998 caused a clear cutting in 1999 and therefore increased the clear-cut class to 6%. The detailed information about land cover areas can be found in Figure 4-6, Table 4-1, and Figure 4-7. Open water did not change dramatically over the years. Only in the year 1976 does a high of 61 ha appear, which could be because of classification error or metrological activities prior to the image being taken by the MSS sensors. Area of clear-cut_1-3 yr pine plantation class ranged from 668 ha to 2467 ha. Since this landscape is dynamic, there is no obvious pattern in terms of land cover area change for this class through the years. The 4-8 yr pine plantation class also changed dramatically in the 25 yr time period. Land cover dynamics are evident in this class as well. The land cover area for this class ranged from 1949 ha (2000) to 4854ha (1976).

PAGE 165

149 The > 8 yr pine plantation class also shows changes over the years. However over time the cyclical fluctuations prove that regeneration of pine trees was taking place. Wetland forest over the time shows consistency. It was expected that no change would be observed, however either owing to classification error or metrological phenomena (rainfall causing flooding immediately before the image was taken) this coverage changed somewhat over the years with 1981 demonstrating the highest area and 1999 the lowest area found.

PAGE 166

150 Figure 4-6. Image classification from 1975 to 2000

PAGE 167

Table 4-1. Look up table for classification. 151151

PAGE 168

152152 Figure 4-7 Area of coverage change based on the years

PAGE 169

153 Evaluation of Classification Accuracy As is shown in Figure 4-8 and in Table 4-2, the Landsat MSS dataset (MSS_PC_TC_30_S_1975, MSS_PC_TC_30_S_1976 and MSS_PC_TC_30_S_1981) give relatively low overall accuracies of respectively 78%, 79.6% and 82.8%. Landsat TM dataset available from 1985 to 1999 (TM_PC_TC_30_S_1982, TM_PC_TC_30_S_1985, TM_PC_TC_30_S_1986, TM_PC_TC_30_S_1987, TM_PC_TC_30_S_1989, TM_PC_TC_30_S_1990, TM_PC_TC_30_S_1991, TM_PC_TC_30_S_1992, TM_PC_TC_30_S_1993, TM_PC_TC_30_S_1994, TM_PC_TC_30_S_1995, TM_PC_TC_30_S_1996, TM_PC_TC_30_S_1997, and TM_PC_TC_30_S_1999) gives better overall accuracies as is seen in Figure 4-8. The Landsat 1998 dataset (TM_PC_TC_30_S_1998) shows relatively low 82.80% overall accuracy. The Landsat ETM+ 2000 images gives 84.80% overall accuracy.

PAGE 170

Table 4-2. Classification accuracy for application site from 1975 to 2000 154

PAGE 171

Figure 4-8. Overall classification for application site from 1975 to 2000 155

PAGE 172

CHAPTER 5 CONCLUSIONS AND RECOMENDATIONS Images from both Landsat MSS and TM datasets were analyzed with different image enhancement techniques. From this analysis, it was found that the smoothing operation performed prior to classification improved the classification accuracy by 7% compared to the original dataset classification. Performing the PCA and TCT to smoothed datasets also improved the classification accuracy by 4% compared to the smoothed datasets only. The first two principal components, PC1 and PC2, added to the first two indexes from TCT, BI and GI, were used to create a new 4-band dataset, MSS_PC_TC_30_S_1986 and TM_PC_TC_30_S_1986. These combinations of images were used to determine the land cover for the application site from the years 1975 to 2000. It was found that the determination of five land cover classes using these techniques produced a highest overall land cover classification accuracy of 88.8%. A new technique was developed to create the forestland cover classes based on their ages using Landsat imagery over the time. Clear-cut areas were visually determined using Landsat images from 1975 to 2000 since there were no ground truth data for the application sites except for the years 1995 and 2000. This technique was found to be capable of determining the forest ages 1 to 3, 4 to 8 and > 8 years pine plantation. Wetland forest was also determined using the Landsat dataset and DOQ dataset satisfactorily. 156

PAGE 173

157 A look up table for land cover classification was investigated and it was found that the 30m Landsat TM dataset that was a composite of the smoothed PC1, PC2, BI, and GI indexes and the 30m Landsat MSS datasets that was a composite of the smoothed PC1, PC2, BI, and GI indexes were the best for forest cover classification accuracy. It was found that a few land cover classes (such as four or five forest cover classes) could be determined with better results using these techniques. However, the efficacy of these techniques to determine a greater number of land cover classes has not been tested. Therefore, using this method for other types of landscape must be tested before using actual land classification. In this study it was found that smoothing operations improved the classification accuracy when it is done before classification. However, this smoothing operation before classification created a selection of homogeneous training sets, which can result in the loss of some detail in the land cover classification. Therefore smoothing is not recommended for other studies unless it is required to improve the visualization and specific classes. Class recognition analysis showed that smoothing before land cover classification for any datasets increased class recognitions for both 30 m and 60 m datasets. It was found that composite of PC1, PC2, BI, and GI images for 30 m dataset also increased class recognition for wetland forest. The other datasets from earth looking satellites (SPOT, IKONOS, SPOT, Quick-bird etc.) should be tested and may be used for forest land cover classification studies. Other vegetation indices such as NDVI, Vegetation Index etc. should be tested for different study areas before being applied to the actual classification.

PAGE 174

APPENDIX A NORMALIZED DIFFERENCE VEGETATION INDEX (NDVI) The normalized difference vegetation index (NDVI) is a non-linear transformation of the visible (RED) and near-infrared bands of satellite information. NDVI is defined as the difference between the visible (RED) and near-infrared (NIR) bands, over their sum. The NDVI is an alternative measure of vegetation amount and condition. It is associated with vegetation canopy characteristics such as biomass, leaf area index and percentage of vegetation cover. Tucker (1979) reported that brightness values from an individual MSS band could be used as vegetation indexes to predicted percent ground cover. Jensen (2000) mentions that brighter pixel the values imply the greater amounts of vegetation. Most of the indices are based on combinations of near infrared and red wavelengths. The most widely used vegetation index is the NDVI, which was developed by Rouse et al. (1973). A number of factors influence variations in the NDVI. NDVI is the traditional vegetation index used by researchers for extracting vegetation abundance from remotely sensed data. It divides the difference between reflectance values in the visible red and near-infrared wavelengths by the overall reflectance in those wavelengths to give an estimate of green vegetation abundance (Tucker, 1979). In essence, the algorithm isolates the dramatic increase in reflectance over the visible red to near infrared wavelengths, and normalizes it by dividing by the overall brightness of each pixel in those wavelengths. Specifically NDVI for general 158

PAGE 175

159 form, Landsat TM, SPOT, and AVHRR are expressed by equations A-1, A-2, A-3, and A-4 respectively. The normalized differential vegetation index (NDVI) is defined by the equation bandredIRbandredIRNDVIMSS__ (A-1) 3434TMTMTMTMNDVITM (A-2) 2323XSXSXSXSNDVISPOT (A-3) bandredIRbandredIRNDVIAVHRR__ (A-4) During the last 4 decades, NDVI has been used for many purposes, e.g. estimating vegetation density and cover; percent vegetation ground cover; leaf area index and crop condition (Wiegand et al. 1991). NDVI has been used extensively to measure vegetation cover characteristics on a broad-scale worldwide, and has been incorporated into many large-scale forest and crop assessment studies (Peterson et al. 1987).

PAGE 176

APPENDIX B TEXTURE ANALYSIS Image grey level differences among vegetation and land cover are often used to create signatures of the cover types in imagery of the earth’s surface. These spectral signatures may be compared to identify the likely cover type of each pixel in the image. This pixel-by-pixel classification approach becomes more problematic as image resolution increases because more individual features are discriminated and the variance in spectral values observed within a cover type increases (Pearlstine, 2000). Individual pixels representing one of features on the earth’s surface may significantly vary in spectral reflectance from a pixel representing one of the other features of the same cover type. This makes it impossible to classify a pixel to a cover type based on its grey level. Jensen (2000) mentioned that grey level (spectral tone) from the various features of a cover type is averaged as images decrease in spatial resolution. Lillesand and Kiefer (1999) said that because the reflectance within a cover type averaged over a multi-meter square pixel is generally more homogeneous across the pixels in the imagery, lower resolution images, such as Landsat and SPOT satellite imagery are often successfully classified from pixel-by-pixel identification of a spectral signature. Texture is the statistical propert of the spatial distribution of gray tones in an image (Haralick et al., 1973). Many researchers have applied texture analyses to improve classification of satellite imagery and aerial photography to forest landscapes (Holopainen and Wang, 1998; Riou and Seyler,1997). 160

PAGE 177

161 Basically, there is no single texture or contextual approach that has emerged in computer texture analysis as superior in its capability to improve land use/land cover classification accuracy. First and second-order statistical textures are the most common approaches. In this study, first-order statistics for texture will be used to achieve the goal. First-order statistical textures known a variance and skew were tested. The equations are presented here from Tomaita and Tsuji (1990): niifiN1*1 (B-1) niifiNVariance12*)(1 (B-2) niifiSkewness13*)( (B-3) i = the observed spectral intensity N = the total number of pixel in window fi = the probability of i occurring in a pixel window, = Mean of the intensities For first-order statistics, basic elements such as the mean, standard deviation, and skew quantify the distribution properties of grey levels in the image. Second-order statistics describe the frequency with which one grey level appears in specific relationship to another grey level on the image through the construction of relative frequency distribution matrices (Haralick et al., 1973). Kushwasha et al. (1994) reported modest improvements in IRS satellite classification of forests using Haralick textures, but Dikshit (1996) obtained better results with first-order statistics (TM images). Both

PAGE 178

162 authors observed better accuracies with a combination of level and texture layers rather than the spectral tone layers alone. The texture values are computed in a 3x3 moving windows. The center cell in a neighborhood is assigned a value that is the computed texture for that neighborhood. The window then is moved over one pixel, and a new value is computed for that neighborhood. The window continues to move across the image assigning texture values to the center cell of the window until the entire image has been scanned. Because analysis windows centered on cells near the edge of an image extend beyond the image, values are assigned to the empty cells in the window by assuming that values just outside the image are the same as values just inside the image and replicating those values to the empty cells. Although texture features have been increasingly incorporated into multi-spectral classifications, no single algorithm combining efficiency and effectiveness has yet to be widely adopted. Jensen (2000) recommended that the texture features derived for one type of application, land-use classification of urban or forest area, may not necessarily be useful when applied to another geographic problem. As many researchers mentioned texture is only a useful addition when classes in the images exhibit differences in spatial tone patterns (Ryherd and Woodcock (1996). They tested a wide range of remote sensed images and results never were degraded when textural information layers where combined with the spectral bands. When textures are used in combination with other bands of information, it is most commonly with the original spectral bands.

PAGE 179

APPENDIX C PRINCIPAL COMPONENT ANALYSIS Determination of Principal Components Principal Component Analysis was used in this research followed by Jensen (1996) and Richards and Jia (1999). Vectors can describe the position of a pixel in multispectral space. These vector’s components are the individual spectral response in each band for remotely sensed datasets. Consider a multispectral space (two bands)with six pixels plotted in Figure C-1, where each pixel in this space is defined by its vector coefficients (DNS). X3 X2 X6 X1 X5 X4 M B2 Figure C-1. Two dimensional multispectral space showing the individual pixel vectors and their mean position, as defined by M, mean vector (adopted from Richards and Jia, 1999). B1 The mean pixel vector is caring to define the average or likely value of the pixels in multispectral vector space (Equation C-2). This role is essential to define the covariance matrix by Equation C-1: 163

PAGE 180

164 KkkXKxM11)( (C-1) M = Mean pixel vector X k = Individual pixel vector K = Number of total vectors (x) = Expected value (C-2) )))(((tXMxXMxX Covariance matrix in X space X t = Vector transpose Mx = Data mean vector in X space X = Individual pixel vector The estimated covariance matrix is given by Equation C-3 XkktkkMXMXK1))((11 (C-3) The covariance matrix is one of the most key statistical concepts in analysis of multispectral remote sensing data. If there is correlation between the responses in a pair of spectral bands, the equivalent diagonal elements in the covariance matrix could be described by the diagonal terms. On the contrary, if there is little correlation, the off-diagonal terms will be close to zero. This can be explained in terms of the correlation matrix R whose elements are related to those of the co-variance matrix (Equations C-4): further information; see Richards and Jia (1999). jjiiijijvvv or (C-4) 11zzR T ij = Element of correlation matrix v ij = Element of the covariance matrix v ii , v jj = Variance of ith and jth bands of images z = Correlation (%)

PAGE 181

165 Consider the two-dimensional sets of data shown in Figure C-1 with correlation between the two bands (band 1 and band 2). It is essential to know that both the co-variance and correlation matrices are symmetrical. An image datasets, in which there is no correlation between any of the multispectral components, must have a diagonal co-variance and correlation matrix. It is essential for the development of the principal components transformation to provide a new coordinate system in the multispectral vector space. This new vector space then can be represented without correlation. In the other words, in the new co-ordinate system the correlation matrix must be diagonal. Figure C-2 illustrates two-dimensional vector space where such a new coordinate system is described. In the new coordinates, if the vector describing the pixel is represented as Y, a linear transformation, G, can be found of the original coordinates (Equation C-5): GXY (C-5) Covariance matrix of the pixel data in Y space is diagonal and covariance matrix for Y space can be defined by equation C-6: YtMyYMyY)))((( (C-6) Covariance matrix in Y space Y My = Data mean vector in X space Y = Individual pixel vector

PAGE 182

166 B2’ Y2 X Y1 B1’ Highly correlated in X space Y N o correlation in Y space B2 X2 Figure C-2. Modified coordinate system in which the pixel vectors have no correlated components. X1 B1 X space My is expressed by Equation C-10: kkkXKGGMxGXYMy11)()( (C-7) Then co-variance matrix in Y space became (Equation C-8): (C-8) YtGMxGXGMxGX)))((( It can be re-organized (equation C-9): YttGMxXMxXG))))(((( YXtGG (C-9)

PAGE 183

167 Where; X Covariance of the pixel data in X space Y Covariance of the pixel data in Y space By definition must be diagonal and G can be recognized as the transposed matrix of eigenvectors of , provided G is an orthogonal matrix. Finally, Y X Y can be then identified as the diagonal matrix of eigenvalues of X (equation C-10): Yn.............21 (C-10) N = Dimensionality of data and is eigenvalue Because is a co-variance matrix and also diagonal, its elements will be the variance of the pixel data in the respective transformed coordinates. It can be arranged such that Y 1 > 2 >> N so that the data exhibits maximum variance in Y1, next largest variance in Y2 and so on, with minimum variance must be in Y N . In order to determine the principal component transformation, it is necessary to find the eigenvalues and eigenvectors. Equation C-10 is to determine the principal (eigen) values. Principal Components Using the two eigenvalues, PCA can be written as the solution to solve eigenvectors as; Equation C-11: 02,1gIX (C-11)

PAGE 184

168 I = Identification Matrix 2,1 = The eigenvalues =Individual eigenvectors 2,1g These eigenvectors can now be used to write the principal component transformation matrix G as (Equation C-12): tggggG22211211 (C-12) In computing the covariance structure of the transformed Y data it can be found that the data is now de-correlated and the covariance between the two principal component bands is (near) zero (Figure C-3) using equations C-5 and C-13 and reorganized the transformation into equation C-13. Y GX (C-13) 21*2122211211XXggggYYt These data show no noticeable correlation between the B1’ and B2’, also called PC1 and PC2 axes. Secondly most of the data spread is in the direction of the first principal component (B1’). In other words, it can be seen that the first axis contains more information than the other axes. If one were to compare the B1’ and B2’ images, B1’ will have range of higher brightness values for each pixel than B2’. Finally, B1’ will show a higher degree of contrast than B2’ does.

PAGE 185

169 B2 B2’ Xk’ X5’ X4’ X3’ X1’ Y X2’ B1’ B1 X space Figure C-3. Principal component axes for the data of figure C-1

PAGE 186

LIST OF REFERENCES Anderson, J.F., Hardy, E.E., Roach, J.T., Witmer, R.E., 1976, “A land use and land cover classification system for use with remote sensor data,” U.S. Geological Survey Professional Paper 964, U.S. Geological Survey, Washington, DC. Anuta, P.E.,. Bartolucci, L.,. Dean, E., . Lozano-Garca, D.F.,. Malaret, E.,. Valdes, J.,. Valenzuela, C., 1983, Landsat-4 data quality evaluation results. Proc. 8th. W.T. Pecora Memorial Remote Sensing Symposium, Sioux Falls, S.D. Oct. 4-7. Apan, A., 1997; “Land cover mapping for tropical forest rehabilitation planning using satellite imagery,” International Journal of Remote Sensing, 18(5): 1029-1049. Arai, K., 1992, “A supervised thematic mapper classification with a purification of training samples,” International Journal of Remote Sensing, 13(11): 2039-2049. Aronoff, S.,1982, “Classification accuracy: a user approach,” Photogrammetric Engineering and Remote Sensing, 48:1299. Benson, A.S., S.D. DeGloria, 1985, “Interpretation of Landsat-4 thematic mapper and multispectral scanner data for forest surveys,” Photogrammetric Engineering and Remote Sensing. 51:1281-1289. Boyd, D. S., Foody, G. M., Curran, P. J., Lucus, R.M., Honzak,M., 1996, “An assessment of radiance in Landsat TM middle and thermal infrared wavebands for the detection of tropical forest regeneration,” International Journal of Remote Sensing, 17: 249261. Binford, M. W., Barnes, G., Gholz, H. L., Smith, S. E., 2002, “ Land-Use and Land-Cover Change: Decadal-Scale Dynamics of Carbon Storage Patterns as a Function of Land Ownership in the Southeastern Lower Coastal Plain Region of the U.S.,” Second annual progress report April 10. Washington DC, USA. Brown, S., 1993, “Tropical forest and the global carbon cycle: the need for sustainable and-use patterns,” Agriculture, Ecosystems and Environment. 46: 3144. Brown, S., Lugo, A. E., 1990, “Tropical secondary forests,” Journal of Tropical Ecology, 6: 1-32. Campbell, J.B., 2002, Introduction to Remote Sensing: New York, The Guilford Press. 170

PAGE 187

171 Collins, J. B. Woodcock, C. E., 1996, “An assessment of several linear change detection techniques for mapping forest mortality using multitemporal Landsat TM data,” Remote Sensing of Environment, 56: 66. Congalton, R.G., Green, K., 1998, Assessing the Accuracy of Remotely Sensed Data:Principles and Practices, Boca Raton, Florida, Lewis Publishers. Coppin, P.R.,. Bauer, M.E., 1994, “Processing of multitemporal Landsat TM imagery to optimize extraction of forest cover change features,” IEEE Transactions on Geosciences and Remote Sensing, 32 (4): 918-927. Crist, E.P. Cicone, R.C., 1984, “Application of the tasseled cap concept to simulated thematic mapper data,” Photogrammetric Engineering and Remote Sensing. 50(3): 343-352. Curran, P.J., Foody, G.M., Lucas, R.M., Honzak, M., Grace, J., 1997, “The carbon balance of tropical forests,” In: Scaling-up: From Cell to Landscape (Eds: Van Gardigen, P., Foody, G.M. and Curran, P.J.), Cambridge University Press, Cambridge. 201-227. Dikshit, O., 1996, "Textural classification for ecological research using TM images.” International Journal of Remote Sensing, 17(5):887-915. Duggin, M. J., Robinove, C. J.,1990, “Assumptions implicit in remote sensing data acquisition and analysis,” International Journal of Remote Sensing, 11:1669. ERDAS Field Guide, 1999, ERDAS Inc., Atlanta, U.S.A. Fearnside, P. M., 1982, “Deforestation in the Brazilian Amazon: how fast is it occurring?” Interciencia 7 (2): 82-85. Fearnside, P.M., 1996, “Amazonian deforestation and global warming: carbon stocks in vegetation replacing Brazil’s Amazon forest,” Forest Ecology and Management, 80: 21-34. Foody, M. G., 2002, “ Status of land cover classification accuracy assessment,” Remote Sensing of Environment, 80:185-201. Genc, L., Smith S. E., Binford M., Dewitt B. Gholz H.,June 2002, “Landsat TM-based forest cover change case study.” Proceeding of the 3 rd International Remote Sensing of Urban Areas Conference, Istanbul, Turkey p.22, 11-13 June 2002. Gholz, H., Clark, K., Cropper, W., 1997, “A comparison of measured and modeled net carbon fluxes for forest ecosystems in northern Florida,” North American Forest Ecology Workshop, Raleigh, NC. Gibson, P.J., Power, C. H., 2000, Introductory remote sensing: digital image processing and applications., London, Routledge.

PAGE 188

172 Gonzalez, R., Woods, R., 1993, Digital Image Processing. Reading, Mass: Addison-Wesley. Hall, F. G., Strebel, D. E., Nickeson, J. E , Goetz, S. J. (1991), “Radiometric rectification: toward a common radiometric response among multi-date, multi-sensor images,” Remote Sensing of Environment, 53:11. Haralick, R. M., Shanmugam, K., Dinstein, I., 1973, "Texture features for image Classification,” IEEE Transactions on Systems, Man, and Cybernetics, 3(6):610-621. Holopainen, M., Wang, G., 1998, “The calibration of digitized aerial photographs for forest stratificatio,” International Journal of Remote Sensing, 19(4):677-696. Horan, J. J., Schwartz, D. S., Love, J. D., 1974, “Partial performance degradation of a remote sensor in a space environment and some possible causes,” Applied. Optic. 13:1230. Houghton, R A., Skole, D L., Nobre, C A., Hackler, J L., Lawrence, K T., Chomentowski, W.H., 2000, “Annual fluxes of carbon from deforestation and regrowth in the Brazilian Amazon,” Nature, 430:301. Hunter, E. L., Power, C. H., 2002, “An assessment of two classification methods for mapping Thames Estuary inertial habitat using CASI data,” International Journal of Remote Sensing, 23: 2989-3008. Jensen, J.R., 1996, Introductory Digital Image Processing: A Remote Sensing Perspective, New Jersey, Upper Saddle River: Prentice-Hall. Jensen, J.R., 2000, Remote Sensing of The Environment, An Earth Resource Perspective. Upper Saddle River, Prentice Hall. Kauth, R.J.and Thomas, G.S.,1976, “The Tasseled Cap A graphic description of the spectral-temporal development of agricultural crops as seen by Landsat,”. Proc. Symposium on Machine Processing of Remotely Sensed Data, 4B-41-4B-50, Purdue University, Indiana. Kauth, R. J, 1979, “Feature extraction applied to agricultural crops as seen by Landsat,” Proc Technical Sessions, The LACIE Symp, 2: 705-721, NASA Johnson Space Center, 23-26 October, Houston, Texas. Kimes, D.S., Nelson, R.F., Skole, D.L., Salas, W.A., 1998, “Accuracies in mapping secondary tropical forest age from sequential satellite imagery,” Remote Sensing of Environment, 65:112-120. King, R. B., 2002, Land cover mapping principles: a retun to interpretation fundamentals International Journal of Remote Sensing, 23(18): 3525-3545.

PAGE 189

173 Knick, S.T., Rotenberry, J.T., Zarriello, T.J., 1997, “Supervised classification of LandsatTematic Mapper imagery in a semi-arid rangeland by nonparametric Discriminate analysis,” Photogrammetric Engineering and Remote Sensing, 63(1): 79-86. Kummer, D. M., Turner II, B. L., 1994, “The human causes of deforestation in southeast Asia,” Bioscience, 44:323328. Kushwaha, S. P. S., Kuntz, S., Oesten, G., 1994, “Applications of image texture in forest classification,” International Journal of Remote Sensing, 15(11):2273-2284. Lillesand, T. M., Kiefer, R.W., 2000, Remote Sensing and Image Interpretation 4 th Edition. New York, New York, John Wiley & Sons. Malila, A. W., Metzler D. M., Rice, P. D., Crist, E.P., 1985, “Charaterization of LATSAT-4 MSS and TM Digital Image Data,” IEEE 178-191. Markham, B. L., BarkerJ. L., “Special Landsat Image Data Quality Analysis (LIDQA) Issue,” , eds., Photogrammetric Engineering & Remote Sensing, 51(9): 1245-1493. Mather P., 1999, Computer Processing of Remotely Sensed Images. New York, John Wiley & Sons, Inc. Metzler, M. D., Malila, W.A.,1985, “Characterization and comparison of Landsat-4 and Landsat-5 thematic mapper data,” Photogrammetric Engineering and Remote Sensing, 51:1315. Meyer, W. B., 1995, “Past and present land use and land cover in the USA,” Consequences, 1(1): 24-33. Ojima, D.S., Galvin, K.A.,. Turner II, B.L., 1994, “The global impact of land use change, ” Bio-Science 44(5):300-304. Olsson, H., 1995, “Radiometric calibration of thematic mapper data for forest change detection,” International Journal of Remote Sensing, 16:81-96. Ott, L., longnecker, M., 2001, Introduction to statistical methods and data analysis. Pacific Grove, California, Thomson Learning Duxbury. Peterson, D. L., Spanner, M. A., Running, S. W., Teuber, K. B. 1987, ”Relationships of Thematic Matter Simulator Data to Leaf Area Index of Temporate Coniferous Forests,” Remote Sensing of the Environment. 22:323-341. Pearlstine. G. L., 2000, Discrimination of an Invasive Plant, Schinus terebinthifolius, from Aerial Digital Imagery. Dissertation University of Florida. Rees, G., 2001, Physical principles of remote sensing. Cambridge, U.K. ; New York, NY, Cambridge University Press.

PAGE 190

174 Richter, R., 1997, “Correction of atmospheric and topographic effects for high spatial resolution satellite imagery,” International Journal of Remote Sensing, 18:1099-1111. Richards, J.A., Jia, X., 1999. Digital Image Processing, New York, New York, Springer-Verlag. Rindfuss, R. R., Stern. C. P., 1998, “Linking remote sensing and social science: The need and the challenges”. In People and Pixel , Eds. Liverman D, et, al., Washington, DC: National Academy Press. pp. 1. Riou, R., Seyler, F., 1997. “Texture analysis of tropical rain forest infrared satelliteimages,” Photogrammetric Engineering and Remote Sensing, 63(5): 515-521. Rosenfield, G. H., Fitzpatrick-Lins, K., 1986, “A coefficient of agreement as a measure of thematic classification accuracy,” Photogrammetric Engineering and Remote Sensing, 52: 223. Rouse, J.W., Haas, R.H., Schell, J.A. Deering, D.W., 1973. “Monitoring vegetation systems in the great plains with ERTS,” Third ERTS Symposium, NASA SP-351, 1(1): 309-317. Ryherd, S., Woodcock, C., 1996, “Combining spectral and texture data in thesegmentation of remotely sensed images,” Photogrammetric Engineering and Remote Sensing, 62(2):181-194. Seyler, F.; Chaplot, V., Muller ; F., Cerri, C. E. P., Bernoux ,M., Ballester, V., Feller ; C., Cerri, C. C. C., 2002, “Pasture mapping by classification of Landsat TM images. Analysis of the spectral behaviour of the pasture class in a real medium-scale environment: the case of the Piracicaba Catchment (12 400 km 2 , Brazil),” International Journal of Remote Sensing, 23 (23):4985. Singh, A., 1986. “Change detection in the tropical forest environment of northeastern India using Landsat,” In Remote Sensing and Tropical Land Management (M. J. Eden and J. T. Parry, Eds.), Wiley, London, pp. 237. Singh, A., Harrison, A., 1985, “Standardized principal components,” International Journal of Remote Sensing, 6:883-896. Skole, D. L., 1994, “Data on global land cover change: acquisition, assessment and analysis,” In Changes in L and Use and L and Cover: a global perspective, edited by W. B. Meyer and B. L. Turner, Cambridge University Press. Skole, D. L., Chomentowski, W. H., Salas, W. A., Nobre, A. D., 1994, “Physical and human dimensions of deforestation in Amazonia,” Bioscience, 44:314-322.

PAGE 191

175 Stehman, S. V., Czaplewski, R. L.,1998, “Design and analysis for thematic map accuracy assessment: fundamental principles,” Remote Sensing of Environment, 64:331– 344. Steininger,M.K., 1996, “Tropical secondary forest regrowth in the Amazon: age, area and change estimation with Thematic Mapper data,” International Journal of Remote Sensing, 17:9-27. Stein,A., Meer ,F., Gorte, B., 1999, Spatial Statistic for Remote Sensing. Volume 1 Dordecht, Netherlands, Kluwer Academic. Swain, P.H., Davis, S.M.,1978, Remote Sensing: The Quantitative Approach. New York, New York, McGraw-Hill. Thome, K., Markham, B., Barker, J., Slater, P., Biggar, S.,1997, “Radiometric calibration of Landsat,” Photogrammetric Engineering and Remote Sensing.63: 853. Tokola, T., Lofman, S., Erkkila, A., 1999, “Relative calibration of multitemporal Landsat data for forest cover change detection,” Remote Sensing of the Environment, 68:1-11. Tomaita, F., Tsuji, S.,1990, Computer Analysis of Visual Textures, Norwell, MA, Kluwer Academic. Tso, B. Mather, P., 2001, Classification methods for remote sensed data. New York, New York, Taylor and Francis. Tucker, C. J., 1979, “Red and photographic infrared linear combinations for monitoring vegetation,” Remote Sensing of Environment, 8: 127-150. Turner II, B. L., 1989 “The human causes of global environmental change,” In:Global change and our common future: Papers from a forum, eds. R.S. Defries and T. Malone,. Washington, DC: National Academy Press. Turner II, B. L., 1991, “Thoughts on linking the physical and the human sciences in the study of global environmental change”. Research and Exploration 7: 133-135. USGS, U.S. Geological Survey, 2002, Digital Orthophoto Quadrangles Source: http://spatialnews.geocomm.com/education/tutorials/doq1/usgs_doq.html#doq16 . Vani, K., Sshanmugavel, S., Marruthachalam, M., 2001, “Fusion of IRS-LISS III and pan images using different resolution ratio,” The 22nd Asian Conference on Remote Sensing Singapore, 5-9 November.

PAGE 192

176 Watson, N., Wilcock, D., 2001, “Preclassification as an aid to the improvement of thematic and spatial sccuracy in land cover maps derived from satellite imagery,” Remote Sensing of the Environment, 75:267. Williams, D.L, Irons J.R, Makham B.L, Nelson R.F, Toll D.L, Latty R.S, Stauffer M.L.,1984 “A statistical evaluation of the advantages of LANDSAT thematic mapper data in comparison to multispectral scanner data,” IEEE Transactions on Geoscience and Remote Sensing, 22(3):294-302. Wolf, P.R. Dewitt, B.A., 2000, Elements of photogrammetry with applications in GIS. Boston , McGraw Hill.

PAGE 193

BIOGRAPHICAL SKETCH Levent Genc attended the University of Uludag in Bursa, Turkey, where he received a Bachelor of Science degree in agricultural engineering in 1993. Levent worked as research and teaching assistant for two years during the his graduate study in the University of Uludag. In 1999, he received a Master of Engineering degree in the Agricultural and Biological Engineering Department at the University of Florida and he entered to Ph.D program at same year. In 2002, he received a Master of Science degree in the Civil and Coastal Engineering Department, Geomatics program, at the University of Florida. Levent received his Ph.D. degree in 2003 after defending his dissertation titled “Comparison of Landsat MSS and TM Imagery for Long term Land Cover Change Assessment.” Upon graduation, Levent plans to teach and do research relating to remote sensing in the area of environmental development in Turkey. 177