<%BANNER%>

Upscaling of soil properties across landscapes of south Florida

Permanent Link: http://ufdc.ufl.edu/UFE0044955/00001

Material Information

Title: Upscaling of soil properties across landscapes of south Florida
Physical Description: 1 online resource (223 p.)
Language: english
Creator: Kim, Jongsung
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2012

Subjects

Subjects / Keywords: digitalsoilmapping -- remotesensing -- scaling
Soil and Water Science -- Dissertations, Academic -- UF
Genre: Soil and Water Science thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Soil nutrients stored in wetland soils are critical toassess the effectiveness of restoration efforts, yet it is challenging toaccurately derive soil heterogeneity. The incorporation of remote sensing (RS)data into digital soil models has shown success to improve soil predictions.However, the effects of multi-resolution imagery on modeling of soil propertiesand classes in aquatic ecosystems are still poorly understood. The objectivesof this study were to develop prediction models for soil properties including totalphosphorus (TP), total nitrogen (TN), total carbon (TC), and soil series utilizingRS images and environmental ancillary data and elucidate the effect ofdifferent spatial resolutions of RS images on inferential modeling of thosesoil properties. The study was conducted in subtropical wetlands: WaterConservation Area-2A (480 km2) and 3A North (722 km2),the Florida Everglades, U.S. The spectral data and derived indices from remotesensing images, which have different spatial resolutions, included: ModerateResolution Imaging Spectroradiometer (MODIS, 250 m), Landsat Enhanced ThermaticMapper Plus (ETM+, 30 m), and Satellite Pour l’Observation de la Terre (SPOT, 10m). Classification Tree, Block kriging, and Random Forest were employed topredict soil series and biogeochemical properties, respectively, using RS imagederived spectral input variables, environmental ancillary data, and soilobservations. Most of the models could explain > 60% of the spatialvariability of the soil properties. This study provided ample evidence thatRS-informed prediction models can successfully infer on multiple biophysicalproperties in soils and soil classes in aquatic ecosystems. The spatialdistributions of soil properties, major controlling environmental factors foreach of the soil properties, and absolute storage amounts for soil TP, TN, andTC were assessed. Interestingly, there was no noticeable distinction amongdifferent spatial resolutions of RS images to develop prediction models forsoil properties. Results provided a better understanding of how fine and coarsegrain resolutions of RS images impact soil modeling, model transferability, andscaling. Also this study showed the potential use of visible-near infrearedspectroscopy for wetland soil property estimations.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Jongsung Kim.
Thesis: Thesis (Ph.D.)--University of Florida, 2012.
Local: Adviser: Grunwald, Sabine.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2014-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2012
System ID: UFE0044955:00001

Permanent Link: http://ufdc.ufl.edu/UFE0044955/00001

Material Information

Title: Upscaling of soil properties across landscapes of south Florida
Physical Description: 1 online resource (223 p.)
Language: english
Creator: Kim, Jongsung
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2012

Subjects

Subjects / Keywords: digitalsoilmapping -- remotesensing -- scaling
Soil and Water Science -- Dissertations, Academic -- UF
Genre: Soil and Water Science thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Soil nutrients stored in wetland soils are critical toassess the effectiveness of restoration efforts, yet it is challenging toaccurately derive soil heterogeneity. The incorporation of remote sensing (RS)data into digital soil models has shown success to improve soil predictions.However, the effects of multi-resolution imagery on modeling of soil propertiesand classes in aquatic ecosystems are still poorly understood. The objectivesof this study were to develop prediction models for soil properties including totalphosphorus (TP), total nitrogen (TN), total carbon (TC), and soil series utilizingRS images and environmental ancillary data and elucidate the effect ofdifferent spatial resolutions of RS images on inferential modeling of thosesoil properties. The study was conducted in subtropical wetlands: WaterConservation Area-2A (480 km2) and 3A North (722 km2),the Florida Everglades, U.S. The spectral data and derived indices from remotesensing images, which have different spatial resolutions, included: ModerateResolution Imaging Spectroradiometer (MODIS, 250 m), Landsat Enhanced ThermaticMapper Plus (ETM+, 30 m), and Satellite Pour l’Observation de la Terre (SPOT, 10m). Classification Tree, Block kriging, and Random Forest were employed topredict soil series and biogeochemical properties, respectively, using RS imagederived spectral input variables, environmental ancillary data, and soilobservations. Most of the models could explain > 60% of the spatialvariability of the soil properties. This study provided ample evidence thatRS-informed prediction models can successfully infer on multiple biophysicalproperties in soils and soil classes in aquatic ecosystems. The spatialdistributions of soil properties, major controlling environmental factors foreach of the soil properties, and absolute storage amounts for soil TP, TN, andTC were assessed. Interestingly, there was no noticeable distinction amongdifferent spatial resolutions of RS images to develop prediction models forsoil properties. Results provided a better understanding of how fine and coarsegrain resolutions of RS images impact soil modeling, model transferability, andscaling. Also this study showed the potential use of visible-near infrearedspectroscopy for wetland soil property estimations.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Jongsung Kim.
Thesis: Thesis (Ph.D.)--University of Florida, 2012.
Local: Adviser: Grunwald, Sabine.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2014-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2012
System ID: UFE0044955:00001


This item has the following downloads:


Full Text

PAGE 1

1 UPSCALING OF SOIL PROPERTIES ACROSS LANDSCAPES OF SOUTH FLORIDA By JONGSUNG KIM A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2012

PAGE 2

2 2012 Jongsung Kim

PAGE 3

3 To my parents and parents in law, sisters, and brother, they supported me all the times To Jungwoo, he has been there all the time in the good and the bad times To Jaden, my joy

PAGE 4

4 ACKNOWLEDGMENTS I would like to take this opportunity to thank my parents for their support and guidance. I also thank the University of Florida, the College of Agriculutral and Life Sciences for providing the resources and funding to conduct this study. I am grateful to all my committee members because of their commitment and patience. Especailly, my advisor, Sabine Grunwald, supported me all the time. She always provided much valuable advice and comments for my research. She always encouraged me and gave positive responses to all my questions. She was not simply an academic advisor but the best mentor of my life. D r.Rosanna G. Rivero from the University of Georgia gave me lots of valuable comments about remote sensing; Dr.Todd Z. Osborne from the Soil and Water Science Department helped me a lot with the Everglades field sampling; Dr.Tim Martin and Dr.Scot Smith fro m the School of Forest Resources and Conservation w ere always nice to me and gave ideas for the dissertation. I thank my working colleagues from the Geographic Information Systems Laboratory Pasicha Chaikaew (Ploy), Xiong Xiong, Baijing Cao (Betty), Wade R oss, Gustavo Vasques, Jinseok Hong, Brent Myers, Nichola Knox, Deoyani Sark h ot, Hoyoung Kwon, Julius Adewopo, Chris Clingensmith, and Brandon Hoover for their ideas, their support, and friendship. I give special thanks to my husband Jungwoo Lee, whose love, friendship, enthusiasm, guidance and support were essential in everything that I have accomplished. Funding for this study was provided from the project Remotesensing Supported Digital Soil Mapping in South Florida by the Cooperative Ecosystem Studies Unit Natural Resocureces Conservation Service. The SPOT images were donated by Planet

PAGE 5

5 Action, a nonprofit ASTRIUM GEO initiative. Also I couldnt finish this research without the internal funding from the GIS research laboratory

PAGE 6

6 TABLE OF CONTENTS page ACKNOWLEDGMENTS .................................................................................................. 4 LIST OF TABLES ............................................................................................................ 9 LIST OF FIGURES ........................................................................................................ 11 ABSTRACT ................................................................................................................... 14 CHAPTER 1 INTRODUCTION .................................................................................................... 16 SoilLands cape Modeling and Geostatistics ........................................................... 16 R emote Sensing in Digital Soil Mapping ................................................................. 18 Wetlands in Florida ................................................................................................. 21 Rationale an d Significance ..................................................................................... 23 2 MULTI SCALE MODELING OF SOIL SERIES USING REMOTE SESNING IN A WETLAND ECOSYSTEM ....................................................................................... 29 Overview ................................................................................................................. 29 Materials and Methods ............................................................................................ 32 Study Area ........................................................................................................ 32 Field Sampling .................................................................................................. 33 Soil Series Description ..................................................................................... 34 Remote S ensing Imagery and Spectral Data Analyses .................................... 34 Geospatial Environmental Ancillary Data ......................................................... 37 Classification Tree Method ............................................................................... 39 Uncertainty Assessment and Verification of Model Output ............................... 41 Results and Discussion ........................................................................................... 43 Distribution of Soil Series within WCA 2A ........................................................ 43 Soil Series Prediction Models Using Spectral Data .......................................... 43 Model I ....................................................................................................... 43 Model II ...................................................................................................... 46 Predictor Variables Providing Inferences on Soil Series .................................. 47 Model I ....................................................................................................... 47 Model II ...................................................................................................... 50 Conclusions ............................................................................................................ 51 Summary ................................................................................................................ 52

PAGE 7

7 3 SOIL PHOSPHORUS AND NITROGEN PREDICTIONS ACROSS SPATIAL ESCALATING SCALES IN AN AQUATIC ECOSYSTEM ....................................... 67 Overview ................................................................................................................. 67 Methods and Materials ............................................................................................ 70 Study Area ........................................................................................................ 70 Field Sampling and Soil Property Measurement .............................................. 71 Spectral Indices Derived from Remote Sensing Imagery ................................. 72 Ancillary Environmental Predictor Data ............................................................ 74 Data Analyses .................................................................................................. 75 Verification of Model Output ............................................................................. 77 Results and Discussion ........................................................................................... 79 Spatial Distribution of Soil Properties: Concentrations and Stocks ................... 79 Soil Properties Prediction Models: Univariate and Multivariate Methods .......... 81 Spatial predictions ...................................................................................... 81 Variable importance of Random Forest models for TP predictions ............ 83 Variable importance of Random Forest models for TN predictions ............ 85 Spatial Resolution Effects of Remote Sensing Images on Soil TP and TN Predictions: Fine and Coarse Resolutions .................................................... 87 Conclusions ............................................................................................................ 91 Summary ................................................................................................................ 93 4 EVALUATING TOTAL CARBON STOCKS IN A SUBTROPICAL WETLAND ...... 112 Overview ............................................................................................................... 112 Methods and Materials .......................................................................................... 114 Study Area ...................................................................................................... 114 Field Sampling and Soil Carbon Measurement .............................................. 115 Environmental Predictor Variables ................................................................. 116 Random Forest ............................................................................................... 118 Results and Discussion ......................................................................................... 119 Spatial Distribution of Total Carbon Stocks .................................................... 119 Soil Total Carbon Stock Predictions using Random Forest ............................ 121 Total Carbon Stocks in Water Conservation Area2A .................................... 125 Conclusions .......................................................................................................... 128 Summary .............................................................................................................. 129 5 EVALUATING MODEL TRANSFERABILITY AND SCALING IN ADJACENT SUBTROPICAL WETLANDS: WATER CONSERVATION AREA 2A AND 3A NORTH, THE EVERGLADES, FLORIDA ............................................................. 141 Overview ............................................................................................................... 141 Methods and Materials .......................................................................................... 143 Study Area ...................................................................................................... 143 Field Sampling and Soil Property Measurements ........................................... 144 Predictors for the Model Development: Spectral and Environmental Variables ..................................................................................................... 144

PAGE 8

8 Soil Prediction Model Development and Transferability Assessment ............. 146 Results and Discussion ......................................................................................... 148 Descript ive Statistics of Soil Properties in Water Conservation Area3A North and Comparison with Water Conservation Area2A .......................... 148 Soil Prope rty Prediction Models using Random Forest in Water Conservation Area3AN .............................................................................. 150 Spatial predictions .................................................................................... 150 Variable imp ortance ................................................................................. 152 Transferability and Scaling Effects of Soil Prediction Models ......................... 154 Application of soil prediction models to other extents .............................. 154 Model transferability and scaling .............................................................. 157 Conclusions .......................................................................................................... 160 Summary .............................................................................................................. 161 6 VISIBLE NEAR INFRARED SPECTROSCOPY FOR SOIL PROPERTY PREDICTIONS OF HIGHLY ORGANIC WETLAND SOILS ................................. 178 Overview ............................................................................................................... 178 Methods and Materials .......................................................................................... 180 Study Area ...................................................................................................... 180 Field Sampling and Laboratory Measurement ................................................ 181 Spectral Scanning and Data Processing ........................................................ 181 Model Developments and Assessments ........................................................ 183 Results and Discussion ......................................................................................... 185 Descriptive Statistics ...................................................................................... 185 Visible Near Infrared Reflectance Spectroscopy Models for Soil Properties in Water Conservation Area2A ................................................................... 186 Conclusions .......................................................................................................... 190 7 SUMMARY AND SYNTHESIS .............................................................................. 197 LIST OF REFERENCES ............................................................................................. 204 BIOGRAPHICAL SKETCH .......................................................................................... 223

PAGE 9

9 LIST OF TABLES Table page 1 1 An overview of digital soil mapping studies that utilized remote sensing data. ... 25 2 1 Summary of observed soil series in Water Conservation Area2A. .................... 54 2 2 Summary of remote sensing spectral indices. .................................................... 55 2 3 Summary of environmental variables employed to build classification tree models to predict soil series. .............................................................................. 56 2 4 Confusion error matrices for single tree soil series prediction models derived from SPOT, Landsat ETM+, and MODIS images, respectively, plus ancillary environmental data, but without bedrock depth .................................................. 58 2 5 Summary of single tree models to predict soil series with and without bedrock depth in the Water Conservation Area2A, Everglades, Florida. .......... 59 2 6 Confusion error matrices for single tree soil series prediction models derived from SPOT, Landsat ETM+, and MODIS images, respectively, plus ancillary environmental data including bedrock depth. ..................................................... 60 2 7 Variable importance of single tree models to predict soil series with and without bedrock depth in the Water Conservation Area2A, Everglades, Florida. ............................................................................................................... 61 3 1 Summary of remote sensing spectral indices. .................................................... 95 3 2 Summary of environmental variables employed to build models to predict soil total phosphorus and nitrogen. ........................................................................... 96 3 3 Descriptive statistics for soil properties in the topsoil observed in Water Conservation Area2A. ....................................................................................... 97 3 4 Summary of model performance assessment for soil total phosphorus (TP) and total nitrogen (TN) concentrations. .............................................................. 98 3 5 Summary of semivariogram parameters of Block Kriging for soil total phosphorus (TP) and total nitrogen (TN) in Water Conservation Area2A. ......... 99 3 6 Summary of the relative importance of predictor variables in Random Forest models predicting soil total phosphorus in Water Conservation Area2A. ........ 100 3 7 Summary of the relative importance of predictor variables in Random Forest models predicting soil total nitrogen in Water Conservation Area2A. .............. 101 3 8 Summary of estimated nutrients stocks in Water Conservation Area2A. ........ 102

PAGE 10

10 3 9 Estimated Shannon entropy ( H) of the predicted soil properties derived using Random Forest with SPOT, Landsat ETM+, and MODIS images derived spectral input variables. ....................................................................... 103 4 1 Descriptive statistics for soil total carbon observed in Water Conservation Area 2A. ........................................................................................................... 131 4 2 Comparison of soil total carbon concentrations in Water Conservation Area2A. .................................................................................................................... 132 4 3 Summary of model performance assessment for soil total carbon stocks. ....... 133 4 4 Predicted total soil carbon stocks using Random Forest in Water Conservation Area2A. ..................................................................................... 134 4 5 Comparison of soil carbon stocks. .................................................................... 135 5 1 Descriptive statistics of measured soil properties in Water Conservation Area 2A and 3A North. ..................................................................................... 164 5 2 Statistical assessments of crossvalidation in Water Conservation Area2A and 3A North and transferability of soil prediction models developed in Water Conservation Areas. ......................................................................................... 165 5 3 Statistical assessments of upscaling of soil prediction models developed in Water Conservation Area2A and 3A North to the combined area. .................. 166 5 4 Statistical assessments of downscaling of soil prediction models developed in the combined area to Water Conservation Area2A and 3A North. .............. 167 5 5 Descriptive statistics of envi ronmental predictor variables in Water Conservation Area2A and 3A North. ............................................................... 168 6 1 Descriptive statistics of measured soil properties in Water Conservation Area 2A. ........................................................................................................... 191 6 2 Summary statistics of the partial least squares regression models for each soil property in Water Conservation Area2A. .................................................. 192

PAGE 11

11 LIST OF FIGURES Figure page 2 1 Distribution of soil series at sampling locations within the Water Conservation Area 2A, Everglades, Florida, U.S. ..................................................................... 62 2 2 Single tree classification model to predict soil series using the SPOT image, but without bedrock depth as predictor, in the Water Conservation Area2A.. ... 63 2 3 Single tree classification model to predict soil series using the Landsat ETM+ image, but without bedrock depth as predictor, in the Water Conservation Area 2A.. ............................................................................................................ 64 2 4 Single tree classification model to predict soil series using the MODIS image, but without bedrock depth as predictor, in the Water Conservation Area2A.. ... 65 2 5 Prediction maps of soil series based on single tree model without geographic coordinates using A) SPOT, B) Landsat ETM+, and C) MODIS image derived spectral input variables with geospatial environmental ancillary data. ................ 66 3 1 Spatial distribution of measured soil concentrations of total phosphorus, TP (left) and total nitrogen, TN (right) within the Water Conservation Area2A. ..... 104 3 2 Prediction maps of soil nutrient concentrations and stocks in th e topsoil using Random Forest with SPOT image derived spectral input variables and environmental predictor variables. .................................................................... 105 3 3 Scatte r plots showing the relationship between observed and predicted soil properties using block kriging (left), Random Forest with SPOT, Landsat ETM+, and MODIS (right) image derived input variables combined with environmental predictor variables. .................................................................... 106 3 4 Prediction maps of soil total phosphorus using Block Kriging and Random Forest models. .................................................................................................. 107 3 5 Prediction maps of soil total nitrogen using Block Kriging and Random Forest models. ............................................................................................................. 108 3 6 Annual total phosphorus (TP) loads for inflow and outflow, and annual flow weighted TP concentrations for inflow and outflow in WCA 2A during the period from May, 2003 to April, 2010. ............................................................... 109 3 7 Variation comparison of predicted soil total phosphorus concentrations. ......... 110 3 8 Variation comparison of predicted soil total nitrogen concentrations. ............... 111

PAGE 12

1 2 4 1 Low water table position and limited accessibility of sampling point by airboats of southern part of Water Conservation Area2A during February, 2009 sampling event. ....................................................................................... 136 4 2 Spatial distribution of measured soil total carbon by depth within the Water Conservation Area2A. ..................................................................................... 137 4 3 Total carbon concentration of the 0 10 cm and the 10 20 cm depth in 2003 and 2009.. ................................................................................................ 138 4 4 Prediction maps of soil total carbon stocks in the 0 10 cm and the 10 20 cm using Random Forest. ................................................................................ 139 4 5 Variable importance of predictor variables in Random Forest models predicting soil total carbon stocks in Water Conservation Area2A. ................. 140 5 1 Location of the Water Conservation Area2A and 3A North within the Everglades, Florida, U.S. .................................................................................. 170 5 2 Spatial distribution of measured soil concentrations of total phosphorus, total nitrogen, and total carbon within the Water Conservation Area3A North. ....... 171 5 3 Total phosphorus (TP) loads of outflow from Water Conservation Area2A and TP loads from Water Conservation Area2A into Water Conservation Area 3A north during the period from May, 2003 to April, 2010. ...................... 172 5 4 Prediction maps of soil properties using Random Forest model with the SPOT and the MODIS image derived spectral input variables combined with environmental predictor variables. .................................................................... 173 5 5 Variable importance of predictor variables in Random Forest models with the SPOT and the MODIS images ......................................................................... 174 5 6 Comparison of prediction maps of soil total phosphorus using Random Forest models with the SPOT and MODIS spectral data in Water Conservation Area3A North. ............................................................................ 175 5 7 Comp arison of prediction maps of soil total phosphorus using Random Forest models with the SPOT and MODIS spectral data in Water Conservation Area2A. ..................................................................................... 176 5 8 Scatter plots showing the relationship between observed and predicted soil total phosphorus using Random Forest models with the SPOT image derived input variables combined with environmental predictor variables. .................... 177 6 1 Raw visible near infrared reflectance spectra of soil samples from Water Conservation Area2A. ..................................................................................... 193

PAGE 13

13 6 2 Scatter plots showing the relationship between observed and predicted soil properties in validation set using partial least squares regression models. ...... 194 6 3 Regression coefficients between soil properties and spectral reflectance used in the partial least squares regression models. ........................................ 196

PAGE 14

14 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 UPSCALING OF SOIL PROPERTIES ACROSS LANDSCAPES OF SOUTH FLORIDA By Jongsung Kim December 2012 Chair: Sabine Grunwald Major: Soil and Water Science Soil nutrients stored in wetland soils are critical to assess the effecti veness of restoration efforts, yet it is challenging to accurately derive soil heterogeneity. The incorporation of remote sensing (RS) data into digital soil models has shown success to improve soil predictions. However, the effects of multi resolution ima gery on modeling of soil properties and classes in aquatic ecosystems are still poorly understood. The objectives of this study were to develop prediction models for soil properties including total phosphorus (TP) total nitrogen (TN) total carbon (TC) and soil series utilizing RS images and environmental ancillary data and elucidate the effect of different spatial resolutions of RS images on inferential modeling of those soil properties. The study was conducted in subtropical wetlands: Water Conservation Area 2A (480 km2) and 3A North ( 722 km2) the Florida Everglades, U.S. The spectral data and derived indices from remote sensing images, which have different spatial resolutions, included: Moderate Resolution Imaging Spectroradiometer ( MODIS 250 m), Lan dsat Enhanced Thermatic Mapper Plus ( ETM+ 30 m), and Satellite Pour lObservation de la Terre ( SPOT 10 m). Classification Tree, Block kriging and Random Forest were employed to

PAGE 15

15 predict soil series and biogeochemical properties, respectively, using RS im age derived spectral input variables, environmental ancillary data, and soil observations. Most of the models could explain > 60% of the spatial variability of the soil properties. This study provided ample evidence that RS informed prediction models can s uccessfully infer on multiple biophysical properties in soils and soil classes in aquatic ecosystems. The spatial distributions of soil properties, major controlling environmental factors for each of the soil properties, and absolute storage amounts for soil TP, TN, and TC were assessed. Interestingly, there was no noticeable distinction among different spatial resolutions of RS images to develop prediction models for soil properties. Results provided a better understanding of how fine and coarse grain resolutions of RS images impact soil modeling, model transferability, and scaling. Also this study showed the potential use of visible near infreared spectroscopy for wetland soil property estimations.

PAGE 16

16 CHAPTER 1 INTRODUCTION Soil is a heterogeneous natural entity formed by the continuous interaction of climate, organisms, relief, and parent materials over time (Arnold, 2006). It is composed of minerals and organic compounds as well as living organisms that interact with the lithosphere (rocks), hydrosphere (water), atmosphere (air), and biosphere (living things). Soil serves as a natural medi um for plant growth and crop production. It sustains biotic activity; it transfers gases, water, stores nutrients and other elements. It also serves as modifier of the atmosphere, regulators of water supplies, recycler of raw materials, and engineering medium (Brady and Weil, 2008). It is important to have accurate information about the soil when we consider its valuable functions and in order to conserve irreplaceable natural resources. Soil Landscape Modeling and Geostatistics Soils are dependent on other environmental properties, which are dynamically changing with time. Landscapes are characterized by a wide variation of the environmental properties, both spatially and temporally. Considering the complexity of the soil landscape continuum it is difficult to quantify the spatial and temporal variability and distribution of soil properties as well as relationships between soils and environmental landscape variables (Grunwald, 2006). However it is impossible to sample all soils exhaustively, i.e. at all locations, to characterize their variability because it is costly, labor and time intensive. Thus, soil sampling can only be done at sparse locations and models are used to estimate properties at unsampled locations. To understand, interpret, and predict soils continuously across a landscape, various soil landscape models have been developed, which are described below.

PAGE 17

17 Dokuchaev introduced the soil factor model in 1880 and Jenny (1941) popularized the factorial model of soil forming factors, which describes soil as a function of five soil forming factors; S = f ( cl o r p t ) (1 1 ) where S is soil property cl is environmental climate, o is organism s including humans r is relief p is parent material and t is time The SCORPAN model was formalized by McBratney et al. (2003). For empirical quantitative descriptions of relationships between soil and other spatially referenced factors, the SCORPAN model includes spatial position and soil itself as a soil forming factor; Sc or Sa [ x,y,~t] = f ( S [ x,y,~t], C [ x,y,~t ], O [ x,y,~t], R [ x,y,~t ], P [ x,y,~t], A [ x,y], N ) ( 1 2) w here Sc is soil class Sa is soil attributes S is soil ( other properties of the soil at a point ) C is climate O is organisms including humans R is relief P is parent material A is age (time factor) N is space x, y is x and y coordinates and t is time Almost all contemporary soil prediction models are rooted in Jennys CLORPT and SCORPAN conceptual models, which facilitate to explain relationships between soil and environmental factors. However, those models do not account for spatial autocorrelations o f soil and environmental factors. Geostatistics has been used widely to quantify spatial autocorrelation and predict variables at unsampled locations. Geostatistics provides a set of statistical tools for analyzing the space/time information of observati ons (Goovaerts, 1997) and it was introduced to soil science by Webster and Butler in 1976 (Webster and Oliver, 2007). The main objectives of using geostatistics in soil science are description and summary

PAGE 18

18 of spatial patterns (Goovaerts and Chiang, 1993; Cahn et al., 1994; Wang et al., 2002; Grunwald et al., 2008; Rossi et al., 2009), prediction (Trangmar et al., 1987; Goovaerts and Chiang, 1993; McGrath and Zhang 2003; Cerri et al., 2004; Rivero et al., 2007a ), and uncertainty assessment (Goovaerts, 1999; Saito and Goovaerts, 2002). It is essential to know the variation of several soil properties for sitespecific agricultural management and other uses (Kerry and Oliver, 2004) and there has been much research analyzing spatial variability of soil properties using geostatistics (Goovaerts and Chiang, 1993; Cahn et al., 1994; Wang et al., 2002; Grunwald et al., 2004; Rivero et al., 2007 a ; Grunwald et al., 2010). Statistical and geostatistical methods have been used widely to implement soil prediction models based on the CLORPT and SCORPAN concepts (Grunwald, 2006), which will be also used in this dissertation research. R emote Sensing in Digital Soil Mapping The development of computers and information technology lead soil scientists to incorporate remote sensi ng into soil prediction models to characterize various SCORPAN factors. The availability of aerial photographs in the late 1920s and remotely sensed imagery in the late 1960s, the advent of global positioning systems (GPS), geophysical soil mapping techni ques, and soil sensors had a positive impact on soil mapping (Grunwald and Lamsal, 2006). Remote sensing is a measurement of the electromagnetic energy reflected or emitted from objects on the Earths surface and the amount of radiation, for any given material, that is reflected at varying wavelength (Jensen, 2005). It means different materials have different reflectance characteristics. Therefore, remote sensing can play a role in the identification, inventory, and mapping of soils that are on the surface of the Earth, whereby the impact of soil grain size, organic matter, and water content on soil spectral reflectance are identified (Jensen,

PAGE 19

19 2000). A remote sensing instrument (e.g., satellite platform) collects information about an object within the instantaneous field of v iew (IFOV) of the sensor system. Usually the collected information is resampled within a fixed pixel size, the so called spatial resolution; a spatial resolution in this dissertation refers to the IFOV T he use of remote sensing techniques to predict soil properties was emphasized by McBratney et al. (2003), because they provide dense information grids across a landscape. In addition, remote sensing can provide information about the O factor in the SCORPAN model (Eq. ( 1 2 ) ), as well as other factors, which assists in predictions of Sa or Sc (Grunwald, 2009). Remote sensing supported mapping of soil properties has been suggested for regional mapping of soils at various spatial scales (Grunwald and Lamsal, 2006) Due to the fact that it supports rapid and inexpensive data collection over large areas, comprehensive research has been focused on predicting soil properties using geostatistical/hybrid methods at various scales and regions with a variety of remote sensing imagery. Peng et al. ( 2003) applied Landsat Thematic Mapper (TM) with 3060m spatial resolution, IKONOS (the Greek word for image) data with 14 m resolution to an agricultural area to predict patterns of soil drainage class variability. Sullivan et al. (2005) predicted total carbon (TC) and clay content in surface soil and found that Cokriging using IKONOS data reduced uncertainty in TC prediction. Chen et al. (2008) applied Advanced Thermal and Land Applications Sensor (ATLAS) to map soil organic carbon (SOC) of an agricultural area in Georgia, U.S. Vrieling et al. (2008) used different satellite imagery for different purposes in a savannah area in Brazil. Moderate Resolution Imaging Spectroradiometer (MODIS, 2501000m resolutions) was used to

PAGE 20

20 assess high erosion risk periods Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER, 1530m resolution) was used to find optimal image timing for erosion risk mapping and QuickBird (0.61 2.88 m resolution) was used to validate erosion risk map. Table 1 1 provides an overview of digital soil mapping studies using a variety of sensors including multispectral (e.g. Landsat, ASTER) and high resolution (e.g. IKONOS, Satellite pour lobservation de la terre SPOT, QuickBird, Indian remote sensing IRS). Various soil pr operties (e.g. TC, salinity, erosion, total phosphorus TP, total nitrogen TN, etc.) were predicted using remote sensing data in various regions and land uses. Several statistical (e.g. Regression, Principal Components Analysis PCA), geostatistical (e .g. Ordinary Point Kriging, Block Kriging), and hybrid (mixed) methods (e.g. PCA Kriging, Cokriging, and Regression Kriging) to relate field observations and spectral data were used in these studies (Table 1 1). Even though the accuracies vary from study t o study, the studies show that the accuracy of soil predictions could be improved by incorporating remote sensing data into models. Although remote sensing provide dense spectral grids over large areas research, gaps still exist to better elucidate on the effect of different spatial resolutions of satellite images on inferential modeling of soil properties. Remote sensing has been used to classify and identify wetlands and map large wetland ecosystem. Rivero et al. (2007 a ) applied Landsat Enhanced Them atic Mapper Plus (ETM+) and ASTER data to a subtropical wetland, Water Conservation Area (WCA) 2A which is located within the Everglades in south Florida. The spectral data and derived indices from remote sensors along with field observations were used to

PAGE 21

21 predict the spatial distribution of floc and soil TP. The authors used hybrid geostatistical prediction methods and confirmed that unique and complex landscape features of wetlands could be captured by remote sensing data. Zhao et al. (2009) used MODIS dat a to monitor vegetation community in an estuarine wetland located in Yangtze River Watershed, China. The authors used various vegetation indices from MODIS to investigate vegetation community and its physical conditions, and they found significant spatial and temporal correlations between vegetation community and the vegetation indices. Rivero et al. (2009) also investigated the relationship between vegetation indices from remote sensing data and floc and soil TP values in an aquatic ecosystem, WCA 2A. The authors found that remote sensing indices could capture subtle changes in the mix of chlorophyll a and carotenoids in vegetation that further relate to TP status in floc and soils. In particular, in wetlands and aquatic systems remote sensing based modelin g of soil properties has much value due to the difficulty of field sampling and limited site access. However, remote sensing based predictive modeling of soil properties is more abundant in upland systems than in aquatic and wetland systems (compare Table 1 1). More research is in need to enhance our understanding of remote sensing based digital soil mapping focusing on key biogeochemical properties that modulate ecosystem function, structure, and resilience. Base soil properties, such as erosion, SOC, TP and TN, have been predicted using remote sensing based modeling (Table 1 1), but other soil properties (e.g. metals), have not been investigated. This will be investigated in this research. Wetlands in Florida Florida receives an average of 1,346 mm of rain fall each year with almost 70 percent of the annual rainfall occur during summer (National Climate Data Center,

PAGE 22

22 2009). Because of seasonal variation of rainfall, high water table, and flat topographic condition wetlands are prominent in Floridas landscape. Wetlands cover about 25% of the state of Florida (Florida Fish and Wildlife Conservation Commission, 2003) and, amongst them, the Everglades is the largest subtropical wetland in the United States. Historically the Everglades was a very nutrient poor ec osystem (oligotrophy) characterized by low surfacewater concentrations of phosphorus (P) and other nutrients with over 90% of water inputs from rainfall (Chimney and Goforth, 2006). However, from the early 1900s, Florida has experienced rapid land use ch anges such as urbanization, and this rapid growth has resulted in a degradation of water resources, loss in environmental quality, and loss and change of ecosystem integrity and function in many wetland ecosystems. In addition, improved drainage for development of the Everglades Agriculture Area (EAA) has changed the nutrients level and ecosystems of downstream wetlands. The Everglades has been impacted by urban development along its boundaries and the construction of an extensive system of levees and canal s that have caused major hydrological changes in the area ( Noe et al., 2001). The encroachment of the urban corridor Miami Fort LauderdaleWest Palm Beach has been the most important anthropogenic intervention on the east boundary, while the EAA has been a source of nutrient loads that are being transported into the Everglades on the north (Rivero et al., 2007 a ). In addition, the construction of roads and canals has contributed to create barriers, inflows, and impoundments in the area, modifying what once w as a dynamic flow through system (Rivero et al., 200 7 a ). These disturbances resulted in widespread changes to the ecosystems of the wetlands. For instance, the spread of Typha has been associated with P enrichment and change in hydroperiod

PAGE 23

23 (Reddy and DeLau ne, 2008) in Water Conservation Areas, Everglades. Rivero et al. (2007a ) suggested using quantitative metrics to establish baselines (or reference values), compare them with natural conditions and conduct long term monitoring of a wetland ecosystem There has been much research studying the Everglades in Florida (Vaithiyanathan and Richardson, 1997; Bruland et al., 2007; Rivero et al., 2007 a ; Rivero et al., 2009). Rationale and Significance Wetlands are heterogeneous ecosystems composed of water, soil, floc/detritus, macrophytes, and periphyton intermixed within complex sloughridge topography. To model environmental properties, especially spatially complex soil properties in wetlands, accura te and precise data are needed. Furthermore, cost is one of the most considerable elements in research. Hence we need to develop time and cost efficient methodologies to derive high quality and inexpensive data for soil analysis. Remote sensing is ideally suited to derive environmental data for remote wetlands areas that are difficult to access. It is an important tool to determine land resources, to observe land surface covers at various spatial scales, and to determine changes that result from natural and anthropogenic processes (Schmid et al., 2008). It has many advantages especially in wetland and soil mapping. For instance, remote sensing based monitoring using temporal sequences of satellite images can capture seasonal patterns in wetlands. Also remote sensing data provide information on surrounding land uses of wetlands and changes over time (Ozesmi et al., 2002). And remote sensing data are available at a spatial resolution (~30 m to cm) magnitude of orders higher than sparse, site specific soil obser vations. In spite of advantages of remote sensing in wetland mapping, many of the studies have been conducted in upland areas such as agricultural

PAGE 24

24 areas (compare T able 1 1). Even though Rivero et al. (2007a 2009) successfully derived predictions of floc and soil TP in a subtropical wetland, there is a lack of research that allows us to predict and analyze spatial variations of various biogeochemical properties in aquatic ecosystems using remote sensing. Be cause of its differential spectral, spatial, and radiometric resolution, each remote sensing dataset derived from different remote sensors (satellite platforms) has different reflectance information. Remote sensing data with higher resolution can capture f eatures of complex aquatic ecosystems and represent them in more detail, more precisely. Hence the spatial resolution could significantly affect prediction accuracy and bias. However, there is a lack of research to compare the effects of spatial resolution to predict biogeochemical properties in wetlands which requires more attention. In the first part of this research study the purpose is to compare several methods (statistical and geostatistical) and data (soil observations and remote sensing images) to predict various biogeochemical properties in floc and soil in a south Florida wetland considering accuracy of predictions, cost, labor, and time. With the comparison of different spatial resolutions of remote sensing data, the ideal spatial resolution will be delineated. To facilitate future widespread use of remote sensing data in wetlands it is essential to test transferability of spectral inferential modeling, which will be tested in two adjacent wetland systems.

PAGE 25

25 Table 1 1. An overview of digital soil mapping studies that utilized remote sensing data. Source (year) Predicted Soil Property a (Unit) Location (climate) Land use (area) Sampling Designs (size)b Remote Sensors (spatial resolution)c Methodsd Prediction Performancee Summaryf Peng et al. (2003) Soil drainage Minnesota, USA (Continental) Agricultural: bare soil condition after crop harvest (57 ha) N/A (49 sites) Landsat TM(30m) IKONOS(4m) DOQ14 aerial photograph(0.6m) N/A Ap:29 ~ 82 Au:41 ~ 82 KE:0.35 ~ 0.62 Patterns of soil drainage class variability was predicted from multiple combinations of remote sensing (RS) data sources of varying spatial/ spectral resolutions: Landsat TM & DOQ 73% accuracy, Landsat TM, IKONOS, and DOQ 76% accuracy, IKONOS 65% accuracy, soil sur vey drainage map55% accuracy Sullivan et al. (2005) TC (%), Clay content (%) Alabama, USA (Subtropical) Agricultural: Coastal plain (52 ha) and Tennessee valley (31 ha) GR (0.2ha) (246/Coastal Plain, 158/Tennessee Valley) IKONOS (4m, 1m) CK, MLR, Fuzzy cmeans clustering RMSE (Te./Co.)g Kriging: TC: 0.11/ 0.22 Clay: 5.03/2.05 CK: TC: 0.04/0.02 Clay: 0.02/4.02 MLR: TC: 0.11/ 0.24 Clay: 3.33/2.18 Mean: TC:0.40~1.27, Clay:9.3~28.6 CK: spatial dependency of data characterized well with IKONOS (NIR) data and uncertainty in TC was reduced compared with Kriging method. MLR: estimation of TC was improved compared with CK and the accuracy of estimation of clay contents in Tennessee area was improved. CK with RS data provided the most accurate result among considered methods Fernandez Buces et al. (2006) Salinity (7 classes) Mexico (Tropical) Basin of Mexico (6,000 ha) SS (86 sites) Landsat ETM+(30m) Airborne photographs(2.6m) Regression R2:0.751~0.826 Combined Spectral Response Index, NDVI were used to predict soil salinity. Correlation coefficients between derived indices and soil salinity were obtained. Huang et al. (2007) TC (g/kg) Michigan, USA (Continental) Glacial till terrain (50 ha) Transect + random (85) Landsat 7 ETM+ (30m, 15m) MLR, SR, PCR R2 (w/NIRS:0.70~0.88, w/Landsat:0.33~0.8 1) Adjusted R2 (w/NIRS:0.68~0.86, w/Landsat:0.31~0.7 8) RMSE (w/NIRS:0.127~0.189, w/Landsat:0.156~0. 269) PRESS (w/NIRS:1.47~2.75, w/Landsat:2.48~6.23) Carbon prediction was performed based on NIRS and Landsat ETM (step wise regression). Based on NIRS, PCs were used w & w/o soil moisture, topography, texture variables :w/terrain slope and elevation data improved the quality of spatial estimates of soil carbon :w/soil moisture improvement were minimal :w/soil texture information improved the accuracy Based on Landsat ETM: PCs (15m) w & w/o soil moisture, topography, texture variables :w/15m resolution performed much better :w/topography remarkably better than only PCs :w/soil moisture & textureimproved the accuracy

PAGE 26

26 Table 11. Continued. Source (year) Predicted Soil Property a (Unit) Location (climate) Land use (area) Sampling Designs (size)b Remote Sensors (spatial resolution)c Methodsd Prediction Performancee Summaryf Lobell et al. (2007) Salinity (3 classes) Mexico (Tropical) Agricultural (13,000 ha) RS in Jan.2005 (~5 ha each, 122) SR in Sep.2005, May 2006 (60 3 depth) ASTER (15m) Landsat TM (30m) Landsat ETM+ (30m) Technique of Lobell et al.(2003) RMSE:0.65 R2:0.65 (reported yield vs. estimated yield with RS) The importance of salinity in wheat production was documented and identification of saline fields. Salinity has small impact on wheat productivity (mapped using EC1) Each RS data was used in different year (did not compared RS in same season), RS was used for yield estimation M.de Asis, Omasa (2007) Erosion (C factor maph: 14 classes) Philippines, Asia (Tropical) Grassland (2,700 ha) N/A (53 sites) Landsat ETM+(30m) QuickBird(2.44m) LSMA Oa:42.23 ~ 70.92 KE:0.28 ~ 0.63 Landsat ETM data was used to model soil erosion: performed a minimum noise fraction transformation on Landsat ETM image/ endmember: QuickBird image & field data (each RS data was used separately) Nield et al. (2007) Gypsic and natric soils Utah, USA (Semiarid) Sodium rich area (22,209 ha) N/A Landsat 7 ETM+ (30m) Normalized difference ratio model Ap:51.9 ~ 92.9 Au:70.9 ~ 100 Landsat data was used to assess the distribution of gypsum rich, sodium rich soil area. SWIR band 7 has more predictive power than band 5 in gypsic soils / SWIR band 5 is better spectral plots than NIR band 4 in natric soils. Gypsic soil areas appeared to have distinct spectral reflectance, but natric soil areas not (predict: 82% vs. observed 52%) Rivero et al. (2007a ) TP (mg/kg) Florida, USA (Subtropical) Wetland: Water Conservation Area 2A (43,281 ha) 111 sites Landsat ETM+1 (30m) ASTER (15m) SR OK CK RK R2:0.39 ~ 0.75 ME:0.9 ~ 27.3 RMSE:134.9 ~ 264.4 Various geostatistical methods were performed with spectral data and derived indices in aquatic ecosystem. CK,RK performed best to predict floc and soil TP when compared to univariate model (OK) Chen et al. (2008) SOC (%) Georgia, USA (Subtropical) Agricultural (114.4 ha) RS (139) ATLAS (2m) WNNS, SR R2 (Group1: 0.87, Group2: 0.91) RMSE (Group1: 0.108, Group2: 0.143) Based on similar image distribution 10 different crop fields were grouped as Group 1 and 2. The image histogram features were extracted from each field image and analysis of field similarity was performed using the WNNS model. Mapping with grouped fields had better R2 and RMSE values than singlefield mapping procedure and it could reduce the number of soil samples (field) and co st.

PAGE 27

27 Table 11. Continued. Source (year) Predicted Soil Property a (Unit) Location (climate) Land use (area) Sampling Designs (size)b Remote Sensors (spatial resolution)c Methodsd Prediction Performancee Summaryf Eldeiry & Garcia (2008) Salinity Colorado, USA (Semiarid) Agricultural (50,000 ha) 68 observation wells (256) IKONOS(4m) Landsat(30m) MLR, Spatial AR, RK R2:0.23~0.38 The correlation between field collected soil salinity data and satellite imagery was tested using MLR, spatial AR. Spatial AR model was able to make some improvements over the MLR model regarding the normality and the homogeneity, some of the spatial autoc orrelation among the residuals was removed Low R2: the model poorly predict soil salinity data Modified ReK (NDVI) was used for both images Modified ReK provided the best estimation Gomez et al. (2008) SOC (%) Australia (Mediterranean) Agricultural ( ) 146 sites Hyperion (30m) PLSR RMSE:0.1 ~ 0.69 R2:0.04 ~ 0.43 RPD:0.98 ~ 1.33 Comparison of soil organic C predictions 1) using the hyperion hyper spectral satellite remote sensor, 2) using fieldcollected visible and near infrared reflectance. The spectral resolution did not change the accuracy of the model and the use of Hyperion hyper spectral data is less accurate than the use of visible and near infrared reflectance. Jaenicke et al. (2008) C storage (Gt) Indonesia (Tropical) Peat domes (6 peat domes: 1 8,800 ~ 734,700 ha per each) 542 (modeling) + 208 (model verification) Landsat ETM+ (30m) OK Estimation of the amount of C storage in peatlands, 3D model was applied with RS data, C storage was estimated using ArcGIS 3D Analyst that was used to calculate peat volume Karnieli et al. (2008) Landcover change Kazakhstan, Central Asia (Arid) Desert (49,800,000 ha) N/A Landsat 2 MSS Landsat 4 TM Landsat 7 ETM+ OK RMSE: 0.0095 in 1975 0.0096 in 1987, 2000 Land cover change and degradation around watering points was mapped in different years (1975, 1987, and 2000). Tasseled Caps brightness index was used for enhancing the contrast of the images. Kheir et al. (2008) Erosion (3 classes) Lebanon, Western Asia (Mediterranean) Forest, Urban, Agricultural (Combined area) (67,600 ha) 172 sites Landsat TM (30m) IRS19 (6m) OASIS Ap:75 ~ 89 Au:73 ~ 100 Ee:17 ~ 25 Ed:11 ~ 25 Oa:82 An erosion risk map at 1:100,000 cartographic scale was produced using structural classification of Landsat TM imagery, accuracy 82% M Pastor et al. (2008) CaCO3 (%), EC (dS/m), SOM (%), Text (%) Spain (Mediterranean) Agricultural, Coastal, Wetland, Salt marsh (area) SS (72 samples) N/A First and second derivatives of spectral data (VNIR) R (|R|>0.750~0.955) Relationship between soil properties and VNIR range was assessed. CaCO3 content: high correlation coefficients were found between 453676nm EC: between 433627nm, SOM: between 349586nm, Silt: 387nm, and Sand: 342 780nm

PAGE 28

28 Table 1 1. Continued. Source (year) Predicted Soil Property a (Unit) Location (climate) Land use (area) Sampling Designs (size) b Remote Sensors (spatial resolution) c Methodsd Prediction Performance e Summaryf Schmid et al. (2008) Text (%), SOM (%), CaCO3 (%) Ethiopia, Africa (Tropical) Grassland (N/A) N/A (98 samples /2001, 93 samples /2006) ASTER (15m) IKONOS (4m) N/A N/A General lithologic and pedologic classifications were conducted using ASTER and IKONOS. Both classification outputs provide useful, complementary information about soil group characteristic and spatial distribution, spectral comparison is best achieved between the soil laboratory curves and ASTER image spectra (ASTER performed better result) Sumfleth et al. (2008) TC (%), TN (%), Text (%), dept (cm) China, Asia (Subtropical Monsoon) Ecological experimental station (1,000 ha) Combined: GS + catenary samping (212 samples) SPOT5(10m) ASTER (15 m) Landsat ETM+ (30 m) MLR, IDW, SK, OK, RK(B,C) R2:0.06~0.41 RMSE:0.039~8.967 The application of regression kriging model "C" resulted in the lowest RMSE for soil variables (TC, TN, Text: the lowest RMSE in OK) Vrieling et al. (2008) Erosion (5 classes) Brazil (Tropical) Savannah (10,000 ha) 331 sites observation MODIS (250m) ASTER(15m) QuickBird (0.7m) N/A Ap:64.6 ~ 84.0 MODIS images were used to know when erosion risk is high. ASTER and slope information was used to assess the optimal image timing for erosion mapping: slope (not very useful), clear relationship between the erosion risk and vegetation cover. QuickBird used to validate erosion risk map West et al. (2008) C change (%) Midwestern region of U.S. around Iowa (11 states) (Continental to Semiarid) Agricultural N/A Landsat TM (30 m) Stepwise integration N/A To estimate change of soil C and to identify regions where changes are likely to occur, field measurements, inventory data, RS data: (1) annual estimates of crop area, (2) area of major crop types using different tillage practices, (3) soil attribute, and (4) soil carbon accumulation and loss were assessed aC carbon ; CaCO3, c alcium carbonate; EC e lectrical conductivity ; SOC soil organic carbon; SOM s oil organic matter ; TC total carbon; Text t exture: sand, silt, and clay fraction; TN t otal nitrogen; TP t otal phosphorus bGR g rid random sampling ; GS g rid sampling ; RS r andom sampling ; SeR semi random sampling ; SR stratified random sampling ; SS, stratified sampling cASTER a dvanced spaceborne thermal emission and reflection radiometer ; ATLAS, a dvanced thermal and land applications sensor ; DOQ d igital orthophoto quadrangle; IRS i ndian remote sensing ; Landsat ETM+ l andsat enhanced thematic mapper plus ; Landsat MSS l andsat multispectral scanner ; Landsat TM l andsat thematic mapper ; MODIS m oder ate resolution imaging spectroradiometer ; SPOT s atellite pour lobservation de la terre. dCK co kriging ; IDW i nverse distance weighting ; LSMA l inear spectral mixture analysis ; MLR m ultiple linear regression; OASIS, o rganization and analysis of spatial structures ; OK o rdinary kriging ; PCR, principal component regression; PLSR p artial least squares regression; RK r egression kriging ; SK, s imple kriging ; S patial AR s patial autoregressive model ; SR s tepwise regression; WNNS w ard neural network system eAp p roducer's accuracy ; Au u ser's accuracy ; Ed d eficit errors or omission ; E e e xcess errors or commission ; KE k appa accuracy ; ME m ean prediction error ; MSE m ean standardized squared prediction error ; Oa o verall accuracy ; PRESS calculated based on true and predicted values obtained from a leaveoneout cross validation; RMSE r oot mean squared error ; RPD, r atio of performance to deviation. fNDVI n ormalized difference vegetation index ; NIR, n ear infrared spectral range; RS r emote sensing; SWIR s hort wave infrared spectral range; VNIR, v isible NIR; WNNS w ard neural network system; N/A n o information available. gTe./Co. RMSE values in Tennessee valley / coastal plain study area. hC factor v egetation factor in soil erosion models (e.g. universal soil loss equation).

PAGE 29

29 CHAPTER 2 MULTI SCALE MODELING OF SOIL SERIES USING REMO TE SESNING IN A WETLAND ECOSYSTEM1Overview Wetland ecosystems are prominent in the southeastern U.S., specifically Florida, due to geologic, topographic, and climatic conditions. They cover about 25% of the State of Florida, and the soils in these subtropical wetlands have shown high variability due to their complex composition of water, soil, floc/detritus, macrophytes, and periphyton intermixed within complex sloughridge topography (Rivero et al., 2007a). Due to the unique values of wetlands encompassing ecosystem services, such as sequestration of soil carbon, provisioning of biodiversity, cycling of nutrients, wildlife habitat, freshwater resources, and more, their sustainability is profoundly important. To characterize spatial patterns of wetland soils in dependence of soil forming factors is critical for the functioning and preservation of them. Remote sensing informed soil prediction models have shown success to improve the predictive power and spatial resolution of predictions in upland systems (Huang et al., 2007; Nield et al., 2007; Vrieling et al., 2008). Yet not much is known if these models also perform well in wetland ecosystems. Grunwalds (2009) comprehensive review confirmed that many digital soil mapping (DSM) studies incorporate the biotic factor (> 39% of studies, case A) in soil prediction models. In her meta analysis, she revealed an increased use of remote sensing in DSM due to its capabilities to directly or indirectly infer on a wide suite of biotic properties. Remote sensing allows rapid and inexpensive data collection over large areas in form of continuous grids. Numerous studies have 1 Accepted by Soil Science Society of America Journal in 2012.

PAGE 30

30 been focused on predicting soil properties utilizing a variety of remote sensing imagery, such as Landsat Thematic Mapper (30 60 m spatial resolution), IKONOS (the Greek word for image, 1 4 m resolution), Moderate Resolution Imaging Spectroradiometer (MODIS) (250 1000 m resolution), Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER, 15 30 m resolution), and QuickBird (0.61 2.88 m resolution). These studies strongly suggested that the accuracy of soil predictions significantly improved compared to traditional soil assessment without remote sensing (Fermandez Buces et al., 2006; Huang et al., 2007; Nield et al., 2007; Vrieling et al., 2008) Although remote sensing based modeling of soil properties has much value in wetland and aquatic ecosystems, where field sampling is challenging and siteaccess limited, the majority of studies have been conducted in upland systems predominantly agricultural areas (Peng et al., 2003; Sullivan et al., 2005; Lobell et al., 2007; Chen et al., 2008; Eldeiry and Garcia, 2008; Gomez et al., 2008; West et al., 2008). Fewer remote sensing supported DSM studies have been conducted in wetland areas (Rivero et al., 2007a; Jaenicke et al., 2008; Rivero et al, 2009; Zhao et al., 2009). For example, Rivero et al. (2007a) applied Landsat Enhanced Thematic Mapper (ETM+) and ASTER data to predict floc and soil phosphorus (P) in a subtropical wetland, Water Conservation Area (WCA) 2A which is located w ithin the Everglades in south Florida, U.S. They demonstrated that spectral data derived from satellite images substantially improved floc and surface soil P predictions using regression kriging, when compared to soil predictions solely based on sitespeci fic soil P values using ordinary kriging.

PAGE 31

31 However, to our knowledge, no study integrated remote sensing data to develop prediction models for soil series (i.e., soil classes) in subtropical wetlands. The Everglades (8,211 km2) are one of the largest and m ost unique freshwater wetlands in the world covering about 5.6% of the State of Florida (Florida Fish and Wildlife Conservation Commission, 2003). Since the Everglades has been highly impacted by hydrologic modifications combined with increased nutrient loading (Noe et al., 2001), comprehensive studies have been focused on soil biogeochemical properties (Vaithiyanathan and Richardson, 1997; Bruland et al., 2007; Grunwald et al., 2008), vegetation (Loveless, 1959; Childers et al., 2003; Corstanje and Reddy, 2006), and periphyton (Browder et al., 1994; McCormick and ODell, 1996; McCormick et al., 1998; Gaiser, 2009), however, not on soil taxonomic classes. The only soil survey in this region dates back to the early 1940s documenting soil types and soil depth (Mowry and Bennett, 1948). It did not provide point observation data, but delimited soil types into crisp polygon shapes creating abrupt boundaries at relatively coarse map scale (about 1:63,350). Moreover, water controls along with the construction of lev ees and canals in the 1950s have led to significant changes in the ecosystem, such as water and soil quality degradation, soil subsidence, and vegetation community in the Everglades (Bruland et al., 2007; Rivero et al., 2007b; Grunwald et al., 2008). Howev er, the historic soil survey does not reflect these changes of soil forming factors over the past decades. For these reasons, our study did not aim to update previous survey data, but it provides a first attempt to collect and map the soil series distribut ion in the area. The main objectives of this study were to (i) develop spectral informed soil taxonomic prediction models and assess their accuracy aiming to improve our understanding of the

PAGE 32

32 distribution of soil series in the northern Everglades; (ii) quantify the relationships between soil classes and environmental covariates derived from remote sensing and geospatial sources; and (iii) compare the effects of spatial resolution (10, 30, and 250 m) of three remote sensing sensors to delineate soil classes. Materials and M ethods Study Area The study was conducted in WCA 2A, which is located in the northern part of the Everglades, Florida (Fig. 1). The study area is approximately 418 km2. The Everglades climate is subtropical, with mean annual precipitation of 1,297 mm (Everglades Depth Estimation Network, EDEN, 2009) and mean annual temperature of 23 C over the past 10 years (National Climatic Data Center, 2009). The majority of the soils in the Everglades, including WCA 2A, are Histosols. Common types of p eat in the region are the Everglades peat and Loxahatchee peat formations that make up the ridge and slough system. The Everglades peat develops on the higher elevated areas that are composed of the remains of sawgrass ( Cladium jamaicense) and these soils are generally brown to black with minimal content of minerals, while the Loxahatchee peat develops on the lower elevated areas that are composed of remains of the roots and rhizomes of water lily ( Nymphaea odorata) and the soils are lighter colored (Lodge, 2005; Bruland et al., 2007). Typical landforms of the Everglades are primarily ridges and sloughs with scattered tree islands of various sizes. Sawgrass has been the dominant vegetation type, specifically in low nutrient zones of WCA 2A, which has been replaced with cattail ( Typha domingensis ) in those areas associated with P enrichment and/or fluctuating water levels (Jensen et al., 1995; Wu et al., 1997; Noe et al., 2001). The ridges consist mainly of s awgrass patches of dense growth surrounded by areas of very

PAGE 33

33 sparse growth. The areas with very sparse growth are typically elongated sloughs which are dominated by spikerush ( Eleocharis spp. ) and periphyton. The topography is nearly level with elevation ranging from around 0.8 to 4.3 m above mean sea level for WCA 2A (USGS, High Accuracy Elevation Data, 2007). Geology is dominated by the Tamiami Formation capped by Pliocene and the Fort Thompson Formation, and the Miami Limestone formed in the Pleistocene (Lodge, 2005). Field Sampling Soil and vegetation sampling was conducted in July and October 2009 and mid March 2010 using airboats. Field sampling consisted of 108 sites spread over the study area (Fig ure 2 1) Pedons were collected and described by soil scientists from the NRCS. Sampling sites were selected based on a stratified random sampling design using historic soil, hydrologic, and ecological data sets identified in 2003. A detailed description of the sampling criteria can be found in Rivero et al. (2007b). Only 48 sampling points of the prese lected sampling sites were revisited because of limited access conditions, such as dense vegetation, lack of accessibility by airboats, and difficult field conditions. The remaining samples were randomly collected nearest to preselected sampling points. A t each site, the status of organic decomposition, depth to bedrock (BR2009), and the presence/absence of mineral substrata were described as primary environmental variables. Site and pedon descriptions including horizon delineation, soil texture, color (hu e, value, and chroma using Munsell color notation), landscape type, landform, and slope as well as x and y coordinates were collected at each sampling point. Each soil sampling location was georeferenced using a global positioning system (Garmin International, Inc., Olathe, KS).

PAGE 34

34 Soil Series Description Parent material within the WCA 2A is exclusively highly decomposed organic material from sawgrass, spikerush, and other herbaceous plants. All 108 pedons were classified as Haplosaprists consisting of highly decomposed (sapric) material. There were seven soil series found in WCA 2A: Gator, Gator with limestone bedrock, Lauderhill, Okeelanta, Pahokee, Terra Ceia, and Terra Ceia with limestone bedrock. They are closely associated with each other and the differences among the series are related to their depth to the limestone bedrock and depth of organic horizon. All of the soil series consisted of very poorly drained organic soils. There were several pedons that follow Gator or Terra Ceia soil series descriptions but included limestone bedrock between the depth of 129 and 200 cm. These pedons did not fit the description of any of the established soil series according to the Official Soil Series Descriptions (OSD; Soil Survey Staff, 2010), but for the purpose of t his study they were grouped into the respective soil series (i.e., Gator and Terra Ceia). Thus, here Gator with limestone bedrock was considered as Gator series and Terra Ceia with limestone bedrock was considered as Terra Ceia series for the purpose t o develop prediction models for soil series. See Table 2 1 for a detailed description of the soil series in WCA 2A. Remote Sensing Imagery and Spectral Data Analyses Three satellite images with different spatial resolutions were selected for the study: MODIS, Landsat ETM+, and SPOT 5 images. Criteria to select suitable satellite images included (i) closeness to field sampling dates, (ii) cloud and noise free images, because clouds may interfere with the analysis, and (iii) seasonality (winter and spring im ages discriminate vegetation signatures better than summer images as reported by Rutchey and Vilchek, 1999 and Rivero et al., 2009).

PAGE 35

35 The MODIS image (ID:MOD13Q1.A2010049.h10v06.005.2010067103334; February 2010) was obtained from Land Processes Distribut ed Active Archive Center (LP DAAC), USGS Earth Resources Observation and Science (EROS) Center. It has 36 spectral bands that have three different spatial resolutions, i.e., 250 m for bands 1 and 2, 500 m for bands 3 to 7, and 1,000 m for bands 8 to 36. A MOD13Q1 that was used for the study provides vegetation indices every 16 days with 250 m spatial resolution. The Landsat ETM+ image (ID:L7101504204220100216; February 2010) was obtained from the USGS EROS Center. Landsat ETM+ has 30 m spatial resolution ( bands 1 to 5 and 7), except band 6 (thermal infrared, 60 m resolution) and the panchromatic band (15 m resolution). The SPOT images (ID: 111711017, 111664010; January 2009) were provided by Planet Action, a nonprofit ASTRIUM GEO initiative. It has 10 m sp atial resolution (bands 1 to 3), except band 4 (short wave infrared, 20 m resolution) and the panchromatic band (2.5 m resolution). All satellite images were imported into the ERDAS Imagine 2010 software (Earth Resource Data Analysis System Inc., Atlanta, GA) and standardized into a single format (*.img) for further processing. They were projected to the Universal Transverse Mercator (UTM) map projection (Zone: 17; Datum: World Geographic System, WGS 84) and subsetted to the study area. The following spectral vegetation indices were derived for different satellite images depending on their spectral bands: (i) from the MODIS image Enhanced Vegetation Index (EVI), Moisture Stress Index (MSI), Normalized Difference Vegetation Index (NDVI), Simple Ratio (S R), and Transformed Vegetation Index (TVI); (ii) from the Landsat ETM+ image Mid Infrared Index (MidIR), MSI, NDVI, Normalized Difference Vegetation Green Index (NDVI green), Normalized Difference Water Index (NDWI),

PAGE 36

36 Reduced Simple Ratio (RSR), SR, and T VI; (iii) from the SPOT image MSI, NDVI, NDVI green, NDWI, RSR, SR, and TVI. A summary of formulas to derive indices is given in Table 2 2. The spectral indices have been widely used to assist in mapping of specific ecosystem attributes by applying algorithms with combinations from different spectral bands and have shown better capability than raw bands of an image to infer on soil properties (Fernndez Buces et al., 2006; Vrieling et al., 2008; Eldeiry et al., 2008). Furthermore, an assumption was made that spatially explicit relationships exist, not only between the spectral values coinciding with the x and y coordinates of pedons, but also between the spectral values within a neighborhood surrounding the x and y coordinates of pedons. Thus, the mean spectral values within neighborhoods surrounding pedon locations were computed using the focal geospatial method nearest neighbor. Spectral indices derived from Landsat ETM+ and SPOT images were aggregated within 7 x 7 cells representing a 210 m x 210 m neig hborhood (Landsat ETM+) and 25 x 25 cells representing a 250 m x 250 m neighborhood (SPOT), respectively, resembling the coarser resolution of spectral data derived from the MODIS image (250 m x 250 m). Mean spectral values within 3 x 3 cells from the SPOT image, i.e., 30 m by 30 m resolution, were included for the comparison of the spectral data effectiveness between the SPOT and Landsat ETM+ images. In addition, mean spectral values within 9 x 9 cells from the SPOT image and 3 x 3 cells from the Landsat E TM+ image, i.e., 90 m by 90 m resolution, were derived to investigate the effects of the intermediate resolution spectral data. A principal components (PC) analysis was performed with Landsat ETM+ (bands 1 to 5 and 7) and SPOT images (bands 1 to 4). A PC analysis can be considered as a

PAGE 37

37 rotation of the original variable coordinate system to new orthogonal axes called principal axes, and this transformation reduces the number of correlated variables into a smaller number of uncorrelated variables. It helps to investigate possible covariations between soil series and transformed spectral values. The tasseled cap (TC) transformation was performed with the Landsat ETM+ image. The TC transformation converts the original bands of an image into a new set of bands to enhance the spectral signatures of soil brightness, vegetation, and moisture content in a Landsat image (Kauth and Thomas, 1976; Price et al., 2002). All of the predictor variables derived from images were combined with sitespecific soil classes through spatial extractions of x and y coordinates for each sampling point with the extraction function in the Spatial Analyst Tool using ArcGIS 9.3 (Environmental Systems Research Institute ESRI, Redlands, CA). Geospatial Environmental Ancillary Data Various exhaustively available geospatial environmental data layers covering the whole study area were compiled including a boundary map, layers of vegetation, elevation, distance to water control structures (WCSs) geophysical properties, and x and y coordinates (Table 2 3). For all environmental ancillary data layers, data values at each soil sampling location were extracted. This resulte d in a data matrix where each of the 108 field sampled soil locations were matched with a data value from each of the environmental ancillary data layers. All geographic information system (GIS) operations for processing of the layers were conducted in Arc GIS 9.3 (ESRI, Redlands, CA). The USGS surveyed the elevations of the Everglades region and released the High Accuracy Elevation Data (USGS, 2007). It provides high vertical accuracy (15 cm) data on an approximate 400 m by 400 m grid and has 2,600 sampling points within

PAGE 38

38 WCA 2A. The spline (regularized) interpolation method was employed on elevation point data to produce a continuous elevation raster. Three elevation data layers which have 10 m, 30 m, and 250 m resolution were produced and used to develop prediction models with SPOT, Landsat ETM+, and MODIS images, respectively. Hydrology is one of the most important factors to control soil formation and wetland biota. L ow velocity sheet flow through the study area is prominent The surface water inflow and outflow from WCA 2A is regulated by the South Florida Water Management District (SFWMD) controlling WCSs A continuous raster based data layer of distances to WCSs was produced using the Euclidean distance function to include the effect of hydrology on soil formation. Geophysical property data served as a proxy to infer on parent material. It included potassium concentration gammaray data, bouguer gravity anomaly data, isostatic residual gravity anomaly data, and magnetic anomaly data (USGS, Digital Data Series DDS 9, 1999). The geophysical property layers have a spatial resolution of 2,000 m by 2,000 m, which is coarser than the remote sensing images. Thus, each layer was reproduced to layer format with a spatial resolution of 10 m, 30 m, and 250 m, re spectively, to match the resolution of the images. The resampling of the pixel values was performed using nearest neighbor. Nearest neighbor uses the input cell value closest to the output cell as the assigned value to the output cell. Two different bedrock depth datasets were used to characterize lithology: Categorized bedrock depth data that was surveyed before 1948 (BR1948) and continuous bedrock depth data (cm) collected at each sampling point in 2009 (BR2009). The BR1948 had four soil depth classes which were less than 91 cm, 91 cm to 152 cm,

PAGE 39

39 152 cm to 244 cm, and more than 244 cm. This class was extracted from the polygon map which was produced by the SFWMD based on the previous soil survey, and then used as one of the predictor variables along with B R2009. Soil series prediction models were developed in two different modes: (i) without bedrock depths (BRs) (Model I) and (ii) with BRs (Model II) to investigate the capabilities of bedrock depth to predict soil series. Continuous raster based x, y, and xy coordinates (the latter representing the interaction effects of northings and eastings) were derived and included as one of the environmental ancillary variables to capture a potential stratification effect on soil series based on geographic coordinates. However, the coordinates over powered all other remaining predictor variables (i.e., hydrology, lithology, and topography) vanquishing other predictors, and as a result, created sharp boundaries among the soil series (maps are not shown). Changes of soi l series in the field are gradual. Thus, another model realization of soil series by excluding the geographic coordinates from predictor variables was created and compared. The accuracy differences of prediction models between with and without geographi c coordinates were less than 2%. Hence, only models without geographic coordinates are presented in this paper. Classification Tree Method Classification trees were employed to build predictive soil series models using CART 6.0 software (Salford Systems San Diego, CA) in four different modes: (i) without spectral data as a control, (ii) with SPOT, (iii) with Landsat ETM+, and (iv) with MODIS images derived spectral data. In addition, g eospatial variables that characterized vegetation, topography, hydrol ogy, lithology, and geographic settings were used as predictor variables to build tree models (Table 2 3). In classification trees the target

PAGE 40

40 variable is categorical (e.g., soil series or soil classes) and predictor variables can be either categorical or n umeric. Tree based methods have advantages for dealing with highly complex data, nonlinear relationships, highorder interactions amongst data, and missing values (Breiman et al., 1984; Death and Fabricius, 2000; Spruill et al., 2002; McBratney et al., 2003 ). Tree structures are generated through recursive partitioning the data into a number of groups and, for the tree construction, each predictor plays a role either as a main splitter or as a surrogate splitter in the tree. The splitting of the set of obs ervations (parent nodes) into two descendent subsets (child nodes) is continued as long as the child nodes become purer than parent nodes (Grunwald et al., 2009). Trees are grown until further splitting is impossible based on user defined settings. Three cases per terminal node were set as a minimum node size for all trees in this study. CART over grows trees and allows pruning of trees. Tree pruning was performed to derive optimized, parsimonious trees selecting the minimum cost tree regardless of size o ption which was assessed by the cross validated relative error computed by CART 6.0. An iteration approach was used to find the optimal tree for soil series by testing three different classification methods including twoing, entropy, and gini for all model s. In CART, variable importance is computed to reflect the relative contribution of each predictor variable in classifying the target variable. The variable importance score is calculated as the total decrease in node impurities measured by entropy. Detail s can be found in Strobl et al. (2009). Single trees build only one treemodel with branches to predict a class, whereas committee trees build hundreds of trees to optimize splitting rules while predicting soil classes or properties (Grunwald et al., 2009). Single tree and committee tree models

PAGE 41

41 using environmental ancillary data and spectral data derived from SPOT, Landsat ETM+, and MODIS images, respectively, were developed and compared to predict the sampled soil series in the study area. The use of committee classification trees have shown better prediction capabilities than single trees in other studies (Prasad et al., 2006; Grunwald et al., 2009); however, the small number of observations (< 10) for two classes (i.e., Gator and Lauderhill) of the t arget variable restricted the building of committee trees for soil series in this study showing lower accuracies (< 40%) than single trees. Hence, only single trees are presented in this paper. Uncertainty Assessment and Verification of Model Output The t rees were evaluated using t enfold cross validation where the dataset is randomly divided into ten partitions. Nine of the partitions were used to develop the model and the remaining partition was used for model evaluation. The procedure was sequentially r epeated for all partitions and then the average over all the partition results was computed. The observed soil series were compared site by site with the series that were predicted from classification trees in form of confusion matrices to assess the uncertainty of predictions. The error matrices summarize agreement (correctly predicted) and disagreement (incorrectly predicted) between observed and predicted soil series in the cells. The producers accuracy, users accuracy, and overall accuracy we re calculated based on the error matrices for all trees as suggested by Brus et al. (2011). The producers accuracy is a percentage of correctly classified cases within the reference class (column), and users accuracy is a percentage of correctly classified cases within the predicted class (row). Each soil series class yielded the producers and users accuracy for each tree, and they were averaged as average

PAGE 42

42 producers accuracy and average users accuracy to compare models. The overall accuracy is a percentage of correctly classified cases within the matrices. It was determined by dividing the total number of correctly classified cases (i.e., the sum of values along the diagonal in the error matrices) by the total number of observed cases. In ad dition, the Kappa coefficient, K (Eq. ( 2 1 ) ), was used as another measure of prediction quality: k i i i k i i i k i ii) x (x N ) x (x x N K1 2 1 1 ( 2 1 ) where k is the number of rows in the matrix, xii is the number of observations in row i and column i (diagonal) and xi+ and x+i are the off diagonal totals for row i and column i respectively, and N is the total number of observations (Congalton, 1991; Jensen, 2005). Landis and Koch (1977) suggested that K values smaller than 0% represent poor accuracy, from 0 to 20% represent slight accuracy, 20 to 40% represent fair accuracy, 40 to 60% represent moderate accuracy, 60 to 80% represent substantial accuracy, and the values greater than 80% represent almost perfect accuracy for categorical data analysis. The metrics to assess prediction quality and accuracy emphasize different aspects because each measure incorporates different information. The Kappa coefficient incorporates not only the major diagonal (correctly classified cases) but also the off diagonal elements (incorrectly classified cases), whereas others incorporate only the major diagonal and exclude the off diagonal elements.

PAGE 43

43 Results and D iscussion Distribution of Soil Series within WCA 2A The distribution of soil series within the WCA 2A is shown in Fig ure 2 1. The most dominant soil series in the study area was Terra Ceia with 57 field observations, about 53% of sampling sites, followed by Pahokee (25), Okeelanta (14), Gator (7), and Lauderhil l (5) observations. The distribution of soil series was related to soil depth. The Terra Ceia and Okeelanta series, which both have deep O horizons, were mainly found in the northern and eastern part of the study area. On the other hand, the Lauderhill ser ies with relatively shallow O horizons was predominantly found in the southern part of the study area. The depth from the soil surface to the bedrock ranged from 64 to 234 cm and about 90% of the observation sites had bedrock within 200 cm in WCA 2A (Table 2 1). About 61% of the field observations (66 sites) matched depth classes of the historic soil survey despite the hydrologic manipulations and nutrient enrichment this freshwater marsh was undergoing since the early 1900s (Porter and Porter, 200 2 ). S oil Series Prediction Models Using Spectral Data Model I The confusion error matrices of the tree models for soil series predictions without incorporated bedrock depth data are shown in Table 2 4. Figures 2 2 to 2 4 show the results of a single tree predicting soil series using the SPOT and Landsat ETM+ images with 14 terminal nodes and the MODIS image with 12 terminal nodes, respectively. All of them showed the dominance of the Lithic group (i.e., Lauderhill and Pahokee) in their left branches and the T erric group (i.e., Gator and Okeelanta) in their right branches. The tree model using the SPOT image correctly predicted 7 Gator (out of 7 observed), 5 Lauderhill (out of 5 observed), 12 Okeelanta (out of 14 observed), 22 Pahokee (out of

PAGE 44

44 25 observed), and 31 Terra Ceia (out of 57 observed) soil series with ancillary environmental predictor variables (Table 2 4). Similar predictions were achieved for the trees derived using Landsat ETM+ and MODIS predictor variables, with few more confusions of soil series a mong Okeelanta, Pahokee, and Terra Ceia in model I. All trees using remote sensing images derived spectral data showed improved accuracy compared to the tree model without spectral data (control). The control tree showed 49.1% overall accuracy (Table 2 5). This suggests that the tree model relying only on geospatial environmental ancillary data had moderate capability to predict soil series in WCA 2A. However, tree models using remote sensing derived spectral data yielded greatly improved overall accurac y when compared to the control tree. The overall accuracy was more than 20% higher in the model derived using the SPOT data when compared to the control model. The best performing model was derived through SPOT, followed by Landsat ETM+, and MODIS tree models as demonstrated by accuracy metrics and the Kappa coefficient (Table 2 5). The tree model to predict soil series using SPOT predictor variables amounted to an average producers accuracy of 85.6%, average users accuracy of 72.5%, and overall accuracy of 71.3%, which were the highest accuracy values among all trees and the tr ee had the smallest relative error in model I. The Kappa coefficient showed the highest accuracy to predict soil series with 61.1% using the SPOT image and geospatial environmental ancillary data, which can be interpreted as substantial accuracy according to Landis and Koch (1977). The other models using Landsat ETM+ and MODIS derived input variables showed a Kappa coefficient of 57.2% and 47.7%, respectively, which can be interpreted as moderate accuracy. Similar results were reported by Kempen et al. (2009) predicting overall map

PAGE 45

45 purity (overall accuracy) of 58% for soil groups in the Netherlands and Wei et al. (2010) predicting average accuracy of 68% for soil series in Iowa, U.S. Our results suggest that the incorporation of remote sensing images improved the prediction capabilities of soil series models in WCA 2A. Both of the pixel specific (i.e., values extracted at each sampling location) and focal (i.e., values calculated within neighborhoods surrounding the sampling points, e.g., NDWI 25 x 25) spectral data derived from remote sensing images played a criteria role for building trees. The pixel specific spectral variables, such as reflectance values of red and near infrared (NIR) wavelength energy which inferred on vegetation, were represented in m odels. Principal component scores which inferred on water and vegetation were also represented in models. Focal NDWI derived from SPOT and Landsat ETM+ images rather than pixel specific NDWI was used as a major splitter in each of the models (Fig ures 2 2 and 2 3). This suggests that spatially explicit relationships exist not only between soil series and point specific spectral values, but also between soil series and surrounding neighborhood values. A specific soil series occupies a certain portion of the l andscape and may show muted relationships with spectral variables within the neighborhood surrounding the soil sampling location. Normalized difference water index as a measurement of liquid water molecules in vegetation incorporates the shortwave infrared (SWIR) band and is more sensitive to the total amount of liquid water in the leaves than other spectral indices such as NDVI (Gao 1996). Our results demonstrate that NDWI has better capability to infer on wetland soils than NDVI, probably as a result of the increased water content on vegetation or increased water depth, as opposed to NDVI that performs better in

PAGE 46

46 drier/upland ecosys tems. A recent study in the Everglades (Wang et al., 2011) showed similar results, in terms of applications of NDWI to capture the spatial and temporal variability, in this case associated with water depth. Other studies have shown predictive power of NDVI rather than NDWI for soil properties in upland areas (Fernandez Buces et al., 2006; Eldeiry and Garcia, 2008; Vrieling et al., 2008). Results further suggest that the finer resolution SPOT image had somewhat better capabilities to infer on soil series when compared to coarser resolution Landsat ETM+ and MODIS images. However, it may be noted that the MODIS derived tree model still showed pronounced ability to predict soil series with overall accuracy of 60.2% and a Kappa coefficient of 47.7%. Previous studies agree that the use of coarser scale resolution images is ideal to capture or classify general characteristics of landscapes and fine scale resolution images are suited to capture small spatial variations of soil properties (Schmid et al., 2008; Vrieli ng et al., 2008). It has been shown that adding fine spatial resolution images increase the soil property prediction accuracy (Sullivan et al., 2005; Huang et al., 2007; Eldeiry and Garcia, 2008). McKenzie and Ryan (1999) stressed the use of fine spatial r esolutions to capture the relationships between soil properties and environmental factors. Implications of scaling and aggregation effects on soil predictions were provided by Grunwald et al. (2011). Model II The confusion error matrices of the tree models for soil series predictions with incorporated bedrock depth data are shown in Table 2 6. Note that the soil series predictions including bedrock depth as an independent variable (Model II) cannot be upscaled to the whole study area, since bedrock depth was only sampled at observation

PAGE 47

47 sites but was not available exhaustively (i.e., at all grid cell locations) throughout the study area. All trees using remote sensing images derived spectral data showed improved accuracy compared to the tree model without spectral data (control). The control tree showed overall accuracy of 68.5% and a Kappa coefficient of 59.3% similar to the SPOT derived model, but substantially lower overall accuracy and Kappa coefficient when compared to Landsat ETM+ and MODIS derived models (Table 2 5). Model II results using the SPOT image and ancillary environmental predictor variables as inputs were similar to model I, except somewhat more confusions occurred between the Gator and Terra Ceia series (Table 2 6). However, substantial improvements were observed predicting the Pahokee and Terra Ceia soil series in model II derived from Landsat ETM+ and MODIS predictor variables when compared to pendants of model I. Furthermore, adding bedrock depth data greatly reduced the relative error when compared to model I. Especially the tree model using the MODIS image showed dramatic changes not only in relative error, but also in the accuracy measures and Kappa coefficients. It showed the best results among all model II trees. Predictor Variabl es Providing Inferences on Soil Series Model I The soil series in this wetland area were controlled by the composition and structure of vegetation species, biomass of vegetation that is decomposed to form soil organic matter, hydroperiods, and hydropatterns. These factors were represented by spectral data and indices; specifically blue, green, red, NIR energy wavelength, NDVI, NDWI, and MSI with the potential to relate to soil series (Fig ure s 2 2 to 2 4). However, the variable importance of spectral derived properties to infer on soil series was

PAGE 48

48 relatively low compared to lithologic, topographic, and hydrologic predictor variables (Table 2 7). The latter properties showed higher variable importance values to predict soil series in WCA 2A discriminating the s ubtle differences among soil series due to ( i) depth of organic horizon (Oa), ( ii) presence/absence of mineral horizon, and ( iii) particle sizes. It is critical to point out that these characteristics are hardly captured directly using passive sensors such as SPOT, Landsat ETM+, and MODIS. Nevertheless, soil class predictions including spectral data in the classification models somewhat improved predictions, suggesting that there is a biotic effect contributing to form different types of soils in this wetland. Although Schmid et al. (2008) derived a lithologic classification soil map using spectral data from remote sensing images (i.e., ASTER and IKONOS) in Ethiopia. This study was conducted only for surface soil (maximum depth of 2 cm) and their study area was mainly composed of bare soil, bare rock, and vegetation which were relatively easy to discriminate spectrally compared to wetland soils. All tree models mainly used the spectral data including vegetation indices and reflectance values as major splitters among nodes, although the contribution scores of spectral data were low. For example, NDWI, RSR, reflectance value of red band, and PC 2 which represents open water areas were used as primary splitters in the tree model using SPOT derived inputs (Fig ure 2 2). This suggests that even though the improvement of predictions was small, spectral data played a major role for root splits in the classification trees. Lithological properties were the most highly ranked predictor variables in all tested models (Table 2 7). Especially isostatic residual gravity anomaly data that

PAGE 49

49 represents the density distribution within the Earths upper crust appeared as a primary split ter dividing soil series between the Lithic and Terric group in most tree models. The Terric group, i.e., Gator and Okeelanta series that have C horizons, had greater isostatic residual gravity values than the Lithic group, i.e., Lauderhill and Pahokee ser ies that are dominated by O horizons. Results suggest that lithological properties reflect the influence of parent material as one of the major soil forming factors. The hydrologic property represented by distance to WCSs was ranked high in all predicti on models indicating its importance. The distance to WCSs reflects the hydrologic pathways within WCA 2A that subsequently control the spatial distribution of nutrients and vegetation communities (Grunwald et al., 2008). Hence, it provided crucial signat ures to distinguish soil series and played an important role as a surrogate splitter. Previous studies in the Everglades have shown that ecosystem features of the wetland can be altered by changes in hydrology (Valk and Warner, 2009; Larsen and Harvey, 201 1). Prediction maps for soil series derived from the developed remote sensing informed tree models are shown in Fig ure 2 5. All of them showed distinct boundaries in the southern part of the study area, which can be explained by the coarse resolution (20 00 m) isostatic residual gravity anomaly data that distinguished the Lithic and the Terric group. All three maps showed similar patterns of soil series across the study area. The Terra Ceia series extensively occurred in the northern area and Pahokee and L auderhill series dominantly occupied the southern area. Maps B and C show the Okeelanta series in the northeastern area, unlike map A, because the former ones confused Terra Ceia and Okeelanta series and misclassified them (Table 2 4). The

PAGE 50

50 continuous raster based soil series prediction maps could be used as a predictor variable to understand the spatial distributions of other soil properties. It also can be used as one of the criteria to stratify sampling sites for future ecological studies. Model II For all prediction models, BR2009 was ranked first with a score of 100% in the variable importance list (Table 2 7). The accuracy of the tree model using MODIS derived input variables amounted to 85.2% overall accuracy and 78.9% Kappa coefficient showing the best performance metrics among all trees in model II (Table 2 5). The relative error of all soil series prediction models greatly decreased using bedrock depth as a predictor. The Lauderhill and Pahokee series were 100% correctly classified for all predictions in model II. This can be explained by the similarity in soil characteristics between the Lauderhill and Pahokee series, except soil depth as a diagnostic criteria discriminating between them. Interestingly the BR1948 (4 classes) did not contribute to any of the prediction models with variable importance lower than 1%. This suggests that the continuous BR2009 dat a showed better capability to derive soil series when compared to the coarse categorical data (BR1948) even though BR1948 and BR2009 showed about 61% agreement. Although the BR2009 data improved the accuracy of the prediction models the data were based on observation points and there was no continuous data layer for bedrock depth available for WCA 2A, which would allow upscaling of the bedrock based prediction models across the whole study area. Lithologic and topographic variables, such as isostati c residual gravity anomaly, magnetic anomaly, bouguer gravity, distance to WCSs and elevation ranked high in all tree models to infer on soil series, whereas remote sensing derived properties which

PAGE 51

51 captured mainly biophysical and phenological properties of vegetation ranked lower. The variable importance scores of spectral data were even lower in model II compared to model I. For example, the variable importance of PC 1 which ranked high among the spectral data in model I (SPOT) was 47.0% (Table 2 7). However, it dropped rapidly to less than 0.06% in model II (data not shown). Moreover, the importance score of NDVI green 25 x 25, that was the only spectral variable ranked among the first seven variables in model I I (SPOT), was less than 1%. Thi s suggests that the variability of soil series can be explained by bedrock/parent material > topographic variables > vegetation properties derived from remote sensing Bodaghabadi et al. (2011) showed similar results suggesting a strong correspondence betw een soil series distribution and topographical properties, although the study was conducted in a hilly piedmont dominant area in Central Iran. Conclusions This study showed that classification trees can be successfully used to predict soil series in WCA 2 A, a prominent wetland of the Everglades, which is a complex and heterogeneous ecosystem in terms of pedologic, topographic, biotic, geologic, and hydrologic factors. The tree models which were developed using spectral data and indices derived from remote sensing images improved prediction power and could explain about > 60% overall accuracy. However, the variable importance of sensor derived properties to predict soil series ranked low suggesting that soil classification was less dependent on biotic/vegetation properties. The overall accuracy increased by about 20% for the SPOT derived tree model I when compared to the control tree (without spectral data inputs). In summary, lithologic, topographic, and geographic properties played a major role to discriminate soil series.

PAGE 52

52 Spectral datasets derived from three remote sensing images with varying spatial resolutions (10, 30, and 250 m) showed different power to predict soil series in the study area. The soil series prediction model using the SPOT image perform ed well as measured by accuracy measures and the Kappa coefficient, followed by the Landsat ETM+, and MODIS images without bedrock depth data (model I). On the other hand, the prediction model using the MODIS image and bedrock depth data (model II) showed the best result. Overall differences among the tree models were small and all of them showed good prediction power in terms of relative error and accuracy metrics. The coarser resolution input variables (e.g., MODIS derived predictor variables at 250 m spa tial resolution) produced rougher patterns of soil series limiting somewhat the ability to capture the underlying finescale variability of soils in WCA 2A. Nevertheless, the grid based soil series maps presented in this study provide a higher spatial resolution than available polygonbased soil series maps in the U.S. through the Soil Survey Geographic Database (SSURGO) Soil Data Mart, NRCS, at a map scale of 1:24,000. As a matter of fact, soil maps of the Everglades, including WCA 2A, were not available since a coarsescale hardcopy map produced in 1948 (Mowry and Bennett, 1948). Given the dramatic hydrologic modifications and needs to assess impacts of external and humaninduced stressors on WCA 2A soil taxonomic information will provide useful informat ion to support future ecological studies. The methodology presented in this study could be transferred to other regions, such as the adjacent wetland WCA 3A in the Everglades, lacking soil taxonomic information. Summary Remote sensing informed soil prediction models have shown success to improve the predictive power and spatial resolution of predictions in upland systems. Yet not

PAGE 53

53 much is known if these models also perform well in wetland ecosystems. The objectives of this study were to (i) develop spectral informed soil taxonomic prediction models and assess their accuracy; (ii) quantify the relationships between soil classes and environmental covariates derived from remote sensing and geospatial sources; and (iii) compar e the effects of spatial resolution (10, 30, and 250 m) of three remote sensing images to delineate soil classes. The study was conducted in a subtropical wetland: Water Conservation Area2A, the Florida Everglades, U.S. Soil series were collected at 108 s ites and three satellite images acquired (i) Satellite Pour lObservation de la Terre (SPOT, 10 m), (ii) Landsat Enhanced Thematic Mapper Plus (ETM+, 30 m), and (iii) Moderate Resolution Imaging Spectroradiometer (MODIS, 250 m). Classification trees were used to predict soil series using spectral data and ancillary environmental datasets. Prediction models derived from spectral data performed better when compared to a control model without spectral inputs. The soil series prediction model derived using SPOT spectral data without bedrock depth and the model derived using MODIS spectral data with bedrock depth showed the best results based on accuracy measures and Kappa coefficient. Results suggest that the variability of soil series can be explained by be drock/parent material > topographic variables > vegetation properties derived from remote sensing

PAGE 54

54 Table 21. Summary of observed soil series in Water Conservation Area2A (2009). Series Family OM thickness a Characteristics Gator Loamy, siliceous, euic, hyperthermic Terric Haplosaprists 41 to 127 cm Loamy material within the control section. Oa (107 127 cm)/Cg1/Cg2/Cg3. Lauderhill Euic, hyperthermic Lithic Haplosaprists 41 to 91 cm Limestone within the control section. Oa1/Oa2/Oa3 (41 64 cm)/2R. Okeelanta Sandy or sandy skeletal, siliceous, euic, hyperthermic Terric Haplosaprists 41 to 127 cm Sandy material within the control section. Oap/Oa (104 127 cm)/C1/C2. Pahokee Euic, hyperthermic Lithic Haplosaprists 66 to 130 cm Limestone within the control section. Oap/Oa1/Oa2 (56 127 cm)/2R. Terra Ceia Euic, hyperthermic Typic Haplosaprists > 130 cm No mineral horizons within the control section. Oap/Oa (127 230 cm). aOM, Organic matter

PAGE 55

55 Table 22. Summary of remote sensing spectral indices. Indices a Formula References EVIb ) 1 ( L Blue Red NIR Red NIR2 1L C C G Huete et al., 1997 Mid infrared index Band7 Landsat Band5 Landsat MidIR MidIR Musick and Pelletier, 1988 MSI NIR MidIR Rock et al., 1986 NDVI Red NIR Red NIR Rouse et al., 1974 NDVI green Green NIR Green NIR Gitelson et al., 1996 NDWI SWIR NIR SWIR NIR Gao, 1996 RSR ) SWIR SWIR SWIR SWIR 1 ( Red NIRmin max min Brown et al., 2000; Chen et al., 2002 SR Red NIR Cohen, 1991; Chen et al., 2002; Colombo et al., 2003 TVI 100 ) 5 0 Red NIR Red NIR (2 / 1 Nellis and Briggs, 1992 aEVI, enhanced vegetation index; MidIR, midinfrared; MSI, moisture stress index; NDVI, normalized difference vegetation index; NDVI green, normalized difference vegetation green index; NDWI, normalized difference water index; NIR, near infrared; RSR, reduc ed simple ratio; SR, simple ratio; S WIR, shortwaveinfrared; TVI, transformed vegetation index. bEmpirical parameters for EVI of MODIS: C1=6.0; C2=7.5; G=2.5; L=1.0

PAGE 56

56 Table 23. Summary of environmental variables employed to build classification tree models to predict soil series. Property Attributes ac Data sources Vegetation Vegetation species Field observations, 2009 2010 Spectral values derived from MODIS U.S. Geological Survey, Land Processes Distributed Active Archive Center, 2010 Reflectance values of each wavelength energy (blue, red, NIR, and MIR); EVI; MSI; NDVI; SR; TVI Spectral values derived from Landsat ETM+ U.S. Geological Survey, Earth Resources Observation and Science Data Center, 2010 Reflectance values of each wavelength energy (blue, green, red, NIR, MIRs, and PAN); MidIR; MSI; NDVI; NDVI green; NDWI; RSR; SR; TVI; 3x3 cells and 7x7 cells for MSI, NDVI, NDVI green, NDWI, RSR, SR, and TVI b ; PC 1 to 4; TC 1 to 6 Spectral values derived from SPOT SPOT Image Corporation, 2009 Reflectance values of each wavelength energy (green, red, NIR, SWIR, and PAN); MSI; NDVI; NDVI green; NDWI; RSR; SR; TVI; 3x3 cells, 9x9 cells, and 25x25 cells for MSI, NDVI, NDVI green, NDWI, RSR, SR, and T VI b ; PC 1 to 4 Topography Elevation (m) U.S. Geological Survey, High Accuracy Elevation Dataset (15 cm vertical accuracy), 2007; derived using spline interpolation method Landform (e.g., slough, ridge etc.) Field observations, 2009 2010 Hydrology Distance to water control structures (km) South Florida Water Management District, Structure, 1997; derived using the Euclidean distance function

PAGE 57

57 Table 23. Continued. Property Attributesac Data sources Lithology Bedrock depth (cm) from 2009 field survey (BR2009) Field observations, 2009 2010 Bedrock depth categorical data from 1948s soil survey map: 1 = less than 91 cm; 2 = 91 cm to 152 cm; 3 = 152 cm to 244 cm; and 4 = more than 244 cm (BR1948) Hydrologic Systems Modeling Division, 1948 Soils Map, pre1996 (Mowry and Bennett, 1948) Potassium concentration U.S. Geological Survey, Digital Data Series DDS 9, Potassium Concentration Data, 1999 Bouguer gravity U.S. Geological Survey, Digital Data Series DDS 9, Bouguer Gravity Anomaly Data, 1999 Isostatic residual gravity anomaly U.S. Geological Survey, Digital Data Series DDS 9, Isostatic Residual Gravity Anomaly Data, 1999 Magnetic anomaly U.S. Geological Survey, Digital Data Series DDS 9, Magnetic Anomaly Data, 1999 Geography x coordinates (m) Field observations, 2009 2010 y coordinates (m) Field observations, 2009 2010 Multiplication of x and y coordinates Field observations, 2009 2010 ; derived using x and y coordinates aETM+, enhanced thematic mapper plus; EVI, enhanced vegetation index; MidIR, midinfrared index; MIR, midinfrared; MODIS, moderate resolution imaging spectroradiometer; MSI, moisture stress index; NDVI, normalized difference vegetation index; NDVI g reen, normali zed difference vegetation green index; NDWI, normalized difference water index; NIR, near infrared; PAN, panchromatic; PCs, principal component scores; RSR, reduced simple ratio; SPOT, satellite pour lobservation de la terre; SR, simple ratio; S WIR, short wave infrared; TC 1 to 6, tasseled cap transformed components; TVI, transformed vegetation index. bMean spectral values within neighborhoods (3 x 3, 7 x 7, 9 x 9, 25 x 25 cells, respectively) surrounding pedon locations using nearest neighbor geospatial method. cAll data are raster based layers except vegetation, landform, and bedrock depth from 2009 field survey.

PAGE 58

58 Table 24. Confusion error matrices for single tree soil series prediction models derived from SPOT, Landsat ETM+, and MODIS images, respectively, plus ancillary environmental data, but without bedrock depth (Model I, total observed data = 108). Tree models Predicted soil series Reference (observed) soil series a Row total (Predicted) Gator N=7 Lauderhill N=5 Okeelanta N=14 Pahokee N=25 Terra Ceia N=57 Control b Gator 5 0 1 6 19 31 Lauderhill 0 5 2 5 3 15 Okeelanta 2 0 11 6 10 29 Pahokee 0 0 0 7 0 7 Terra Ceia 0 0 0 1 25 26 SPOT Gator 7 0 2 2 5 16 Lauderhill 0 5 0 0 0 5 Okeelanta 0 0 12 1 10 23 Pahokee 0 0 0 22 11 33 Terra Ceia 0 0 0 0 31 31 Landsat ETM+ Gator 7 0 0 1 5 13 Lauderhill 0 5 0 2 2 9 Okeelanta 0 0 14 1 16 31 Pahokee 0 0 0 20 7 27 Terra Ceia 0 0 0 1 27 28 MODIS Gator 7 0 1 2 6 16 Lauderhill 0 5 0 4 2 11 Okeelanta 0 0 13 5 15 33 Pahokee 0 0 0 12 6 18 Terra Ceia 0 0 0 2 28 30 aNumbers in the major diagonal indicate correctly classified soil series cases. bTree model without spectral data.

PAGE 59

59 Table 25. Summary of single tree models to predict soil series with and without bedrock depth in the Water Conservation Area2A, Everglades, Florida. Tree Models Methods Relative error Average producers accuracy Average users accuracy Overall accuracy K coefficient of agreement Without Bedrock (Model I) Control a Gini 0.75 64.4 56.7 49.1 36.2 SPOT Twoing 0.69 85.6 72.5 71.3 61.1 Landsat ETM+ Gini 0.80 85.5 65.0 67.6 57.2 MODIS Entropy 0.83 78.0 57.7 60.2 47.7 With Bedrock (Model II) Control a Gini 0.42 84.5 74.2 68.5 59.3 SPOT Twoing 0.45 87.3 76.3 69.4 60.7 Landsat ETM+ Gini 0.55 78.7 73.2 76.9 66.0 MODIS Entropy 0.35 93.3 82.5 85.2 78.9 aTree model without spectral data.

PAGE 60

60 Table 26. Confusion error matrices for single tree soil series prediction models derived from SPOT, Landsat ETM+, and MODIS images, respectively, plus ancillary environmental data including bedrock depth (Model II, total observed data = 108). Tree models Predicted soil series Reference (Observed) data a Row total (Predicted) Gator N=7 Lauderhill N=5 Okeelanta N=14 Pahokee N=25 Terra Ceia N=57 Control b Gator 6 0 1 0 19 26 Lauderhill 0 5 0 0 0 5 Okeelanta 1 0 13 0 13 27 Pahokee 0 0 0 25 0 25 Terra Ceia 0 0 0 0 25 25 SPOT Gator 7 0 1 0 23 31 Lauderhill 0 5 0 0 0 5 Okeelanta 0 0 13 0 9 22 Pahokee 0 0 0 25 0 25 Terra Ceia 0 0 0 0 25 25 Landsat ETM+ Gator 5 0 1 0 6 12 Lauderhill 0 5 0 0 0 5 Okeelanta 1 0 7 0 10 18 Pahokee 0 0 0 25 0 25 Terra Ceia 1 0 6 0 41 48 MODIS Gator 7 0 1 0 4 12 Lauderhill 0 5 0 0 0 5 Okeelanta 0 0 13 0 11 24 Pahokee 0 0 0 25 0 25 Terra Ceia 0 0 0 0 42 42 aNumbers in the major diagonal indicate correctly classified soil series cases. bTree model without spectral data.

PAGE 61

61 Table 27. Variable importance of single tree models to predict soil series with and without bedrock depth in the Water Conservation Area2A, Everglades, Floridaa. Rank Control b SPOT Landsat ETM+ MODIS Without Bedrock (Model I) 1 Elevation (100) Isostatic gravity (100) Isostatic gravity (100) Magnetic anomaly (100) 2 Magnetic anomaly (90.4) Magnetic anomaly (98.1) Magnetic anomaly (98.4) Elevation (83.9) 3 Isostatic gravity (73.6) Bouguer gravity (88.9) Elevation (93.3) Distance to WCSs (77.5) 4 Distance to WCSs (67.5) Elevation (82.9) Bouguer gravity (92.4) Bouguer gravity (72.6) 5 Bouguer gravity (64.5) Distance to WCSs (75.1) Distance to WCSs (92.0) Isostatic gravity (72) 6 Potassium concentration (13.5) PC1 (47.0) PC1 (51.5) NIR (45.0) 7 Vegetation species (9.3) TVI 25X25 (3.8) MidIR (8.8) Red (23.7) With Bedrock (Model II) 1 BR2009 (100) BR2009 (100) BR2009 (100) BR2009 (100) 2 Elevation (76.5) Magnetic anomaly (43.9) Magnetic anomaly (51.5) Elevation (72.8) 3 Distance to WCSs (64.6) Isostatic gravity (43.7) Isostatic gravity (51.4) Magnetic anomaly (66.5) 4 Magnetic anomaly (64.1) Bouguer gravity (37.8) Distance to WCSs (48.6) Distance to WCSs (64.7) 5 Bouguer gravity (55.7) Distance to WCSs (35.3) Elevation (47.6) Isostatic gravity (56.0) 6 Isostatic gravity (53.2) Elevation (34.4) Bouguer gravity (47.3) Bouguer gravity (47.4) 7 Potassium concentration (10.7) NDVI green 25X25 (0.8) PAN (1.7) Red (16.2) aBR2009, bedrock depth surveyed in 2009; MIR, midinfrared; NDVI green, normalized difference vegetation green index; NIR, near infrared; PAN, panchromatic ; PCs, principal component scores; TVI, transformed vegetation index; WCSs, water control structures. bTree model without spectral data.

PAGE 62

62 Figure 21. Distribution of soil series at sampling locations within the Water Conservation Area2A, Everglades, Florida, U.S.

PAGE 63

63 Figure 22. Single tree classification model (14 terminal nodes) to predict soil series using the SPOT image, but without bedrock depth as predictor, in the Water Conservation Area2A. The tree shows the splitting criteria on top of each node. Terminal nodes (TNodes) are shown in gray and other nodes in white.

PAGE 64

64 Fig ure 2 3. Single tree classification model (14 terminal nodes) to predict soil series using the Landsat ETM+ image, but without bedrock depth as predictor, in the Water Conservation Area2A. The tree shows the splitting criteria on top of each node. Terminal nodes (TNodes) are shown in gray and other nodes in white.

PAGE 65

65 Fig ure 2 4. Single tree classification model (12 terminal nodes) to predict soil series using the MODIS image, but without bedrock depth as predictor, in the Water Conservation Area2A. The tree shows the splitting criteria on top of each node. Terminal nodes (TNodes) are shown in gray and other nodes in white.

PAGE 66

66 Fig ure 2 5. Prediction maps of soil s eries based on single tree model without geographic coordinates using A) SPOT, B) Landsat ETM+, and C) MODIS image derived spectral input variables with geospatial environmental ancillary data ( Stripes in map B were caused by permanent scanline corrector failure (SLC off) of Landsat Enhanced Thematic Mapper Plus sensor since May 2003).

PAGE 67

67 CHAPTER 3 SOIL PHOSPHORUS AND NITROGEN PREDICTIONS ACROSS SPATIAL ESCALATING SCALES IN AN AQUATIC ECOSYSTEM Overview There is an increased use of remote sens ing (RS) in digital soil mapping (DSM) due to its capabilities to directly or indirectly infer on biotic properties (Grunwald, 2009) Prominent conceptual models, such as the SCORPAN model with soil (S), climate (C), organism (O), relief (R), parent material (P), age (A), and space (N) factors (McBratney et al., 2003) and the STEP AWBH model with soil (S), topog raphy (T), ecology (E), parent material (P), atmosphere (A), water (W), biota (B), and human (H) factors (Grunwald et al., 2011) use soil environmental factors to predict soil properties or classes. These empirical factorial models are rooted in the paradi gm that combinations of soil forming factors allow to infer on a specific soil property of interest using various statistical and geostatistical methods. Remote sensing data, which provide dense spectral grids over a large area, have been used widely to populate the O factor in the SCORPAN model or B factor in the STEP AWBH model, because of their cost effectiveness, high spatial resolutions capturing the variability of landscape features, and capabilities to deduct various potential soil forming proper ties such as biomass, crop stress, leaf area index, chlorophyll content, and more (Tucker, 1979; Mirik et al., 2005; Huang et al., 2007; Schmid et al., 2008; Rivero et al., 2009). Mulder et al. (2011) comprehensively reviewed the use of RS as a primary data source (e.g., in sparsely vegetated areas) or secondary data source (e.g., in densely vegetated areas or urban areas) for soil properties mapping. Considerable research has been conducted on the use of RS images to assess soil properties, and many of the studies have shown accuracy improvement of soil property prediction models utilizing a

PAGE 68

68 variety of RS imagery compared to traditional soil assessment without RS (Peng et al., 2003; Sullivan et al., 2005; Simbahan et al., 2006; Selige et al., 2006; Rivero et al., 2007a; Schmid et al., 2008) However, the effect of different spatial resolutions of RS images on developing digital prediction models for soil properties is relatively unknown. The implications of varying grain sizes (resolutions) on soil predictions are profound and entail possible scaling effects (Vasques et al., 2012), which impact the prediction capability of DSM. The selection of an appropriate spatial resolution of RS images could save costs, reduce the computational load, and capture important landscape features. Aquatic ecosystems are characterized by a high diversity of biological and nonbiological features. Subtropical wetlan d soils, especially, have shown higher complexity than upland soils due to the intricate composition of water, detritus, macrophytes, and periphyton. There are increasing demands for accurate and precise data to predict biogeochemical properties in these heterogeneous wetland soils, and a RS based modeling approach has much value for wetlands, where field sampling is challenging and siteaccess is limited. However, fewer RS supported DSM studies have been conducted in wetland areas when compared to upland s ystems. Moreover, RS supported prediction models have shown good prediction results in homogenous landscapes, but they have shown limitations in heterogeneous areas for certain methods such as multiple linear regression (Odeha et al., 1994; Mulder et al., 2011) Therefore, a critical question i s: Does a remote sensing image provide sufficient information for the prediction of soil biogeochemical properties in a heterogeneous wetlands ecosystem?

PAGE 69

69 The Everglades in Florida, U.S., which have heterogeneous environmental features, have experienced a cascade of ecological changes because of the extensive water drainage and anthropogenic nutrient enrichment from the adjacent Everglades Agricultural Area. Alterations of the ecosystem components are interrelated and form a highly complex freshwater system. The changes of periphyton and vegetation community, caused by high phosphorus (P) loading, have especially affected the biogeochemical processes in the Everglades (Noe and Childers, 2007). The impacts of the eutrophication on biota and biogeochemi stry, such as the alteration of the nitrogen (N) cycle, of this aquatic ecosystem have been comprehensively studied (Inglett et al., 2004, 2011), culminating in restoration stretching over more than two decades (Chimney and Goforth, 2006). One of the resto ration goals of the Comprehensive Everglades Restoration Plan (CERP), which was approved in the Water Resource Development Act of 2000, is maintaining or reducing long term average soil total phosphorus (TP) less than 400 mg kg1. The control of concentration of elements (e.g., P, N) is the primary goal for most restoration plans. However, the quantification of an actual amount of storage of nutrients preserved in wetland soils is critical to assess the effectiveness of restoration efforts because of the legacy effect of nutrients in soils. Therefore, the objectives of this study were to (i) assess the differences of the spatial distribution of soil TP and total nitrogen (TN) concentrations and stocks, (ii) develop prediction models for soil TP and TN utilizing RS images and environmental ancillary data at escalating grain resolution (scale), and (iii) elucidate the effects of different spatial resolutions of RS images on inferential modeling of soil properties in a subtropical wetland.

PAGE 70

70 Methods and Materials Study Area The Everglades is divided by levees and canals into hydrologic compartments that include the Water Conservation Areas (WCAs) and the Everglades National Park. The study area, WCA 2A, covers about 418 km2 and is located in the northern part of the Everglades, Florida, U.S. The climate of the study area is subtropical, with 1,297 mm (Everglades Depth Estimation Network, EDEN, 2009) and 23 C (National Climatic Data Center, 2009) mean annual precipitation and temperature, respectively. His torically, the Everglades was an oligotrophic ecosystem characterized by low surface water concentrations of P and other nutrients with over 90% of water inputs from rainfall (Chimney and Goforth, 2006) However, extensive drainages and the construction of levees and canals have resulted in a degradation of water resources of the Everglades. Changes in hydrology combined with increased nutrient loading have led to significant changes in the wetlands ecosystem structure and function (Noe et al., 2001; Bruland et al., 2007; Rivero et al., 2007b) The soil and plant communities in WCA 2A are influenced by hydroperiod (wet and dry periods), water depth, f ire, and other regime (Bernhardt and Willard, 2009) The dominant soils in WCA 2A are highly decomposed Haplosaprists (Histosols). Typical landforms are primarily ridges consisting of sawgrass ( Cladium jamaicense ) and openwater sloughs dominated by water lily ( Nymphaea odorata) with scattered tree islands of various sizes. However, the vegetation patterns have been affected by nutrient influx, especially P, which has been associated with the expansion of cattail ( Typha domingensis ) in the study area (Jensen et al., 1995; Wu et al., 1997) Elevatio n ranges

PAGE 71

71 from 0.8 to 4.3 m above mean sea level in WCA 2A (United States Geological Survey, USGS, High Accuracy Elevation Data, HAED, 2007). Field Sampling and Soil Property Measurement A total of 108 soil cores spread over the whole study area were collec ted in July and October 2009 and mid March 2010 using airboats. Soil collection was based on a random stratified sampling design using historic soil, hydrologic, and ecological sites identified in 2003 (compare Rivero et al., 2007b). Because of dense veget ation, lack of site accessibility by airboats, and other field conditions, only 48 sampling points of the pre selected sampling sites were revisited. The remaining samples were randomly collected nearest to preselected sampling points. The soil samples were collected using a 10 cm diameter stainless steel coring device to a depth of 10 cm beneath the soil surface at each point. A detailed description of the sampling design and the method of soil core collection can be found in Rivero et al. (2007b). Sampli ng locations within the study area are shown in Figure 3 1. Any living organic materials were removed from samples. The samples were ovendried at 70 C, and bulk density (BD) was measured. The soil samples were analyzed for TP utilizing the association o f official analytical chemists (AOAC) 978.01 method and TN utilizing the LECO combustion method (LECO TruSpec, MI, USA). The amount of TP and TN stored per unit land area (g m2) for a given depth was calculated using the measured BD and the measured nutri ent concentrations on a mass basis (mg kg1 for TP and g kg1 for TN, respectively) at each sampling point according to: c z N Nconc stock (3 1)

PAGE 72

72 where Nstock is nutrient stocks (g m2), Nconc. is nutrient concentrations (mg kg1 for TP; g kg1 for TN), is BD of the soil sample at each point (g cm3), z is depth (10 cm), and c (0.01 for TP; 10 for TN) is a constant for unit conversion. Spectral Indices Derived from Remote Sensing Imagery Three satellite images were selected to investigate the effects of different spatial resolutions on developing prediction models for soil properties: Moderate Resolution Imaging Spectroradiometer (MODIS), Landsat Enhanced Thematic Mapper Plus (ETM+), and Satellite Pour lObservation de la Terre (SPOT) images. T he closeness to field sampling dates, cloud and noise (including stripes of Landsat ETM+) images, and seasonality (e.g., winter images are preferred over summer images; compare Rivero et al., 2009) were considered to select suitable satellite images for the study. The MODIS image (MOD13Q1; February 2010) obtained from Land Processes Distributed Active Archive Center, USGS Earth Resources Observation and Science (EROS) Center has blue, red, near infrared (NIR), and midinfrared (MIR) bands with 250 m spatial resolution. The Landsat ETM+ image (path: 15/ row: 42; February 2010) obtained from the USGS EROS Center has band 1 ( blue ; 450 515 nm) band 2 ( green; 525 605 nm) band 3 ( red ; 630 690 nm) band 4 ( NIR; 750 900 nm) band 5 (MIR; 1,550 1,750 nm), and band 7 ( MIR ; 2,090 2,350 nm) with 30 m spatial resolution and a band 6 ( thermal infrared; 10,400 12,500 nm) with 60 m spatial resolution. The SPOT image (January 2009) donated by Planet Action, a nonprofit ASTRIUM GEO initiative, has band 1 ( green; 500 590 nm) band 2 ( red ; 610 680 nm) and band 3 ( NIR; 780 890 nm) with 10 m spatial resolution and a band 4 (s hortwaveinfrared (SWIR) ; 1,580 1,750 nm) with 20 m spatial resolution. Both the Landsat ETM+ and SPOT images have a panchromatic band with 15 m and 2.5 m

PAGE 73

73 spatial resolution, respectively. Although the sensors have different spectral resolutions, the ranges of the bands of each image are not dramatically different. For instance, the spectral range of the red band of MODIS is from 620 to 67 0 n m, Landsat ETM+ is 610 to 68 0 n m, a nd SPOT is 630 to 69 0 n m. Thus, this allowed the study to focus on testing the effects of different spatial resolutions among the sensors. All images were projected to the Universal Transverse Mercator (UTM) map projection (Zone: 17; Datum: World Geographic System, WGS 84) and geometrically rectified with USGS digital orthophoto quadrangles (DOQQ) using ERDAS Imagine 2010 (Earth Resource Data Analysis System Inc., Atlanta, GA). Root mean square errors (RMSE) assessing geometrical accuracy were less than 0.5 pixels for all images. Each of the remote sensing images bands contains some amount of spectral noise, and the noise can be effectively removed by taking the ratio of the reflectance of different spectral bands to derive socalled spectral indices. The spectral indices have been widely used and have shown improved capability to infer on soil properties (Chen, 1996; Huete et al., 1997; Fern ndez Buces et al., 2006; Eldeiry et al., 2008; Vrieling et al., 2008). The derived spectral vegetation indices for different satellite images, depending on their spectral bands, were the following: (i) from MODIS Enhanced Vegetation Index (EVI), Moistur e Stress Index (MSI), Normalized Difference Vegetation Index (NDVI), Simple Ratio (SR), and Transformed Vegetation Index (TVI); (ii) from Landsat ETM+ Mid infrared index (MidIR), MSI, NDVI, Normalized Difference Vegetation Green Index (NDVI green), Norma lized difference water index (NDWI), Reduced simple ratio (RSR), SR, and TVI; (iii) from SPOT MSI, NDVI, NDVI green, NDWI, RSR, SR, and TVI. The formulas to derive indices are given in Table 3 1. A

PAGE 74

74 principal component analysis (PCA) was performed to der ive principal component (PC) scores for the SPOT (bands 1 to 4), Landsat ETM+ (bands 1 to 5 and 7), and the MODIS (bands 1 to 4) images. By rotating the original variable coordinate system to new orthogonal axes (principal axes), a PCA reduces the number o f correlated variables into a smaller number of uncorrelated variables. Tasseled cap (TC) transformation was performed with the Landsat ETM+ image to enhance the spectral signatures of soil brightness, greenness, and wetness (Kauth and Thomas, 1976). Furth ermore, the mean spectral values within neighborhoods were computed using the nearest neighbor focal geospatial method and were included as input variables (Table 3 2). For instance, the mean spectral values of 25 x 25 pixel counts of SPOT (10 m) represent ing a 250 m x 250 m neighborhood were included for comparison of the spatial resolution effect between SPOT and MODIS images. Table 3 2 shows the summary of input variables for the development of prediction models. Derived spectral indices and reflectance values from each image were extracted for each sampling point (i.e., x and y locations) and combined with sitespecific soil properties using ArcGIS 10 (Environmental Systems Research Institute ESRI Inc., Redlands, CA, USA). Ancillary Environmental Pre dictor Data Various environmental properties including topography, hydrology, and lithology were considered as factors to infer on spatially explicit nutrient distributions in WCA 2A. The environmental data layers included the elevation, distance to water control structures (WCSs), and geophysical properties. The continuous, raster based elevation data layer was produced using point observation data surveyed by USGS for the Everglades region (USGS, HAED, 2007). The point observation data had a vertical accu racy of 15 cm, and it contained 2,600 points within WCA 2A. The distance to

PAGE 75

75 WCSs was derived using the Euclidean distance function representing hydrology effects associated with soil nutrient distributions. Coarse resolution (2,000 m by 2,000 m) geophys ical property data layers were delineated to infer on parent material. The geophysical data layers included potassium concentration gammaray data, bouguer gravity anomaly data, isostatic residual gravity anomaly data, and magnetic anomaly data (USGS, Digital Data Series DDS 9, 1999). Bouguer gravity anomaly and isostatic residual gravity anomaly data contain similar information about the density distribution of the Earths upper crust, and magnetic anomaly data reflects the distribution of iron (Fe) minera ls in t he rocks of the Earths crust. All of the environmental predictor data layers were prepared as layers with spatial resolutions of 10 m, 30 m, and 250 m, respectively, to match the resolution of the satellite images to develop prediction models. All of the geographic information system (GIS) data processing and data value extractions from each environmental data layer were done in the same fashion as spectral indices above using ArcGIS 10 (ESRI Inc., Redlands, CA, USA). Data Analyses Block Kriging (BK ) and three different Random Forest (RF) models were built to predict soil TP and TN in WCA 2A. The RF models were developed with (i) SPOT (RFSPOT), (ii) Landsat ETM+ (RFETM+), and (iii) MODIS (RFMODIS) images derived spectral input variables complemented with ancillary environmental predictor data. Block Kriging is a geostatistical method that estimates the value of a random variable at unsampled locations using a block as its support, and it enables prediction of the block average using the point values (Heuvelink and Pebesma, 1999). It uses only field observation data. In contrast, RF is a regression and data mining method relating input

PAGE 76

76 variables (described in Table 3 2) to output or predictor variables (soil TP and TN, respectively). In classification and regression trees (CART), tree structures are generated through recursively partitioning the data into a number of groups (Breiman, 1984) The RF is an ensemble (or set, collection) of CART like tree s that are built on bootstrap samples of the data, using a randomized subset of predictor variables at each tree (Breiman, 2001) A large number of tr ees are grown to maximum size without pruning (Breiman, 2001; Prasad et al., 2006) and they are aggregated by averaging the trees to give one single prediction; thus, the RF has shown low bias and low variance (Daz Uriarte and de Andrs, 2006; Wiesmeier et al., 2011) The RF has several advantages over other statistical methods in terms of its ability to handle skewed data and outliers, dealing with a large number of predictor vari ables, and modeling of interaction effects among the variables (Breiman, 2001; Liaw and Wiener, 2002; Strobl et al., 2009; Seni and Elder, 2010) Moreover, the RF algorithm uses out of bag data (i.e., the data are not in the bootstrap sample) to estimate the error at each bootstrap iteration. Henc e, it practically eliminates the need of an independent validation process and validation dataset (Grimm et al., 2008) A major disadvantage of RF is a limited interpretation of the forest (black box). The procedure does not allow a user to i nvestigate the structure of each tree. Instead, a variable importance is computed to reflect the contribution of each predictor variable to build a forest in RF. The variable importance score for each predictor variable is estimated by looking at how much the prediction error increases when the variable is altered, while all other variables are the same (Liaw and Wiener, 2002)

PAGE 77

77 The RF is not excessively sensitive to the changes of model parameters such as the overall number of trees in the forest (ntree) and the number of randomly preselected predictor variables for each split (mtry) unless they are very lo w (Liaw and Wiener, 2002; Strobl et al., 2009) In this study an iteration approach was used to develop a best fit model for soil TP and TN by varying the ntree and mtry parameters within expected ranges. As a result, 1000 of ntree and 7 of mtry was chosen for the study based on model accuracy (RMSE) and percent of variance explained. Mor e complex models such as Regression Kriging could not be built since there were no significant spatial autocorrelations in TP and TN residuals of the RF models. Block Kriging was performed using the ISATIS geostatistical analysis package (ISATIS v8.02, Ge ovariances Inc., France) with a block size of 10 m by 10 m. The RF analysis was performed using the randomForest package (Liaw and Wiener, 2002) and the raster package (Hijmans and van Etten, 2012) was used to produce a map in the R statistical language (R Development Core Team, 2012). Maps of each RF model were produced using the same spatial resoluti on as the respective RS images. For instance, the prediction map of the RFSPOT model had 10 m spatial resolution, because SPOT images have a 10 m grid resolution. Verification of Model Output The coefficient of determination (R2) and the root mean squared error (RMSE) were derived to assess the quality of the soil TP and TN prediction models using cross validation. Moreover, entropy was analyzed to elucidate the differences of spatial variations of predicted soil TP and TN among the prediction maps derived from the RF models in different modes (RFSPOT, RFETM+, and RFMODIS). Entropy is a metric developed from information theory, which was originally established for signal processing. In the

PAGE 78

78 context of this study, entropy was used to quantify information that was contained in RF prediction maps. The relationship between the information theory in signal processing applied to soil maping is well explained in Bishop et al. (2001). The Shannons entropy is given by p k k kH1) log( (3 2) where H is the amount of information (entropy), p is a categorical size of a random variable x, and k is a cell probability of a variable x with k = 1, 2, p k = 1) ( Shannon, 2001) The which is based on a probability mass function, is unknown in practice. Thus, the maximum likelihood (ML) esti mator was used to estimate entropy (Eq. ( 3 3)) in this study. p k k kH1) log( (3 3) where H and k are entropy and pixel probability, respectively, estimated using ML method from observed pixel counts ( Hausser and Strimmer, 2009) An area of 750 m by 750 m was chosen from high and low concentration areas of soil TP and TN for the comparison of entropy. The high concentration area was selected from the dense cattail area near the S 10 structures for soil TP. Typically, this area has shown high soil TP concentrations (DeBusk et al., 1994; Richardson et al., 1997; Wu et al., 1997; Grunwald et al., 2004; Rivero et al., 2009). The low concentration area was selected from the interior of the marsh where TP concentrations were shown to be smaller than 400 mg kg1 based on the RFMODIS model result. High (> 29 g kg1) and low (< 27 g kg1) concentration areas of soil TN were delineated in the same fashion as the soil TP. Each pixel value (raster) within the 750 m by 750 m area was converted to a

PAGE 79

79 point value using ArcGIS 10 (ESRI Inc., Redlands, CA, USA) to estimate entropy. For example, a total of 5,625 pixels in the RFSPOT prediction map w ithin 750 m by 750 m area were converted to points, and it yielded 5,625 point values. The point values were grouped using a bin size of 100 for TP (e.g., 100 to 200, 200 to 300, 300 to 400 mg kg1, etc.) and 1 for TN (e.g., 21 to 22, 22 to 23, 23 to 24 g kg1, etc.) based on the observation ranges and this resulted in 16 bins for TP (i.e., p = 16 for TP since TP ranges from 165.4 to 1694 mg kg1 in Table 3 3) and 18 bins for TN (i.e., p = 18 for TN) in this study. The point counts of each bin were used to estimate H and k in Eq. (3 3 ). As a measure of a complexity or a measure of information in an image, entropy can quantify how prediction maps differ from each other in terms of its spatial variability. The entropy analysis was performed using the entropy package ( Hausser and Strimmer, 2009) in R. Results and Discussion Spatial Distribution of Soil Properties: Concentrations and Stocks Des criptive statistics of soil properties are shown in Table 3 3. The mean of the BD in WCA 2A was 0.11 g cm3, which indicates high organic matter content. Slough areas had a mean BD of 0.13 g cm3, slightly higher than ridge areas with a mean BD of 0.10 g c m3 in WCA 2A. This was expected since the vegetation communities were different in sloughs and ridges. In general, the BD showed a homogeneous distribution throughout the study area. Figure 3 1 shows the spatial distribution of observed TP and TN in WCA 2A. Total P showed large variability (observed concentration data range of 1,529 mg kg1), which was also reported by other similar studies in WCA 2A (DeBusk et al., 1994;

PAGE 80

80 Grunwald et al., 2004; Rivero et al., 2007b) The mean soil TP concentration was 601.4 mg kg1 and higher than TP means observed in 2003 ( 550.7 mg kg1) in the same study area (Rivero et al., 2007b) However, it was lower when compared to the mean values of soil TP observed in 1990 (DeBusk et al., 1994) and 1998 (Grunwald et al., 2004) in WCA 2A. The spatial distributions of soil TP concentrations in WCA 2A were somewhat resembling those presented by (DeBusk et al., 1994; Grunwald et al., 2004, 2008; Rivero et al., 2007a, b) for precious years. In general, high concentrations of TP were found near S 10 structures and low concentrations were found in the interior marsh. Total N concentrations showed smaller var iability across WCA 2A when compared to TP. The mean of TN was 29.7 g kg1. It was slightly higher in 2009 when compared to mean TN values of 28.0 g kg1 observed in 1990 (DeBusk et al., 1994) and 27.1 g kg1 (Rivero et al., 2007b) The spatial observations of TN were more heterogeneous across WCA 2A compared to soil TP. The spatial distribution of nutrient concentrations and the one of nutrient stocks can vary according to the measured BD at each sampling point. Stocks of soil TP and TN stored in the study area were estimated since stocks reflect natural pedons and ecological perspective more realistically than concentrations according to Bruland et al. (2007) and Grimm et al. (2008). The observed mean value of soil TP and TN stocks was 6.4 g m2 and 323.8 g m2, respectively, in 2009 (Table 3 3) It was slightly elevated when compared to 6.1 g m2 for TP and 298.1 g m2 for TN reported in 2003 by Rivero et al. (2007b) The median value of 5.1 g m2 of soil TP was also slightly elevated when compared to 4.8 g m2 that was reported by Rivero et al. (2007b).

PAGE 81

81 Soil nutrient concentrations and stocks were s patially correlated (R2 of 0.63) due to low and homogeneous patterns of BDs throughout the study area. It is because soils in WCA 2A were composed of highly decomposed organic matter with minimal mineral content and less calcite than other hydrologic units of the Everglades, such as the Everglades National Park (Gleason and Spackman, 1984) Prediction maps for soil TP and TN in concentration and stock units using RF with SPOT imagederived spectral input variables and other environmental predictor variables for the study area are shown in Figure 3 2. Generally, high concentration areas showed high stocks. However, stocks showed somewhat more contrast between low and high values than concentrations because of the larger variability of both TP and TN. The Everglades have shown a wide range of BD and organic matter due to the composition of the organic peat and inorganic marl soils (Osborne et al., 2011a, b). Bruland et al. (2007) found areas which had increased concentrations of soil TP that showed decreased soil TP stocks between 1992 and 200 3 in WCA 3, adjacent to WCA 2A. Also, Grimm et al. (2008) found that some of their study area, Barro Colorado Island in Panama a tropical forest, had low carbon concentrations but high carbon stocks because of high BD caused by nonsoil components such as pebbles. It suggests that stocks have a value to be considered for restoration assessment, although our study area did not show a dramatic difference between nutrient concentrations and stocks. Soil Properties Prediction Models: Univariate and Multivaria te Methods Spatial predictions The statistical assessment of the models for soil TP and TN concentrations using BK and RF with SPOT, Landsat ETM+, and MODIS derived input variables combined with environmental predictor variables are shown in Table 3 4. The observed versus

PAGE 82

82 predicted values for the BK and RF models are shown in Figure 3 3. Block Kriging that used only field observation data showed better TP prediction results with a R2 of 0.60 than TN with a R2 of 0.30. Soil TP showed stronger spatial autocor relation than soil TN (Table 3 5). In comparison, TP in WCA 2A showed strong spatial autocorrelation patterns and high TP concentrations in proximity of WCSs. Soil TN had a less pronounced spatial structure in the semivariogram and showed weak spatial dependence (nugget to sill ratio > 60%) that can be explained by the biochemical behavior of N, which has various gaseous forms and complex cycle processes creating a heterogeneous patchwork of hotspots. Figures 3 4 and 3 5 show the prediction maps of soil TP and TN using BK and RF models. Overall, the spatial prediction results are consistent with findings documented in Grunwald et al. (2004) and Rivero et al. (2007b). However, the area of high soil TP concentrations expan ded further into the southern direction of S 10 structures and east south direction of the S 7 pump station. For example, Rivero et al. (2007b) reported that surface soils approximately 5 km south of S 10 structures exceeded 450 mg kg1, but our results showed the expansion of the elevated TP area up to approximately 8.5 km south of S 10 structures. Moreover, soils about 8 km south of S 10 structures exceeded 500 mg kg1, which is the established threshold for P impacted status in CERP. Also, elevated soil TP concentrations were observed in the northern part of WCA 2A. The expansion of areas high in TP could be explained by increased P mobilizations. Phosphorus in the system is mineralized because of organic matter decomposition or fire and translocated downstream (Osborne et al., 2011a) The direction of the mobilized P translocation is facilitated by an over all north to south flow of

PAGE 83

83 surface water in WCA 2A. Another possible explanation of the expansion of P impacted areas is continuous P loading into the system. Although the TP concentrations of inflow water into WCA 2A showed a decreasing trend from 2003 to 2010, the TP loading did not (Figure 3 6). Moreover, there have been excessive P loadings into the marsh, especially in 2004 and 2009 (> 40 metric tons). Phosphorus loading inflow exceeded P outflow significantly in 2008 and 2009 suggesting that P accumul ated within WCA 2A during that time preceding our field sampling. The RF models using various spectral and environmental data greatly increased the prediction results for all soil properties (R2 > 0.90) when compared to BK. All RF based prediction maps of soil TP have high concentrations in the northeast and west, and low concentrations in the interior of the study area similar to BK predictions (Figure 3 4). The prediction maps of RF models also show the pronounced area of P impact about 8 km south of the S 10 structures due to P loading inflow into the wetland. The status of P affects the N cycle, and previous studies observed that denitrification rates increased as soils become more P enriched in the Everglades (Craft and Richardson, 1993; Inglett et al., 2011) This can explain the predicted N patterns that are low in the northern portion and high in the southeastern portion of the study area. Variable importance of Random Forest models for TP predictions The RS image derived spectral variables ranked high for most of the RF models compared to lithologic, topographic, and hydrologic predictor variables for TP predictions (Table 3 6). Among various predictor variables, TVI, SR, NDVI green, and NDVI ranked high in the RFSPOT model. All these indices use the same spectral wavelength regions (i.e., red and NIR), except NDVI green. The high rank of spectral indices that use red and NIR bands was expected since the major signals on vegetation

PAGE 84

84 conditions are carried in the red and NIR reflectance (Rouse et al., 1974; Cohen, 1991; Chen, 1996). It is interesting that TVI showed strong predictive power for soil TP (Table 3 6). Chen (1996) argued that it amplif ies noise of remote sensing images when a spectral index takes the square or square root of the reflectance in the calculation (e.g., TVI takes the square root in the calculation, Table 3 1). The high rank of TVI could be explained by previous fires in WCA 2A. Fire is one of the major stressor in natural ecosystems and it has altered the vegetation community, nutrient forms, and nutrient storages in the study area. Our results suggest that the spatial variation of the vegetation caused by fires, directly or indirectly, could be captured by TVI. A previous study conducted by Nellis and Briggs (1992) showed that high variability of TVI inferred the spatial variation of above ground biomass with different burning frequency in Kansas. The simple ratio also showed good predictive power for soil TP. The simple ratio, the first true vegetation index developed in 1969, has not been widely used to predict soil properties when compared to the NDVI, despite the fact that they are both functionally equivalent in terms of carrying information on vegetation vigor (Chen, 1996). Our results indicate that SR was preferred over NDVI for soil TP predictions in the RFSPOT model. Tasseled cap transformed variable 2 (TC 2) ranked first in the RFETM+ model (Table 3 6). The TC 2 contains greenness information of landscape features, specifically vegetation and periphyton. The greenness carries enhanced spectral signatures of the presence/absence and density of green vegetation of the Landsat ETM+ image. It showed high reflectance values in a densely vegetated area (e.g., northeast area of WCA 2A). Unlike the TC 2, TC 1 (brightness) and TC 3 (wetness) did

PAGE 85

85 not perform well for soil TP predictions. The TC 1 and TC 3 showed less variation than TC 2 due to the high moisture content of the soils throughout the study area. The variation of wetness was small because the whole study area was covered by water, except ridges of tree islands. The high water coverage resulted in high brightness values throughout WCA 2A and a small variation of brig htness. The NDVI green index ranked high in the RFETM+ model. This index is similar to NDVI except that NDVI green encompasses green wavelength reflectance instead of red reflectance. It has been shown that the spectral reflectance around 520 nm (i.e., green band) is more sensitive to the chlorophyll concentrations in a wide range of chlorophyll variations, and it estimates the concentration more precisely than using the red band (Gitelson et al., 1996; Rivero et al., 2009) Distance to WCSs also ranked as an important variable in the RFETM+ and RFMODIS models among other predictor variables (Table 3 6). The EVI ranked as the best predictor variable for the RFMODIS model (Table 3 6). This index provides improved sensitivity in high biomass areas with less atmospheric in fluence than NDVI (Jiang et al., 2 008) Principle components variables that had the ability to discriminate vegetated areas, which included red and NIR bands, also ranked high. The major difference between RFMODIS and other RF models was the ranking of nonRS derived input variables. For instance, elevation ranked high in the RFMODIS model, but not in RFSPOT and RFETM+ models. Variable importance of Random Forest models for TN predictions Numerous of the RS image derived spectral variables ranked as important predictors in the RF model s for soil TN predictions. However, also nonRS derived input variables showed substantial predictive capabilities in all RF models for TN predictions. For example, elevation and distance to WCSs ranked high in all RF models to predict

PAGE 86

86 soil TN. In gene ral, the elevation is high (> 3.5 m) in the northeast of the study area near WCSs and low (< 2.5 m) in the southern portion. This generates low velocity sheet flow across the study area. Nitrogen contents were higher in lower elevation areas due to inhibit ion of decomposition by longer inundation periods (Sumfleth and Duttmann, 2008). Distance to WCSs also indirectly related to hydrology. Overall, the RS imagederived spectral indices that contain water contents and hydrologic factors were ranked high for the soil TN predictions in the RFSPOT and RFETM+ models (Table 3 7). For example, PC 1, MSI, and NDWI ranked high in the RFSPOT model, and RSR, midinfrared index, and NDWI ranked high in the RFETM+ model. All these predictor variables incorporate the short to mid infrared wavelength region, which is sensitive to the amount of water in plants (Gao, 1996; Jensen, 2005) The PC 1, which had high loadings of NIR and SWIR bands in the RFSPOT model, discriminated vegetated and open water areas; and the PC 2, which had high loadings of NIR and MIR bands, in the RFMODIS model represented open water areas of the study area. The high rank of NDWI in the RFSPOT and RFETM+ models to predict soil TN is consistent with the findings reported by Wang et al. (2011) They found a strong relationship between 15 N and NDWI derived from SPOT images in the Everglades, although their study was only focused on the tree islands and not on the freshwater marsh area. The variable importance of the RFMODIS model showed the EVI as the best predictor variable for soil TN pr edictions, mirroring the TP predictions. The MSI and NDWI were not ranked as an important predictor variable in the RFMODIS model, unlike the RFSPOT and RFETM+ models.

PAGE 87

87 One interesting finding is the limited prediction capability of NDVI, which is the most widely used vegetation index. The variable importance of NDVI to build soil TP and TN prediction models was not as high as expected, although it has shown good predictive power for soil properties in previous studies (Fernndez Buces et al., 2006; Vrieling et al., 2008; Eldeiry et al., 2008; Rivero et al., 2009). In summary, our results suggest that multivariate models (RF models) incorporating spectral and environmental ancillary data are more accurate in predicting continuous variation of soil properties (TP and TN) in WCA 2A than univariate models (BK models) only relying on field observations. Spatial Resolution Effects of Remote Sensing Images on Soil TP and TN Predictions: Fine and Coarse Resolutions The RFSPOT and RFETM+ models showed slightly better prediction results, as measured by R2 and RMSE, than the RFMODIS model (Table 3 4). However, the differences of R2 to estimate soil TP and TN were marginal. The R2 differed only 3% for TP predictions (R2: 0.90 to 0.93) and 4% for TN predictions (R2: 0.91 to 0.95) among all RF models. This was also true for the differences of RMSE among the models. The RFMODIS model showed slightly higher RMSE than the RFSPOT and RFETM+ models for both soil TP and TN predictions, but still the differences were small compared to the differences of the spatial resolutions of RS images. The prediction raster of RF models for soil TP showed similar patterns except for differences in the roughness of RFMODIS grids (Figure 3 4). The tree islands heads with high TP concentrations w ere clearly shown in prediction maps of RFSPOT and RFETM+ models, but not in maps derived from the RFMODIS model. The RFSPOT and RFETM+ models showed similar TN predictions, but the RFMODIS model showed substantially lower concentrations in the interior of the

PAGE 88

88 marsh (Figure 3 5). Interestingly, the MODIS based soil prediction models nearly achieved the same fit (R2) and absolute deviations (RMSE) in predictions as the higher resolution RS soil prediction models derived from SPOT and Landsat ETM+ data. This suggests that the vegetationwater floc properties depicted from space, even at the coarser spatial resolution of 250 m (MODIS), were able to resemble the heterogeneity of landscape features in WCA 2A. The high correlations between spectral and soil proper ties indicate that spatial patterns of soils and above ground features coincided well. This interpretation is supported by the fact that spatial autocorrelations derived from semivariogram analysis for BK were 9,065 m for soil TP and 7,608 m for soil TN ex ceeding clearly the 250 m resolution of MODIS (Table 3 5). Long spatial autocorrelations for soil P and N were reported in various land uses such as grassland and farmland (Schloeder et al., 2001; Sumfleth and Duttmann, 2008; Wang et al., 2009; Wei et al., 2009; Fu et al., 2010 ; Grunwald et al., 2010). Subtropical wetlands have also shown long (> 250 m) spatial autocorrelation range for soil P and N (Corstanje et al., 2006; Rivero et al., 2007b) For example, Grunwald and Reddy (2008) reported the spatial autocorrelation range of 7,240 m for soil TP and 1,007 m for soil TN in the Blue Cypress Marsh Conservation Area in east central Florida. Our results suggest that the RF model using the MODIS derived spectral indices could explain the spatial distributions of soil TP and TN in WCA 2A because soil TP and TN had a long spatial autocorrelation range greater than the MODIS spatial resolution. Our results are consistent with the finding by Eldeiry and Garcia (2008) who reported less than 0.1 % of R2 difference for soil salinity predictions derived from IKONOS (4 m spatial resolution) and Landsat ETM+ data in southeastern Colorado. Similar results regarding the e ffects

PAGE 89

89 of spatial resolution were documented not only for soil properties, but also for others such as energy flux. For instance, Kustas et al. (2004) compared different spatial resolutions of spectral input variables derived from Landsat thematic mapper and ETM+ images for evapotranspirat ion flux mapping. They aggregated pixels of the images to test resolution effects of spectral input variables and found ~ 240 m spatial resolution of input variables provided a similar amount of spatial details than the original resolution (30 m) for the e nergy flux mapping in Iowa. Total stocks of soil TP and TN in WCA 2A were estimated and compared among the developed RF models to identify possible deviations among the models. Table 3 8 shows the estimated nutrient stocks for each model. Block Kriging showed slightly higher TP stocks and lower TN stocks than the RF models. About 2,408 mt of TP was stored in soils in WCA 2A based on the RFSPOT model. The RFETM+ and RFMODIS models showed about 2,519 mt and 2,614 mt of soil TP stocks, respectively. The RFSPO T and RFMODIS models showed a similar amount of soil TN stocks (~142,959 mt), which were slightly higher than the RFETM+ model. The estimated stocks of soil TP and TN showed less than 8 % and 3 % differences among the RF models, respectively. Total soil TP stocks of a previous study conducted by Rivero et al. (2007b) representing soil mapping in 2003 was estimated using reported soil TP concentrations and BD for the comparison. It showed a similar amount of TP storage to our results. However, Reddy et al. ( 2011) reported a much greater amount of TP storage in WCA 2A (Table 3 8). A lack of information about model development such as interpolation methods, parameters, and accuracy assessments in Reddy et al. (2011) limited further comparisons and investigation of the differences.

PAGE 90

90 Figures 3 7 and 3 8 show the selected high and low TP and TN concentration areas for the entropy ( H ) analysis, respectively. It shows the differences of predicted soil TP variations among the RF models. For example, the RFSPOT model prediction map shows smoother variations of soil TP concentrations than the RFMODIS model. The RFMODIS model prediction map shows only few grid cells. The differences of soil TP and TN variations among the RF models were quantified by H Table 3 9 shows H values in each of the high and low concentration areas for soil TP and TN. In general, H values tended to be higher in high concentration areas than in low concentration areas. The H values showed a distinct difference between high and low concentration areas for soil TP predictions. The RFSPOT model showed an H of 1.87, which was the highest value among the RF models for soil TP predictions, followed by the RFETM+ (1.65) and RFMODIS (0.99) models in high concentration area. The RFSPOT model showed a higher H than the RFMODIS model in the low concentration area for soil TP predictions and for soil TN predictions. Entropy values also tended to be higher in high concentration areas than in low concentration areas for soil TN predictions, although the dif ferences of the H values between high and low soil TN concentration areas were not as large as soil TP predictions (Table 3 9). The H describes the amount of information, and the information was estimated based on the probability distribution of pixel values according to Eq. (3). Hence, the H refers to a spatial variation of a soil property in a soil prediction map. Therefore, our results suggest that the RFSPOT model (high H ) contains more information about the spatial variability of soil TP and TN than the RFETM+ and RFMODIS models (lower H ) within the same area, although the R2 and RMSE were similar among the prediction models. This can be explained by the fact that the RFMODIS model reduces

PAGE 91

91 the information in a map by data smoothing when compared to models derived using higher resolution RS images. Interestingly, the differences of the H values in the low concentration areas among models were smaller than the high concentration areas. For instance, the entropy of the RFSPOT and RFMODIS models in the low TP concentration area were 0.49 and 0.34, respectively. This suggests that the areas with low TP concentrations are characterized by similarities rather than differences among the models. Our results suggest that the RFSPOT model has more randomness of pixel values (concentrations) than the RFMODIS model, according to Hausser and Strimmer (2009). It further indicates that the prediction map of the RFSPOT model has more complexity (heterogeneity) than the coarser resolution RF models. The entropy analysis was applied and shown to be useful in analyzing the diversity of the world pedosphere (Ibez et al., 1998), to characterize particle soil distribution (Martn et al., 2005), and to assess the uncertainty of a soil map (Bishop et al., 2001), althoug h there are subtle differences on the interpretation of the entropy (Martn and Rey, 2000). In general, entropy is useful as a tool to quantify the effects of spatial resolution on soil property predictions in terms of its variability and complexity. Co nclusions This study showed high performance to predict TP and TN concentrations and stocks in wetland soils using spectral indices derived from satellite RS images and geospatial environmental properties. The RF models with varying spatial resolutions of RS images showed substantial improvement of prediction accuracy when compared to BK to predict soil TP and TN in WCA 2A. The variable importance of the RF models for soil TP and TN suggests that soil TP is highly dependent on biotic/vegetation properties

PAGE 92

92 t hat can be inferred by RS and soil TN is controlled by a combination of RS derived and others (elevation and distance to WCSs). The RFSPOT and RFETM+ showed similar power to predict soil TP and TN greater than the RFMODIS model. Remote sensing based va riables of importance included 10 (RFSPOT), 8 (RFETM+), and 7 (RFMODIS) variables to predict soil TP. Models that predicted soil TN showed 6 (RFSPOT), 8 (RFETM+), and 8 (RFMODIS) RS based variables as top predictors. Although spectral data from different s ensors covered a similar spectral range, there was no consistency in terms of one or more specific spectral variables correlating best with soil TP and/or TN. This suggests that the spectral inference capabilities of different spectral data (SPOT, ETM+, an d MODIS) may have been confounded by the differences in their spatial (grid) resolution. The spectral data were able to infer on the chlorophyll and carotenoids content (e.g., NDVI green), water status (e.g., NDWI, MSI), and composition of vegetation (e.g. TVI), which were strongly correlated with soil biogeochemical properties. The entropy analysis quantified differences in the spatial variations and complexities in soil TP and TN predictions. It showed that the RF models derived from fine resolution RS i mages (i.e., SPOT and Landsat ETM+) contain more information than the RF model derived from coarse resolution MODIS image, although the R2 and RMSE were similar among the prediction models. Digital soil mapping using RS images has great potential to reduc e the cost, labor, and time as well as improving the accuracy of predictions when compared to traditional soil survey and assessment. Findings from this study suggest that both fine and coarse resolution RF models performed well to quantify soil TP and TN in this

PAGE 93

93 wetland. Although the accuracy was marginally different between fine and coarse resolution models the heterogeneity of soil nutrients could only be described using the former one. This means that coarser resolution models are still useful in studies where only regional patterns are of interest. Noteworthy is that SPOT images are not freely available (except for this study where the SPOT images were donated by Planet Action), whereas Landsat ETM+ and MODIS images are available at no cost. Specificall y the availability of MODIS images at high temporal resolution alludes to the possibility of sequential mapping of this wetland to generate spacetime soil nutrient projections, although retrieving suitable RS images are still limited by satellite orbital constraints and climatic conditions in the region. This could be beneficial to support monitoring of restoration success and changes in nutrient status in this wetland. The methodological approach of RS informed DSM can be transferred to other aquatic syst ems. Summary Soil nutrients stored in wetland soils are critical to assess the effectiveness of restoration efforts, yet challenging to accurately derive soil heterogeneity. The incorporation of remote sensing (RS) data into digital soil models has shown success to imp rove soil predictions. However, the effects of multi resolution imagery on modeling of biogeochemical soil properties in aquatic ecosystems are still poorly understood. The objectives of this study were to (i) assess the spatial distribution of soil total phosphorus (TP) and soil total nitrogen (TN) concentrations and stocks, (ii) develop prediction models for TP and TN utilizing RS images and environmental ancillary data, and (iii) elucidate the effect of different spatial resolutions of RS images on infer ential modeling of those soil properties. Soil cores were collected (n=108) from the top 10 cm in a

PAGE 94

94 subtropical wetland: Water Conservation Area2A, the Florida Everglades, U.S. The spectral data and derived indices from remote sensing images, which have different spatial resolutions, included: MODIS (250 m), Landsat ETM+ (30 m), and SPOT (10 m). Block kriging (BK) and Random Forest (RF) were employed to predict soil TP and TN using RS image derived spectral input variables, environmental ancillary data, and soil observations. The RF models showed R2 between 0.90 to 0.93 and root mean square error (RMSE) between 100.4 to 115.9 mg kg 1 for soil TP and 1.45 to 1.52 g kg 1 for soil TN. These RF models performed much better than BK prediction models. Soil TP was mainly predicted from RS derived spectral indices that infer on biotic/vegetation characteristics, whereas soil TN was predicted using a combination of biotic/vegetation, topographic, and hydrologic variables. Results suggest that the spectral data inform ed soil prediction models have excellent predictive capabilities in this aquatic ecosystem. Interestingly, there was no noticeable distinction among different spatial resolutions of RS images to develop prediction models for soil TP and TN. However, the variability and complexity of soil TP and TN variations were much better expressed with the finer resolution RFSPOT model than the coarser resolution RFMODIS model.

PAGE 95

95 Table 31. Summary of remote sensing spectral indices. Indices a Formula References EVIb Huete et al., 1997 Mid infrared index Musick and Pelletier, 1988 MSI Rock et al., 1986 NDVI Rouse et al., 1974 NDVI green Gitelson et al., 1996 NDWI Gao, 1996 RSR Brown et al., 2000; Chen et al., 2002 SR Cohen, 1991; Chen et al., 2002; Colombo et al., 2003 TVI Nellis and Briggs, 1992 aEVI, enhanced vegetation index; MidIR, midinfrared; MSI, moisture stress index; NDVI, normalized difference vegetation index; NDVI green, normalized difference vegetation green index; NDWI, normalized difference water index; NIR, near infrared; RSR, reduc ed simple ratio; SR, simple ratio; SWIR, shortwaveinfrared; TVI, transformed vegetation index. bEmpirical parameters for EVI of MODIS: C1=6.0; C2=7.5; G=2.5; L=1.0 ) 1 ( L Blue Red NIR Red NIR2 1L C C G Band7 Landsat Band5 Landsat MidIR MidIR NIR MidIR Red NIR Red NIR Green NIR Green NIR SWIR NIR SWIR NIR ) SWIR SWIR SWIR SWIR 1 ( Red NIRmin max min Red NIR 100 ) 5 0 Red NIR Red NIR (2 / 1

PAGE 96

96 Table 3 2. Summary of environmental variables employed to build models to predict soil total phosphorus and nitrogen. Property Attributes a Spatial resolutions Data sources Vegetation Spectral values derived from MODIS: Reflectance values of blue, red, NIR, and MIR; EVI; MSI; NDVI; SR; TVI; PC 1 to 3 250 m by 250 m USGS, Land Processes Distributed Active Archive Center, 2010 Spectral values derived from Landsat ETM+: Reflectance values of blue, green, red, NIR, MIRs, and PAN; MidIR; MSI; NDVI; NDVI green; NDWI; RSR; SR; TVI; Mean w/ 3x3 cells and 7x7 cells for MSI, NDVI, NDVI green, NDWI, RSR, SR, and TVI b ; PC 1 to 3; TC 1 to 3 30 m by 30 m USGS, Earth Resources Observation and Science Data Center, 2010 Spectral values derived from SPOT: Reflectance values of green, red, NIR, SWIR, and PAN; MSI; NDVI; NDVI green; NDWI; RSR; SR; TVI; Mean w/ 3x3 cells, 9x9 cells, and 25x25 cells for MSI, NDVI, NDVI green, NDWI, RSR, SR, and TVI b ; PC 1 to 3 10 m by 10 m SPOT Image Corporation, 2009 Topography Elevation (m) Point observation data; 10 m by 10 m, 30 m by 30 m, and 250 m by 250 m rasterized elevation layers were derived using Splines USGS, High Accuracy Elevation Dataset (15 cm vertical accuracy), 2007 Hydrology Distance to water control structures (km) 10 m by 10 m, 30 m by 30 m, and 250 m by 250 m layers were derived using Euclidean distance function South Florida Water Management District, 1997 Lithology Potassium concentration 2,000 m by 2,000 m USGS, Digital Data Series DDS 9, 1999; Surveyed in 1993 as National geophysical data grids work Bouguer gravity 2,000 m by 2,000 m Isostatic residual gravity anomaly 2,000 m by 2,000 m Magnetic anomaly 2,000 m by 2,000 m a ETM+, enhanced thematic mapper plus; EVI, enhanced vegetation index; MidIR, midinfrared index; MIR, midinfrared; MODIS, moderate resolution imaging spectroradiometer; MSI, moisture stress index; NDVI, normalized difference vegetation index; NDVI green, normalized difference vegetation green index; NDWI, normalized difference water index; NIR, near infrared; PAN, panchromatic; PCs, principal component scores; RSR, reduced simple ratio; SPOT, satellite pour lobservation de la terre; SR, simple ratio; SWIR shortwaveinfrared; TC 1 to 3, tasseled cap transformed components; TVI, transformed vegetation index. b Mean spectral values within neighborhoods (3x3, 7x7, 9x9, 25x25 cells, respectively) using nearest neighbor geospatial method.

PAGE 97

97 Table 3 3 Descriptive statistics for soil properties in the topsoil (010 cm) observed in Water Conservation Area2A (n=108). Property a Unit Mean Min. Max. Median Std. deviation Skewness BD g cm 3 0.11 0.05 0.24 0.10 0.04 1.30 TP conc. mg kg 1 601.4 165.4 1,694 478.8 336.7 1.25 TP stock g m 2 6.4 1.6 24.9 5.1 4.2 2.31 TN conc. g kg 1 29.7 21.3 38.2 30.0 3.35 0.19 TN stock g m 2 323.8 147.0 808.8 293.0 121.0 1.82 aBD, bulk density; TNconc., total nitrogen concentration; TNstock, total nitrogen stock; TPconc., total phosphorus concentration; TPstock, total phosphorus stock.

PAGE 98

98 Table 3 4 Summary of model performance assessment for soil total phosphorus (TP) and total nitrogen (TN) concentrationsa. Property Statistical measure b BK RF SPOT RF ETM+ RF MODIS TP conc. (mg kg 1 ) R 2 0.60 0.93 0.93 0.90 RMSE 213.7 102.0 100.4 115.9 TN conc. (g kg 1 ) R 2 0.30 0.95 0.94 0.91 RMSE 2.88 1.45 1.45 1.52 a BK, block kriging; RFSPOT,EMT+,MODIS, Random Forest models using SPOT, Landsat ETM+, and MODIS images derived input variables, respectively. b R2, coefficient of determination; RMSE, root mean squared error.

PAGE 99

99 Table 3 5. Summary of semivariogram parameters of Block Kriging for soil total phosphorus (TP) and total nitrogen (TN) in Water Conservation Area2A. Property Transformation Model Nugget Sill Range (m) TP Natural log Gaussian 0.13 0.33 9,065 TN Gaussian 8.12 11.97 7,608

PAGE 100

100 Table 3 6. Summary of the relative importance of predictor variables in Random Forest models predicting soil total phosphorus in Water Conservation Area2A a. Rank Random Forest SPOT Random Forest ETM+ Random Forest MODIS 1 TVI (100) Tasseled cap transformed data 2 (100) Enhanced vegetation index (1 00) 2 Simple ratio (97.7) NDVI green 3 X 3 (90.8) Distance to WCSs (57.1) 3 Simple ratio 9 X 9 (91.8) NDVI green (58.5) Principal components score 3 (53.5) 4 TVI 25 X 25 (89.0) TVI 3 X 3 (50.9) Near infrared reflectance value (49.8) 5 NDVI green 25 X 25 (85.7) NDVI green 7 X 7 (50.7) Principal components score 1 (47.0) 6 Simple ratio 25 X 25 (79.7) Distance to WCSs (40.2) Elevation (30.7) 7 NDVI (79.4) NDVI 7 X 7 (39.4) Bouguer gravity (23.9) 8 TVI 3 X 3 (75.6) Principal components score 3 (39.4) Principal components score 2 (18.3) 9 Reduced simple ratio 25 X 25 (71.8) Simple ratio 3 X 3 (37.6) Mid infrared reflectance value (17.6) 10 Simple ratio 3 X 3 (64.4) Bouguer gravity (36.4) TVI (15.2) aNDVI, normalized difference vegetation index; NDVI green, normalized difference vegetation green index; TVI, transformed vegetation index; WCSs, water control structures.

PAGE 101

101 Table 3 7. Summary of the relative importance of predictor variables i n Random Forest models predicting soil total nitrogen in Water Conservation Area2A a. Rank Random Forest SPOT Random Forest ETM+ Random Forest MODIS 1 Principal components score 1 (100) Reduced simple ratio 3 X 3 (100) Enhanced vegetation index (100) 2 Elevation (92.0) Reduced simple ratio (85.6) Elevation (99.5) 3 Moisture stress index (80.5) Mid infrared index (77.6) Principal components score 3 (94.0) 4 NDWI 25 X 25 (74.2) NDWI (68.6) Red reflectance value (86.2) 5 Panchromatic values (73.9) Principal components score 2 (66.1) Mid infrared reflectance value (84.0) 6 Isostatic gravity (73.2) Elevation (65.2) Distance to WCSs (75.4) 7 Bouguer gravity (72.5) Blue reflectance value (64.8) Principal components score 1 (67.0) 8 Distance to WCSs (72.5) Distance to WCSs (62.7) Simple ratio (66.4) 9 Moisture stress index 25 X 25 (71.3) Moisture stress index (61.6) Near infrared reflectance value (64.6) 10 Short wave infrared reflectance value (71.0) Principal components score 1 (60.6) Principal components score 2 (63.9) aNDWI, normalized difference water index; WCSs, water control structures

PAGE 102

102 Table 3 8. Summary of estimated nutrients stocks in Water Conservation Area2A (top 10 cm). Source Cell size Total phosphorus (mt a ) Total nit rogen (mt) Block Kriging 10 m X 10 m 2,749.6 137,908 Random Forest SPOT 10 m X 10 m 2,408.2 141,296 Random Forest ETM+ 30 m X 30 m 2,519.7 138,312 Random Forest MODIS 250 m X 250 m 2,614.1 142,959 Rivero et al. (2007b) 2,530.5 b 124,524 b Reddy et al. (2011) 3,381 c aMetric tons bEstimated value based on the reported mean of TP and TN and the BD measured in 2003. cPhosphorus storage in floc (30% of total phosphorus storage) and top 10 cm soil.

PAGE 103

103 Table 3 9. Estimated Shannon entropy ( H) of the predicted soil properties derived using Random Forest with SPOT, Landsat ETM+, and MODIS images derived spectral input variables a. Soil Property Area RF SPOT RF ETM+ RF MODIS Total phosphorus High concentration area 1.87 1.65 0.99 Low concentration area 0.49 0.33 0.34 Total nitrogen High concentration area 1.47 0.86 0.35 Low concentration area 0.74 0.81 0.69 aRFSPOT,EMT+,MODIS, Random Forest models using SPOT, Landsat ETM+, and MODIS images derived input variables, respectively.

PAGE 104

104 Figure 3 1. Spatial distribution of measured soil concentrations of total phosphorus, TP (left) and total nitrogen, TN (right) within the Water Conservation Area2A.

PAGE 105

105 Figure 3 2. Prediction maps of soil nutrient concentrations and stocks in the topsoil (010 cm) using Random Forest with SPOT image derived spectral input variables and environmental predictor variables: a) total phosphorus concentrations (mg kg1), b) total phosphor us stocks (g m2), c) total nitrogen concentrations (g kg1), and d) total nitrogen stocks (g m2).

PAGE 106

106 Figure 3 3. Scatter plots showing the relationship between observed and predicted soil properties using block kriging (left), Random Forest with SPOT, Landsat ETM+, and MODIS (right) image derived input variables combined with environmental predictor variables: a) total phosphorus (mg kg1) and b) total nitrogen (g kg1): the dotted line indicates a ratio of 1:1.

PAGE 107

107 Figure 3 4. Prediction maps of soil total phosphorus, TP (mg kg1) using Block Kriging and Random Forest models: a) Block Kriging with block size 10 by 10 m, b) Random Forest model with the SPOT image, c) Random Forest model with the Landsat ETM+ image, and d) Random Forest model with the MODIS image derived spectral input variables combined with environmental predictor variables. Stripes in map c' were caused by permanent scanline corrector failure (SLC off) of Landsat Enhanced Thematic Mapper Plus sensor since May 2003.

PAGE 108

108 Fi gure 3 5. Prediction maps of soil total nitrogen, TN (g kg1) using Block Kriging and Random Forest models: a) Block Kriging with block size 10 by 10 m, b) Random Forest model with the SPOT image, c) Random Forest model with the Landsat ETM+ image, and d) Random Forest model with the MODIS image derived spectral input variables combined with environmental predictor variables. Stripes in map c' were caused by permanent scanline corrector failure (SLC off) of Landsat Enhanced Thematic Mapper Plus sensor since May 2003.

PAGE 109

109 Figure 3 6. Annual total phosphorus (TP) loads (kg) for inflow (gray filled bar) and outflow (empty bar), and annual flow weighted TP concentrations (ppb) for inflow (solid line) and outflow (dashed line) in WCA 2A during the period from M ay, 2003 to April, 2010 (Data obtained from the South Florida Water Management District). 0 5 10 15 20 25 30 35 40 45 0 10000 20000 30000 40000 50000 60000 2003 2004 2005 2006 2007 2008 2009 2010 Phosphorus Concentration (ppb) Phosphorus Load (kg) Year P loads (kg) Inflow P loads (kg) Outflow P concentrations (ppb) Inflow P concentrations (ppb) Outflow

PAGE 110

110 Figure 3 7. Variation comparison of predicted soil total phosphorus concentrations: a) high concentration area, b) low concentration area.

PAGE 111

111 Figure 3 8. Variation comparison of predicted soil total nitrogen concentrations: a) high concentration area, b) low concentration area.

PAGE 112

112 CHAPTER 4 EVALUATING TOTAL CAR BON STOCKS IN A SUBT ROPICAL WETLAND Overview There is an increasing interest in understanding the role of soil carbon (C) in biogeochemical cycling since it is directly related to food security, climate change, biome energy, and ecosystem services (Bai et al., 2008; Conant et al., 2011). Climate change accelerated by anthropogenic emi ssions of greenhouse gases (GHG), especially carbon dioxide (CO2), into the atmosphere has received great attention Consequently, the role of soils to sequester atmospheric CO2 has been gaining importance in offsetting the anthropogenic C emissions (Forst er et al., 2007; Lal and Folett, 2009). Carbon fluxes between soil, biotic, and atmospheric pools are dynamic in space and through time and depend on many environmental and anthropogenic drivers (Grunwald et al., 2011; Vasques et al., 2012). Quantifying C sources, sinks, and ecosystem processes that modulate the global C system is critical in order to identify imbalances and counteract global climate change. Globally soils contain about 1,400 to 2,300 Pg of C; this is almost two times greater than the amount in the atmosphere and three times greater than the amount in terrestrial biomass, and about 20 to 30 % of the soil C is stored in wetlands (Mitsch and Gosselink, 2007; Lal, 2008). The role of wetlands is critical in global biogeochemical cycles, especial ly C dynamics, since wetland soils both take up and release GHG. Wetland soils also play an important role in regulation of climate change, although wetlands cover only 3 % of the global land area (Bridgham et al., 1995; Raich and Potter, 1995; Whiting and Chanton, 2001). High productivity and a slow decomposition rate make wetlands optimum natural environments for storing C (Mitsch et al., 2012).

PAGE 113

113 On the other hand, wetlands are also considered a significant source of naturally produced GHG, especially methane. Carbon release from permafrost soils into the atmosphere under changing climate have been documented (Schuur et al., 2008; Mack et al., 2011). These risks of accelerated C pulses into the atmosphere due to changing global climate are not limited to hi gh latitudes, but also pose a risk to subtropical and tropical areas, specifically in the southeastern U.S. Temperature in the southeastern U.S. is expected to increase by 2.5 5 C on average by the end of the century (Karl et al., 2009) and precipitation is projected to decrease by approximately 20 % during summer months in the southeast (IPCC, 2007) which may lead to large potential losses of C from wetland soils shifting them from a C sink to a source. The assessment of soil C stocks in wetlands and i ts loss are extremely important since large stocks of C become exposed to aerobic conditions when wetlands drain, and this climate sensitivity cannot be modeled with data dominated by upland ecosystems (Davidson, 2010). To assess the future impacts of soil C change under climate change projections it is critical to quantify their historic and actual soil C stocks in wetlands. Grimm et al. (2008) comprehensively reviewed soil organic C stocks in tropical areas, and Follett et al. (2009) extensively measured C stocks from 30 different sites across the U.S. However, no wetlands were included. Digital soil mapping (DSM) has shown improved accuracy to assess spatial patterns and stocks of soil C compare to traditional soil survey (Thompson and Kolka, 2005; Simbahan et al., 2006; Grimm et al., 2008; Vasques et al., 2010; Wiesmeier et al., 2011). Moreover, remote sensing (RS) informed soil prediction models have been shown to improve the predictive power of DSM (Sullivan et al., 2005; Selige et al., 2006;

PAGE 114

114 Dang et al ., 2011; Minasny and Hartemink, 2011; Mulder et al., 2011) because of its capabilities to directly or indirectly infer on biotic properties (Grunwald, 2009). Hence, RS informed DSM has much value in wetlands, where field sampling is challenging and site ac cess limited. For instance, Rivero et al. (2009) and Kim et al. (2012) showed successful results of RS supported DSM for soil total phosphorus and taxonomic classes in a subtropical wetland, Water Conservation Area (WCA) 2A which is located within the Grea ter Everglades. The Greater Everglades is one of the largest wetlands on Earth and it stores enormous amounts of C in soils and biomass. Although it has been impacted by nutrient influx, climate change, water level changes, and other environmental stressors (Noe et al., 2001; Bernhardt and Willard, 2009) it has sustained to preserve large amount of C in soils. Additional losses of C from this system, in the form of CO2 or methane, would possibly accelerate climate change not only at regional, but also at g lobal scale. Thus, it is essential to assess the amount of C stored within the ecosystem and improve our understanding how C storage responds to changes in climate, water levels, and nutrient status, all of which impact C cycling and losses/gains of C. Our objectives were to (i) develop C stock prediction models using RS images and environmental ancillary data, (ii) identify the largest predictive environmental factors controlling the spatial distribution of soil C, and (iii) assess the amount of C stored i n the investigated subtropical wetland and evaluate its role as a C sink or a C source. Methods and Materials Study Area The Everglades, located in Florida in the United States, is one of the largest wetlands in the world. It was fragmented into several hydrologic units, including WCAs

PAGE 115

115 and the Everglades National Park (ENP) in the 1960s. The study was conducted in WCA 2A, th e smallest hydrologic unit, which covers about 418 km2. The WCA 2A is located in the northern part of the Everglades. The climate is subtropical characterized by an average annual precipitation of 1,297 mm (Everglades Depth Estimation Network, EDEN, 2009) and an average annual temperature of 23 C (National Climatic Data Center, 2009). The t opography is flat averaging 3.0 m above sea level (United States Geological Survey, USGS, High Accuracy Elevation Data, HAED, 2007), and this generates slow sheet flow f rom the northeast to southwest of the study area. Soils are mainly Histosols dominated by Saprists (suborder) which are the most highly decomposed organic soil materials. This wetland is primarily composed of ridges with sawgrass ( Cladium jamaicense) and c attail ( Typha domingensis ), sloughs with mostly waterlily ( Nymphaea odorata), and tree islands. Field Sampling and Soil Carbon Measurement We collected a total of 108 soil cores spread over the whole study area using a 10 cm diameter stainless steel cori ng tube at two depths (0 10 cm and 10 20 cm) in February 2009, October 2009, and March 2010. We used a stratified random sampling design using historic soil, hydrologic, and ecological sites identified in 2003 (compare Rivero et al., 2007). We could revisit only 48 preselected sampling sites because dense vegetation and a low water table made it difficult to access the sites via airboats. Water table positions of the southern part of the study area were abnormally low (i.e., below soil surface) in Febr uary 2009 due to a prolonged drought from 2007 (Figure 4 1). The remaining sampling sites were adjusted to areas as close as possible to original sampling sites. Soil samples collected in the field were packed with doublezipped plastic bag and transported to the laboratory.

PAGE 116

116 All samples were ovendried at 70 C for 72 hours to prevent any organic C loss. Weights of the soil samples before and after drying were recorded to measure bulk density (BD) of each sample. Soil samples were ground, ball milled, a nd sieved using a prior to total C (TC) analysis. Soil TC was analyzed with a Shimadzu SSM 5000A TC analyzer (Shimadzu Corporation, Columbia, MD, USA) using 15 25 mg samples and a combustion temperature of 900 C. The amount of TC stored per unit land area (kg m2) for a given depth was calculated using the measured TC concentrations on a mass basis (g kg1) and the measured BD at each sampling point according to: c z TC TCconc stock (4 1) where TCstock is TC stocks (kg m2), TCconc. is TC concentrations (g kg1), is the BD of the soil sample at each point (g cm3), z is the depth (10 cm), and c is a constant for unit conversion (0.01). We performed all statistical analyses using SAS statistical software (SAS Institute Inc., Cary, NC). A t test was applied to test for differences of soil TC concentrations between 2003 and 2009 by soil depth. A significant level of p = 0.05 was used. Environmental Predictor Variables We used topographical, hydrological, and lithological properties as environmental predictor variables to predict soil TC stocks. In addition, we included various spectral indices derived from satellite RS images as a proxy of various properties related to vegetation. Elevation data surveyed by USGS using 400 m grid space sampling (15 cm vertical accuracy) represented topography. Distance to water control structures (WCSs)

PAGE 117

117 derived using the Euclidean distance function represented hydrology effects since the nutrient enriched inflow comes from WCSs. G eophysical data incl uding potassium concentration gammaray, bouguer gravity anomaly, isostatic residual gravity anomaly, and magnetic anomaly data which were surveyed and released by USGS represented lithological properties as environmental predictor variables for soil TC pr edictions. Bouguer gravity anomaly and isostatic residual gravity anomaly data contain similar information about the density distribution of the Earths upper crust, and magnetic anomaly data reflects the distribution of iron (Fe) minerals in he rocks of t he Earths crust. Spectral indices, such as the Normalized Difference Vegetation Index (NDVI), Moisture Stress Index (MSI), and Simple Ratio (SR), enhance spectral signatures of the Earths surface properties as well as reduce spectral noise. They have shown strong correlations to various soil properties (Nield et al., 2007; Vrieling et al., 2008; Rivero et al., 2009, Kim et al., 2012). We derived various spectral indices from three different RS images including Satellite Pour lObservation de la Terre (SPO T), Landsat Enhanced Thematic Mapper Plus (ETM+), and Moderate Resolution Imaging Spectroradiometer (MODIS). The SPOT image which has band 1 (green; 500 590 nm), band 2 (red; 610 680 nm), and band 3 (NIR; 780 890 nm) with 10 m spatial resolution and a band 4 (shortwaveinfrared (SWIR); 1,580 1,750 nm) with 20 m spatial resolution were taken on January 2009. The Landsat ETM+ image which has band 1 (blue; 450 515 nm), band 2 (green; 525 605 nm), band 3 (red; 630 690 nm), band 4 (NIR; 750 900 n m), band 5 (MIR; 1,550 1,750 nm), and band 7 (MIR; 2,090 2,350 nm) with 30 m spatial resolution and a band 6 (thermal infrared; 10,400 12,500 nm) with 60 m spatial

PAGE 118

118 resolution were taken on February 2010. T he MODIS image which ha s a 250 m spatial reso lution were taken on February 2010. The SPOT images were donated by Planet Action, a nonprofit ASTRIUM GEO initiative, and the Landsat ETM+ and MODIS images were released by USGS. We also employed a principal component analysis to derive possible relatio nships which were muted between soil C and original spectral signatures for all RS images. Furthermore, we utilized a tasseledcap transformation for the Landsat ETM+ image to enhance the spectral signatures of soil brightness, greenness, and wetness (Kaut h and Thomas, 1976). All image processes were conducted using ERDAS Imagine 2010 software (Earth Resource Data Analysis System Inc., Atlanta, GA, USA). We extracted spectral data from each sensor, various spectral indices, principal component (PC) variables, and tasseledcap transformed variables at each sampling point. All of the geographic information system (GIS) data processing work was performed using ArcGIS 10 (Environmental Systems Research Institute ESRI Inc., Redlands, CA, USA). Details of spectr al indices and environmental predictor variables were described in Chapter 3. Random Forest We applied Random Forest (RF) to develop soil C stock models and identify environmental variables that control the spatial distribution of soil C stocks. The RF is one of widely used machine learning methods which are commonly used for data exploration. It is a recursive partitioning method particularly well suited to small observation sets and large predictor variables (Strobl et al., 2009). The RF is an ensemble of classification or regression trees that are calculated on random subsets of the data (bootstrap samples) using a randomized subset of predictor variables at each

PAGE 119

119 tree (Breiman, 2001) Numerous trees are grown within the algorithm, and one single prediction is given by averaging the trees. I t has shown low bias and low variance in previous studies (Daz Uriarte and de Andrs, 2006; Wiesmeier et al., 2011) The RF estimates its own errors (i.e., mean of squared residuals, percent variance explained) bas ed on an out of bag dataset and produces a variable importance to reflect the contribution of each predictor variable to build a forest. Details of the RF algorithm can be found in Breiman (2001), Liaw and Wiener (2002), and Strobl et al. (2009). We deve loped three different RF models using input variables derived from the SPOT (i.e., RFSPOT model), Landsat ETM+ (i.e., RFETM+ model), and MODIS images (i.e., RFMODIS model). One strength of the RF is that it is user friendly in terms of input parameters. It has only two parameters: the number of trees in the forest and the number of variables in the random subset at each node. For this study, we used 1,000 as the overall number of trees i n the forest and 7 as the number of randomly preselected predictor variables for each split based on the models error rates and percent variable explained. The RF analysis was performed using the randomForest package (Liaw and Wiener, 2002) and the raster package (Hijmans and van Etten, 2012) was used to produce prediction maps in the R statistical l anguage (R Development Core Team, 2012). The model performance was assessed using oneleaveout cross validation. Results and Discussion Spatial Distribution of Total Carbon Stocks Descriptive statistics of TC concentrations and stocks are shown in Table 4 1. The 0 10 cm soils showed a mean concentration of 410.5 g kg1, and 10 20 cm showed a mean concentration of 432.9 g kg1. The mean of the BD in WCA 2A was

PAGE 120

120 0.11 g cm3 for 0 10 cm soils and 0.12 g cm3 for 10 20 cm soils, which indicates high or ganic matter content for both layers. The observed mean value of soil TC stocks was 4.5 kg m2 for 0 10 cm soils and 5.2 kg m2 for 10 20 cm soils (Table 4 1). The spatial distributions of measured soil TC concentrations and stocks by depth are shown i n Figure 4 2. In general, TC concentrations and stocks showed similar spatial patterns, except the northeastern and southern portions of the study area. These areas showed high soil TC concentrations but low TC stocks because of dense vegetation. Total C concentrations in both soil layers decreased when compared to previously documented studies conducted in WCA 2A (Table 4 2; DeBusk et al., 1994; Wright and Reddy, 2001; Rivero et al., 2007). All three studies in Table 4 2 used the same methods and procedur es to measure BD and TC concentrations as in this study. Thus, had no effect on explaining different TC concentrations shown in Table 4 2. However, these studies differed in terms of sampling site selection, sampling density, and seasonality of sampling. F urthermore, there was a significant ( p < 0.05) decrease of soil TC concentrations between 2003 and 2009 in both soil layers (Figure 4 3). Total C concentrations were significantly higher ( p < 0.05) in the 10 20 cm depth than the 0 10 cm depth in 2009. Previous studies conducted in WCA 2A showed increasing TC concentrations with depth (DeBusk et al., 1994; Wright and Reddy, 2007; Rivero et al., 2007), and the same trend was observed in WCA 1 (Corstanje et al., 2006), the ENP area (Osborne et al., 2011), and WCA 3B (Bruland et al., 2006) all located within the Greater Everglades ecosystem. However, WCA 3A showed a reversed trend: decreasing TC concentrations with depth (Bruland et al., 2006). Total C stocks were significantly higher ( p < 0.05) in the 10 20 cm depth than the 0 10 cm depth. No

PAGE 121

121 previous studies have been reported vertical distribution of TC stocks in the Everglades region, only soil C concentrations were documented. In contrast to our results, Wantzen et al. (2012) reported decreasing soil organic C stocks with depth (10 cm increment) in predominantly herbcovered wetlands in Brazil. Most studies on soil C stocks conducted in other land uses, such as a forest and a grassland, have observed decreasing stock values with depth (Jobbgy and Jac kson, 2000; Grimm et al., 2008; Vasques et al., 2010; Wantzen et al., 2012). For instance, Grimm et al. (2008) found the highest stocks in the upper 10 cm soils in a tropical forest located in Panama, and Vasques et al. (2010) observed decreasing soil TC t o the depth of 180 cm in the Santa Fe River watershed in Florida. Hughes et al. (1999) found increasing soil organic C stocks to the depth of 20 cm but decreasing stocks from 20 cm to 100 cm in a tall evergreen secondary forest in Mexico. Soil Total Ca rbon Stock Predictions using Random Forest The statistical assessment of the RF models for soil TC stocks by depth with the SPOT, the Landsat ETM+, and the MODIS images derived input variables combined with environmental predictor variables are shown in Ta ble 4 3. All RF models showed high R2 values ranging from 0.85 to 0.92 and small root mean squared errors (RMSEs) ranging from 0.75 to 0.86 kg m2 for TC stock models. Random Forest has shown high performance in predicting soil C in other ecosystems (Grimm et al., 2008; Wiesmeier et al., 2011). For instance, Wiesmeier et al. (2011) predicted SOC stocks up to 1 m depth in a semi arid ecosystem in China, and a RF model yielded an R2 of 0.74. Spatial prediction maps for soil TC stocks using the RF models with RS images derived spectral input variables and other environmental predictor variables for the study area are shown in Figure 4 4. Soil TC stocks showed higher variability in the 0

PAGE 122

122 10 cm than the 10 20 cm depth in prediction maps. All of the RF predic tion maps showed consistently low TC stocks in the lower northeastern portion of the study area in 0 10 cm soils, which spatially coincide with soils enriched in soil phosphorus (compare Chapter 3). One possible explanation is that high C losses have occ urred within this local region as a result of the phosphorus enriched inflow from the WCSs. Because of phosphorus enriched inflow from the adjacent Everglades Agricultural Area (EAA), typically, this area has shown high soil total phosphorus concentrations and dense vegetation communities dominated by cattail (Grunwald et al., 2004, 2008; Rivero et al., 2007). It is known that cattail detritus are better electron donors in methane production than sawgrass because of their cellulose and lignin contents. This was underpinned by Gerard and Chanton (1993) and Wright and Reddy (2007) showing significantly increased methane production in phosphorus enriched sites of WCA 2A. Pinney et al. (2000) also reported high methane release within cattail communities in a constructed wetland in Arizona, U.S. This loss of C might produce profoundly low C stocks in phosphorus enriched areas. The prediction maps of the RFSPOT and RFETM+ models clearly showed low TC stocks in several tree islands in WCA 2A, which occur on ridges and are less inundated than the surrounding marsh. Although we did not collect any samples from tree islands, our prediction results showed some agreement with previously observed TC values in tree islands with low or similar TC values when compared to mar sh soils (Hanan and Ross, 2010). The prediction maps also showed TC variations caused by the sloughridge system mostly in the central and south of WCA 2A. Rivero et al. (2007) suspected that the sloughridge system might affect TC variations, and our results clearly showed

PAGE 123

123 this on prediction maps of WCA 2A. The RFMODIS model had too coarse spatial resolution to capture either the tree islands or the sloughridge system in WCA 2A as shown in Figure 4 4. The 10 20 cm depth soils showed less spatial variability than 0 10 cm depth soils. High TC stocks in 10 20 cm depth dominantly occupied the boundary of the study area, and the interior of WCA 2A showed relatively low and homogeneous spatial distribution of TC stocks when compared to the 0 10 cm soils The boundary areas usually have deep water depth because of canals, and this creates longer hydroperiods than the interior of the marsh. The soils under long hydroperiods have more chance to retain organic matter and have shown a great ability to obtain high TC (Osborne et al., 2011). There was less spatial variability of soil TC stocks in the 10 20 cm soils due to less pronounced effects of tree islands and the sloughridge features. This was expected since the organic matter in subsurface and deep soi ls are highly decomposed, and as a result, they have a more homogeneous distribution than surface soils. Our results suggest that the RF models with RS images derived spectral indices are robust not only to predict soil TC stocks in the top soil but also t o predict subsoil C stocks, although RS images can only observe the earths surface features. Impacts of each predictor variable varied among the RF models (Figure 4 5). Principal component variable 1 (PC 1) and distance to WCSs ranked as the most imp ortant variables in the RFSPOT model to predict soil TC stocks for 0 10 cm and 10 20 cm, respectively. Principal component variable 1 in the RFSPOT model was derived from NIR and SWIR bands of the SPOT image. This variable explained about 75 % of spectral signatures of WCA 2A captured by the SPOT image, and it represented

PAGE 124

124 different land covers (i.e., vegetation and water) of the study area. Distance to WCSs contained combined effects of nutrient loadings and inflow across the study area. For the RFETM+ model, the mid infrared index and tasseled cap transformed variable 2 ranked as the best predictor variables to predict soil TC stocks for 0 10 cm and 10 20 cm, respectively. The midinfrared index showed strong correlation with soil moisture (Musick and Pelletier, 1988), while tasseled cap transformed variable 2 contained greenness (vegetation) information (Kauth and Thomas, 1976). For the R FMODIS model, distance to WCSs and the NIR reflectance variable ranked high to infer on soil TC stocks for 0 10 cm and 10 20 cm, respectively. The near infrared reflectance variable is very responsive to the amount of vegetation biomass present. Although we could not find one major common environmental predictor variable that controls the spatial distribution of C stocks, either among the models or between soil depths in WCA 2A as indicated by variable importance rankings (Figure 4 5), most of the highly ranked importance variables carried information relating to hydrology (e.g., PC 1, distance to WCSs, and midinfrared index) and/or vegetation (e.g., tasseled cap transformed variable 2, near infrared reflectance). Other studies reported to pographical factors as important variables for soil C predictions (Minasny et al., 2006; Grimm et al., 2008; Wiesmeier et al., 2011). However, our results showed a low ranking of elevation in the prediction models. It suggests that topographical variables are less important in controlling the distribution of C stocks in our study area, which is dominated by smooth/flat topography and macrophytes commonly found in subtropical wetlands in Florida.

PAGE 125

125 Total Carbon Stocks in Water Conservation Area 2A The RFSPOT and RFMODIS models showed similar total amount of TC stocks, which are 1.96 mega tons (Mt) for 0 10 cm and 2.24 Mt and 2.23 Mt for 10 20 cm soils, respectively (Table 4 4). The RFETM+ model yielded slightly lower amounts of TC stocks: 1.92 Mt for 0 1 0 cm and 2.19 Mt for 10 20 cm soils in WCA 2A. The lower amounts of TC stocks for the RFETM+ model were expected since the no data areas, expressed as blank stripes in the prediction map, were not counted as a part of the study area in the RFETM+ model (Figure 4 4). Overall, all of the RF models yielded similar amounts of TC stocks. For comparison, TC stocks using observed mean and median values weighted by the total area of WCA 2A were estimated, and they showed slightly lower TC stocks than the RF mod els (Table 4 4). Based on the prediction results, about 4.2 Mt (+/ 0.05) of C was stored in the top 20 cm of the study area. Although we only estimated TC stocks in the top 20 cm soils, a contribution of deep soils to global soil C stocks has been propos ed to be much greater than surface soils. For instance, Jobbgy and Jackson (2000) estimated that more than 50 to 70 % of soil organic C is stored at depths greater than 20 cm. Schmidt et al. (2011) estimated that more than half of the global soil C stocks are stored in deep soils. This might be especially true in wetland soils consisting of deep layers of peat, although most of the studies that measured soil C stocks in deep soils were conducted in upland areas and did not include any wetland areas. The av erage depth of soils in WCA 2A was 154 cm according to Kim et al. (2012), and the average depth of O horizon (consisting mainly of organic material) was 136.4 cm. Since the soils in our study area were classified as Haplosaprists, the O horizon contained m ost of the soil organic matter, and typically, organic matter contains about 45 % to 50 % C (Kayranli et al., 2010). Consequently, the

PAGE 126

126 O horizon contained the majority of soil organic C in WCA 2A. Therefore, the estimated 4.2 Mt of C (0 20 cm depth) might be only a small portion of the total amount of C stored in this wetland, and a substantial greater amount of C is expected to be stored in Haplosaprists soils below 20 cm. According to Bhatti and Tarnocais (2009) reported data, only 14 % of soil C stock s were stored in the top 30 cm of Organic soils (Histosols in U.S. taxonomic classification) in Canada. Water Conservation Area2A only accounts for 6.5 % of the total area of the Everglades region, and previous studies reported higher soil organic matter and soil organic C concentrations in other units, such as WCA 1 and WCA 3 (Corstanje et al., 2006; Osborne et al., 2011). Moreover, the C content of ENP soils is much greater than in WCAs because of a high portion of inorganic C in marl soils (Osborne et al., 2011). Table 4 5 shows soil C stocks in wetlands located in various regions. For better comparison, TC stocks per unit area (kg m2) were calculated since the extent differed among studies. The RF models in WCA 2A yielded about 9.7 kg soil TC stocks p er unit area in the top 20 cm. It was higher than 7.3 kg m2, which is the estimated median of soil TC stocks stored in all of the U.S. wetlands in the top 20 cm using the State Soil Geographical database according to Guo et al. (2006). Canadian wetlands s howed 124 kg m2 throughout the whole soil profile (Roulet, 2000). Comparisons with other wetlands show that soil C stocks in WCA 2A are higher than other wetlands, except boreal wetlands (Table 4 5). According to Post et al. (1982) and Post and Kwon (2000) soil C stocks follow a geographic gradient where soil C stocks are primarily controlled by precipitation (water table position) and soil temperature. A water table position in wetland soils can

PAGE 127

127 signi ficantly alter the C balance. Specifically water table position is related to methane emissions from a wetland. Comprehensive studies conducted in various wetlands have shown a strong relationship between water table position and methane emissions where hi gher water table positions in wetland soils increased methane emissions (Harris et al., 1982; Augustin et al., 1998; Fiedler and Sommer, 2000; Altor and Mitsch, 2008; Berryman et al., 2009). High water table positions increase anaerobic zones, and hence, l ead to elevated methane emissions from a wetland due to methanogenic bacteria which are obligate anaerobes. Studies conducted in the Everglades region have shown methane emission rates ranging from 40 to 140 mg m2 d1 depending on the location of observat ion points, vegetation, and seasons (Gerard and Chanton, 1993; Wright and Reddy, 2007). This range is much broader than the average emission rate (5 to 80 mg m2 d1) of North America peat lands reported by Blodau (2002). Methane emissions are also related to soil temperatures (Augustin et al., 1998). Previous studies have shown higher methane emissions from wetlands located in tropical/subtropical climate regions than in others such as temperate or boreal zones (Sorrell and Boon, 1992; Melack et al., 2004; Mitsch et al., 2012). In contrast, WCA 2A has also acted as a C sink. A previous study conducted in WCA 2A reported a soil C sequestration rate of 86 387 g C m2 yr1 (Reddy et al., 1993), and it was much greater than the estimated average soil C seques tration rate of North America wetlands, 29 g C m2 year1 reported by Gorham (1991). Guo et al. (2006) reported that wetland soils sequester a much greater amount of C than other ecosystems, and numerous studies have shown that a wetland acts as a C sink r ather than a C source when it is managed appropriately (Armentano and Menges, 1986; Lal, 2003; Koelbener et al., 2010; Mitsch et al., 2012).

PAGE 128

128 Our results of the comparison between 2003 and 2009 showed a significant decrease of soil TC concentrations (Figur e 4 3), although previous studies showed that WCA 2A has sequestered C into soils (Reddy et al., 1993; Craft and Richardson et al., 1993; Craft and Richardson et al., 1998). This might be explained by the water table positions and C oxidation. Water table positions (i.e., inflow rate at WCSs) of WCA 2A, unlike other natural wetlands, are controlled by the South Florida Water Management District (SFWMD) to prevent floods/droughts and other ecosystem services (e.g., water birds). However, South Florida has ex perienced a prolonged drought since 2007, and the SFWMD decreased the inflow rate into WCA 2A. Water table positions below the soil surface have exposed highly organic soils of the wetland to the atmosphere during the drought (Figure 4 1), and it might hav e released a considerable amount of C stored in this wetlands soils and decreased soil TC concentrations. This extreme climate event (i.e., prolonged draught) would make this wetland lose its role as a C sink. Bellamy et al. (2005) reported that C loss fr om peat soils is much faster than other soil types. They also found that the rate of C loss increased with original organic C content. Ross (2011) also reported the largest net C loss in Histosols (Saprists) in the St. Johns River Watershed, Florida. Since soils in WCA 2A are dominated by high organic matter (Saprists) it is at risk to lose large amounts of C under prolonged drought conditions potentially intensified through global climate warming. Conclusions Our study showed successful model development to predict soil TC stocks in WCA 2A using RS images derived spectral indices and environmental ancillary variables. The RF models with varying spatial resolutions of RS images yielded similar amounts of TC stocks. Although we could not find a specific envi ronmental or spectral

PAGE 129

129 variable that controls the spatial distribution of soil TC stocks in the area throughout depths, several variables related to hydrology and vegetation ranked high in the variable importance of the prediction models. Our results showed that WCA 2A stored a tremendous amount of C in soils and hydrologic regime, such as water table position, was the most important factor controlling C sequestration or loss in the wetland. It suggests that inflow rate controlled by the SFWMD should be cons idered for its effect on C cycling in the region and further global C cycling. Disruptions in hydrological regimes of the wetland to maintain the upper basins (i.e., EAA) water level can significantly impact its function as a long term C sink. Our results showed a significant decrease of TC concentrations between 2003 and 2009. While these results point out trends, suggesting that C losses from soils were pronounced, they need further investigation. However, one important conclusion is that appropriate management of the wetland is critical to maintain its potential ability to sequester C into soils. Large changes in temperature and precipitation in the Everglades region are predicted to occur over the next several decades. Thus, sequentially reproducible as sessment methods are necessary for measuring and monitoring the impacts of those changes in C storage capacity of the wetland. Our methodology (RS supported DSM) suits the need well and it will provide useful information to preserve the wetlands role as a C sink. Summary Wetland soils are able to exhibit both consumption and production of greenhouse gases and they play an important role in regulation of global carbon (C) cycles. Yet it is challenging to accurately evaluate the actual amount of C storag es and sequestration or emission rates in wetlands. The incorporation of remote sensing (RS) data into digital

PAGE 130

130 soil models has great potential to assess C dynamics in wetland soils. Our objectives were to (i) develop C stock prediction models utilizing RS images and environmental ancillary data, (ii) identify the largest predictive environmental factors controlling the spatial distribution of soil C, and (iii) assess the amount of C stored in the investigated wetland and evaluate its role as a C sink or a C source. We collected a total of 108 soil cores at two soil depths (0 10 cm and 10 20 cm) in a C rich ecosystem: Water Conservation Area2A (WCA 2A), Florida, U.S. Random Forest models to predict soil C were developed using field observation data, envi ronmental ancillary data, and spectral data derived from RS images including SPOT (10 m), Landsat ETM+ (30 m), and MODIS (250 m). The RF models showed high performance to predict TC stocks with a R2 between 0.85 to 0.92 and a root mean squared error between 0.75 to 0.86 kg m2. The variable importance of the RF models revealed that hydrology was the major environmental factor controlling the spatial distribution of soil C stocks in WCA 2A. Our results showed that WCA 2A stores about 4.2 mega tons of C in th e top 20 cm soils.

PAGE 131

131 Table 41 Descriptive statistics for soil total carbon observed in Water Conservation Area 2A (n=108). Property Depth Mean Min. Max. Median Std. deviation Concentrations (g kg 1 ) 0 10 cm 410.5 262.8 492.3 413.2 35.6 10 20 cm 432.9 325.5 486.7 438.1 32.5 Stocks (kg m 2 ) 0 10 cm 4.5 2.2 10.8 4.0 1.6 10 20 cm 5.2 1.3 11.2 4.7 1.6

PAGE 132

132 Table 42. Comparison of soil total carbon concentrations (g kg1) in Water Conservation Area 2A. Year (Season) Depth Mean Median Sample number Reference 1990 (wet) 0 10 10 20 20 30 420 450 458 434 459 479 74 74 74 DeBusk et al., 1994 1996 (wet) 1997 (dry) 0 10 10 30 0 10 10 30 433.0 451.7 414.7 433.0 n/a a n/a n/a n/a 8 8 8 8 Wright and Reddy, 2001 2003 (dry) 0 10 10 20 426.7 444.1 435.4 454.9 111 110 Rivero et al., 2007 a n/a, non available

PAGE 133

133 Table 4 3 Summary of model performance assessment for soil total carbon stocks ( kg m2) a. Depth Statistical measure b RF SPOT RF ETM+ RF MODIS 0 10 cm R 2 0.85 0.89 0.89 RMSE 0.75 0.77 0.80 10 20 cm R 2 0.92 0.90 0.87 RMSE 0.79 0.78 0.86 aRFSPOT,EMT+,MODIS, Random Forest models using SPOT, Landsat ETM+, and MODIS images derived input variables, respectively. bR2, coefficient of determination; RMSE, root mean squared error.

PAGE 134

134 Table 4 4. Predicted total soil carbon stocks (Mta, mega tons) using Random Forest in Water Conservation Area2Ab. Depth Simple Estimation c RF SPOT RF ETM+ RF MODIS Mean Median 0 10 cm 1.88 1.67 1.96 1.92 1.96 10 20 cm 2.17 1.96 2.24 2.19 2.23 aMt, mega tons (109 kg). bRFSPOT,EMT+,MODIS, Random Forest models using SPOT, Landsat ETM+, and MODIS images derived input variables, respectively. cSimple estimation using observed mean and median values of TC stocks multiplied by the total area of WCA 2A (41,773 m2), respectively.

PAGE 135

135 Table 4 5. Comparison of soil carbon stocks Property a Climate (Area) Land use Depth (cm) Total stocks (Mt) b Stocks per unit area (kg m -2 ) Standardized stocks by depth (kg m-2) c Reference TC Subtropical (Florida) Wetland (marsh) 0 20 4.2 9.7 9.7 This study SOC Semi arid (China) Wetland (marsh) 0 100 5,680 33.5 6.7 Wiesmeier et al. (2011) TC Semi arid (China) Wetland (marsh) 0 100 6,350 37.4 7.5 SOC Boreal (Canada) Wetland (bog, fen, marsh) 0 30 14,900 18.7 12.5 Bhatti and Tarnocai (2009) SOC Tropical (Brazil) Wetland A (riparian) 0 30 n/a 11.8 7.9 Wantzen et al. (2012) Wetland B (riparian) 0 30 n/a 8.6 5.7 Wetland C (riparian) 0 30 n/a 6.8 4.5 TC Temperate (Ohio) Wetland A (swamp) 0 24 n/a 14.7 12.3 Bernal and Mitsch (2008) Wetland B (riverine) 0 24 n/a 9.0 7.5 Tropical (Costa Rica) Wetland C (swamp) 0 24 n/a 6.9 5.8 Wetland D (slough) 0 24 n/a 8.0 6.7 Wetland E (riverine) 0 24 n/a 6.8 5.7 TC U.S.A. Wetland 0 20 2,310 d 7.3 7.3 Guo et al. (2006) a SOC, soil organic carbon; TC, total carbon b Mt, mega tons (109 kg). c Depth standardized density (kg m-2) within top 20 cm under the assumption that soil carbon is uniformly distributed with depth. d Median value.

PAGE 136

136 Figure 41. Low water table position and limited accessibility of sampling point by airboats of southern part of Water Conservation Area2A during February, 2009 sampling event.

PAGE 137

137 Figure 42. Spatial distribution of measured soil total carbon: a) concentrations (g kg1) and b) stocks (kg m2) by depth within the Water Conservation Area2A.

PAGE 138

138 Figure 43. Total carbon concentration (g kg1) of the 0 10 cm and the 10 20 cm depth in 2003 (gray bar) and 2009 (white bar). The different letters demote significant differences between the years and depths ( p < 0.05). The data shown in the gray columns were adapted from Rivero et al. (2006).

PAGE 139

139 Figure 44. Prediction maps of soil total carbon stocks (kg m2) in the 0 10 cm (above) and the 10 20 cm (below) using Random Forest: a) Random Forest with the SPOT image, b) Random Forest with the Landsat ETM+ image, and c) Random Forest with the MODIS image derived input variables combined w ith environmental predictor variables. Stripes in map b' were caused by permanent scanline corrector failure (SLC off) of Landsat Enhanced Thematic Mapper Plus sensor since May 2003.

PAGE 140

140 Figure 45. Variable importance of predictor variables i n Random Forest models predicting soil total carbon stocks in Water Conservation Area2A (X axis: increased node purity, Y axis: predictor variable; see Table 3 2 in Chapter 3 for environmental input variables abbreviations).

PAGE 141

141 CHAPTER 5 EVALUATING MODEL TRA NSFERABILITY AND SCALING IN ADJACENT SUBTROPICAL WETLANDS : WATER CONSERVATION AREA 2A AND 3A NORTH THE EVERGLADES, FLOR IDA Overview There is an increasing demand for global models in soil science that provide information about the spatial variation of soi l properties. The accumulation of a massive amount of available datasets and advanced computer technologies has enhanced the capability to predict soil properties on a global scale. The problem, however, is that soil observations or measurements are sparse and made on relatively small areas, and assessments of soils may require larger regions (McBratney, 1998; Urban, 2005). Empirical or mechanistic soil models developed on a fine scale (e.g., experimental plot, laboratory, or small specific study area) may or may not perform well at a coarse scale (e.g., watershed, region, or global). This unavoidably involves extrapolation or the transfer of information across scales (referred to as scaling). Because of the heterogeneity and non linearity of soils, adding or multiplying soil property data identified at fine scale to coarse scale is misleading and the effects of scaling must be considered (Grunwald et al., 2011). Scaling is related to manipulations of grain and extent (Bl 1998; Schneider, 2001; Grunwald et al. 2011), and Thrush et al. (1997) emphasized the influence of the extent (place) of the model. For this reason, the application of a model developed at one extent to another extent is necessary to understand scaling e ffects. Chaplot et al. (2003) investigated model transferability across a region that has different landform factors, and they found greatly decreased prediction power for the hydromorphic index in France. Thompson et al. (2006) applied a model developed s pecifically for one area to another area, both

PAGE 142

142 areas showing the same landform factors, and they showed limited model transferability in predicting soil organic matter content in Kentucky, U.S. More recently, Vasques et al. (2012b) comprehensively investig ated model transferability across three different extents for soil carbon (C) in Florida, U.S. However, research gaps still exist to better elucidate how soil prediction models change with increasing size of the study area size (i.e., how the models behave across escalating spatial scales) and how soil prediction models developed in one region transfer to another one with similar ecosystem characteristics. Digital soil mapping (DSM) explains and quantifies the spatial distribution of soil properties using soil forming factors. Grunwald (2009) comprehensively reviewed various approaches of DSM and soil mapping and documented the increase of its applications. Additionally, she noted an increase in the use of remote sensing (RS) in DSM. Remote sensing informed digital soil models potentially have the capability to overcome the extent limitations, because they provide dense information to infer on landscape features over large areas. Additionally, the benefits of the RS informed soil models for wetlands are greater than the upland models specifically considering the efforts of field sampling. Therefore, we hypothesized that the soil prediction models developed in one specific aquatic area are transferable to another area with similar environmental landscape characteristics. Our specific objectives were to (i) develop prediction models for soil total phosphorus (TP), total nitrogen (TN), and total carbon (TC) utilizing RS images with different grain sizes, (ii) identify the environmental variables controlling the s patial distribution of soil TP, TN, and TC, (iii) investigate the model transferability of soil prediction models to an adjacent region, and (iv) analyze the

PAGE 143

143 behavior of soil models across escalating spatial extent (upscaling) and across declining extent (downscaling). Methods and M aterials Study Area The Everglades, located in south Florida in the United States, has been impacted by urban and agricultural development along its boundaries. The extensive constructions of levees and canals to provide flood control and water supply to agricultural and urban areas have modified the natural hydrology patterns in the area ( Noe et al., 2001; Bernhardt and Willard, 2009). The levees and canals fragmented the Everglades into several hydrologic units, including Water Conservation Areas (WCAs) and the Everglades National Park (ENP). The study was conducted in WCA 2A, which covers about 418 km2, and WCA 3A north (3AN), which covers about 722 km2. These adjacent wetlands are located in the northern part of the Everg lades (Figure 5 1). The WCA 3AN was divided into two zones by the Miami canal. The climate and soils in the area are similar to WCA 2A as described in Chapter 2 and 3. The hydrology of WCA 3A is managed by the South Florida Water Management District (SFWMD ) controlling water control structures (WCSs), the same as with WCA 2A. About 36% of the hydrologic inputs, excluding precipitation, comes from WCS 2A through the WCSs (i.e., S 11A, S 11B, and S 11C in Figure 5 2; SFWMD DBHYDRO data). The t opography of WCA 3AN is flat with an average 2.5 m above sea level (United States Geological Survey, USGS, High Accuracy Elevation Data, HAED, 2007). This wetland is primarily composed of ridges with sawgrass ( Cladium jamaicense) alternating with deep sloughs and tree isl ands.

PAGE 144

144 Field S ampling and Soil Property Measurements A subset (66 sites) of sampling sites identified in 2003 (147 sites) was collected from WCA 3AN (compare Bruland et al., 2006). A stratified random sampling design using historic soil, hydrologic, and ecological sites was used in 2003. A field sampling was conducted in July 2010 by airboats, and soil cores were collected using a 10 cm diameter stainless steel coring tube. The low water table and dense vegetation limited site accessibility by airboats; hence no samples were collected from the northeast part of WCA 3AN (Figure 5 2). Soil samples collected in the field were packed with doublezipped plastic bags and transported to the laboratory. All samples were ovendried at 70 C for 72 hours, and all procedures for soil property measurements were done as described in Chapter 3 and 4. Bulk density (BD) was measured. The soil samples were analyzed for TP utilizing the association of official analytical chemists (AOAC) 978.01 method, for TN utilizing the LECO combustion method (LECO TruSpec, MI, USA), and for TC utilizing a combustion method with a Shimadzu SSM 5000A TC analyzer (Shimadzu Corporation, Columbia, MD, USA). Predictors for the Model Development: Spectral and Environmental Variables Spectral data derived from RS images and environmental ancillary variables such as topographical, hydrological, and lithological properties were used as predictors for development of the soil property models. All of the RS image processing and geographic information system (GIS) data processing work was performed using ERDAS Imagine 2010 software (Earth Resource Data Analysis System Inc., Atlanta, GA, USA) and ArcGIS 10 (Environmental Systems Research Institute ESRI Inc., Redlands, CA, USA), respectively.

PAGE 145

145 Satellite Pour lObservation de la Terre (SPOT), which has a 10 m spatial resolution, and Moderate Resolution Imaging Spectroradiometer (MODIS), which has a 250 m spatial resolution, were selected to build soil property prediction models for WCA 3AN and test the model transferability. Detailed descriptions about the SPOT image (January 2009) and the MODIS image (February 2010) used in this study can be found in Chapter 2. Both of the images were cloud free and had good area coverage for WCA 3AN, although the SPOT image did not cover the whole study area. About 83% of the area was covered by the SPOT image. Landsat Enhanced Thematic Mapper Plus (ETM+), which has a 30 m spatial resolution, could not be used in this chapter. Because of a scanline corrector failure from May 2003, the Landsat ETM+ image had stripes (i.e., no data) over the entire W CA 3AN area, and this limited the use of the Landsat ETM+ image. Spectral indices such as Normalized Difference Vegetation Index (NDVI) have been widely used to assist in the mapping of soil properties and they have shown strong correlations with soil properties (Huete et al., 1997; Fernndez Buces et al., 2006; Rivero et al., 2009; Kim et al., 2012). Various spectral indices were derived from the SPOT and MODIS images and used as predictor variables to develop soil prediction models in WCA 3AN. A principal component analysis (PCA) was performed with bands 1 to 4 for the SPOT and MODIS images. A PCA reduces the number of correlated variables into a smaller number of uncorrelated variables and helps to derive possible relationships that were muted between a t arget (i.e., a specific soil property) and original spectral signatures of the images.

PAGE 146

146 Elevation data surveyed by USGS using 400 m grid space sampling (15 cm vertical accuracy) and distance to WCSs derived using the Euclidean distance function represen ted the topography and hydrology of the area, respectively. The geophysical data including potassium concentration gammaray, bouguer gravity anomaly, isostatic residual gravity anomaly, and magnetic anomaly data surveyed by USGS were included to represent lithology of the area. Details of spectral indices and environmental predictor variables were described in Chapter 3. Soil P rediction Model Development and Transferability Assessment Random Forest (RF) was employed to develop soil prediction models and id entify environmental variables that control the spatial distribution of soil TP, TN, and TC in WCA 3AN. Among the numerous available methods, the RF was chosen because it is wellsuited to small observation sets and large sets of predictor variables (Strobl et al., 2009). Also it has shown low bias and low variance (Daz Uriarte and de Andrs, 2006; Wiesmeier et al., 2011), which are important to maintain stability for model transfer. The RF is an ensemble of classification or regression trees that are calc ulated on random subsets of the data (bootstrap samples) using a randomized subset of predictor variables at each tree. The RF builds hundreds or thousands of trees and produces one single prediction by averaging the trees (Breiman, 2001) Details of the RF algorithm and descriptions can be found in Chapter 3 and 4. To test how transferable the models developed in one area are to another area, the RF mo dels using RS images derived spectral variables and environmental predictor variables were developed within their specific geographic domain and then transferred to each other (objective iii). First, the RF models developed in WCA 2A in Chapter 3 (for TP a nd TN) and Chapter 4 (for TC) using the SPOT (RFSPOT_WCA2A) and MODIS

PAGE 147

147 (RFMODIS_WCA2A) images were transferred to WCA 3AN. Second, the RF models developed in WCA 3AN using the SPOT (RFSPOT_WCA3AN) and MODIS (RFMODIS_WCA3AN) images and ancillary environmental variables were transferred to WCA 2A. To implement objective (iv), the overall RF models using all data from both areas (WCA 2A and WCA 3AN) were developed with the SPOT (RFSPOT_ALL) and MODIS (RFMODIS_ALL) images and applied to WCA 2A and WCA 3AN (downs caling), and vice versa (upscaling). All RF models were complemented by including various additional environmental covariates as predictor variables (compare Chapters 3 and 4). For the statistical model assessments using cross validation, the coefficient of determination (R2), the root mean squared error (RMSE), and the residual prediction deviation (RPD) were derived according to: n i i n i iy y y y R1 2 1 2 2 ( 5 1) n y y RMSEn i i i 1 2 ( 5 2) ) 1 ( 11 2 n n RMSE n y y RPDn i i ( 5 3) where iy are the predicted values, y is the mean of observed values, iy are the observed values, and n is the number of predicted/observed values. The RF analysis was performed using the randomForest package (Liaw and Wiener, 2002) in the R statis tical language (R Development Core Team, 2012). Maps of each RF model were produced using the same spatial resolution as the respective RS images using the raster package (Hijmans and van Etten, 2012). As a result, the

PAGE 148

148 prediction maps of the RFSPOT and R FMODIS models had 10 m and 250 m spatial resolution, respectively, due to the inherent grain resolution of the satellite images derived from SPOT and MODIS. Results and Discussion Descriptive Statistics of Soil Properties in Water Conservation Area 3A Nor th and Comparison with Water Conservation Area 2A Descriptive statistics of soil properties in WCA 2A and WCA 3AN are shown in Table 5 1. The mean of the BD in WCA 3AN was 0.15 g cm3, which was a little higher than the mean in WCA 2A. It indicates that soils in WCA 3AN contain less organic matter than soils in WCA 2A. Specifically, the soils in the northwestern part of WCA 3AN showed BDs larger than 0.24 g cm3, which was the maximum BD in WCA 2A. This was expected since this area has experienced chronic overdrainage (Craft and Richardson, 1993; David, 1996), and much of the organic matter in the area may have been oxidized or lost because of the drainage and low water table positions. Figure 5 2 shows the spatial distribution of observed soil TP, TN, an d TC concentrations in WCA 3AN. Soil TP in WCA 3AN showed a mean concentration of 574.9 m g kg1 with a median of 559.1 mg kg1. In WCA 3AN TP showed much smaller variability (observed concentration data range of 621.6 mg kg1) when compared to WCA 2A (Tabl e 5 1). The mean soil TP concentration in 2009 was higher than the TP means observed in 1992 (461 m g kg1) and 2003 (523 m g kg1) in WCA 3AN (Bruland et al., 2007), although sampling sizes and locations varied by year. This may be explained by continuous P loading from the Everglades Agricultural Areas (EAAs) through S 8 WCS and WCA 2A through S 11A, S 11B, and S 11C entering WCA 3AN. More than half of the P loads of outflow from WCA 2A have entered into WCA 3AN (Figure 5 3),

PAGE 149

149 and this is about 16% of the total P loads of WCA 3, including WCA 3AN, WCA 3AS, and WCA 3B (SFWMD DBHYDRO data). The spatial distribution of soil TP concentrations in WCA 3AN was not pronounced as WCA 2A. As shown in chapter 3, soil TP in WCA 2A had concentrated P enrichment area near WCSs, whereas WCA 3AN showed a relatively homogeneous distribution of soil TP across the area. Total N and TC showed wider ranges in WCA 3AN when compared to WCA 2A. Soil TN showed a mean concentration of 30.5 g kg1 with a median of 32.0 g kg1, and soil TC showed a mean concentration of 388.3 g kg1 with a median of 412.3 g kg1 in WCA 3AN (Table 5 1). Interestingly, the median and maximum values of soil TN and TC were similar in WCA 2A and WCA 3AN, but showed contrasting minimum TN and TC values. Soil s amples from the northwestern part of WCA 3AN had less organic matter, which indicates that the soils in this area have more of a chance to lose highly organic nutrients, such as C and N, because of oxidation. Phosphorus is relatively more stable than C and N because P could be bound with iron (Fe), aluminum (Al), calcium (Ca), and magnesium (Mg). The mean value of TN observed in 2009 was comparable with a mean value (29.1 g kg1) observed in 1992 (Reddy et al., 1998), but higher than the mean value (24.6 g kg1) observed in 2003 (Bruland et al., 2006). The mean value of TC was lower than a mean value (412 g kg1) observed in 1992 (Reddy et al., 1998), but higher than a previously observed mean value (353 g kg1) in 2003 in WCA 3AN (Bruland et al., 2006).

PAGE 150

150 Soil Property Prediction Models using Random Forest in Water Conservation Area 3AN Spatial predictions The statistical assessment of the RF models for soil TP, TN, and TC concentrations using the SPOT and MODIS derived input variables combined with e nvironmental predictor variables are shown in Table 5 2 (i.e., RFWCA3AN). The RF models showed high performance in predicting soil properties in WCA 3AN. The R2 values ranged from 0.87 to 0.91, and the RMSE values were smaller than 59.6 mg kg1 for soil TP 2.9 g kg1 for soil TN, and 34.6 g kg1 for soil TC models. The results of Chapters 3 and 4 also showed high performance of the RF models in predicting soil properties in WCA 2A. Spatial prediction maps for soil TP, TN, and TC concentrations using the R F models with RS images derived spectral input variables and other environmental predictor variables for WCA 3AN are shown in Figure 5 4. The SPOT image that was used for the predictions could not cover the whole area of WCA 3AN, and this limited the spatial prediction in the west of the study area. In general, both of the RFSPOT WCA3AN and RFMODIS WCA3AN models showed similar spatial distributions for soil TP, TN, and TC. Soil TP prediction maps showed high concentration areas in the northwest and the central part of WCA 3AN. Also, the TP prediction maps showed continuously high concentration areas along the Miami canal. Phosphorus enriched water from the Everglades Agricultural Area enters into WCA 3AN through the S 8 and G 404 WCSs, and it has created the high concentration areas around these WCSs and the northwestern part of WCA 3AN. Moreover, the P enriched water flows from north to south along the Miami canal, and the S 339 WCS delivers water from the Miami canal

PAGE 151

151 to the interior of WCA 3AN. This has created high P concentration areas around the S 339 WCS, especially the central part of the study area. High concentrations of the east boundary of the study area could also be explained by the P enriched inflows from WCA 2A through the S 11 WCSs. Overal l, the spatial prediction results for soil TP concentrations in WCA 3AN are consistent with findings documented in Bruland et al. (2006), although the spatial prediction methods differed from each other. However, predicted soil TP concentrations increased throughout the whole study area when compared to the Bruland et al.s study conducted in 2003. They estimated about 25% of WCA 3AN had elevated soil TP levels, i.e., soil TP concentrations greater than 500 mg kg1, but our results showed about 84% of WCA 3 AN had elevated TP levels based on the RFMODIS WCA3AN model. Specifically, the north and east boundary areas and the interior of the marsh showed increased TP concentrations. As discussed in Chapter 3, the expansions could be explained by increased P mobil izations and continuous P loading into the system, the same as in WCA 2A. In particular, there have been excessive P loadings into WCA 3AN in May 2005 through April 2006. About 33.4 metric tons of P has been loaded through the S 8, G 404, and S 11s WCSs during this period according to the SFWMD (Figure 5 3). The spatial distribution of soil TN and TC concentrations showed similar patterns in terms of high and low concentration areas, except in the northeast part of the area. Overall, predicted soil TN and T C concentrations were low in the west part and high in the south part of WCA 3AN. Similar to the results given in Chapter 3 and 4 in WCA 2A, high P concentration areas showed generally low TN and TC concentrations in WCA 3AN. The P enriched environment has a higher likelihood of N and C loss from the

PAGE 152

152 system through increasing denitrification rates and methane release (Craft and Richardson, 1993; Gerard and Chanton, 1993; Pinney et al., 2000; Wright and Reddy, 2007). Our prediction results are similar to predicted results using the Ordinary Kriging method documented by Bruland et al. (2006). Variable importance The impact of each predictor variable varied among the soil properties and the RF models (Figure 5 5). For soil TP predictions, the RS im age derived spectral variables, such as TVI, NDVI green, and SR ranked high in the RFSPOT WCA3AN model. In contrast, nonRS derived input variables, such as magnetic anomaly, distance to WCSs, and elevation ranked high in the RFMODIS WCA3AN model. Th e rankings of the variable importance of the RF models in WCA 3AN are similar to the results of the RF models developed in WCA 2A for soil TP predictions in Chapter 3. For the RFSPOT WCA3AN model, most of the highly ranked spectral indices use red and NIR spectral bands, and these bands carry the major signals on vegetation conditions (Rouse et al., 1974; Cohen, 1991; Chen, 1996). Previous studies have shown that vegetation is a good indicator of soil phosphorus status of the Everglades (Noe et al., 2002; J uston and DeBusk, 2006; Noe and Childers, 2007). The high rank of TVI was expected since WCA 3AN has experienced various fires in the 1970s and 1999 (Bruland et al., 2006). Fire, which is one of the major stressors in natural ecosystems, plays a critical r ole in nutrient cycles of many ecosystems, and especially the Everglades is an especially fire frequented ecosystem (Beckage et al., 2005). Nellis and Briggs (1992) showed a strong relationship between TVI and above ground biomass with different burning frequencies in Kansas. For the RFMODIS WCA3AN model, magnetic anomaly, which reflects the distribution of Fe minerals in the rocks of the Earths crust, ranked as the

PAGE 153

153 best predictor variable for soil TP predictions. This may be explained by the Feand A l bound P, which accounted for about 23% of TP in WCA 3A (Reddy et al., 1998). Distance to WCSs in WCA 3AN reflects indirectly the P loadings into the system, and previous studies conducted in WCA 2A have shown a strong relationship between soil TP and W CSs (Rivero et al., 2007; Chapter 3). The EVI, which provides improved sensitivity in high biomass areas with less atmospheric influence than other spectral indices for MODIS images, ranked high among the spectral indices in the RFMODIS WCA3AN model. For s oil TN and TC predictions, nonRS derived input variables ranked high for both of the RFSPOT WCA3AN and RFMODIS WCA3AN models. Specifically, bouguer gravity anomaly and isostatic residual gravity anomaly, which contain similar information about the density distribution of the Earths upper crust, ranked high. It is interesting to note the great decrease of relative importance, which is expressed in %, of the input variables in both of the RFSPOT WCA3AN and RFMODIS WCA3AN models after the second important variable. The high rank of the geophysical predictor variables, which have a 2,000 m spatial resolution, created distinct squareshaped features on the soil TN and TC prediction maps in Figure 5 4. In general, bouguer gravity anomaly and isostatic res idual gravity anomaly are high in the east and low in the west of the study area. The low values of the anomalies may reflect severe soil subsidence in the western part of WCA 3AN caused by extensive drainage, and, as a result, the area had more chances t o lose soil TN and TC due to peat oxidation under prolonged overdrainage.

PAGE 154

154 Transferability and Scaling Effects of Soil Prediction Models Application of soil prediction models to other extents Random Forest models developed in Water Conservation Area 2A (RFSPOT, MODIS WCA2A) applied to WCA 3AN (Case I) The RFSPOT and RFMODIS models developed in WCA 2A were applied to WCA 3AN. The RF models developed in WCA 2A can be found in Chapter 3 for soil TP and TN, and Chapter 4 for soil TC predictions. Statistical a ssessments for transferability of soil prediction models are shown in Table 5 2 (Case I). The RFSPOT WCA2A models showed an R2 of 0.93 for TP, 0.95 for TN, and 0.94 for TC, which indicate high performance, but the prediction power greatly decreased when the model was transferred to WCA 3AN. For instance, the model showed a R2 of 0.15 for soil TP, and nearly zero for soil TN and TC predictions. The RMSE values for each soil property also greatly increased, and the RPD values decreased when the models were tr ansferred to WCA 3AN. Although the prediction results for soil TP of the RFWCA 2A models were not good (R2 of 0.15 for RFSPOT and 0.003 for RFMODIS) in WCA 3AN, the prediction maps were derived for the comparison (Figure 5 6). In Figure 5 6, each a) map shows the spatial prediction results of soil TP using the RFWCA3AN models (i.e., the original model developed in WCA 3AN), and each b) map shows the spatial prediction results of soil TP using the RFWCA2A models (i.e., the prediction models developed in WCA 2A and applied in WCA 3AN). Interestingly, the spatial patterns of soil TP concentrations in the prediction maps, especially the RFSPOT model, are similar despite the low transferability. The transferred RFSPOT WCA2A model over predicted in high TP concentration areas and under predicted in low TP concentration areas in WCA 3AN. The RFMODIS WCA2A models showed worse results than the RFSPOT WCA2A models for soil TP, TN, and TC predictions when transferred to

PAGE 155

155 WCA 3AN (Table 5 2). This indicates that the models developed in WCA 2A have limited transferability to WCA 3AN. Random Forest models developed in Water Conservation Area 3A North (RFSPOT, MODIS WCA3AN) applied to WCA 2A (Case II) The RFSPOT and RFMODIS models developed in WCA 3AN were applied to WCA 2A. Statistical assessments for transferability of soil prediction models are shown in Table 5 2 (Case II). The RFSPOT WCA3AN models showed high performance for predicting soil TP, TN, and TC concentrations in WCA 3AN as measured by the R2 and RMSE, but the prediction power of the models for soil TN and TC greatly decreased when the models were transferred to WCA 2A. The RF models for soil TP showed some level of transferability to WCA 2A. The RFSPOT WCA3AN and RFMODIS WCA3AN showed an R2 of 0.40 and 0.38, respectively. The RFSPOT, MODIS WCA3AN models for soil TP were applied to WCA 2A, and the prediction maps were derived (Figure 5 7). As with Figure 5 6, each a) map in Figure 5 7 shows the spatial prediction results of soi l TP using the RFWCA2A models developed in WCA 2A (i.e., the original model), and each b) map shows the spatial prediction results of soil TP using the RFWCA3AN models developed in WCA 3AN and applied to WCA 2A (i.e., transferred results). Both TP predic tion maps of the RFSPOT and RFMODIS models had similar spatial patterns in terms of high and low concentrations, but the transferred models showed much less contrast. The transferred RF models predicted much lower in high TP concentration areas and much hi gher in low TP concentration areas. This is the opposite of the transferred results of the models developed in WCA 2A and applied to WCA 3AN. This may be explained by the differences of the observed data range of soil TP in each subarea. As shown in Table

PAGE 156

156 5 1, the range of soil TP in WCA 2A (165 to 1,694 mg kg1) is much wider than in WCA 3AN with 289 to 911 mg kg1. Figure 5 8 shows the observed versus predicted TP values of each transferred model. The models developed using a narrow range of observation data tend to narrow down the prediction range of the target variable like those shown in Figure 5 8b. This also can explain the higher RMSEs when the RFWCA 3AN models were applied to WCA 2A for soil TP predictions. Random Forest models developed in t he combined Water Conservation Area 2A and 3A North Area (RFSPOT, MODIS ALL) applied to separate WCA 2A and WCA 3A North, respectively, and vice versa. The RFWCA 2A and RFWCA 3AN models were applied to the combined area (upscaling), and the RFSPOT and RFM ODIS models developed in the pooled area (i.e., WCA 2A and WCA 3AN) were applied to WCA 2A and WCA 3AN, respectively (downscaling). Statistical assessments for upand downscaling of soil prediction models are shown in Tables 5 3 and 5 4, respectively. In all cases, prediction abilities of the RF models developed in subareas decreased when the models were applied to the combined area (Table 5 3). For instance, the RFSPOT models showed an R2 of 0.94 and 0.89 for soil TC predictions in WC A 2A and WCA 3AN, respectively, and the R2 values were greatly decreased to 0.24 and 0.46, respectively, when the models were upscaled to the combined area. Upscaling of the RFWCA 2A models for soil TP showed much better results with an R2 of 0.82 and 0.73 when compared to the upscaling of the RFWCA 3AN models, which showed an R2 of 0.36 and 0.34. In contrast, the upscaled RFWCA 3AN models for soil TN and TC showed better results when compared to the upscaled RFWCA 2A models. This can be explained by the ra nges of observed soil TP, TN, and TC concentrations in each subarea. The range of

PAGE 157

157 soil TP of WCA 3AN was much narrower than WCA 2A, and could not cover the range of the pooled dataset (i.e., observation data of WCA 2A and WCA 3AN). The narrow range of soi l TP in WCA 3AN limited upscaling performance of the models. Similarly, the ranges of soil TN and TC of WCA 3AN were wider than in WCA 2A, and the adequate coverage of the observation values yielded better results of the RFWCA 3AN models. Usually, increasi ng variability is a concomitant of an increase of extent, and because of this a model should have enough information on variation of input variables for upscaling (McBratney, 1998). The RFSPOT, MODIS ALL models showed high performance for predicti ng soil TP, TN, and TC concentrations in the pooled area, and they also showed high capability for model downscaling to the subareas (Table 5 4). The models showed high R2 values (greater than 0.75) and low RMSE values for soil TP, TN, and TC. In most cas es, the model also showed good RPD values ( Change et al. (2001). This was expected because the overall models using all data from both WCA 2A and WCA 3AN cover not only all observation data ranges but also span across both wetlands. In contrast to our results, Thompson et al. (2006) reported an overall model could not produce any significant relationships between terrain attributes and soil properties, such as soil organic C in the Pennyroyal region in Kentuck y. Vasques et al. (2012b) also reported a decline in prediction results when they applied a soil prediction model developed in the whole State of Florida to smaller extents, spatially nested within Florida, for soil TC. Model transferability and scaling O ur results show that model transferability and scaling varies depending on both the attribute space (ie., the target soil property and environmental covariates (predictor

PAGE 158

158 variables)) and geographic space domains (i.e., smaller subregions of WCA 2A and WCA 3AN, respectively; or the pooled wetland area). Overall, the RF models for soil TP showed better transferability than soil TN and TC models. This can be explained by the variable importance rankings of the RF models. The RF models developed in WCA 2A and the ones developed in WCA 3AN had different variable importance rankings for the same soil properties (i.e., TP, TN, and TC, respectively). The highly ranked predictor variables in the RF models for soil TP predictions were similar in WCA 2A and WCA 3AN. F or instance, several spectral indices derived from the SPOT image, such as TVI, NDVI green, and SR, were consistently reliable predictors of soil TP predictions in both wetlands. The spectral indices ranked high in the RFSPOT WCA2A model and the RFSPOT WCA 3AN model (compare Chapter 3). On the other hand, the highly ranked predictor variables in the RF models for soil TN and TC predictions were different in WCA 2A and WCA 3AN. For instance, PC variables, elevation, and EVI ranked high in the RFWCA 2A model s, whereas isostatic residual gravity anomaly and bouguer gravity anomaly ranked high in the RFWCA 3AN models for soil TN and TC predictions. Table 5 5 shows the descriptive statistics of environmental predictor variables in WCA 2A and WCA 3AN. Overall the least differences are seen in the spectral indices derived from the RS images, both the SPOT and MODIS images. The predictor variables based on the RS images, such as TVI, NDVI, NDVI green, and SR showed similar distributions in WCA 2A and WCA 3AN. T he greatest differences were found in elevation and geophysical properties. The elevation in WCA 2A was higher than in WCA 3AN, and this was expected since WCA 3AN experienced chronic overdrainage and reported severe soil subsidence since 1960 (Bruland et al., 2007; Osborne et al., 2011). The differences of

PAGE 159

159 geophysical properties, such as magnetic anomaly, bouguer gravity anomaly, and isostatic residual gravity anomaly, in WCA 2A and WCA 3AN are greater than any other predictor variables. Overall, the controlling factors of soil TP distributions were similar in WCA 2A and WCA 3AN, and the similarity allowed some level of model transferability to each other. However, the controlling factors of soil TN and TC distributions were different in WCA 2A and WC A 3AN, and this dissimilarity might limit the transferability of the models. Lagacherie and Volts (2000) noted that prediction abilities of models, especially based on landform attributes, are limited for upscaling because of inconsistency of the soil landscape relationships. Our results suggest that, out of numerous predictor variables, only a few control the soil prediction models, which are developed based on the relationship between soil property and environmental variables, and the importance of the v ariable may change at different extent. Thrush et al. (1997) noted that processes operating at a large scale are not always the same as those operating at a small scale, or vice versa. Vasques et al. (2012a, b) also noted that selections of major predictor variables in models are related to differences in variances and correlations between soil properties and environmental properties, and the relationships vary by spatial extents. Our results also suggest that the extents should have a certain level of simi larity in terms of variability of input variables to increase the transferability of the soil prediction models. Although WCA 2A and WCA 3AN have similar parent material and climatic factors (e.g., precipitation and temperature), human introduced anthropog enic changes occurred in the Everglades regions, such as the constructions of canals and levees, the operations of WCSs, and nutrient loadings into the system. Those anthropogenic

PAGE 160

160 changes may have localized each hydrologic unit to have different soil formi ng factors operating influencing the genesis of soil TP, TN, and TC. For instance, hydrology and nutrient influx are different in the adjacent wetlands WCA 2A and WCA 3AN. The WCA 3AN has been overdrained when compared to WCA 2A, and WCA 2A has had higher nutrient loading when compared to WCA 3AN. As a result, soil prediction models developed in one hydrologic unit (WCA 2A or WCA 3AN) may not be readily transferable to another hydrologic unit, even if they are in close proximity geographically. Therefore, t he development of readily transferable soil prediction models for the localized hydrologic units, which have been separately operated for many years, may be a challenge. Conclusions The RF models developed in WCA 2A and WCA 3AN using RS image derived spec tral indices and environmental ancillary variables were transferred among each other and upscaled to the combined area, and vice versa. The prediction abilities were greatly decreased when the models transferred to another area and upscaled, although the m odel performance of the RFWCA 2A, RFWCA 3AN, and RFALL models showed good prediction power for soil TP, TN, and TC. Only the prediction models for soil TP, which have spectral indices derived from RS images as the major controlling factor, showed some level of model transferability, but not for soil TN and TC. These findings suggest that RS images can provide a sufficient amount of variability to predict soil TP in the Everglades. It is necessary to develop universally or regionally applicable soil models. For this reason, understanding model transferability and scaling behaviors is a critical step in approaching global soil prediction models. Understanding model transferability and

PAGE 161

161 scaling has important benefits to help land resources management or decision makers to save cost and time. Findings from this study suggest that local customized soil models (developed separately in WCA 2A and WCA 3AN) and globally developed soil models ( using the pooled dataset from both wetlands) performed similarly and exceptionally well, but model transferability (from one wetland to another) was severely constrain ed. Also uncertainties of soil TP, TN, and TC were greatly increased when the models developed in one area were applied to another area and upscaled to a larger area. This suggests that the extent is less important for successful modeling of soil TP, TN, and TC as long as the domain variability of soil and environmental parameters (biotic, hydrologic, topographic, etc.) is incorporated into the model. In empirical regression models, like the ones presented in this study, the relationships between soil and environmental covariates impose constraints to transfer or scale soil models. The spatiall y explicit relationships between predictor (environmental covariates) and target variables (soil TP, TN, and TC) are interlinked with the heterogeneity in soil and environmental properties and have profound effects on the sensitivity and universality of a given soil model. Summary There is an increasing demand for global models in soil science, and understanding the model transferability is necessary in development of global soil prediction models. However, research gaps still exist on how a model developed in a specific area transfers to another area and how model predictor variables change. The objectives of this study were to (i) develop prediction models for soil total phosphorus (TP), total nitrogen (TN), and total carbon (TC) utilizing remote sensing (RS) images with different spatial resolutions, (ii) identify the environmental variables controlling the

PAGE 162

162 spatial distribution of soil TP, TN, and TC, and (iii) investigate the model transferability among two similar f reshwater ecosystems. Soil cores were collected (n=66) from the top 10 cm in Water Conservation Area (WCA) 3A North (3AN), the Florida Everglades, U.S. Random Forest (RF) models to predict soil properties were developed using topographic, hydrologic, and g eophysical properties. In addition, spectral data derived from remote sensing (RS) images including SPOT (10 m) and MODIS (250 m) were used as predictor variables. The RF soil prediction models developed in WCA 2A (Chapter 3 and 4) were applied to WCA 3AN, and the models developed in WCA 3AN were applied to WCA 2A to test model transferability. In addition, RF soil predictions models developed in subregions were upscaled to the whole study region (i.e., WCA 2A 2A combined with WCA 3AN); and WCA 3AN whole wetland region), and RF soil prediction models were downscaled from the whole wetland region to subregions (i.e., whole wetland region 2A; and whole wetland region 3AN). Results showed that model transferability and scaling varies depending on attribute domain space (soil properties and environmental covariates) and geographic domain space (i.e., the extent used to develop models). The RF models for soil TP developed in WCA 3AN showed moderate transferability with an R2 between 0.38 and 0.40, when the models were applied to WCA 2A. However, no transferability was discovered for soil TN and TC predictions. In contrast, the only RF model that showed weak transferability from WCA 2A to WCA 3AN (R2 = 0.15) was derived from SPOT predictor variables. Constraints that limit transferability of soil prediction models among the two wetlands were found to be based on differences in hydrologic management and subsequent soil genesis. The limited transferability stands

PAGE 163

163 in sharp c ontrast to the excellent soil prediction models developed for subregions using MODIS and SPOT images along with ancillary environmental predictors: (i) R2 of 0.88 and 0.94 (TC), 0.90 and 0.93 (TP), and 0.90 and 0.95 (TN) in WCA 2A; and (ii) R2 of 0.87 and 0.89 (TC), 0.91 and 0.90 (TP), and 0.87 and 0.89 (TN) in WCA 3AN. Upscaling of soil prediction models from subregion to whole region scale (i.e., WCA 2A 2 from 0.23 to 0.82; and WCA 3AN 2 from 0.34 to 0.55) degraded model performances somewhat. In contrast, downscaling of soil prediction models from the whole region to subregions (i.e., whole region 2A with R2 from 0.90 to 0.95; and whole region 3AN with R2 from 0.75 to 0.91) was successful.

PAGE 164

164 Table 5 1. Descriptive statistics of measured soil properties in Water Conservation Area2A and 3A Northa. Property b Water Conservation Area 2A Water Conservation Area 3A North Mean Min Median Max Std. dev. C.V. Mean Min Median Max Std. dev. C.V. BD (g cm 3 ) 0.11 0.05 0.10 0.24 0.04 0.33 0.15 0.04 0.13 0.52 0.07 0.50 pH 6.96 6.00 6.90 8.00 0.55 0.08 6.95 6.20 6.89 7.81 0.47 0.07 TP (mg kg 1 ) 605.6 165.4 478.8 1694.0 337.7 0.56 574.9 289.3 559.1 910.9 142.3 0.25 TN (g kg 1 ) 29.8 21.3 30.1 38.2 3.38 0.11 30.5 7.6 32.0 41.2 6.7 0.22 TC (g kg 1 ) 410.8 262.8 414.1 492.3 35.9 0.09 388.3 61.8 412.3 484.0 80.4 0.21 aC.V., coefficient of variation; Std. dev., standard deviation. bBD, bulk density; TC, total carbon; TN, total nitrogen; TP, total phosphorus.

PAGE 165

165 Table 5 2. Statistical assessments of crossvalidation in Water Conservation Area2A and 3A North an d transferability of soil prediction models developed in Water Conservation Areas. Propertya Case I b Case II b Developed in WCA 2A (RF WCA2A ) Applied to WCA 3AN Developed in WCA 3AN (RF WCA3AN ) Applied to WCA 2A R 2 RMSE RPD R 2 RMSE RPD R 2 RMSE RPD R 2 RMSE RPD RF SPOT c TP (mg kg 1 ) 0.93 102.27 3.29 0.15 147.07 0.96 0.90 56.06 2.52 0.40 303.12 1.11 TN (g kg 1 ) 0.95 1.46 2.29 0.05 6.57 1.02 0.90 2.88 2.32 0.04 3.50 0.96 TC (g kg 1 ) 0.94 16.08 2.22 0.02 83.92 0.95 0.89 34.57 2.31 0.00 46.52 0.77 RF MODIS c TP (mg kg 1 ) 0.90 115.90 2.90 0.00 205.79 0.69 0.91 59.56 2.37 0.38 308.13 1.09 TN (g kg 1 ) 0.92 1.50 2.24 0.02 6.63 1.00 0.87 2.87 2.34 0.01 3.54 0.95 TC (g kg 1 ) 0.88 16.86 2.12 0.00 81.48 0.98 0.87 34.04 2.34 0.00 59.50 0.60 aTC, total carbon; TN, total nitrogen; TP, total phosphorus. bR2, coefficient of determination; RMSE, root mean squared error; RPD, ratio of performance to deviation. cRFSPOT,MODIS, Random Forest models using the SPOT and the MODIS spectral data derived input variabl es, respectively.

PAGE 166

166 Table 5 3. Statistical assessments of upscaling of soil prediction models developed in Water Conservation Area2A and 3A North to the combined area. Propertya Developed in WCA 2A b Developed in WCA 3AN b RF SPOT WCA2A RF MODIS WCA2A RF SPOT WCA3AN RF MODIS WCA3AN R 2 RMSE RPD R 2 RMSE RPD R 2 RMSE RPD R 2 RMSE RPD Applied to the combined area TP (mg kg 1 ) 0.82 121.55 2.30 0.73 156.65 1.79 0.36 241.90 1.16 0.34 246.21 1.14 TN (g kg 1 ) 0.34 3.90 1.18 0.32 3.90 1.18 0.55 3.21 1.44 0.54 3.20 1.44 TC (g kg 1 ) 0.24 46.17 1.12 0.23 45.68 1.13 0.46 41.28 1.25 0.39 50.20 1.03 aTC, total carbon; TN, total nitrogen; TP, total phosphorus. bR2, coefficient of determination; RMSE, root mean squared error; RPD, ratio of performance to deviation; RFSPOT,MODIS, Random Forest models using the SPOT and the MODIS spectral data derived input variables, respectively.

PAGE 167

167 Table 5 4. Statistical assessments of downscaling of soil prediction models developed in the combined area to Water Conservation Area2A and 3A North. Propertya Developed in the combined area Applied to WCA 2A Applied to WCA 3AN R 2 RMSE RPD R 2 RMSE RPD R 2 RMSE RPD RF SPOT c TP (mg kg 1 ) 0.93 102.27 3.29 0.93 102.37 3.28 0.87 57.93 2.44 TN (g kg 1 ) 0.95 1.46 2.29 0.95 1.45 2.32 0.81 3.59 1.86 TC (g kg 1 ) 0.94 16.08 2.22 0.93 16.07 2.22 0.78 45.38 1.76 RF MODIS c TP (mg kg 1 ) 0.90 115.90 2.90 0.90 116.19 2.89 0.91 58.67 2.41 TN (g kg 1 ) 0.92 1.50 2.24 0.92 1.49 2.26 0.75 3.79 1.76 TC (g kg 1 ) 0.88 16.86 2.12 0.92 15.94 2.24 0.75 45.23 1.76 aTC, total carbon; TN, total nitrogen; TP, total phosphorus; gray filled cells show the original model result and white cells show the transferred model result. bR2, coefficient of determination; RMSE, root mean squared error; RPD, ratio of performance to deviation. cRFSPOT,MODIS, Random Forest models using the SPOT and the MODIS spectral data derived input variables, respectively.

PAGE 168

168 Table 5 5. Descriptive statistics of environmental predictor variables in Water Conservation Area2A and 3A Northa. Property Water Conservation Area 2A Water Conservation Area 3A North Mean Min Median Max Std. dev. C.V. Mean Min Median Max Std. dev. C.V. Spectral indices b SPOT Green 79 50 77 127 11.8 0.15 79 48 80 107 10.2 0.13 Red 80 63 77 122 9.4 0.12 80 67 78 95 6.3 0.08 NIR 92 54 91 129 16.1 0.17 91 68 91 112 8.3 0.09 SWIR 90 30 95 136 23.9 0.26 106 43 109 151 24.7 0.23 PAN 83 63 81 136 12.8 0.15 84 62 82 112 9.2 0.11 NDVI 0.07 0.08 0.06 0.25 0.09 0.07 0.04 0.07 0.17 0.05 NDVI green 0.07 0.07 0.08 0.25 0.08 0.07 0.05 0.07 0.26 0.06 NDWI 0.02 0.17 0.01 0.54 0.13 0.06 0.24 0.07 0.36 0.12 RSR 0.77 0.56 0.76 1.02 0.10 0.13 0.66 0.47 0.69 0.96 0.10 0.15 SR 1.16 0.84 1.12 1.68 0.21 0.18 1.15 0.93 1.14 1.42 0.11 0.10 TVI 74.96 64.46 74.64 86.74 5.80 0.08 75.22 67.95 75.25 82.13 3.31 0.04 MSI 0.98 0.30 0.99 1.40 0.22 0.22 1.16 0.47 1.16 1.64 0.26 0.22 MODIS Blue 278 167 264 610 81.03 0.29 304 168 280 556 86.57 0.28 Red 548 362 507 1176 150.07 0.27 611 398 557 1157 172.39 0.28 NIR 1661 1093 1563 2670 419.83 0.25 1652 1197 1625 2235 247.08 0.15 MIR 701 401 684 1116 165.77 0.24 811 319 797 1260 220.16 0.27 NDVI 0.50 0.24 0.52 0.69 0.09 0.47 0.28 0.47 0.57 0.06 EVI 2.14 1.09 1.97 3.85 0.65 1.76 0.85 1.82 2.40 0.30 SR 3.11 1.64 3.11 5.35 0.74 0.24 2.80 1.76 2.77 3.65 0.42 0.15 TVI 99.80 56.19 100.68 108.87 4.58 0.05 98.27 88.04 98.45 103.46 3.26 0.03 MSI 0.43 0.26 0.42 0.66 0.07 0.17 0.49 0.18 0.53 0.70 0.12 0.23

PAGE 169

169 Table 55. Continued. Property Water Conservation Area 2A Water Conservation Area 3A North Mean Min Median Max Std. dev. C.V. Mean Min Median Max Std. dev. C.V. Topography Elevation 3.03 2.41 3.05 3.55 0.26 0.08 2.57 1.86 2.59 2.94 0.24 0.09 Hydrology c Dist WCSs 7,419 885 7,007 17,385 3,850 0.52 19,229 4,437 19,740 40,145 8,350 0.43 Geophysical properties Potassium 0.02 0.01 0.02 0.04 0.01 0.35 0.02 0.01 0.02 0.05 0.01 0.39 Magnetic 12.20 82.66 8.36 101.06 42.53 106.06 15.39 114.79 237.94 80.16 Bouguer gravity 29.25 24.55 29.46 31.05 1.42 0.05 19.39 13.65 19.14 24.03 2.31 0.12 Isostatic gravity 15.01 11.17 14.84 17.36 1.68 0.11 7.31 2.41 7.27 10.72 1.91 0.26 aC.V., coefficient of variation; Std. dev., standard deviation. bEVI, enhanced vegetation index; MIR, midinfrared; MODIS, moderate resolution imaging spectroradiometer; MSI, moisture stress index; NDVI, normalized difference vegetation index; NDVI green, normalized different vegetation green index; NDWI, normalized difference water index; NIR, near infrared; PAN, panchromatic; RSR, reduced simple ratio; SPOT, satellite pour lobservation de la terre; SR, simple ratio; SWIR, shortwaveinfrared; TVI, transformed vegetation index. cDist WCSs, distance to water control structures.

PAGE 170

170 Figure 5 1. Location of the Water Conservation Area2A and 3A N orth within the Everglades, Florida, U.S.

PAGE 171

171 Figure 5 2. Spatial distribution of measured soil concentrations of total phosphorus, TP (mg kg1), total nitrogen, TN (g kg1), and total carbon, TC (g kg1) within the Water Conservation Area3A N orth.

PAGE 172

172 Figure 5 3. Total phosphorus (TP) loads (kg) of outflow fr om Water Conservation Area2A (gray filled bar) and TP loads from Water Conservation Area2A into Water Conservation Area3A north (empty bar) during the period from May, 2003 to April, 2010 (Data obtained from the South Florida Water Management District).

PAGE 173

173 Figure 5 4. Prediction maps of soil properties using Random Forest model with the SPOT (left) and the MODIS (right) image derived spectral input variables combined with environmental predictor variables: a) total phosphorus, TP (mg kg1), b) total nitrogen, TN (g kg1), and c) total carbon, TC (g kg1) in Water Conservation Area3A North.

PAGE 174

174 Figure 5 5. Variable importance of predictor variables in Random Forest models with the SPOT (above) and the MODIS (bottom) images to predict: a) t otal phosphorus, TP (mg kg1), b) total nitrogen, TN (g kg1), and c) total carbon, TC (g kg1) in Water Conservation Area3A N orth (X axis: relative importance, Y axis: predictor variable; see Table 3 2 in Chapter 3 for environmental input variables abbreviations).

PAGE 175

175 Figure 5 6. Comparison of prediction maps of soil total phosphorus, TP (mg kg1) using Random Forest models with the SPOT (above) and MODIS (below) spectral data in Water Conservation Area3A North (WCA 3AN): a) Random Forest models developed in WCA 3AN (same as Figure 5 4 a) and b) Random Forest models developed in WCA 2A and applied to WCA 3AN.

PAGE 176

176 Figure 5 7. Comparison of prediction maps of soil total phosphorus, TP (mg kg1) using Random Forest models with the SPOT (above) and MODIS (below) spectral data in Water Conservation Area (WCA) 2A: a) Random Forest models developed in WC A 2A and b) Random Forest models developed in WCA 3AN.

PAGE 177

177 Figure 58. Scatter plots showing the relationship between observed and predicted soil total phosphorus, TP (mg kg1) using Random Forest models with the SPOT image derived input variables combined with environmental predictor variables: a) Random Forest model developed in Water Conservation Aare (WCA) 2A and applied to WCA 3AN, b) Random Forest model developed in WCA 3AN and applied to WCA 2A, c) Rando m Forest model developed in WCA 2A, and d) Random Forest model developed in WCA 3AN. a) b ) c ) d )

PAGE 178

178 CHAPTER 6 VISIBLE NEAR INFRARED SPECTROSCOPY FOR SOIL PROP ERTY PREDICTIONS OF HIGHLY ORGANIC WE TLAND SOILS Overview There is an increase in demand for large amounts of i nexpensive soil data collection in order to develop soil property prediction models T he development of chemometrics and various multivariate methods (e.g., tree and regression analysis) have advanced the use of diffuse reflectance spectroscopy (DRS) in soil science (Brown et al., 2006; Mevik and Wehrens, 2007). Diffuse reflectance spectroscopy (e.g., visible, near infrared, and mid infrared spectroscopy ) could provide less expensive, rapid, and nondestructive data collection when compared to conventional l aboratory methods. Furthermore, DRS has the potential for simultaneous characterization of various soil properties (McBratney et al., 2006; Viscarra Rossel et al., 2006; Vasques et al., 2010). Visible and near infrared wavelength regions, which range from 350 to 2500 nm, contain useful information such as overtones (i.e., half, one third, or one fourth of the wavelength of the fundamental features) and combinations that have predictive ability for various soil properties, although the fundamental features o f most soil constituents can be found in the mid to thermal infrared regions (McCarty et al., 2002; Brown et al., 2006). Visiblenear infrared reflectance spectroscopy (VNIRS) has been applied in predictions of physical, chemical, and biological soil properties (Change et al., 2001; Change and Laird, 2002; McCarty et al., 2002; Shepherd and Walsh, 2002; Cozzolino and Morn, 2003; Cohen et al., 2005; Brown et al., 2006; Stevens et al., 2006; Vasques et al., 2009, 2010), and the results of the studies have shown that it is possible for the VNIRS to enhance the effectiveness of soil data collection and to improve model

PAGE 179

179 accuracy. However, the diversity of soil samples analyzed in most studies was limited to upland ecosystems, such as agricultural areas and pastures. For instance, McCarty et al. (2002) applied the DRS approach to develop soil prediction models for soil carbon (C) using soil samples collected from 9 states and 14 geographically diverse locations in the central United States. In their study, the sam ples covered various soil temperature and moisture regions, but did not include permanently inundated, highly organic wetland soils. Similarly, Brown et al. (2006) developed a global soil characterization model with VNIRS using 4,184 soil samples from the U.S. and 36 different countries in Africa, Asia, South America, and Europe, but only five samples obtained from O horizon (i.e., dominated by organic materials). However, inclusion of highly organic soils from wetland areas is critical in terms of the development of soil spectral libraries to enhance the capabilities of spectroscopy methods for soil properties. Despite the advantages and numerous applications of VNIRS, only a few studies were conducted in wetland systems. Bouchard et al. (2003) measured the decomposition rate of salt marsh litters using near infrared reflectance spectroscopy in France; and Cohen et al. (2005) developed soil prediction models for various soil properties using VNIRS for riparian wetland soils in Pensacola, Florida. Typical ch aracteristics of wetland soils, such as great amounts of organic matter and the coexistence of oxidized and reduced conditions, which affect soil color (i.e., hue, chroma, and value) by changing the valence state of iron and manganese, may hamper the appli cation of VNIRS in wetland soils (Cohen et al., 2005). The soils in wetland areas are characterized by a high diversity of biological and nonbiological features. The Everglades in Florida, U.S., which have been impacted by

PAGE 180

180 nutrient influx and modifi ed hydrologic regimes ( Noe and Childers, 2007), have shown a much wider range of soil biogeochemical properties when compared to upland soils ( DeBusk et al., 1994; Grunwald et al., 2004) because of heterogeneous compositions of water, soil, detritus, macrophytes, and periphyton. For instance, anthropogenic modifications of the wetland hydrology and nutrient loading have created a phosphorus (P) enriched area, and as a result, the Everglades are composed of highly nutrient impacted soils. Although Cohen et al. (2005) successfully applied the VNIRS to measure various properties in riparian wetland soils, further work on highly organic peat soils is necessary to improve the capability of the VNIRS to develop soil prediction models. This could help to enhance our understanding of spectral behaviors of the wetland soils. Therefore, the objective of this study was to estimate various soil properties including soil pH, total P (TP), total C (TC), total nitrogen (TN), MehlichI calcium (Ca), and MehlichI magnesium ( Mg) of highly organic wetland soils in the Everglades using VNIRS. Methods and Ma terials Study A rea The study was conducted in Water Conservation Area (WCA) 2A, in the Florida Everglades U.S. The majority of the soils in WCA 2A are Histosols consisting of highly decomposed organic material and most of the soils are inundated year round. The soils encompass the Everglades and Loxahatchee peat formations in the region. The Everglades peat is composed of the remains of sawgrass ( Cladium jamaice nse ), and the soils are generally brown to black with minimal mineral content, while the Loxahatchee peat are composed of remains of the roots and rhizomes of water lily

PAGE 181

181 ( Nymphaea odorata), and the soils are lighter colored (Lodge, 2005; Bruland et al., 2007). A complete description of WCA 2A was presented in Chapter 2. Field S ampling and L aboratory M easurement Soil samples were collected in July and October 2009 and mid March 2010 using airboats. Field sampling consisted of 108 sites spread over the stud y area. At each site, soil cores were collected using a 10 cm diameter stainless steel coring tube at two depths: 0 10 cm and 10 20 cm, totaling 216 samples. Details of field sampling were described in Chapters 2 and 3. All samples were ovendried at 70 C for 72 hours for laboratory analyses, except for soil pH measurement. Soil pH was determined by a pH meter on 10 g of wet soil after equilibrating it with 10 ml of deionized water, i.e., 1:1 (soil:liquid) ratio, based on the Soil Survey Laboratory Met hods (Soil Survey Staff, 2004). The soil samples were analyzed for TP utilizing the association of official analytical chemists (AOAC) 978.01 method, for TN utilizing the LECO combustion method (LECO TruSpec, MI, USA), and for TC utilizing combustion with a Shimadzu SSM 5000A TC analyzer (Shimadzu Corporation, Columbia, MD, USA). Mehlich I extractable calcium (Ca) and magnesium (Mg) were analyzed using the Method 200.7 (United States Environmental Protection Agency, 2001). Spectral Scanning and Data Proce ssing A total of 211 soil samples were prepared for spectral scanning. Samples were sieved through a 2mm mesh, and then ovendried for 12 h at 45 C to avoid any effects of moisture before scanning Each soil sample was scanned in the VNIR region, i.e., w avelength range of 350 to 2500 nm using a QualitySpec Pro spectroradiometer (Analytical Spectral Devices, Inc., Boulder, CO) The spectroradiometer collects the

PAGE 182

182 reflected spectral signature in 1nm intervals, based on the average of 100 internal readings per wavelength using three internal spectrometers, which are wavelength ranges of 350 to 1000 nm 1000 to 1800 nm, and 1800 to 2500 nm. Each soil sample was measured four times, with replications rotated 90. White Spectralon ( LabSphere, North Sutton, NH ) was used as a reference, and it was scanned prior to the first scan and at every 10 samples ( 4 0 scans). A n average spectral curve was calculated for each sample based on the four replicate scans (Figure 6 1) The averaged reflectance spectra were composed of 2151 bands. The averaged reflectance spectra were further processed to develop soil property prediction models. Two preprocessing transformation methods were employed for model development and prediction c apability comparison: (i) a Savitzky Golay 3rdorder polynomial transformation averaging 4 bands for each of the left and right sides, named hereafter SG, and (ii) a first derivative transformation with 2ndorder Savitzky Golay smoothing, named hereafter SG 1st D (Savitzky and Golay, 1964). Both transformed spectra were then averaged across a 10nm window to reduce dimensionality by a factor of 10. This resulted in the reduction of the spectra to 215 reflectance values for SG transformation and 213 refl ectance values for SG 1st D transformation. Previous studies showed that there was no significant loss of information because of reducing dimensionality (Shepherd and Walsh, 2002). Although Cohen et al. (2005) argued that regions of 350 to 400 nm, 970 to 1010 nm, and 1780 to 1920 nm needed to be omitted for VNIR analysis due to a low signal to noise ratio caused by splicing between internal spectrometers, these wavelength regions also contain the major absorbance areas associated with soil organic constitue nts. For

PAGE 183

183 instance, the region around 1900 nm is one of absorbance regions of OH groups (Vasques et al., 2009). To avoid any loss of spectral bands which have potential power in predicting soil properties, the regions were not omitted for this study. All pr eprocessing w as performed using the Unscrambler v9.2 (CAMO Software Inc., Woodbridge, NJ ) Model Developments and Assessments Partial least squares regression (PLSR) was applied to develop prediction models for soil pH, TP, TN, TC, and MehlichI extractable Ca and Mg. Partial least squares regression has shown widespread success for spectral data analysis in predicting soil properties (McCarty et al., 2002; Reeves III et al., 2006; Stevens et al., 2006; Vasques et al., 2008, 2009; Zornoza et al., 2008). The power of PLSR is that it can deal with large numbers of predictor variables that are highly collinear, and reduce the data with minimal loss of original information (Helland, 1990; Mevik and Wehrens, 2007). Partial least squares regression, i.e., a reg ression analysis, assumes normal distribution of a target variable. Hence, normality tests were performed for each soil property to fulfill the assumption. Soil pH, TN, and MehlichI Mg showed approximate normal distributions. Total P, TC, and MehlichI Ca showed skewed distributions, thus, they were transformed using the natural logarithm, and prediction models were developed with natural logarithm transformed data, i.e., lnTP, lnTC, and lnCa. The whole dataset (n = 211) was split randomly into 147 samples (70%) for calibration and 64 samples (30%) for validation for each soil property. The splitting was done separately for each soil variable, and hence, the validation sets of each soil property were not identical. The sample distributions and means between the calibration and

PAGE 184

184 validation sets were compared using ShapiroWilk test and S tudent s t test respectively, to confirm that the validation set resembled the calibration set of a given soil property The PLSR models were developed using a calibration set with leaveoneout cross validation and evaluated using a validation set. The optimum number of partial least squares (PLS) factors, which are linear combinations of the predictors that explain both response and predictor variation (Stevens et al., 2006), were chosen based on a root mean squared error (RMSE) and coefficient of determination (R2) of calibration set. The standard error of prediction (SEP) and the ratio of performance to deviation (RPD) were also compared to select the best performance models for each soil property. The SEP, which is the most used indicator in spectroscopy studies (BellonMaurel and McBratney, 2011), is an average prediction error on the validation set (Eq. ( 6 1 ) ), and the RPD is one of the indices that standardize the SEP usi ng the standard deviation (SD) of the validation set (Eq. ( 6 2 ) ). 1 ) (1 2 n y y SEPn i i i (6 1) SEP SD RPD (6 2) where i is the predicted value for sample i yi is the observed value for sample i and n is the number of samples in the validation set. Chang et al. (2001) suggested three categories for the model reliability based on the RPD values: (i) category A: excellent models, with RPD > 2, (ii) category B: fair models, with 1.4 RPD 2, and (iii) category C: nonreliable models, with RPD < 1.4. Lastly, the ratio to performance to inter quartile distance (RPIQ) suggested by BellonMaurel et al. (2010) was also used. The RPIQ is the same as RPD, except SD is replaced by inter quartile distance, i.e.,

PAGE 185

185 difference between 75% quartile and 25% quartile of a data range. All model developments were performed using the Unscrambler v9.2 (CAMO Software Inc., Woodbridge, NJ ) and statistical analyses were performed using the R statistical language ( R Development Core Team, 2012). Results and Discussion Descriptive S tatistics Descriptive statistics of soil properties in WCA 2A are shown in Table 6 1. Both original measurements and natural logarithm transformed data are shown for soil TP, TC, and MehlichI Ca. Soil pH varied from 5.6 to 8.0, with a mean and median of 6.8 for the whole dataset, which was similar to pH values reported by DeBusk et al. (1994). The ubiquitous calcium carbonates (CaCO3) released from limestone bedrocks in the region make the soils less acidic than other wetland soils. For instance, riparian wetland soils in Pensacola Bay in Florida showed a soil pH range from 3.3 to 7.2 (Cohen et al., 2005), and peat soils in northern Minnesota showed a soil pH of less than 4.0 (Updegraff et al., 1995 ). Gleason et al. (1974) reported that the Everglades peat and Loxahatchee peat, which are the main formations in WCA 2A, tend to have near neutral pH and slightly lower pH, respectively. MehlichI extractable Ca showed a wide range from 3,580 to 28,500 mg kg1, with a mean of 10,907 mg kg1 and median of 9,850 mg kg1 for the whole dataset. MehlichI extractable Mg varied from 279.6 to 4,030 mg kg 1, with a mean of 2,044 mg kg1 and median of 1,906 mg kg1. No studies were reported previously for Mehlich I extractable Ca and Mg in WCA 2A, thus no comparisons were conducted. Complete descriptions of soil TP, TN, and TC were presented in Chapters 3 and 4. In most cases, except for MehlichI Mg, the range of soil properties in the validation set was smaller th an the range of soil properties in the calibration set.

PAGE 186

186 Visible Near Infrared Reflectance Spectroscopy Models for Soil Properties in Water Conservation Area2A Summary statistics of the PLSR models for soil properties are shown in Table 6 2. The PLSR models for lnTP and MehlichI Mg performed better with SG 1st D preprocessing transformation, and other soil properties performed slightly better using the SG transformation method. The better treatment between SG and SG 1st D transformations was consi dered to be the one with low SEP in the calibration mode, as suggested by Stevens et al. (2006). The differences in the model prediction results between SG and SG 1st D transformations were less than 5%. The number of PLS factors varied by soil propert ies, and the numbers indicate reduced predictor variables to explain the variability of the soil properties. For instance, the number of PLS factors for soil TN was 16, and it means that the PLSR model reduced the number of predictor variables from 215 ref lectance bands (i.e., preprocessed with SG transformation) to 16 factors, which explained about 80% of the variability of the calibration set for soil TN. Similarly, 213 reflectance bands (i.e., preprocessed with SG 1st D transformation) were reduced t o 9 PLS factors for soil lnTP, and the 9 PLS factors explained about 70% of the variability of the calibration and validation sets for lnTP. Among the soil properties, the prediction models for soil pH showed the best performance with an R2 of 0.78 for cal ibration and validation using the SG transformation. Previous studies have shown good prediction power for soil pH (Chang et al., 2001; Shepherd and Walsh, 2002, 2004; Cohen et al., 2005) although pH per se cannot be sensed directly from soil samples using VNIR Soil TN showed an R2 of 0.80 for calibration, which was the highest calibration accuracy, and slightly decreased to 0.73 for validation. Natural logarithm transformed TP showed an R2 of 0.70 for

PAGE 187

187 validation using the SG 1st D transformation. Overall the PLSR models showed similar prediction power for the calibration and validation modes although there were slight decreases of the R2 values and slight increases of the RMSE values in the validation set (Table 6 2). The RPD values of the validation mode for all soil properties, except soil pH and TN, were smaller than 1.4, which can be grouped as Category C according to Change et al. (2001), indicating a nonreliable model. Soil pH and TN showed RPD values greater than 1.4 in the validation set, whic h indicate fair models. However, as suggested by Reeves III and Smith (2009) and BellonMaurel et al. (2010), the use of fixed RPD thresholds as an only indicator could be misleading in the interpretation of the predictability of the models. In fact, Reeves III and Smith (2009) noted that many studies found calibrations were useful with considerably lower RPD values than the fixed thresholds. Moreover, ViscarraRossel and McBratney (2008) classified a good model as having an R2 from 0.61 to 0.8, and hence, the PLSR models developed in WCA 2A for soil lnTP and lnTC showed some level of predictability based on the classification of Viscarra Rossel and McBratney (2008) However, it is important to note that all these classification thresholds are somewhat arbit rary. The PLSR model for soil pH showed the highest RPIQ. The RPIQ, which us es inter quartile distance instead of SD, could be a better indicator to describe the spread of soil property population (Bellon Maurel et al., 2010) G enerally, high RPIQ values indicate better models than models with low values. The scatter plots showing the relationship between observed (y axis) and predicted values (x axis) of the PLSR models for the validation set are shown in Figure 6 2. The model f or lnCa greatly underestimated most of the high observation

PAGE 188

188 values as shown in Figure 6 2e. Similarly, great under and overestimations were observed for MehlichI Mg when compared to other soil properties. Overall, the prediction accuracies of the PLSR models for soil properties in WCA 2A were not as good when compared to other studies. For instance, comprehensive studies were conducted to predict soil C using VNIRS, and most of the studies showed high prediction results with RPD values greater than 2.0 (C hange et al., 2001; Change and Laird, 2002; Stevens et al., 2006; Vasques et al., 2008, 2009, 2010; Reeves III et al., 2011). However, it is important to note that all of the studies were conducted with upland soils, which have less organic matters, more m inerals, and more contrasting soil colors than wetland soils. Soil color is important in the visible spectral range that allows to distinguish light colored from dark brown/black soils. Since the soil color in wetland soils is relatively homogenous it does not provide good predictive power. Although Cohen et al. (2005) showed high prediction results with an RPD value of 5.74 for TC in riparian wetland soils, comparisons between Cohen et al.s 2005 study and our study are quite challenging because there are clear differences in the characteristics of different wetland soil s. For instance, soil samples in Cohen et al.s 2005 study showed a mean organic matter content of 29% and a mean value of soil TC of 143.6 g kg1. In contrast, the mean of soil TC in WCA 2A was much higher, as shown in Table 6 1. Also, a previous study reported that WCA 2A soils contain more than 87% of organic matter (Rivero et al., 2007). High organic matter content and TC concentrations indirectly indicate that the soil colors of WCA 2A a re very dark (e.g., dark brown to black). Pronounced dark soils and high organic matter content might limit the predictability of VNIRS. Change et al. (2001) reported that organic matter content had a large effect on

PAGE 189

189 the accuracy of spectroscopy and it aff ected model prediction results for many soil properties. For these reasons (i.e., high organic matter and dark soil), relatively low prediction results of other soil properties could be explained. The regression coefficients between soil properties and spectral reflectance used in the PLSR models are shown in Figure 6 3. The important wavelengths varied by soil properties. For instance, the important wavelengths were around 400, 1900, 2000, and 2200 nm for soil pH, and around 900, 1900, 2000, and after 2200 nm for lnTP. Around 1900 and 2000 nm were shown as important predictors for the PLSR models of lnTP, TN, and lnTC. Our results show that VNIRS can be used for the predictions of various properties of highly organic wetland soils. Also, the prediction results showed that the models developed with VNIRS for WCA 2A soils could possibility be improved by using different transformation methods or different multivariate methods. Vasques et al. (2008) comprehensively tested and compared prediction results amo ng various preprocessing transformations and various multivariate methods. They found that the methods, not only for preprocessing but also for multivariate techniques, affect prediction results, although PLSR performed better than other methods. Brown et al. (2006) reported notably improved validation results using the boosted regression tree method rather than PLSR. Though the models showed some level of prediction ability of VNIRS, our results also showed some limitations of VNIRS for wetland soils. Furt her investigations of other spectral wavelength regions would be needed to overcome the limitations and to increase the capability of DRS for wetland soils. For instance, the use of the mid infrared (MIR) wavelength region could possibly improve prediction power for

PAGE 190

190 wetland soils. McCarty et al. (2002) reported that the MIR region, which ranges from 2500 to about 25000 nm, performed significantly better than visible and near infrared regions when developing prediction models for soil C. ViscarraRossel et al. (2006) also found improved accuracies with the spectral signature of the MIR region for various soil properties when compared to visible and near infrared regions. Conclusions This study demonstrated the use of VNIRS to assess soil properties in wetland ecosystems. The soil prediction models developed in WCA 2A showed moderate prediction ability for soil pH, lnTP, TN, and lnTC, but low prediction power for MehlichCa and Mg. The development of soil spectral reflectance libraries is essential for the application of VNIRS to estimate soil properties, and the spectral reflectance data of WCA 2A soils could be the most useful information for wetland systems because the data ranges of WCA 2A soils exceeded typical ranges of soil properties in Florida. Further the collected data in this study could be combined into the Florida soil spectral library of previous studies, providing a much broader coverage of soil properties. This will enhance the applicability of VNIRS to predict various soil properties. Our findi ngs show the potential use of VNIRS as a cost effective and rapid monitoring method of wetland soils, although further research is required to verify the models through applications in other wetland areas. It is particularly beneficial in the Everglades region, which is highly heterogeneous and has been affected by various anthropogenic changes.

PAGE 191

191 Table 61 Descriptive statistics of measured soil properties in Water Conservation Area2Aa. Property b Set c Mean Min. Median Max. SD SE C.V. Skewness pH Cal. 6.8 5.6 6.8 8.0 0.54 0.04 0.08 0.22 Val. 6.9 6.0 6.9 8.0 0.51 0.06 0.07 0.34 TP (mg kg 1 ) Cal. 525.6 27.6 384 2290 373.3 30.8 0.71 1.70 Val. 434 77.6 386 1660 293.4 36.7 0.67 1.78 ln TP d (ln mg kg 1 ) Cal. 6.0 3.3 6.0 7.7 0.68 0.06 0.11 0.32 Val. 5.9 4.3 6.0 7.4 0.65 0.08 0.11 0.15 TN (g kg 1 ) Cal. 29.7 15.2 29.9 39.5 3.8 0.31 0.13 0.40 Val. 30.3 23.9 30.1 37.5 3.3 0.41 0.11 0.32 TC (g kg 1 ) Cal. 421.6 262.8 430.2 492.3 38.4 3.17 0.09 1.31 Val. 422.9 342.6 428.4 480.6 31.3 3.91 0.07 1.35 ln TC d (ln g kg 1 ) Cal. 6.0 5.6 6.1 6.2 0.10 0.01 0.02 0.74 Val. 6.0 5.8 6.1 6.2 0.08 0.01 0.01 0.53 Mehlich I Ca (mg kg 1 ) Cal. 10850 3580 9820 28500 3763 310.4 0.35 2.07 Val. 11030 5400 9960 25600 3969 496.1 0.36 2.14 Mehlich I ln Ca d (ln mg kg 1 ) Cal. 9.2 8.2 9.2 10.3 0.29 0.02 0.03 0.87 Val. 9.3 8.6 9.2 10.2 0.29 0.04 0.03 1.32 Mehlich I Mg (mg kg 1 ) Cal. 2039 596.8 1910 4030 648.7 53.5 0.32 0.64 Val. 2056 279.6 1890 4020 714.8 59.4 0.35 0.32 aSD standard deviation; SE, standard error; C.V., coefficient of variation bCa, calcium; Mg, magnesium; TC, total carbon; TN, total nitrogen; TP, total phosphorus cCal., calibration set (n=147); Val., validation set (n=64) dNatural logarithm transformed soil properties.

PAGE 192

192 Table 62. Summary statistics of the partial least squares regression models for different soil propert ies in Water Conservation Area2Aa. Property b Preprocessing transformationc Number of PLS factors Calibration (n=147) Validation (n=64) R 2 RMSE R 2 RMSE SEP RPD RPIQ pH SG 11 0.78 0.34 0.78 0.34 0.34 1.52 2.23 ln TP (ln mg kg 1 ) SG 1 st D 9 0.69 0.49 0.70 0.47 0.47 1.39 1.61 TN (g kg 1 ) SG 16 0.80 2.29 0.73 2.24 2.25 1.47 1.87 ln TC (ln g kg 1 ) SG 11 0.75 0.07 0.60 0.06 0.06 1.25 1.51 Mehlich I ln Ca (mg kg 1 ) SG 3 0.37 0.27 0.35 0.27 0.27 1.06 0.94 Mehlich I Mg (mg kg 1 ) SG 1 st D 5 0.51 555.2 0.50 617.1 620.0 1.15 1.62 a PLS, partial least squares; R2, coefficient of determination; RMSE, root mean squared error; RPD, ratio of performance to deviation; RPIQ, ratio to performance to inter quartile distance ; SEP, standard error of prediction. b C a, calcium; Mg, magnesium; TC, total carbon; TN, total nitrogen; TP, total phosphorus. c SG, savitzky golay 3rdorder polynomial transformation averaging 4 bands for each of the left and right sides; SG 1st D, first derivative transformation with 2ndorder Savitzky Golay smoothing.

PAGE 193

193 Figure 61. Raw visible near infrared reflectance spectra of soil samples from Water Conservation Area2A (n = 211).

PAGE 194

194 Figure 62. Scatter plots showing the relationship between observed and predicted soil properties in validation set using partial least squares regression models: a) pH, using Savitzky Golay smoothing, b) natural log transformed total phosphorus, lnTP (ln mg kg1), using first derivative transformation with second order Savitzky Golay method, c) total nitrogen, TN (g kg1), using Savitzky Golay smoothing, d) natural log transformed total carbon, lnTC (ln g kg1), using Savitzky Golay smoothing, e) natural log transformed M ehlich I extractable calcium, ln MehlichI Ca (ln mg kg1), using Savitzky Golay smoothing, and f) M ehlich I extractable magnesium, MehlichI Mg, using first derivative transformation with second order Savitzky Golay method. a) b ) c ) d ) e ) f )

PAGE 195

195 Figure 63. Regression coefficients between soil properties and spectral reflectance used in the partial least squares regression models: a) pH, using Savitzky Golay smoothing, b) natural log transformed total phosphorus, lnTP (ln mg kg1), using first derivative transformation with second order Savitzky Golay method, c) total nitrogen, TN (g kg1), using Savitzky Golay smoothing, d) natural log transformed total carbon, lnTC (ln g kg1), using Savitzky Golay smoothing, e) natural log transformed M ehlich I extractable calcium, ln Mehlich I Ca (ln mg kg1), using Savitzky Golay smoothing, and f) M ehlich I extractable magnesium, MehlichI Mg, using first derivative transformation with second order Savitzky Golay method. a) pH b ) lnTP c ) TN

PAGE 196

196 Figure 63. Continued. d ) lnTC e ) Mehlich I lnCa f ) Mehlich I Mg

PAGE 197

197 CHAPTER 7 SUMMARY AND SYNTHESI S Wetland ecosystems are prominent in Florida because of geologic, topographic, and climatic conditions, and they cover approximately 25% of the State. Due to the unique values of wetlands encompassing ecosystem services, such as sequestration of soil carbon provisioning of biodiversity, cycling of nutrients, serving as wildlife habitat, providing freshwater resources, and more, their sustainability is profoundly important. To characterize spatial patterns of wetland soils in dependence of soil forming factors is critical for the functioning and preservation of them. For this reason, there are increasing demands for accurate and precise data to predict biogeochemical properties in wetland soils. Remote sensing (RS) data, which provide dense spectral grids ov er a large area, have been used widely for soil predictions, because of their cost effectiveness, high spatial resolutions capturing the variability of landscape features, and capabilities to deduct various potential soil forming properties. Remote sensing informed soil prediction models have shown success to improve the predictive power and spatial resolution of predictions in upland systems. This study was motivated by the fact that few RS supported digital soil mapping (DSM) studies have been conducted i n wetland areas, and the potential of RS informed soil modeling has not been realized yet. This is specifically important in wetlands where field sampling is challenging and siteaccess is limited. This study provided evidence of the usefulness of RS imag es to estimate various soil properties in Water Conservation Area (WCA) 2A and 3A North (3AN) Everglades, Florida, USA The RS images included Moderate Resolution Imaging Spectroradiometer

PAGE 198

198 (MODIS), Landsat Enhanced Thematic Mapper Plus (ETM+), and Satelli te Pour lObservation de la Terre (SPOT) images. In Chatper 2 RS supported modeling produced the first soil taxonomic map exist in WCA 2A C lassification trees were used to predict soil series using spectral data and ancillary environmental datasets. The tree models predicting soil series, which were developed using spectral data and indices derived from RS images could explain > 60% of the overall variability However, the variable importance of sensor derived properties to predict soil series ranked low suggesting that soil classification was less dependent on biotic/vegetation properties. In contrast, lithologic, topographic, and geographic properties ranked high indicating that they played a major role to discriminate soil series. Results suggest that the variability of soil series can be explained by bedrock/parent material > topographic variables > vegetation properties derived from remote sensing Spectral datasets derived from three remote sensing images with varying spatial resolutions (10, 30, an d 250 m) showed different power to predict soil series in WCA 2A. The soil series prediction model using the SPOT image showed the highest prediction power followed by the Landsat ETM+, and MODIS images without bedrock depth data. On the other hand, the prediction model using the MODIS image showed the highest prediction power with bedrock depth data. Overall differences among the tree models were small and all of them showed good prediction power in terms of relative error and accuracy metrics. The coars er resolution input variables (e.g., MODIS derived predictor variables at 250 m spatial resolution) produced rougher

PAGE 199

199 patterns of soil series limiting somewhat the ability to capture the underlying finescale variability of soils in WCA 2A. Soil nutrients s tored in wetland soils are critical to assess the effectiveness of restoration efforts, yet challenging to accurately derive soil heterogeneity. In chapter 3, univariate Block Kriging (BK) and multivariate Random Forest (RF) methods were used to assess the spatial distributions of soil total phosphorus (TP) and total nitrogen (TN) in WCA 2A. The RF models with varying spatial resolutions of RS images showed substantial improvement of prediction accuracy when compared to BK to predict soil TP and TN in WCA 2 A. The variable importance of the RF models for soil TP and TN suggests that soil TP is highly dependent on biotic/vegetation properties that can be inferred by RS, whereas soil TN was predicted using a combination of biotic/vegetation, topographic, and hy drologic variables. Results suggest that the spectral data informed soil prediction models have excellent predictive capabilities in this aquatic ecosystem. Interestingly, there was no noticeable distinction among different spatial resolutions of RS images to develop prediction models for soil TP and TN in terms of coefficient of determination ( R2) and root mean square error ( RMSE ) However, the entropy analysis quantified differences in the spatial variations and complexities in soil TP and TN predictions. It showed that the RF models derived from fine resolution RS images (i.e., SPOT and Landsat ETM+) contain more information than the RF model derived from the coarse resolution MODIS image, although the R2 and RMSE were similar among the prediction models. Another key nutrient is soil carbon (C). Wetland soils are able to exhibit both consumption and production of greenhouse gases and they play an important role in

PAGE 200

200 regulation of global C cycles. Although it is challenging to accurately evaluate the actual amount of C storages and sequestration or emission rates in wetlands, the incorporation of RS data into digital soil models has great potential to assess C dynamics in wetland soils. In chapter 4, soil total carbon (TC) stocks were assessed using RF with s pectral indices derived from RS images and environmental ancillary data. The RF models showed high performance to predict TC stocks with a R2 between 0.85 to 0.92 and a RMSE between 0.75 to 0.86 kg m2. The variable importance of the RF models revealed tha t hydrology was the major environmental factor controlling the spatial distribution of soil C stocks in WCA 2A. Our results showed that WCA 2A stores about 4.2 mega tons of C in the top 20 cm soils with about 9.7 kg soil TC stocks per unit area (m2) One i mportant conclusion is that appropriate management of the wetland is critical to maintain its potential ability to sequester C into soils. For successful model developments and applications, it is essential to understand model transferability and scaling effects. It is necessary to elucidate how a model developed in a specific area transfers to another area and how model predictor variables change. In chapter 5, the RF soil prediction models developed in WCA 2A (Chapters 3 and 4) were applied to WCA 3AN, a nd the models developed in WCA 3AN were applied to WCA 2A to test model transferability. In addition, RF soil predictions models developed in subregions were upscaled to the whole study region (i.e., WCA 2A 2A combined with W CA 3AN); and WCA 3AN whole wetland region), and RF soil prediction models were downscaled from the whole wetland region to subregions (i.e., whole wetland region 2A; and whole wetland region 3AN). Results showed that model transferability and scaling

PAGE 201

201 varies depending on attribute domain space (soil properties and environmental covariates) and geographic domain space (i.e., the extent used to develop models). The RF models for soil TP developed in WCA 3AN showed moderate transferability with an R2 between 0.38 and 0.40, when the models were applied to WCA 2A. However, no transferability was discovered for soil TN and TC predictions. In contrast, the only RF model that showed weak transferability from WCA 2A to WCA 3AN (R2 = 0.15) was derived fr om SPOT predictor variables. Constraints that limit transferability of soil prediction models among the two wetlands were found to be based on differences in hydrologic management and subsequent soil genesis. The limited transferability stands in sharp contrast to the excellent soil prediction models developed for subregions using MODIS and SPOT images along with ancillary environmental predictors. Upscaling of soil prediction models from subregion to whole region scale (i.e., WCA 2A region with R2 from 0.23 to 0.82; and WCA 3AN 2 from 0.34 to 0.55) degraded model performances somewhat. In contrast, downscaling of soil prediction models from the whole region to subregions (i.e., whole region 2A with R2 from 0.90 to 0.95; and whole region 3AN with R2 from 0.75 to 0.91) was successful. Findings from this study suggest that local customized soil models (developed separately in WCA 2A and WCA 3AN) and globally developed soil models (using the pooled dataset from both wetlands) performed similarly and exceptionally well, but model transferability (from one wetland to another) was severely constraint. Also uncertainties of soil TP, TN, and TC were greatly increased when the models developed in one area were applied to anot her area and upscaled to a larger area. This suggests that the extent is less important for successful modeling of soil TP, TN, and

PAGE 202

202 TC as long as the domain variability of soil and environmental parameters (biotic, hydrologic, topographic, etc.) is incorporated into the model. In empirical regression models, like the ones presented in this study, the relationships between soil and environmental covariates impose constraints to transfer or scale soil models. The spatially explicit relationships between predi ctor (environmental covariates) and target variables (soil TP, TN, and TC) are interlinked with the heterogeneity in soil and environmental properties and have profound effects on the sensitivity and universality of a given soil model. This dissertation also showed the potential use of visible and near infrared reflectance spectroscopy (VNIRS) to predict various soil properties including soil pH, TP, TN, TC, Mehlich I calcium (Ca) and magnesium (Mg) in chapter 6. The partial least squares regression (PLS R) models for soil properties in WCA 2A showed intermediate prediction results (R2 ranges from 0.60 to 0.78 for the validation set), except MehlichI Ca (R2 of 0.35) and Mg (R2 of 0.50). Although the prediction results were not as high as other studies usi ng VNIRS, it is important to note that other studies were conducted with upland soils, which have less organic matters, more minerals, and more contrasting soil colors than wetland soils. Our results show that the models developed with VNIRS for WCA 2A soi ls have the potential to be improved by using different transformation methods or different multivariate methods, and demonstrate applicability of VNIRS for the predictions of various properties of highly organic wetland soils. Overall, this study provided ample evidence that RS informed prediction models can successfully infer on multiple biophysical properties in soils and soil classes in aquatic ecosystems. The spatial distributions of soil properties, major controlling

PAGE 203

203 environmental factors for each soi l properties, and absolute storage amounts for soil TP, TN, and TC were assessed. Results provided a better understanding of how fin e versus coarse grain resolutions of RS images impact s oil modeling, model transferability and scaling effects. Also this study showed the potential use of VNIRS for wetland soil property estimations. Large changes in temperature and precipitation in the Everglades region are predicted to occur over the next several decades. Thus, sequentially reproducible assessment methods are necessary for measuring and monitoring the impacts of those changes in soil nutrients of the wetland. The methodology (RS supported DSM) could be beneficial to support measuring and monitoring these changes and has the c apability to provide useful information to preserve the wetlands values and functions.

PAGE 204

204 LIST OF REFERENCES Altor, A.E., and W.J. Mitsch. 2008. Methane and carbon dioxide dynamics in wetland mesocosms: effects of hydrology and soils. Ecol Appl 18:1307 1320. Armentano, T.V., and E.S. Menges. 1986. Patterns of change in the carbon balance of organic soil wetlands of the temperate zone. J Ecol 74:755 774. Arnold, R.W. 2006. Soil survey and soil classification. p. 3759. In S. Grunwald (ed.) Environmental soill andscape m odeling. G eographic i nformation t echnologies and p edometrics. CRC Press, New York. Augustin, J., W. Merbach, and J. Rogasik. 1998. Factors influencing nitrous oxide and methane emissions from minerotrophic fens in Northeast Germany. B iol. Fert Soils 28:1 4. Bagheri Bodaghabadi, M., M.H. Salehi, J.A.Martinez Casasnovas, J. Mohammadi, N. Toomanian, and I.E. Borujeni. 2011. Using canonical correspondence analysis to identify the most important DEM attributes for digital soil mapping appl ications. Catena 86:66 74. Bai, Z.G., D.L. Dent, L. Olsson, and M.E. Schaepman. 2008. Proxy global assessment of land degradation. Soil Use Manage. 24:223 234. Beckage, B., W.J. Platt, and B. Panko. 2005. A climatebased approach to the restoration of firedependent ecosystems. Restor Ecol 13:429 431. Bellamy, P.H., P.J. Loveland, R.I. Bradley, R.M. Lark, and G.J.D. Kirk. 2005. Carbon losses from all soils across England and Wales 1978 2003. Nature 437:245 248. Bernal, B., and W.J. Mitsch. 2008. A comparison of soil carbon pools and profiles in wetlands in Costa Rica and Ohio. Ecol Eng 34:311 323. Bernhardt, C.E., and D.A. Willard. 2009. Response of the Everglades ridge and slough landscape to c limate variability and 20thcentury water management. Ecol Appl 19:1723 1738. Berryman, E.M., R.T. Venterea, J.M. Baker, P.R. Bloom, and B. Elf. 2009. Phosphorus and greenhouse gas dynamics in a drained calcareous wetland soil in Minnesota. J. Environ. Qual 38:2147 2158. Bhatti, J.S., and C. Tarnocai. 2009. Influence of climate and land use change on carbon in agriculture, forest, and peatland ecosystems across Canada. In R. Lal and R.F. Follett (eds .) Soil carbon sequestration and the g reenhouse e ffec t. SSSA Special Publication 57, Madison, WI.

PAGE 205

205 Bishop, T.F.A., A.B. McBratney, and B.M. Whelan. 2001. Measuring the quality of digital soil maps using information criteria. Geoderma 103:95 111. Blodau, C. 2002. Carbon cycling in peatlands A review of processes and controls. Environ. Rev 10:111 134. Bl Hydrol P rocess. 9:251 290. Brady, N.C., and R.R. Weil. 2008. The n ature and p roperties of soils. Pearson Prentice Hall, Ohio. Breiman, L. 1984. Classification and regression trees. Chapman & Hall CRC Press, New York Breiman, L. 2001. Random forests. Mach Learn 45:5 32. Breiman, L. J.H. Friedman, R.A. Olshen, and C.J. Stone. 1984. Classi regression trees. Wads w orth International Group, Pacific Grove, California. Bridgham, S.D., C.A. Johnston, J. Pastor, and K. Updegraff. 1995. Potential f eedbacks of Northern Wetlands on climate c hange. BioScience 45: 262 274. Browder, J.A., P.J. Gleason, and D.R. Swift. 1994. Per iphyton in the Everglades: Spatial variation, environmental correlates, and ecological implications. p. 379 418. In S.M. Davis and J.C. Ogden (eds.) Everglades: The ecosystem and its restoration. St. Lucie Press, Boca Raton, FL. Brown, D.J., K.D. Shepherd M.G. Walsh, M.D. Mays, and T.G. Reinsch. 2006. Global soil characterization with VNIR diffuse reflectance spectroscopy. Geoderma 132:273 290. Brown, L., J.M. Chen, S.G. Leblanc, and J. Cihlar. 2000. A shortwave infrared modification to the simple ratio f or LAI retrieval in Boreal forests: A n image and model analysis. Remote Sens. Environ. 71:16 25. Bruland, G.L., S. Grunwald, .Z. Osborne, K.R. Reddy, and S. Newman. 2006. Spatial distribution of soil properties in Water Conservation Area 3 of the Everglades. Soil Sci Soc Am J. 70:1662 1676. Bruland, G.L., T.Z. Osborne, K.R. Reddy, S. Grunwald, S. Newman, and W.F. DeBusk. 2007. Recent changes in soil total phosphorus in the Everglades: W ater conservation area 3. Environ. Monit. Assess. 129:379 395. Brus, D.J., B. Kempen, and G.B.M. Heuvelink. 2011. Sampling for validation of digital soil maps. Eur. J. Soil Sci. 62:394 407.

PAGE 206

206 Cahn, M.D., J.W. Hummel, and B.H. Brouer. 1994. Spatial Analysis of Soil Fertility for Site Specific Crop Management. Soil Sci. Soc. Am J. 58:1240 1248. Cerri, C.E.P., M. Bernoux, V. Chaplot, B. Volkoff, R.L. Victoria, J.M. Melillo, K. Paustian, and C.C. Cerri. 2004. Assessment of soil property spatial variation in an Amazon pasture: basis for selecting an agronomic experimental area. Ge oderma 123:54 68. Chaplot, V., C. Walter, and P. Curmi. 2003. Testing quantitative soil landscape models for predicting the soil hydromorphic index at a regional scale. Soil Sci 168:445 454. Chen, D. Y., J.F. Huang, and T.J. Jackson. 2005. Vegetation water content estimation for corn and soybeans using spectral indices derived from MODIS near and shortwave infrared bands. Remote Sens. Environ. 98:225 236. Chen, F., D.E. Kissel, L.T. West, W. Adkins, D. Rickman, and J.C. Luvall. 2008. Mapping Soil Organic Carbon Concentration for Multiple Fields with Image Similarity Analysis. Soil Sci. Soc. Am. J. 72:186 193. Chen, J.M., G. Pavlic, L. Brown, J. Cihlar, S.G. Leblanc, H.P. White, R.J. Hall, D.R. Peddle, D.J. King, J.A. Trofymow, E. Swift, J. Van der Sanden, and P.K.E. Pellikka. 2002. Derivation and validation of Canadawide coarseresolution leaf area index maps using highresolution satellite imagery and ground measurements. Remote Sens. Environ. 80:165 184. Childers, D.L., R.F. Doren, R. Jones, G.B. Noe, M. Rugge, and L.J. Scinto. 2003. Decadal changes in vegetation and soil phosphorus pattern across the Everglades Landscape. J. Environ. Qual. 32:344 362. Chimney, M.J., and G. Goforth. 2006. History and description of the Everglades nutrient removal project a subtropical constructed wetland in south Florida (USA). Ecol Eng 27:268 278. Chmura, G.L., S.C. Anisfeld, D.R. Cahoon, and J.C. Jynch. 2003. Global carbon sequestration in tidal, saline wetland soils. Global biogeochem C y 17:1111 1122. Cohen, W.B. 1991. Response of vegetation indices to changes in three measures of leaf water stress. Photogramm. Eng. Remote Sens. 57:195 202. Cohen, M.J., J.P. Prenger, and W.F. DeBusk. 2005. Visiblenear infrared reflectance spectroscopy for rapid, nondes tructive assessment of wetland soil quality. J. Environ. Qual. 34:1422 1434.

PAGE 207

207 Conant, R.T., M.G. Ryan, G.I. gren, H. Birge, E.A. Davidson, P. Eliasson, S. Evans, S. Frey, C.P. Giardina, F.M. Hopkins, R. Hyvnen, M.U.F. Kirschbaum, J.M. Lavallee, J. Leifeld, W.J. Parton, J.M. Steinweg, M.D. Wallenstein, J..M. Wetterstedt, and M.A. Bradford. 2011. Temperature and soil organic matter decomposition rates synthesis of current knowledge and a way forward. Glob. Change Biol 17:3392 3404. Colombo, R., D. Bellin geri, D. Fasolini, and C.M. Marino. 2003. Retrieval of leaf area index in different vegetation types using high resolution satellite data. Remote Sens. Environ. 86:120 131. Congalton, R. G. 1991. A review of assessing the accuracy of classifications of rem otely sensed data. Remote Sens. Environ. 37: 35 46. Corstanje, R., and K.R. Reddy. 2006. Microbial indicators of nutrient enrichment: A mesocosm study. Soil Sci. Soc. Am. J. 70:1652 1661. Corstanje, R., S. Grunwald, K.R. Reddy, T.Z. Osborne, and S. Newman 2006. Assessment of the spatial distribution of soil properties in a northern Everglades marsh. J Environ. Qual 35:938 949. Cozzolino, D., and A. Morn. 2003. The potential of near infrared reflectance spectroscopy to analyze soil chemical and physical characteristics. J. Agr Sci 140:65 71. Craft, C.B., and C.J. Richardson. 1993. Peat accretion and N, P, and organic C accumulation in nutrient enriched and unenriched Evergaldes peatlands. Ecol Appl 3:446 458. Craft, C.B., and C.J. Richardson. 1998. R ecent and long term organic soil accretion and nutrient accumulation in the Everglades. Soil Sci. Soc. Am. J. 62:834 843. Dang, Y.P., M.J. Pringle, M. Schmidt, R.C. Dalal, and A. Apan. 2011. Identifying the spatial variability of soil constraints using mul ti year remote sensing. Field Crop. Res 123:248 258. David, P.C. 1996. Changes in plant communities relative to hydrologic conditions in the Florida Everglades. Wetlands 16:15 23. Death, G., and K.E. Fabricius. 2000. Classification and regression trees: a powerful yet simple technique for ecological data analysis. Ecology 81:3178 3192. DeBusk, W.F., K.R. Reddy, M.S. Koch, and Y. Wang. 1994. Spatial distribution of soil nutrients in a northern Everglades marsh: Water Conservation Area 2A. Soil Sci. Soc. A m. J. 58:543 552.

PAGE 208

208 Daz Uriarte, R., and S.A. de Andrs. 2006. Gene selection and classification of microarray data using random forest. BMC Bioinformatics 7 doi:10.1186/147121057 3. Eldeiry, A.A., and L.A. Garcia. 2008. Detecting soil salinity in Alfalf a f ields using spatial m odeling and r emote s ensing. Soil Sci. Soc. Am. J. 72:201 211. Everglades Depth Estimation Network, EDEN, 2009. Available at http://sofia.usgs.gov/eden/meteorologic.php (verified 22 Jan. 2010). USGS US Army Corps of Engineers, Jacksonville, FL. Fernndez Buces, N, C. Siebe, S. Cram, and J.L. Palacio. 2006. Mapping soil salinity using a combined spectral response index for bare soil and vegetation: a case study in the former lake Texcoco, Mexico. J. Arid Environ. 65:644 667. Fiedler, S., and M. Sommer. 2000. Methane emissions, groundwater levels and redox potentials of common wetland soils in a temperate humid climate. Global Biogeochem Cy 14:1081 1093. Florida Fish and Wildlife Conservation Commission. 2003. Florida vegetation and land cover data derived from 2003 Landsat ETM+ imagery. FFWCC, Tallahassee, FL. Follett, R.F., J.M. Kimble, E.G. Pruessner, S. Samson Liebig, and S. Waltman. 2009. Soil organic carbon stocks with depth and land use at various U.S. sites. In R. Lal and R.F. Follett (eds.) Soil carbon sequestration and the g reenhouse ef fect. S SSA Special Publication 57, Madison, WI. Forster, P., V. Ramaswamy, P. Artaxo, T. Berntsen, R. Betts, D.W. Fahey, J. Hayw ood, J. Lean, D.C. Lowe, G. Myhre, J. Nganga, R. Prinn, G. Raga, M. Schulz and R. Van Dorland. 2007. Changes in a tmospheric constituents and in r adiative f orcing. In S. Solomon, D. Qin, M. Manning, Z. Chen, M. Marquis, K.B. Averyt, M.Tignor and H.L. Miller (eds.) Climate c hange 2007: The p hysical science b asis. Contribution of w orking g roup I to the f ourth a ssessment r eport of the i ntergovernmental p anel on climate change Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA. Fu, W., H. Tunney, and C. Zhang. 2010. Spatial variation of soil nutrients in a dairy farm and its implications for sitespecific fertilizer application. Soil Till. Res 106:185 193. Gaiser, E. 2009. Periphyton as an indicator of restoration in the Florida Everglades. Ecol. Indic. 9:S37 S45. Galbraith, J.M., P.J.A. Kleinman, and R.B. Bryant. 2003. Sources of uncertainty affecting soil organic carbon estimates in northern New York. Soil Sci. Soc. Am. J. 67:1206 1212.

PAGE 209

209 Gao, B.C. 1996. NDWI A normalized difference w ater index for remote sensing of vegetation liquid water from space. Remote Sens. Environ. 58:257 266. Gao, X., A.R. Huete, Wenge Ni, and T. Miura. 2000. Optical biophysical relationships of vegetation spectra without background contamination. Remote Sens. Environ. 74:609 620. Gerard, G., and J. Chanton. 1993. Quantification of methane oxidation in the rhizosphere of emergent aquatic macrophytes: defining upper limits. Biogeochemistry 23:79 97. Gitelson, A.A., Y.J. Kaufman, and M.N. Merzlyak. 1996. Use of a green channel in remote sensing of global vegetation from EOS MODIS. Remote Sens. Environ. 58:289 298. Gitelson, A.A., and M.N. Merzlyak. 1994. Quantitative estimation of chlorophyll a using reflectance spectra: Experiments with autumn ches tnut and maple leaves. J Photoch. Photobio. B 22:247 252. Gleason, P.J., A.D. Cohen, P. Stone, W.G. Smith, H.K. Brooks, R. Goodrick, and W. Spackman. 1974. The environmental significance of Holocene sediments from the Everglades and saline tidal plains. p. 297 351. In P.J. Gleason (ed.) Environments of south Florida, present and past. Coral Gables, FL. Gleason, P.J., and W.S. Spackman Jr. 1984. Calcareous periphyton and water chemistry in the Everglades. In P.J. Gleason ( e d), Environments of s outh F lorida p resent and past Coral Gables, FL. Gomez, C., R.A. Viscarra Rossel, and A.B. McBratney. 2008. Soil organic carbon prediction by hyperspectral remote sensing and field vis NIR spectroscopy: an Australian case study. Geoderma 146:403 411. Goovaerts, P. 1997. Geostatistics for n atural r esources e valuation. Oxford University Press, New York. Goovaerts, P. 1999. Geostatistics in soil science: Stateof the art and perspectives. Geoderma 89:1 45. Goovaerts, P., and C.N. Chiang. 1993. Temporal p ersistence of s pat ial p atterns for m ineralizable n itrogen and s elected soil p roperties. Soil Sci. Soc. Am. J. 57:372 381. Gorham, E. 1991. Northern peatlands: role in the carbon cycle and probable responses to climatic warming. Ecol. Appl 1:182 195. Grimm, R., T. Behrens, M. Mrker, and H. Elsenbeer. 2008. Soil organic carbon concentrations and stocks on Barro Colorado Island Digital soil mapping using Random Forests analysis. Geoderma 146:102 113.

PAGE 210

210 Grunwald, S. 2006. What d o w e r eally know about the s pacet ime c ontinuum o f soill andscapes? p. 336. In S. Grunwald (ed.). Environmental soill andscape m odeling. Geographic i nformation t echnologies and p edometrics. CRC Press, New York. Grunwald, S. 2009. Multi criteria characterization of recent digital soil mapping and modeling approaches. Geoderma 152:195 207. Grunwald, S. and S. Lamsal. 2006. The i mpact of e merging g eographic i nformation t echnology on soill andscape m odeling. p. 127 154. In S. Grunwald (ed.). Environmental soill andscape m odeling Geographic i nformatio n t echnologies and p edometrics. CRC Press, New York. Grunwald, S., T.Z. Osborne, and K.R. Reddy. 2008. Temporal trajectories of phosphorus and pedopatterns mapped in water conservation area 2, Everglades, Florida, USA. Geoderma 146:1 13. Grunwald S., S.H. Daroub, T.A. Lang and O.A. Diaz. 2009. Treebased modeling of complex interactions of phosphorus loadings and environmental factors. Sci. Total Environ. 407:3772 3783. Grunwald S., J.A. Thompson, and J.L. Boettinger. 2011. Digital soil mapping and modelin g at continental scales finding solutions for global issues. Soil Sci. Soc. Am. J. 75:1201 1213. Grunwald, S., and K.R. Reddy. 2008. Spatial behavior of phosphorus and nitrogen in a subtropical wetland. Soil Sci. Soc. Am. J. 72:1174 1183. Grunwald, S., K.R. Reddy, S. Newman, and W.F. DeBusk. 2004. Spatial variability, distribution and uncertainty assessment of soil phosphorus in a south Florida wetland. Environmetrics 15:811 825. Grunwald S., G.M. Vasques, N.B. Comerford, G.L. Bruland, and C.M. Bliss. 20 10. Regional modeling of carbon, nitrogen and phosphorus geospatial patterns. In G. Hanrahan (ed.). Modeling of pollutants in complex environmental systems Vol. II. ILM Publ., Hertfordshire, UK Guo, Y., P. Gong, R. Amundson, and Q. Yu. 2006. Analysis of f actors controlling soil carbon in the conterminous United States. Soil Sci. Soc. Am. J. 70:601 612. Hanan, E.J., and M.S. Ross. 2010. Across scale patterning of plant soilwater interactions surrounding tree islands in Southern Everglades landscapes. Lands cape Ecology 25:463 476. Harris, R.C., D.I. Sebacher, and F.P. Day. 1982. Methane flux in the Great Dismal Swamp. Nature 297:673 674.

PAGE 211

211 Hausser, J., and K. Strimmer. 2009. Entropy inference and the James Stein estimator, with application to nonlinear gene as sociation networks. J Mach Learn Res 10:1469 1484. Helland, I.S. 1990. Partial least squares regression and statistical models. Scand. J. Stat. 17:97 114. Heuvelink, G.B.M., and E.J. Pebesma. 1999. Spatial aggregation and soil process modeling. Geoderm a 89:47 65. Huang, X., S. Senthilkumar, A. Kravchenko, K. Thelen, and J. Qi. 2007. Total carbon mapping in glacial till soils using near infrared spectroscopy, Landsat imagery and topographical information. Geoderma 141:34 42. Huete, A.R., H.Q. Liu, K. Batchily, and W. van Leeuwen. 1997. A comparison of vegetation indices over a global set of TM images for EOS MODIS. Remote Sens. Environ. 59:440 451. Hughes, R.F., J.B. Kauffman, and V.J. Jaramillo. 1999. Biomass, carbon, and nutrient dynamics of secondary f orests in a humid tropical region of Mexico. Ecology 80:1892 1907. Ibez, J.J., S. De Alba, A. Lobo, and V. Zucarello. 1998. Pedodiversity and global soil patterns at coarse scales (with discussion). Geoderma 83:171 192. Inglett, P.W., K.R. Reddy, and P.V. McCormick. 2004. Periphyton tissue chemistry and nitrogenase activity in a nutrient impaced Everglades ecosystem. Biogeochemistry 67:213 233. Inglett, P.W., V.H. Rivera Monroy, and J.R. Wozniak. 2011. Biogeochemistry of nitrogen across the Everglades landscape. Crit Rev Env Sci Tec 41:187 216. Jaenicke, J., J.O. Rieley, C. Mott, P. Kimman, and F. Siegert. 2008. Determination of the amount of carbon stored in Indonesian peatlands. Geoderma 147:151 158. Jensen, J.R. 2000. Remote sensing of the e nvir onment: an E arth resource perspective. Prentice Hall, New Jersey. Jensen, J.R. 2005. Introductory d igital i mage p rocessing: a remote sensing perspective. Prentice Hall, New Jersey. Jensen, J.R., K. Rutchey, M.S. Koch, and S. Narumalani. 1995. Inland wetland change detection in the Everglades water conservation area 2A using a time series of normalized remotely sensed data. Photogram m Eng. Remote Sens. 61:199 209. Jiang, Z., A.R. Huete, K. Didan, and T. Miura. 2008. Development of a twoband enhanced vegetation index without a blue band. Remote Sens Environ. 112:3833 3845.

PAGE 212

2 12 Jobbgy, E.G., and R.B. Jackson. 2000. The vertical distribution of soil organic carbon and its relation to climate and vegetation. Ecol Appl 10:423 436. Juston, J., and T.A. DeBusk. 20 06. Phosphorus mass load and outflow concentration relationships in stormwater treatment areas for Everglades restoration. Ecol Eng 26:206 223. Karl, T.R., J.M. Melillo, and T.C. Peterson. 2009. Global climate change impacts in the United States. Cambridge University Press. New York, USA. Karnieli, A., U. Gilad, M. Ponzet, T. Svoray, R. Mirzadinov, and O. Fedorina. 2008. Assessing land cover change and degradation in the Central Asian deserts using satellite image processing and geostatistical methods. J Arid Environ. 72:2093 2105. Kauth, R.J., and G.S. Thomas. 1976. The tasseled cap A graphic description of the spectral temporal development of agricultural crops as seen by Landsat. p.4B41 4B51. In Proc. of the Symposium on Machine Processing of Remotel y Sensed Data, Purdue University, West Lafayette, IN. June 29 July 1, 1976. IEEE Kayranli, B., M. Scholz, A. Mustafa, and Hedmark. 2010. Carbon storage and fluxes within freshwater wetlands: a critical review. Wetlands 30:111 124. Kempen, B., D.J. Bru s, G.B.M. Heuvelink, and J.J. Stoorvogel. 2009. Updating the 1:50,000 Dutch soil map using legacy soil data: a multinomial logistic regression approach. Geoderma 151:311 326. Kerry, R., and M.A. Oliver. 2004. Average variograms to guide soil sampling. Int J Appl Earth Obs 5:307 325. Kim, J., S. Grunwald, R. Rivero, and R. Robbins. 2012. Multi scale modeling of soil series using remote sensing in a wetland ecosystem. Soil Sci. Soc. Am. J. In press doi:10.2136/sssaj2012.0043 Koelbener, A., L. Strm, P.J. Edwards, and H.O. Venterink. 2010. Plant species from mesotrophic wetlands cause relatively high methane emissions from peat soil. Plant Soil 326:147 158. Kulmatiski, A., D.J. Vogt, T.G. Siccama, J.P. Tilley, K. Kolesinskas, T.W. Wickwire, and B.C. Larson. 2004. Landscape determinants of soil carbon and nitrogen storage in southern New England. Soil Sci. Soc. Am. J. 68:2014 2022. Kustas, W.P., F. Li, T.J. Jackson, J.H. Prueger, J.I. MacPherson, and M. Wolde. 2004. Effects of remote sensing pixel resolution on modeled energy flux variability of croplands in Iowa. Remote Sens Environ. 92:535 547.

PAGE 213

213 Lagacherie, P., and M. Voltz. 2000. Predicting soil properties over a region using sample information from a mapped reference area and digital elevation data: a con ditional probability approach. Geoderma 97:187 208. Lal, R. 2003. Global potential of soil carbon sequestration to mitigate the greenhouse effect. Crit R ev P lant S ci 22: 151 184. Lal, R. 2008. Carbon sequestration. Philos T Roy Soc B 363:815 830. Lal, R., and R.F. Follett. 2009. Soils and climate change. In R. Lal and R.F. Follett (eds.) Soil carbon sequestration and the g reenhouse e ffect. SSSA Special Publication 57, Madison, WI. Landis, J.R., and G. Koch. 1977. The measurement of observer agreement for categorical data. Biometrics 33:159 174. Larsen, L.G., and J.W. Harvey. 2011. Modeling of hydroecological feedbacks predicts distinct classes of landscape pattern, process, and restoration potential in shallow aquatic ecosystems. Geomorphology 126:279 296. Liaw, A., and M. Wiener. 2002. Classification and regression by random forest. R news 2:18 22. Lobell, D.B., J.I. Ortiz Monasterio, F.C. Gurrola, and L. Valenzuela. 2007. Identification of saline soils with multiyear remote sensing of crop yields. So il Sci. Soc. Am. J. 71:777 783. Lodge, T. E. 2005. The Everglades handbook: understanding the ecosystem CRC Press, Boca Raton, FL. Loveless, C.M. 1959. A study of the vegetation in the Florida Everglades. Ecology 40:1 9. M Pastor, I., J. N Pedreno, I. Gomez, and M. Koch. 2008. Identifying optimal spectral bands to assess soil properties with VNIR radiometry in semi arid soils. Geoderma 147:126 132. Mack, M.C., M.S. BretHarte, T.N. Hollingsworth, R.R. Jandt, E.A.G. Schuur, G.R. Shaver, and D.L. Verbyla. 2011. Carbon loss from an unprecedented Arctic tundra wildfire. Nature 475:489 492. Martn, M.A., and J M. Rey. 2000. On the role of Shannon's entropy as a measure of heterogeneity. Geoderma 98:1 3. Martn, M.A., J M. Rey, a nd F.J. Taguas. 2005. An entropy based heterogeneity index for mass size distributions in Earth science. Ecol Model 182:221 228.

PAGE 214

214 McBratney, A.B. 1998. Some considerations on methods for spatially aggregating and disaggregating soil information. Nutr Cyc l Agroecosys 50:51 62. McBratney, A.B., M.L Mendonca Santos, and B. Minasny. 2003. On digital soil mapping. Geoderma 117:3 52. McCormick, P.V., and M.B. ODell. 1996. Quantifying periphyton responses to phosphorus in the Florida Everglades: A synoptic ex perimental approach. J. N. Am. Benthol. Soc. 15:450 468. McCormick, P.V., R.B.E. Shuford III, J.G. Backus, and W.C. Kennedy. 1998. Spatial and seasonal patterns of periphyton biomass and productivity in the northern Everglades, Flordia, U.S.A. Hydrobiologi a 362:185 208. McCarty, G.W., J.B. Reeves III, V.B. Reeves, R.F. Follett, and J.M. Kimble. 2002. Mid infrared and near infrared diffuse reflectance spectroscopy for soil carbon measurement. Soil Sci. Soc. Am. J. 171:94 102. McGrath, D., and Zhang, C. 2003. Spatial distribution of soil organic carbon concentrations in grassland of Ireland. Appl Geochem 18: 1629 1639. McKenzie, N.J., and P.J. Ryan. 1999. Spatial prediction of soil properties using environmental correlation. G eoderma 89:67 94. Melack, J.M., L.L. Hess, M. Gastil, B.R. Forsberg, S.L. Hamilton, I.B.T. Lima, and E.M.L.M. Novo. 2004. Regionalization of methane emissions in the Amazon Basin with microwave remote sensing. Glob. Change Biol 10:530 544. Mevik, B H., a nd R. Wehrens. 2007. The pls package: principal component and partial least squares regression in R. J Stat Softw 18:1 23. Minasny, B., A.B. McBratney, M.L. Mendona Santos, I.O.A. Odeh, and B. Guyon. 2006. Prediction and digital mapping of soil carbon storage in the Lower Namoi Valley. Aust J Soil Res 44:233 244. Minasny, B., and A.E. Hartemink. 2011. Predicting soil properties in the tropics. EarthSci Rev 106:52 62. Mirik, M., J.E. Norland, R.L. Crabtree, and M.E. Biondini. 2005. Hyperspectral one meter resolution remote sensing in Yellowstone National Park, Wyoming: II. Biomass. Rangeland Ecol Manag 58:459 465. Mitsch, W.J., and J.G. Gosselink 2007. Wetlands (Fourth ed.). John Wiley & Sons, Inc., New Jersey, USA. Mitsch, W.J., B. Bernal, A.M Nahlik, Mander, L. Zhang, C.J. Anderson, S.E. Jrgensen, and H. Brix. 2012. Wetlands, carbon, and climate change. Landscape Ecol I n press doi : 10.1007/s10980012 97588.

PAGE 215

215 Mowry, H., and H.H. Bennett. 1948. Soils, geology, and water control in the Ever glades region. University of Florida Agricultural Experiment Station Bulletin 442, Gainesville, FL. Mulder, V.L., S. de Bruin, M.E. Schaepman, and T.R. Mayr. 2011. The use of remote sensing in soil and terrain mapping A review. Geoderma 162:1 19. Musick, H.B., and R.E. Pelletier. 1988. Response to soil moisture of spectral indexes derived from bidirectional reflectance in thematic mapper wavebands. Remote Sens. Environ. 25:167 184. Mutanga, O., and A.K. Skidmore. 2004. Narrow band vegetation indices overc ome the saturation problem in biomass estimation. Int J. Remote Sens 25:3999 4014. M de Asis, A., and K. Omasa. 2007. Estimation of vegetation parameter for modeling soil erosion using liner spectral m ixture a nalysis of Landsat ETM data. J Photogrammetry Remote Sens 62: 309 324. National Climatic Data Center. 2009. National Oceanic and Atmospheric Administration. Available at http://www.ncdc.noaa.gov/oa/ncdc.html (verified 6 Jan. 2010). USD C NCDC, Asheville, NC. Nellis, M.D., and J.M. Briggs. 1992. Transformed vegetation index for measuring spatial variation in drought impacted biomass on Konza Prairie, Kansas. T Kansas Academy Sci 95:93 99. Nield, S.J., J.L. Boettinger, and R.D. Ramsey. 2007. Digitally m apping g ypsic and n atric soil a reas u sing Landsat ETM d ata. Soil Sci. Soc. Am. J. 71:245 252. Noe, G. B., D. L. Childers, and R. D. Jones. 2001. Phosphorus biogeochemistry and the impact of phosphorus enrichment: Why is the Everglades so unique? Ecosystems 4:603 624. Noe, G.B., and D.L. Childers. 2007. Phosphorus budgets in Everglades wetland ecosystems: the effects of hydrology and nutrient enrichment. Wetlands Ecol Manage. 15:189 205. Noe, G.B., D.L. Childers, A.L. Edwards, E. Gaiser, K. Jayachandran, D. Lee, J. Meeder, J. Richards, L.J. Scinto, J.C. Trexler, and R.D. Jones. 2002. Short term changes in phosphorus storage in an oligotrophic Everglades wetland ecosystem receiving experimental nutrient enrichment. Biogeochemistry 59:239 26 7. Odeh, I.O.A., A.B. McBratney, and D.J. Chittleborough. 1994. Spatial prediction of soil properties from landform attributes derived from a digital elevation model. Geoderma 63:197 214.

PAGE 216

216 Osborne, T.Z., G. Bruland, S. Newman, K. Reddy, and S. Grunwald. 2011a. Spatial distributions and ecopartitioning of soil biogeochemical properties in the Everglades National Park. Environ. Monit Assess. 183:395 408. Osborne, T.Z., S. Newman, D.J. Scheidt, P.I. Kalla, G. Bruland, M.J. Cohen, L.J. Scinto, and L.R. Ell is. 2011b. Landscape patterns of significant soil nutrients and contaminants in the Greater Everglades ecosystem: Past, present, and future. Crit Rev Env Sci Tec 41:121 148. Ozesmi, S.L., and M.E. Bauer. 2002. Satellite remote sensing of wetlands. Wet lands Ecol Manage 10:381 402. Peng, W., D.B. Wheeler, J.C. Bell, and M.G. Krusemark. 2003. Delineating patterns of soil drainage class on bare soils using remote sensing analyses. Geoderma 115:261 279. Pinney, M.L., P.K.M. Westerhoff, and L. Bakerm. 2000. Transformations in dissolved organic carbon through constructed wetlands. Water Res 34:1897 1911. Porter, J.W., and K.G. Porter. 2002. The Everglades, Florida Bay, and coral reefs of the Florida Keys: An ecosystem sourcebook. CRC Press, Boca Raton, FL. Post, W.M., and K.C. Kwon. 2000. Soil carbon sequestration and landuse change: processes and potential. Glob Change Biol 6: 3 17 327. Post, W.M., W.R. Emanuel, P.J. Zinke, and A.G. Stangenberger. 1982. Soil carbon pools and world life zones. Nature 298 : 1 56 159. Prasad, A.M., L.R. Iverson, and A. Liaw. 2006. Newer classification and regression tree techniques: bagging and random forests for ecological prediction. Ecosystems 9:181 199. Price, K.P., X. Guo, and J.M. Stiles. 2002. Optimal Landsat TM band comb inations and vegetation indices for discrimination of six grassland types in eastern Kansas. Int. J. Remote Sens. 23:5031 5042. Raich, J.W. and C.S. Potter. 1995. Global patterns of carbon dioxide emissions from soils. Glob al Biogeochem Cy 1:23 36. Reddy, K.R. and R.D. DeLaune. 2008. Biogeochemistry of w etlands: science and a pplications, CRC Press Taylor & Francis Group. Florida. Reddy, K.R., S. Newman, T.Z. Osborne, J.R. White, and H.C. Fitz. 2011. Phosphorus cycling in the greater Everglades ecosys tem: Legacy phosphorus implications for management and restoration. Crit Rev Env Sci Tec 41:149 186. Reddy, K.R., W.F. DeBusk, R.D. DeLaune, and M.S. Koch. 1993. Long term nutrient accumulation rates in the Everglades. Soil Sci. Soc. Am. J. 57:1147 11 55.

PAGE 217

217 Reddy, K.R., Y. Wang, W.F. DeBusk, M.M. Fisher, and S. Newman. 1998. Forms of soil phosphorus in selected hydrologic units of the Florida Everglades. Soil Sci. Soc. Am. J. 62:1134 1147. Reeves III, J.B., R.F. Follett, G.W. McCarty, and J.M. Kimble. 20 06. Can near or midinfrared diffuse reflectance spectroscopy be used to determine soil carbon pool? Commun Soil Sci. Plan 37:2307 2325. Richardson, C.J., S. Qian, C.B. Craft, and R.G. Qualls. 1997. Predictive models for phosphorus retention in wetlands. Wetlands Ecol Manage 4:159 175. Rivero, R.G., S. Grunwald, and G.L. Bruland. 2007a. Incorporation of spectral data into multivariate geostatistical models to map soil phosphorus variability in a Florida Wetland. Geoderma 140:428 443. Rivero, R.G., S. Gr unwald, T.Z. Osborne, K.R. Reddy, and S. Newman. 2007b. Characterization of the spatial distribution of soil properties in Water Conservation Area 2A, Everglades, Florida. Soil Sci J. 172:149 166. Rivero, R.G., S. Grunwald, M.W. Binford, and T.Z. Osborne. 2009. Integrating spectral indices into prediction models of soil phosphorus in a subtropical wetland. Remote Sens. Environ. 113:2389 2403. Rock, B.N., J.E. Vogelmann, D.L. Williams, A.F. Vogelmann, and T. Hoshizaki. 1986. Remote detection of forest damage. Bioscience 36:439 445. Rossi, J., A. Govaerts, B. De Vos, B. Verhist, A. Vervoort, J. Poesen, B. Muys, and J. Deckers. 2009. Spatial structures of soil organic carbon in tropical forests A case study of Southeasten Tanzania. Catena 77:19 27. Roulet, N.T. 2000. Peatlands, carbon storage, greenhouse gases, and the Kyoto protocol: prospects and significance for Canada. Wetlands 20:605 615. Rouse, J.W., R.H. Haas, J.A. Schell, and D.W. Deering. 1974. Monitoring vegetation systems in the Great Plains with ERTS. Third ERTS symposium, NASA, 1:309 317. Rutchey, K., and L. Vilchek. 1999. Air photointerpretation and satellite imagery analysis techniques for mapping cattail coverage in a northern Everglades impoundment. Photogram m Eng. Remote Sens. 65:185 191. Saito, H. and P. Goovaerts. 2002. Accounting for measurement error in uncertainty modeling and decision making using indicator kriging and pfield simulation: application to a dioxin contaminated site. Environmetrics 13:555 567. Schneider, D.C. 2001. The rise of the concept of scale in Ecology. BioScience 51:545 553.

PAGE 218

218 Schmid, T., M. Koch, M. Di Blasi, and M. Hagos. 2008. Spatial and spectral analysis of soil surface properties for an archaeological area in Aksum, Ethiopia, applying high and medium resoluti on data. Catena 75:93 101. Schmidt, M.W.I., M.S. Torn, S. Abiven, T. Dittmar, G. Guggenberger, I.A. Janssens, M. Kleber, I. Kgel Knabner, J. Lehmann, D.A.C. Manning, P. Nannipieri, D.P. Rasse, S. Weiner, and S.E. Trumbore. 2011. Persistence of soil organic matter as an ecosystem property. Nature 478:49 56. Schuur, E.A.G., J. Bockheim, J.G. Canadell, E. Euskirchen, C.B. Field, S.V. Goryachkin, S. Hagemann, P. Kuhry, P.M. Lafleur, H. Lee, G. Mazhitova, F.E. Nelson, A. Rinke, V.E. Romanovsky, N. Shiklom anov, C. Tarnocai, S. Venevsky, J.G. Vogel, and S.A. Zimov. 2008. Vulnerability of p ermafrost c arbon to climate change: Implications for the g lobal c arbon cycle. BioScience 58:701 714. Selige, T., J. Bhner, and U. Schmidhalter. 2006. High resolution topsoil mapping using hyperspectral image and field data in multivariate regression modeling procedures. Geoderma 136:235 244. Seni, G., and J.F. Elder. 2010. Ensemble methods in data mining: improving accuracy through combining predictions. Synthesis lectures on data mining and knowledge discovery 2:1 126. Shannon, C.E. 2001. A mathematical theory of communication. SIGMOBILE Mobile Computing and Communications Rev 5:3 55. Shepherd, K.D., and M.G. Walsh. 2002. Development of reflectance spectral libraries for c haracterization of soil properties. Soil Sci. Soc. Am. J. 66:988 998. Simbahan, G.C., A. Dobermann, P. Goovaerts, J. Ping, and M.L. Haddix. 2006. Fineresolution mapping of soil organic carbon based on multivariate secondary data. Geoderma 132:471 489. Soi l Survey Staff. 2004. Soil survey field and laboratory methods manual. Soil survey investigations report No.51. R.Burt (ed.). United States Department of Agriculture, Natural Resources Conservation Service. Lincoln, Nebraska. Soil Survey Staff, Natural Res ources Conservation Service, United States Department of Agriculture. 2010. Official Soil Series Descriptions. Available at http://soils.usda.gov/technical/classification/osd/index.html. ( verified 6 Apr., 2010). USDA NRCS, Urbandale, IA. Sorrell, B.K., and P.I. Boon. 2006. Biogeochemistry of billabong sediments. II. Seasonal variations in methane production. Freshwater Biol 27:435 445. South Florida Water Management District. DBHY DRO environmental data. Available at http://www.sfwmd.gov/portal/page/portal/xweb%20environmental%20monitoring/d bhydro%20application (v erified 12 June, 2012).

PAGE 219

219 Spruill, T.B., W.J. Showers, and S.S. Howe. 2002. Application of classificationtree methods to identify nitrate sources in ground water. J. Environ. Qual. 31:1538 1549. Stevens, A., B. van Wesemael, G. Vandenschrick, S. Tour, and B. Tychon. 2006. Detenction of carbon stock change in agricultural soils using spectroscopic techniques. Soil Sci. Soc. Am. J. 70:844 850. Strobl, C., J. Malley, and G. Tutz. 2009. An introduction to recursive partitioning: Rationale, application, and characteristic s of classification and regression trees, bagging, and random forests. Psychol. Methods 14:323 348. Sullivan, D.G., J.N. Shaw, and D. Rickman. 2005. IKONOS imagery to estimate surface soil property variability in two Alabama physiographies. Soil Sci. Soc. Am. J. 69:1789 1798. Sumfleth, K., and R. Duttmann. 2008. Prediction of soil property distribution in paddy soil landscapes using terrain data and satellite information as indicators. Ecol Indic 8:485 501. Thompson, J.A., E.M. Pena Yewtukhiw, and J.H. Gr ove. 2006. Soil landscape modeling across a physiographic region: topographic patterns and model transportability. Geoderma 133:57 70. Thompson, J.A., and R.K. Kolka. 2005. Soil carbon storage estimation in a forested watershed using quantitative soil landscape modeling. Soil Sci. Soc. Am. J. 69:1086 1093. Thrush, S.F., D.C. Schneider, P. Legendre, R.B. Whitlatch, P.K. Dayton, J.E. Hweitt, A.H. Hines, V.J. Cummings, S.M. Lawrie, J. Grant, R.D. Pridmore, S.J. Thuner, and B.H. McArdle. 1997. Scaling up from experiments to complex ecological systems: where to next? J Exp Mar. Biol. Ecol 216:243 254. Trangmar, B.B., R.S. Yost, M.K. Wade, G. Hehara, and M. Sudjadi. 1987. Spatial v ariation of soil p roperties and r ice y ield on r ecently cleared l and. Soil Sci. So c. Am. J. 51:668 674. Tucker, C.J. 1979. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens Environ. 8:127 150. United States Geological Survey. 1999. Digital Data Series DDS 9: Terrestrial Gamma Radioactivity. Available at http://energy.cr.usgs.gov/radon/DDS 9.html ( verified 3 May, 2011). USGS, Reston, VA. United States Geological Survey. 2007. High Accuracy Elevation Data Water Conservation Areas and Great er Everglades Region. Available at http://sofia.usgs.gov/projects/index.php?project_url=elev_data ( verified 7 June, 2010). USGS, Reston, VA.

PAGE 220

220 Updegraff, K., J. Pastor, S.D. Bridgha m, and C.A. Johnston. 1995. Environmental and substrate controls over carbon and nitrogen mineralization in northern wetlands. Ecol Appl 5:151 163. Urban, D.L. 2005. Modeling ecological processes across scales. Ecology 86:1996 2006. Vaithiyanathan, P., and C.J. Richardson. 1997. Nutrient profiles in the Everglades: examination along the eutrophication gradient. Sci. Total Environ. 205:81 95. Valk, A.G., and B.G. Warner. 2009. The development of patterned mosaic landscapes: an overview. Plant Ecol. 200:1 7. Vasques, G.M., S. Grunwald, and J.O. Sickman, 2008. Comparison of multivariate methods for inferential modeling of soil carbon using visible/near infrared spectra. Geoderma 146:14 25. Vasques, G.M., S. Grunwald, and J.O. Sickman. 2009. Modeling of soil organic carbon fractions using visiblenear infrared spectroscopy. Soil Sci. Soc. Am. J. 73:176 184. Vasques, G.M., S. Grunwald, N.B. Comerford, and J.O. Sickman. 2010. Regional modeling of soil carbon at multiple depths within a subtropical watershed. Geoderma 156:326 336. Vasques, G.M., S. Grunwald, and D.B. Myers. 2012a. Associations between soil carbon and ecological landscape variables at escalating spatial s cales in Florida, USA. Landscape Ecol. 27:355 367. Vasques, G.M., S. Grunwald, and D.B. Myers. 2012b. Influence of the spatial extent and resolution of input data on soil carbon models in Florida, USA. J Geophys Res In press Viscarra Rossel, R.A., D.J .J. Walvoort, A.B. McBratney, L.J. Janik, and J.O. Skjemstad. 2006. Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties. Geoderma 131:59 75. Vrieling, A., S.M.de Jong, G. Sterk, and S.C. Rodrigues. 2008. Timing of erosion and satellite data: A multi resolution approach to soil erosion risk mapping. Int. J. Appl. Earth Obs. 10: 2 67 281. Wang, H., C.A.S. Hall, J.D. Cormell, and M.H.P. Hall. 2002. Spatial dependence and the rel ationship of soil organic carbon and soil moisture in the Luquillo Experimental Forest, Puerto Rico. Landscape Ecol 17:671 684.

PAGE 221

221 Wang, X., D.O. Fuller, L.S.L.O. Sternberg, and F. Miralles Wilhelm. 2011. Foliar nutrient and water content in subtropical tree islands: A new chemohydrodynamic link Environ. 115:923 930. Wang, Y., X. Zhang, and C. Huang. 2009. Spatial variability of soil total nitrogen and soil total phosphorus under differ ent land uses in a small watershed on the Loess Plateau, China. Geoderma 150:141 149. Wantzen, K.M., E.G. Couto, E.E. Mund, R.S.S. Amorim, A. Siqueira, K. Tielbrger, and M. Seifan. 2012. Soil carbon stocks in stream vally ecosystems in the Brazilian Cerra do agroscape. Agr Ecosys. Environ. 151:70 79. Webster, R. and M.A. Oliver. 2007. Geostatistics for Environmental Scientist (Second ed.). John Wiley & Sons Ltd. England. Wei, S., A. McBratney, J. Hempel, B. Minasny, B. Malone, T. DAvello, L. Burras, and J Thompson. 2010. Digital harmonisation of adjacent analogue soil survey areas 4 Iowa counties. p.17. In R.J. Gilkes and N. Prakongkep (eds.) Proc. World congress of s oil science: Soil solutions for a changing world, 19th, Brisbane, Australia. 16 Augus t 2010. IUSS. Wei, Y., Y. Bai, J. Jin, F. Zhang, L. Zhang, and X. Liu. 2009. Spatial variability of soil chemical properties in the reclaiming marine foreland to Yellow Sea of China. Agr Sci China 8:1103 1111. West, T.O., C.C. Brandt, B.S. Wilson, C.M. H ellwinckel, D.D. Tyler, G. Marland, D.G. De La Torre Ugarte, J.A. Larson, and R.G. Nelson. 2008. Estimating r egional changes in soil carbon with h igh spatial r esolution. Soil Sci. Soc. Am. J. 72:285 294. Whiting, G.J., and J.P. Chanton. 2001. Greenhouse carbon balance of wetlands: methane emission versus carbon sequestration. Tellus 53B:521 528. Wiesmeier, M., F. Barthold, B. Blank, and I. Kgel Knabner. 2011. Digital mapping of soil organic matter stocks using Random Forest modeling in a semi arid steppe e cosystem. Plant Soil 340:7 24. Wright, A.L. and K.R. Reddy. 2007. Substrateinduced respiration for phosphorus enriched and oligotrophic peat soils in an Everglades wetland. Soil Sci. Soc. Am. J. 71:1579 1583. Wu, Y., F. H. Sklar, and K. Rutchey. 1997. Ana lysis and simulations of fragmentation patterns in the Everglades. Ecol. Appl. 7:268 276. Zhao, B., Y. Yan, H. Guo, M. He, Y. Gu, and B. Li. 2009. Monitoring rapid vegetation succession in estuarine wetland using time series MODIS based indicators: An appl ication in the Yangtze River Delta area. Ecol. Indic. 9:346 356.

PAGE 222

222 Zornoza, R., G. Guerrero, J. Mataix Solera, K.M. Scow, V. Arcenegui, and J. Mataix Beneyto. 2008. Near infrared spectroscopy for determination of various physical, chemical and biochemical pr operties in Mediterranean soils. Soil Biol Biochem 40:1923 1930.

PAGE 223

223 BIOGRAPHICAL SKETCH Jongsung Kim was born and raised in Daejeon, South Korea. She earned a Master of Engineering degree in 2006 from Chungnam National Universiy in South Korea. She worked for Korea Ministery of Environmental Department She got an opportunity to study at the Muroran Institute of Technology in Japan as a government research student funded by Japan Ministry of Education, Culture, Sports, Science and Technology unti l 2008. She married Jungwoo Lee, who was studying at the Civil and Coastal Engineering, University of Florida in 2008 and came to Gainesville, Florida. She started the Ph.D. program at the Soil and Water Science Department, University of Florida. Jongsung Kim and Jungwoo Lee have a son, Jaden Kim Lee.