Integrating Geographic Information Technologies for Land Change Analysis and Modeling in an Urban Area Ting Liu

Florida State University Libraries

Electronic Theses, Treatises and Dissertations The Graduate School

2014 Integrating Geographic Information Technologies for Land Change Analysis and Modeling in an Urban Area Ting Liu

Follow this and additional works at the FSU Digital Library. For more information, please contact [email protected] FLORIDA STATE UNIVERSITY

COLLEGE OF SOCIAL SCIENCES AND PUBLIC POLICY

INTEGRATING GEOGRAPHIC INFORMATION TECHNOLOGIES FOR

LAND CHANGE ANALYSIS AND MODELING IN AN URBAN AREA

TING LIU

A Dissertation submitted to the Department of Geography in partial fulfillment of the requirements for the degree of Doctor of Philosophy

Degree Awarded: Summer Semester, 2014 Ting Liu defended this dissertation on July 1, 2014. The members of the supervisory committee were:

Xiaojun Yang Professor Directing Dissertation

Timothy S. Chapin University Representative

Joseph Pierce Committee Member

J. Anthony Stallins Committee Member

Tingting Zhao Committee Member

The Graduate School has verified and approved the above-named committee members, and certifies that the dissertation has been approved in accordance with university requirements.

ACKNOWLEDGMENTS

Many thanks to my committee, Dr. Xiaojun Yang, Dr. Joseph Pierce, Dr. Tony Stallins, Dr. Tingting Zhao, and Dr. Timothy Chapin for their advising and support on my dissertation research. The Department and University gave me the opportunity to pursue my Ph.D. degree and supported my studies and research. This impact on my academic career is too great to be expressed in words.

iii

TABLE OF CONTENTS

List of Tables ...... vii List of Figures ...... ix Abstract ...... xi CHAPTER ONE - INTRODUCTION ...... 1 1.1 Background ...... 1 1.1.1 Land change and the emergence of land change science ...... 1 1.1.2 Geographic information technologies and land change science ...... 2 1.2 Objectives ...... 3 1.3 Study area...... 4 1.4 Organization of the dissertation ...... 6 CHAPTER TWO - MAPPING LAND USE/COVER IN AN URBAN AREA WITH STRATIFIED CLASSIFICATION AND MULTIPLE ENDMEMBER SPECTRAL MIXTURE ANALYSIS ...... 8 2.1 Introduction ...... 8 2.2 Research methods ...... 11 2.2.1 Data collection and preprocessing ...... 11 2.2.2 Land classification scheme ...... 14 2.2.3 Landscape partition ...... 14 2.2.4 Multiple endmember spectral mixture analysis (MESMA) ...... 18 2.2.5 Supervised classification ...... 21 2.2.6 Thematic accuracy assessment ...... 22 2.3 Results and discussion ...... 23 2.4 Conclusions ...... 32 CHAPTER THREE - MONITORING LAND CHANGES IN AN URBAN AREA USING SATELLITE IMAGERY, GIS AND LANDSCAPE METRICS ...... 34 3.1 Introduction ...... 34 3.2 Research methods ...... 37 3.2.1 Data collection and preprocessing ...... 37 3.2.2 Image processing ...... 40 3.2.3 Urban land change analysis ...... 44 3.3 Results and discussion ...... 51 3.4 Conclusions ...... 56

CHAPTER FOUR - A SCALE-DEPENDENT ANALYSIS OF THE FACTORS DRIVING URBAN LAND USE CHANGES ...... 58 4.1 Introduction ...... 58 4.2 Research methods ...... 61 4.2.1 Data preparation ...... 61 4.2.2 Statistical analysis ...... 71 4.3 Results of driving force analysis ...... 73 4.3.1 Residential land use ...... 73 4.3.2 Commercial/industrial land use ...... 74 4.4 Discussion ...... 75 4.4.1 Effects of aggregation levels ...... 75 4.4.2 Effects of spatial extents ...... 77 4.5 Conclusions ...... 79 CHAPTER FIVE - LAND CHANGE MODELING: STATUS AND PROSPECT ...... 81 5.1 Introduction ...... 81 5.2 Land change modeling approaches ...... 84 5.2.1 Statistical regression model ...... 84 5.2.2 Artificial neural networks ...... 85 5.2.3 Markov chain modeling ...... 86 5.2.4 Cellular automata ...... 86 5.2.5 Economic models ...... 87 5.2.6 Agent-based models ...... 88 5.3 Major issues in land change modeling ...... 89 5.3.1 Coupling human-environment systems...... 89 5.3.2 Scale dependency and multilevel interactions ...... 90 5.3.3 Temporal dynamics and complexity ...... 91 5.4 Integrated land change modeling for global environmental changes ...... 91 5.5 Conclusions ...... 93 CHAPTER SIX - SIMULATING RESIDENTIAL LAND USE CHANGE THROUGH AN AGENT-BASED MODELING APPROACH ...... 95 6.1 Introduction ...... 95 6.2 Research methods ...... 98 6.2.1 Model conceptualization ...... 98

6.2.2 Model implementation ...... 103 6.3 Results and discussion ...... 109 6.4 Conclusions ...... 112 CHAPTER SEVEN - SUMMARY AND CONCLUSIONS ...... 113 7.1 Land use and land cover mapping ...... 113 7.2 Land change analysis ...... 114 7.3 Driving forces of land change ...... 114 7.4 Land change modeling ...... 115 7.5 Future studies ...... 115 REFERENCES ...... 117 BIOGRAPHICAL SKETCH ...... 135

LIST OF TABLES

Table 2.1. Land cover classification scheme, training sample size, number of training clusters and validation data size ...... 15

Table 2.2. The error matrix for the land cover map produced using stratified classification and multiple endmember spectral mixture analysis (MESMA) ...... 25

Table 2.3. The error matrix for the land cover map produced using maximum likelihood classifier (MLC) on each of the urban and rural subsets ...... 26

Table 2.4. The error matrix for the land cover map produced using maximum likelihood classifier (MLC) on the entire image ...... 27

Table 2.5. Comparison of the classification accuracies by different approaches ...... 28

Table 2.6. Comparison of the classification results by different approaches ...... 28

Table 3.1. List of the Landsat Thematic Mapper (TM) scenes used ...... 38

Table 3.2. Thematic accuracy assessment for the 2000 land use/cover map produced from Landsat Thematic Mapper (TM) data ...... 45

Table 3.3. Thematic accuracy assessment for the 2010 land use/cover map produced from Landsat TM data ...... 46

Table 3.4. Land use/cover changes between 2000 and 2010 for the 29 counties under the Atlanta Regional Commission (ARC) ...... 47

Table 3.5. Selected landscape metrics for the three urban land classes in 2000 and 2010 ...... 50

Table 3.6. Land use/cover conversion between 2000 and 2010 ...... 51

Table 4.1. Description of land use types and candidate explanatory factors used in the analysis. 63

Table 4.2. A sample table prepared for residential land use in 2000 at the county level, 20-county region ...... 64

Table 4.3. A sample table prepared for residential land use in 2010 at the county level, 20-county region ...... 65

Table 4.4. A sample table prepared for residential land use changes between 2000 and 2010 at the county level, 20-county region ...... 66

Table 4.5. Results of the stepwise regression models explaining the proportion of residential land use at different aggregation levels for the 20-county region in 2010 ...... 77

vii

Table 4.6. Results of the stepwise regression models explaining the proportion of residential land use over different spatial extents at tract level (n=948) in 2010 ...... 78

Table 6.1. Agent preferences (weights) experimented for calculating land utility for resident agents using Equation 6.1 ...... 106

Table 6.2. Model validation using Kappa coefficient based on cell-by-cell comparison of the simulated and the actual resident development in 2010 ...... 109

viii

LIST OF FIGURES

Figure 1.1. The conceptual framework of the dissertation research ...... 3

Figure 1.2. Location of the study area ...... 5

Figure 2.1. Flowchart of the working procedural route that includes three major components: landscape partition, sub-pixel analysis, and supervised classification ...... 12

Figure 2.2. GPS-based field survey routes with a total length of 850 km, including the two trips in December 2010 and May 2011 ...... 13

Figure 2.3. The urban subset in false color display (with the near infrared, red, and green channel as the red, green, and blue guns, respectively) clipped using the urban mask defined with a road- density analysis ...... 17

Figure 2.4. Spectral reflectance of endmenbers used for the multiple endmember spectral mixture analysis (MESMA) ...... 19

Figure 2.5. Land cover map produced from Landsat Thematic Mapper imagery by using stratified classification and multiple endmember spectral mixture analysis (MESMA) techniques ...... 23

Figure 2.6. Comparison of the land cover maps produced by different approaches ...... 29

Figure 2.7. Comparison between the per-pixel analysis and the sub-pixel analysis...... 30

Figure 3.1. Flowchart of the working procedural route ...... 37

Figure 3.2. Land use/cover maps for 2000 and 2010, which were derived from Landsat Thematic Mapper (TM) imagery ...... 47

Figure 3.3. Spatial growth of the three urban land classes during the period of 2000 and 2010: (a) Residential land; (b) Commercial/industrial land; and (c) Other urban land...... 48

Figure 3.4. Land use/cover statistics for the three geographic areas in the Atlanta metropolitan area (the 10 counties, the 20 counties, and the 29 counties): (a) Land use/cover classes in 2000; (b) Land use/cover classes in 2010; (c) Land use/cover changes during the period of 2000- 2010...... 49

Figure 3.5. Land use/cover conversion for the 29 counties in the Atlanta metropolitan area during the period 2000-2010 ...... 52

Figure 4.1. Three dimensions of forces driving urban land use change (modified from Turner et al. 1995) ...... 59

Figure 4.2. A flowchart of the working procedure route ...... 62

Figure 4.3. Dasymetric map of population density by block groups 2000 ...... 68

Figure 4.4. The elevation and slope layers to be associated with the distribution of land uses for the extraction of topographic measures at each set of areal units ...... 69

Figure 4.5. The layers of distance surface to be associated with the distribution of land uses for the extraction of location measures at each set of areal units ...... 70

Figure 4.6. Correlation coefficient (in absolute value) between commercial land use proportions and five explanatory variables at three aggregation levels for the 20-county region for 2010 .....76

Figure 5.1. Number of publications on land change modeling between 1994 and 2013 based on the search criteria ...... 84

Figure 6.1. A conceptual model to simulate residential development proposed in this dissertation research ...... 98

Figure 6.2. Spatial patterns of residential land use growth in Gwinnett County, 2000-2010...... 104

Figure 6.3. The composite spatial criteria for calculating the land utility for resident agents (Equation 6.1): (a) factor of home sales price; (b) land accessibility (combined proximity to major roads and urban centers); and (c) land attractiveness (combined proximity to parks and water bodies)...... 106

Figure 6.4. Development probability surface for developer agents calculated based on Equation 6.4...... 107

Figure 6.5. The land suitability surface representing the land regulation for government agents produced based on Equation 6.5 ...... 108

Figure 6.6. Simulated residential land use growth patterns in Gwinnett County, Georgia for 2010 using the proposed agent-based model ...... 110

ABSTRACT

Land changes are complex and dynamic processes that involve the human and natural systems interacting over space and time to reshape the earth‟s surface. As a fundamental form of global environmental changes, land changes also hold wide-ranging significance for the functioning of the earth‟s ecosystem and the human society. However, understanding land change dynamics remains a major challenge for global environmental change and sustainability research. The primary objective of this dissertation research is to investigate the feasibility and applicability of integrating various geographic information technologies to improve the understanding of land change dynamics in a complex urban environment. Specifically, the following dimensions of land change science are examined: land change observation and monitoring, driving force analysis, and spatially-explicit modeling. Firstly, a stratified classification approach combined with sub-pixel analysis is developed to map various land use and land cover types in the heterogeneous urban area from medium-resolution satellite imagery. Secondly, remote sensing, GIS and landscape metrics are used in combination to characterize both the spatial characteristics and the nature of urban land changes. Thirdly, a multi-scale analysis is performed to explore the biophysical and socioeconomic driving factors of urban land use change at different spatial aggregation levels and across different spatial extents. Fourthly, given a wide array of existing land change modeling approaches, the theoretical and methodological foundation of these modeling techniques are reviewed and the outstanding issues are discussed in the context of global environmental change research. Lastly, an agent-based model is developed that is coupled with GIS-based spatial data analysis to simulate the residential development decision-making processes and the emergent land use patterns. Overall, this dissertation research has demonstrated the usefulness of integrating various geographic information technologies, such as remote sensing, GIS, and spatial modeling, in land change research. The technological integration also provides the foundation for the coupling of human and environment sciences in understanding land change as a coupled system. An interdisciplinary effort is needed towards more comprehensive research in land change that integrates theories, methods, and techniques in human, environmental, and geographic information sciences.

CHAPTER ONE

INTRODUCTION

1.1 Background

1.1.1 Land change and the emergence of land change science

Rapid human population growth, along with the increasing demand for food, water, energy, and other benefits, has substantially altered the earth system and its capacity to sustain life, especially since the second half of the twentieth century (Turner et al. 2007; U.S. Global Change Research Program 2014). Land use and land cover change is one of the well-documented global environmental changes, which involves human and natural systems interacting over space and time to reshape the earth‟s surface. To date, as much as 50% of the earth‟s ice-free land surface has been transformed or degraded (Vitousek 1994; Haberl et al. 2007). Only between 2000 and 2010, approximately 13 million hectares of land area (about the area of Greece) were converted each year to other land cover types (FAO 2010). Moreover, land changes have far- reaching influences on the structure and function of the earth‟s ecosystems, with equally significant implications for the human society (Steffen et al. 2004). On the one hand, land changes affect the ecosystems in several ways, such as reducing native habitat and species, accelerating soil decomposition, disrupting freshwater resources and quality, as well as leading to additional greenhouse gas release (Turner, Moss, and Skole 1993; Camill 2010). For example, deforestation is thought to contribute to nearly 20% of the global carbon dioxide release (1.5-2 billion tons of carbon) (Camill 2010). On the other hand, rapid urbanization and the concentration of human populations into large metropolises have altered the city‟s cultures, politics, and economics, which are just beginning to be fully recognized as a significant global problem. Given the increasing recognition of its global importance, land change science has recently emerged as a fundamental component of global environmental change and sustainability science (Gutman et al. 2004; Rindfuss et al. 2004; Turner et al. 2007). This interdisciplinary field seeks to understand land cover and land use dynamics through integrating the human, environmental, and geographical information-remote sensing sciences. Research in land change science has been dedicated to enhance our understanding of land changes through: (i)

1 observation and monitoring of land change dynamics, (ii) understanding land change as a coupled system in exploring the causes, impacts and consequences, (iii) spatially explicit modeling of land changes, and (iv) assessing system outcomes, such as vulnerability, resilience and sustainability (Lambin and Geist 2006; Turner et al. 2007). Despite the progress made by various research communities, comprehensive understanding of the land change process remains challenging due to the complex interactions between the functioning of ecosystems and the human systems across multiple scales.

1.1.2 Geographic information technologies and land change science

Towards a better understanding of land change dynamics as coupled human- environmental systems, it is necessary to integrate theories, concepts, and methods from multiple disciplines, such as resource economics, institution governance, landscape ecology, and biogeography. Geographic information technologies, including geographic information systems (GIS), remote sensing, spatial modeling and other geospatial techniques, together offer great opportunities for the interdisciplinary integration in land change science. Specifically, geographic information technologies can be used to support research in the major components of land change science. Firstly, the recent progress of observing and monitoring land changes largely relies on the technological and methodological advances in remote sensing. Increasing data availability at finer resolutions and the development of advanced image processing techniques allow the observation of land change across different regions and at different scales. Other geospatial data and technologies may also be incorporated to improve land use/cover mapping and change analysis (Rindfuss et al. 2004). Secondly, geographic information technologies, such as GIS, provide platforms that can help integrate spatial data that characterizing the patterns and processes of the human and biophysical subsystems for analyzing the causes, impacts and consequences of land change. In social-demographic analysis, data are usually collected at some levels of aggregation, while direct measurement and remote sensing techniques have been more commonly used to extract biophysical variables (Jensen 1983). Thirdly, various geospatial modeling techniques have been developed and employed for land change simulation, prediction and informing decision-making. The modeling approaches range from statistical modeling that takes the coupled system as a whole to agent-based models that can represent individual decision-making. Advances in remote sensing-based land change observation and GIS-based spatial data integration and analysis further spur the development of 2 land change models (National Research Council 2013). Finally, the synthesis and assessment of system outcomes are based on more comprehensive understanding of land change as a coupled human-environment system in the first three dimensions. Integrated modeling approaches, typically land change modeling with other ecological models, have also been used to inform practices on land use and resource management.

1.2 Objectives

The goal of this dissertation is to explore the potential of integrating various geographic information technologies to improve the understanding of land changes in an urban area as coupled human-environmental systems. In particular, this dissertation research examines three specific components of land change science, namely, observation and monitoring, understanding the causes, and modeling of land changes in an urban area. A conceptual framework underlying this dissertation research is illustrated in Figure 1.1.

Figure 1.1. The conceptual framework of the dissertation research.

The specific objectives include: 1. To develop an image analysis method that can improve land use and land cover mapping in an urban area from medium-resolution remotely sensed data; 3

2. To characterize the spatial characteristics and the nature of land changes, especially for different urban land uses, in a complex urban environment; 3. To examine the biophysical and socioeconomic factors driving urban land use changes at multiple spatial scales (i.e., resolution and extent); 4. To review recent progresses of land change modeling techniques and related theoretical and methodological issues; and 5. To construct an agent-based model to simulate the decision-making processes and the emergent patterns of residential land use in a suburban area.

1.3 Study area

Atlanta, Georgia is selected as the study area. Atlanta has been a fast-growing large metropolis in the United States over the past four decades as it emerged as the premier commercial, industrial, and transportation center of the southeast. Since the 1970s, Atlanta has experienced a rapid growth in both population and spatial extent. Population had increased 30- 40% each decade between 1970 and 2000. The rampant suburbanization process has significantly expanded the urban area outward on the fringe of cities mainly at the cost of forest and cropland (Yang and Lo 2002). The rapid growth and suburbanization trend in Atlanta is expected to continue to the twenty-first century through at least 2030 (Rao 2007). According to the U.S. Census Bureau, the population in the 28-county Atlanta MSA had increased 24% between 2000 and 2010, which was the second-highest among the nation‟s largest metro areas. Atlanta has also been reported by the Metropolitan Research Center at the University of Utah as the most sprawling large metro area in the United States as of 2010. On the one hand, Atlanta has been recognized as one of the few typical postmodern metropolises in North America which exhibit unique and complex urban forms (Dear and Flusty 1998; Dear 2000; Hall 2000). The suburban sprawl radiates from the urban center along highways in a multinucleated form, with the concentrations of economic activities at the intersections of urban beltways and hub-and- spoke lateral roads that led to the formation of edge cities (Garreau 1991). On the other hand, a range of environmental and societal issues have emerged as the consequences of the far-reaching suburban sprawl in Atlanta, such as deforestation, habitat fragmentation, air quality degradation, environmental and social inequity (Bullard, Johnson, and Torres 2000; Miller 2012). Given its restless growth and complex patterns, Atlanta becomes an ideal site to develop and test the methodology for studying land changes in the complex urban areas. 4

Geographically, the study area consists of 29 counties designed by the Atlanta Regional Commission (ARC) to accommodate various planning purposes (Figure 1.2), including Hall County and the 28-county “Atlanta-Sandy Springs-Marietta” Metropolitan Statistical Area (MSA) as designated by the U.S. Office of Management and Budget (OMB) in 2010.

Figure 1.2. Location of the study area. It covers 29 counties designated by the Atlanta Regional Commission. Note the geographic regions of the Atlanta Regional Commission with 10 counties and 20 counties are also shown. Gwinnett County is highlighted as it is selected to implementing agent-based model (see Chapter 6).

Georgia‟s capital and largest city, City of Atlanta, resides in the center of the region. The study area also includes many smaller cities and emerging suburbs and exurbs. The total area is approximately 23,072 km2, which strides across three Landsat Thematic Mapper (TM) image scenes. Physiographically, the Atlanta metropolitan area is mostly located in the low foothills of the southern Appalachian Mountains in north Georgia. The northwestern portion (approximately 7% of the total area) lies in the Appalachians and tends to be higher in elevation and significantly hillier than the southeast. The average elevation is approximately 300 m above sea level. The relatively high mean elevation has contributed to the temperate climate in the Atlanta area. The Chattahoochee River, a tributary of the Apalachicola River, flows through the study site. The entire metropolis is characterized by a complex mosaic of urban and suburban landscapes, with a combination of diverse land use and land cover types at varying spatial scales. Note that given the consideration of computation efficiency, the agent-based modelling of residential land use change (objective 5/Chapter 6) is implemented on Gwinnett County only (Figure 1.2). It is a suburban county, part of the 10-county core in the Atlanta metropolitan area. Gwinnett County is adjacent to Fulton County where downtown Atlanta is located. There are two major reasons leading to this choice. Firstly, Gwinnett County has led the Atlanta region‟s residential construction since 2000, with more than 15% of all new units built in the entire Atlanta metropolitan region. According to the ARC PLAN 2040 (ARC 2010), growth in Gwinnett County would be one of the principle drivers of the region‟s overall population growth over the next thirty years. Secondly, the diverse physiographic environment across Gwinnett County has made it an ideal site to study the complex interactions between human and nature. Gwinnett County is located in the Piedmont physiographic province, characterized by rolling hills and clay soils underlain by crystalline rock. It is also situated at the headwaters of multiple watersheds, including the Chattahoochee River.

1.4 Organization of the dissertation

This dissertation is organized into seven chapters. Chapter 1 (this chapter) introduces the research background, the study area, and the research objectives. Chapter 2-6 focuses on the first three research dimensions in land change science through integrating various geographic information technologies. Specifically, Chapter 2 discusses an image processing method combined stratified classification with sub-pixel analysis that can improve urban land use and 6 land cover mapping from medium-resolution remotely sensed data (e.g., Landsat Thematic Mapper). Chapter 3 adopts the land use/cover mapping techniques described in Chapter 2 into a post-classification change detection protocol in order to characterize both the spatial characteristics and the nature of land changes through the combined use of satellite imagery, GIS and landscape metrics. Chapter 4 further explores the potential biophysical and socioeconomic factors driving urban land use changes at multiple spatial aggregation levels and over different spatial extents. The effect of scale on the observed relationship is also examined. Chapter 5 reviews various land change modeling approaches and discusses some outstanding theoretical and methodological issues. Examples of integrating land change models with environmental models for integrated global environmental change studies are also included. Chapter 6 describes an agent-based modeling approach to simulate the interactions between the residential development decision-making processes and the emergent land use patterns. Finally, Chapter 7 concludes this dissertation by summarizing the major findings of this research. The theoretical and methodological implications, limitations and future research are also discussed.

CHAPTER TWO

MAPPING LAND USE/COVER IN AN URBAN AREA WITH STRATIFIED CLASSIFICATION AND MULTIPLE ENDMEMBER SPECTRAL MIXTURE ANALYSIS

This chapter discussed a method that can help improve land use/cover mapping in the complex urban areas from medium-resolution remote sensor imagery. Central to the proposed method was the combined use of stratified classification and multiple endmember spectral mixture analysis (MESMA) techniques. Specifically, the entire landscape was firstly partitioned into rural and urban subsets using road network density so that each subset can be processed independently to minimize the spectral confusion between some urban features and agricultural land covers. Secondly, the urban built-up areas and vegetation covers and were carefully extracted using information at the sub-pixel level by using the MESMA technique for the urban subset in order to account for small, fragmented land patches that would be misclassified otherwise. Thirdly, a separate supervised classification protocol was adopted to the rural subset and the vegetation covers extracted from the urban subset. Finally, the classified outcomes from the two subsets were combined to produce a complete map. It was found that this method has substantially outperformed two related ones that use the same supervised protocol to the entire area directly or to the rural subset and the urban subset without being MESMA processed.

2.1 Introduction

Urban area is characterized by a complex mosaic of heterogeneous land use and land cover types, including built-up land for various urban uses, water bodies, agricultural land and all types of spontaneous and cultivated vegetation. Developing methods that can improve land use/cover mapping in urban areas has become increasingly important for both planning and academic communities. On the one hand, the growth of urban population and human demand of land resources asks for operational approaches that can help inventory the amount and distribution of various land use/cover types for urban land use planning and environmental management purposes (Sukopp, Numata, and Huber 1995; Pauleit and Duhme 2000). On the other hand, land use/cover mapping provides a useful source of information for the scientific research in land use/cover change analysis (Yang and Lo 2002; Xiao et al. 2006; Shalaby and

Tateishi 2007), land change modeling (Clarke and Gaydos 1998; Weng 2002), and environmental impacts assessment (Carlson and Arthur 2000; Pauleit, Ennos, and Golding 2005). Land use/cover in urban areas can be inventoried through ground surveys or remote sensing. While ground surveys are often restricted by their logistical constraints, remote sensing makes direct observations across large areas, allowing land use/cover to be mapped in a timely and cost-effective fashion. Both visual interpretation and computer-based digital classification can be used to extract land use/cover information from remote sensor data. Conventional methods are usually based on visual interpretation of color infrared aerial photographs (e.g., Nowak et al. 1996). But the applicability of aerial photographs can be constrained by their limited spectral information and relatively small geographic coverage that may not be efficient for large area mapping (Nichol and Lee 2005). Recent advances in high-resolution satellite remote sensing and object-oriented classification techniques have allowed large-scale land use/cover mapping to be conducted in a relatively effective way (e.g., Nichol and Lee 2005; Tooke et al. 2009; Zhang, Feng, and Jiang 2010; Myint and Stow 2011; Van Delm and Gulinck 2011). However, the high cost of data acquisition can undermine the feasibility of high- resolution satellite imagery for large area mapping (Mumby and Edwards 2002). In addition, high-resolution data could be quite good for ground feature extraction at the local scale but this capability may not always be necessary for some applications that aim to map regional-scale land use/cover patterns in urban areas (Van de Voorde, Vlaeminck, and Canters 2008; Xie, Sha, and Yu 2008). In contrast, medium-resolution data, such as the images acquired by the US Landsat program and French SPOT satellites, have been widely available over a quite long time period, and can be most valuable for large-area mapping (Xie, Sha, and Yu 2008). Nevertheless, mapping land use/cover in urban areas with medium-resolution images can be challenging due to the spectral similarity between various land cover types and the presence of complex urban features (Welch 1982; Forster 1983; Gao and Skillcorn 1998; Herold, Gardner, and Roberts 2003). Over the years, various strategies have been developed to improve urban land use/cover mapping from medium-resolution remote sensor data, which were largely built upon the use of per-pixel based hard classifiers (e.g., Banzhaf, Grescho, and Kindler 2009; Rafiee, Mahiny, and Khorasani 2009; Zhou and Wang 2011). While these efforts have demonstrated a varying level of success, they tended to be more problematic when mapping land use/cover types in the urban

9 environment due to the presence of a large proportion of mixed pixels caused by the high spectral and spatial variability in urban areas, especially in low-density residential areas (Myint 2006; Powell et al. 2007; Van de Voorde, Vlaeminck, and Canters 2008; Franke et al. 2009). To deal with the mixed pixel problem associated with the use of medium to coarse resolution remote sensor data, several strategies have been developed largely through the use of a soft classifier to partition the proportions of each pixel between classes. Examples of these sub-pixel analysis strategies include linear spectral mixture analysis (Smith et al. 1990; Settle and Drake 1993; Small 2001; Rashed et al. 2003; Wu and Murray 2003), fuzzy set possibilities (Eastman 1997), fuzzy c-mean partitioning (Foody 2000), and Bayesian probabilities (Wang 1990; Foody et al. 1992). Among them, spectral mixture analysis (SMA) appears to be most promising, which has been successfully applied to model the heterogeneous urban land cover composition (e.g., Small 2001; Lu and Weng 2004; Wu 2004; Small and Lu 2006). Spectral mixture analysis (SMA) assumes that each pixel within the image can be modeled as a linear combination of a few spectrally pure land cover components, known as endmembers (Robert et al. 1998). Some researchers have applied the vegetation-impervious surface-soil (V-I-S) model proposed by Ridd (1995) to extract sub-pixel components in urban environments (e.g., Rashed et al. 2003; Lu and Weng 2004; Powell et al. 2007). SMA aims to estimate the sub-pixel fraction of endmembers that best model the recorded spectrum within a pixel (Powell 2011). However, because spectral mixture analysis uses an invariant set of endmembers to map the entire landscape and the spectrum for each endmember is assumed to be constant across the image scene, this technique can be quite limited when mapping heterogeneous urban landscapes (Lu and Weng 2004; Wu 2004; Song 2005). As an extension of the simple SMA technique, multiple endmember spectral mixture analysis (MESMA) allows the number and type of endmembers to vary for each pixel and thus can account for both the spatial and spectral variability of the complex urban landscape (Robert et al. 1998). Nevertheless, most of the existing simple SMA or MESMA applications have restricted on the extraction of land cover composition information rather than thematic cover types (e.g., Robert et al. 1998; Rashed et al. 2003; Powell et al. 2007; Franke et al. 2009; Myint and Okin 2009). The objective of this study was to identify a method for mapping land use/cover in the urban environment from medium-resolution remote sensor data. This method was developed through the combined use of stratified classification and MESMA techniques. The stratified

10 classification strategy was employed to tackle the problem of spectral confusion between some urban features and agricultural land uses that has been considered as a major challenge in land cover mapping (e.g., Seto et al. 2002; Yang and Lo 2002). In this study, the stratified classification was implemented by clipping the entire landscape into rural and urban subsets and then processing each subset independently (Myint and Okin 2009). On the other hand, the MESMA technique was used to extract vegetation covers and urban built-up areas at the sub- pixel level for the urban subset in order to account for many small, fragmented land patches that would be misclassified otherwise. A separate supervised classification protocol was further implemented to classify the rural subset and all vegetation covers extracted from the urban subset through the MESMA technique. And the classified results from the two subsets were combined to produce a complete map. This method was applied to produce a land use/cover map from a medium-resolution image covering a large metropolitan area. For comparison purposes, the performance of the proposed method was assessed with respect to two related methods: one applying the same supervised classification protocol to the entire area directly and the other using a separate supervised protocol to the rural subset and the urban subset without being MESMA processed. The following sections will describe the research methodology, evaluate the performance, and discuss some possible limitations of this work.

2.2 Research methods

The working procedural route is illustrated in Figure 2.1, which includes several major components. The following sections will describe the procedure of data collection and preprocessing, land classification scheme, landscape partition, multiple endmember spectral mixture analysis (MESMA), supervised classification, and thematic accuracy assessment.

2.2.1 Data collection and preprocessing

The primary data used in this study were Landsat Thematic Mapper (TM) imagery. Image acquisition date is important for land use/cover mapping due to the influence of phenological characteristics of vegetation. A limitation of using Landsat data for large-area mapping is that there is a 7-day interval between the acquisition dates of two adjacent paths. Another issue related to imagery acquisition is the availability of cloud-free data. Given these considerations, three TM scenes were acquired from the USGS EROS Data Center, which can cover the entire test site. Specifically, they include two scenes of Path 19 (Row 36 and 37)

11 acquired on 19 May 2007 and one Path 18 (Row 37) scene acquired on 16 August 2007. Note that the path 18 scene only covers a very small portion of the study site. And all the six visible and reflected infrared bands were actually used but the thermal band was excluded due to its coarse spatial resolution.

Figure 2.1. Flowchart of the working procedural route that includes three major components: landscape partition, sub-pixel analysis, and supervised classification. Note: NDVI = normalized difference vegetation index, MESMA = multiple endmember spectral mixture analysis, and MLC = maximum likelihood classifier.

In addition to primary data, two ancillary data sets were collected to assist the remote sensing work. These include a street centerline dataset for use in creating the urban mask and a set of high-resolution imagery for assisting the thematic accuracy assessment. The street centerline dataset contains all paved roads across local, state, and interstate levels, which was

12 produced in 2007 through a joint effort by Georgia Department of Transportation (GDOT), Carl Vinson Institute's Office of Information Technology Outreach Services (ITOS), Atlanta Regional Commission (ARC), and the ARC county governments. The high resolution imagery dataset includes orthoimagery with 1m spatial resolution that was collected from USGS EROS Data Center. This dataset was acquired in April 2008. Finally, two GPS-guided field surveys were conducted in December 2010 and May 2011. The field itineraries went across the entire Atlanta metropolitan area (Figure 2.2). The routes were designed based on the author‟s knowledge about the study site and a preliminary examination of the satellite images. During the field trips, GPS points and observed various land cover types were recorded. The field data were used in combination with Quickbird images in Google Earth™ (http://earth.google.com) for designing a land cover classification scheme and for developing training and validation data that will be discussed in the next several sections.

Figure 2.2. GPS-based field survey routes with a total length of 850 km, including the two trips in December 2010 and May 2011. The boundaries of the 29 counties are shown. 13

Several image preprocessing procedures were conducted, namely, atmospheric correction, mosaicking, georeferencing, and subsetting. Firstly, to account for possible atmospheric effects, the digital number of the three TM images were converted into surface reflectance using the Chavez (1996) COST model. Then, these reflectance images were mosaicked and georeferenced to the UTM (Universal Transverse Mercator) map projection (Zone 16), NAD83 datum and GRS1980 spheroid. To address the differences in spectral reflectance between the two adjacent Landsat scenes (Path18/Row37 and Path19/Row37), the overlaying area of the two TM scenes that covers less than 2% of the entire study site were spectrally matched. For georeferencing, fifteen ground control points were selected and nearest- neighbor resampling was applied with a first-degree polynomial fit. The average root mean squared error (RMSE) was 0.094 pixels. Finally, the boundary of the 29-county Atlanta region was used to subset the image to fit the test site. Note that in order to account for the possible georeferencing error, a 200m buffer was created immediately surrounding the 29-county polygon for the image subsetting. The street centerline data were also geometrically rectified with reference to the mosaicked image.

2.2.2 Land classification scheme

Based on a preliminary examination of the satellite images and field works, a land cover classification scheme have been designed based on a modification from the Anderson scheme (Anderson et al. 1976). The classification system includes 11 major categories of land cover: evergreen forest, deciduous forest, mixed forest, vegetated wetland, grass, pasture, cropland, fallow/barren land, low density urban, high density urban, and water. The definition for each class is summarized in Table 2.1.

2.2.3 Landscape partition

A preliminary examination of the spectral signals of the major land cover types in the test site indicates that spectral confusion exists between several thematic classes, especially between urban impervious surfaces and agricultural land that was fallowed or at certain growth stages. This spectral confusion problem has been identified in earlier literatures as a major challenge in land cover mapping from remote sensor imagery (e.g., Seto et al. 2002; Yang and Lo 2002). To tackle this problem, a stratified classification strategy was adopted that initially separated the entire landscape into rural and urban subsets and then processed each subset independently.

Table 2.1. Land cover classification scheme, training sample size, number of training clusters and validation data size.

Training size Number of Validation Land cover (pixels) training clusters No Description size class Urban Rural Urban Rural (pixels) subset subset subset subset Evergreen Mostly coniferous forests with leaves in all seasons, including vegetative 1 123 143 1 1 50 forest species such as pine, cedar, and cypress.

Deciduous Forests with a majority of the trees losing their foliage seasonally, 2 131 126 1 1 53 forest including vegetative species such as maple, oak, and birch.

Coniferous and deciduous species, mixed with a variety of shrubs, brushes, 3 Mixed forest 103 217 1 4 65 and young trees.

Vegetated Hardwood, mixed forest and shrubs, distributing along rivers and around 4 104 72 1 1 46 wetland lakes.

A variety of natural and man-made grass land types, mainly found in golf 5 Grass 88 68 4 1 63 courses, parks, and other recreational areas.

Vegetation cover used for grazing of livestock, normally consisting of 6 Pasture 0 264 0 3 45 grasses, forbs, shrubs or their mixture. Various types of crops in rural areas, normally with unique patterns 7 Cropland 0 303 0 5 48 representing different growth stages or different crop types.

Fallow/barren Areas with sparse vegetation covers, mainly found in the rural area and 8 95 110 2 2 49 land urban transitional land. Low density Approximately 20-70% impervious surface, typically residential land and 9 93 54 1 1 41 urban local roads.

Areas with more than 70% impervious surface, including commercial and High density industrial constructions, as well as large transportation facilities such as 10 218 52 4 1 44 urban parking lots, highways, and airports; it may also includes residential area located near the city core area. 11 Water All areas of open water, including lakes, streams, rivers, reservoirs. 121 125 2 2 46

In this way, the best available pattern recognition strategies could be implemented to each subset. This section focuses on the specific procedures for landscape partition while other processing procedures will be discussed in the next several sections. In this study, landscape partition was implemented by using road density analysis modified from Yang and Liu (2005). This method is based on the location rent model that considers urban land use patterns representing the competition for the most accessible location between different land-use types (Kaplan, Wheeler, and Holloway 2008). As a result, residential, commercial, and industrial land uses tend to locate near the urban centers, while agricultural lands are normally distant from the urban centers. And a preliminary examination of the land use/cover patterns from satellite imagery and field surveys have confirmed the validity of this hypothesis in the test site. Specifically, the entire landscape was partitioned by using an urban mask that was created through a three-step procedure. Firstly, a road intersection density surface was generated using road intersections extracted from the ARC 2007 street centerline data. Each cell in the continuous density surface has a value, which is influenced by the assigned search radius value. A larger search radius would result in a more continuous surface while a smaller one could lead to a more discrete surface. The morphological evolution of the Atlanta metropolitan area has shown some typical patterns of a postmodern city that is characterized by a major urban center being surrounded by some emerging suburban centers (Yang 2002). In order to better delineate the urban built-up area, two road density surfaces were generated using search radius values of 2000 and 14,000 ft, respectively. The larger search radius was used to single out the major urban center from the surrounding area, while the smaller search radius was used to separate the smaller suburban centers from the rural land covers. The two search radius values were determined iteratively to better delineate the spatial distribution of urban and rural land covers in the study site. Secondly, a threshold value was used to separate the urban area from the rural part for each road intersection density surface. The threshold value (in intersections per square kilometer) was determined iteratively through visual interpretation of each surface. A value of 30 was used as the threshold for the surface generated with the smaller search radius and 40 was used for the other surface. The urban area generated from the large radius surface includes the major urban center and only a portion of the surrounding suburban centers that are smaller than their actual

16 size. On the other hand, the urban area generated from the smaller radius surface covers most of the smaller suburban centers but only part of the major urban center. Therefore, the urban area derived from each surface was unionized to single out both the urban center and the surrounding suburban centers. Lastly, a road buffer layer was created with distance values varying by road levels. The purpose of doing so was to account for some possible spatial errors within the original road dataset so that most built-up area can be singled out from rural features by using the mask. Specifically, a 328 ft (100m) buffer was created along the state highways and a 164 ft (50m) buffer for roads at other levels. In this way, an initial urban mask was created by combining the delineated urban area with the road buffer layer. Some large urban features located at the areas with relatively low road density (e.g., airport) were not well delineated by the initial mask, which were then manually digitized and appended to the initial mask to generate the final urban mask. Any area outside the mask was defined as the rural subset. Figure 2.3 illustrates the urban subset image extracted by using the urban mask.

2.2.4 Multiple endmember spectral mixture analysis (MESMA)

As discussed before, the mixed pixel problem associated with the use of medium resolution data (such as Landsat TM images) has been a major challenge for land cover mapping in urban areas. To deal with this problem, the multiple endmember spectral mixture analysis (MESMA) was implemented to the urban subset. The goal was to ensure the large amount of isolated land patches in the urban subset to be correctly mapped. Note that the MESMA was chosen over the simple SMA (Smith et al. 1990; Settle and Drake 1993; Small 2001; Rashed et al. 2003; Wu and Murray 2003) here because the former allows the number and type of endmembers to vary for each pixel (Roberts et al. 1998; Dennison et al. 2003; Dennison and Roberts 2003a, 2003b; Dennison, Halligan, and Roberts 2004; Powell et al. 2007; Powell and Roberts 2010) and thus can account for both the spatial and spectral variability of the complex urban landscape. The following sections will document the specific procedures for MESMA implementation. 2.2.4.1. Endmember selection. Building an endmember library is critical for estimating land cover fractions. The Vegetation-Impervious Surface-Soil (V-I-S) model (Ridd 1995) was applied for endmember selection to represent the major land cover components in the urban environment. In addition, shade is commonly present in most urban areas and thus is included as an additional endmember (Smith et al. 1990; Dennison, Halligan, and Roberts 2004). For a reflectance image, the shade endmember should have a zero value in all bands (Powell 2011). Therefore, the four-endmember model was used to estimate the sub-pixel V-I-S components for the urban subset. Because reference spectra are not available, image endmembers were collected. The goal was to select spectrally “pure” pixels which can account for the variability of spectra for each land cover type in the V-I-S model. Firstly, the pixel purity index (PPI) (Boardman, Kruse, and Green 1995) was calculated for the urban subset image to help select the candidate image endmembers. And pixels with high PPI scores were selected, which represent the purest pixels within the image. In selecting candidate endmembers, the PPI displayed in an N-dimensional visualizer was dynamically linked with the original urban subset and a spectral profile viewer. The selected endmembers were grouped into different land cover types in the V-I-S model according to their spectral signatures at all bands. The most representative endmember for each subclass was identified using count-based endmember selection (CoB) (Roberts et al. 2003) and

18 endmember average root mean square error (EAR) (Dennison and Roberts 2003a). The two methods were implemented simultaneously to account for their sensitivity to different criteria (Powell β011). The “optimal” set of endmembers for each V-I-S component was selected iteratively by adding low in_CoB, high out_CoB, or low-EAR endmembers to the library and assessing the model performance based on root mean square (RMS) error images and visual comparison with high resolution orthoimagery. 2.2.4.2. Implementation of MESMA. With the endmember library, all the possible combinations of endmember spectra can be considered in order to build the candidate SMA models. For the three non-shade endmembers, the impervious surface category included 17 distinct endmember spectra, the vegetation category contained four spectra, and the soil category was divided into bright and dark soil classes (Figure 2.4). Therefore, a total of 136 (17×4×2) SMA models were used to model each pixel within the urban subset.

Figure 2.4. Spectral reflectance of endmenbers used for the multiple endmember spectral mixture analysis (MESMA). In total, 23 endmembers were selected, including 4 vegetation, 17 impervious surfaces, and 2 soil. The six bands on the x-axis correspond to Landsat TM bands 1-5, and 7.

The model constraints were determined through iterative experimentations. The decision was based on a visual inspection of the distribution of land cover fractions. For example, only impervious surface fractions were expected in built-up areas. Therefore, the constraints were selected to minimize the presence of impervious surface in other land-cover patches such as forest, pasture, and cropland. Specifically, the non-shade fractions were constrained between - 0.05 and 1.05; the maximum RMS error allowed was set to 0.025; and the maximum allowable shade fraction was set to 0.5 so that the water bodies can be excluded by making them remain unmodeled (Powell 2011). And the model that met all the constraints and had the lowest RMS error was selected to estimate the V-I-S and shade fractions for each pixel within the urban subset. The output was an image with six “bands” representing the fractions of impervious surface, vegetation, soil, and shade, model number, and RMS error, respectively. Pixels that cannot be modeled by the candidate models with the above constraints remained unmodeled in the output. Because the shade fraction is a variation of brightness (Adams and Gillespie 2006), but not a property of the land covers, land cover fractions at each pixel was normalized by dividing each land cover fraction by the sum of total non-shade fractions. Note that the derived shade-normalized fraction images were not the final product. Rather, they were used to extract land covers in the urban subset that will be discussed in the next section. In this case, a quantitative accuracy assessment procedure was conducted for the final land cover map rather than for the fraction images (Roberts et al. 2002; Lu, Moran, and Batistella 2003; Weng and Lu 2009). 2.2.4.3. Extraction of vegetation covers and built-up areas. The derived V-I-S fraction images were first used to extract all vegetation covers in the urban subset by defining a threshold value for each fraction image. And the urban image was further separated into vegetation and non-vegetation subsets that can help reduce potential classification errors when a pixel-based classifier was applied to the image later (see the next section). The goal here was to preserve all the sparse vegetation pixels that would be classified as non-vegetated classes otherwise. In order to determine a vegetation cover threshold, different combinations of land cover fractions were tested using a decision tree classifier. The combination of impervious surface fraction less than 0.475 and the vegetation fraction higher than 0.4 were determined to exclude the „pure‟ impervious surface pixels and preserve the pixels with relative high percentages of vegetation cover. Then, the combination of the impervious surface fraction approaching 0 and the

20 vegetation cover higher than 0.22 were used to extract the sparsely vegetated areas. And these two vegetation parts were merged to produce the distribution of vegetation covers in the urban subset. The high density and low density urban classes were also identified by thresholding the impervious surface fractions using a decision tree classifier. The thresholds were determined by testing a range of fractions and then choosing the one that best separated the two urban classes based on visual inspection. Given the difficulty of using the simple threshold technique to discriminate different land use/cover types from the fraction images, a very popular statistical classifier, namely, maximum likelihood classifier (MLC), was further used to help identify vegetation types from the sub-pixel fractional information. This part of work will be discussed in the next section.

2.2.5 Supervised classification

The maximum likelihood classifier (MLC) was used to classify each of the rural subset and the urban vegetation image derived through MESMA. This classifier was chosen due to its robustness and popularity in remote sensing. A separate training procedure was conducted for each subset. Note that some thematic (or information) classes contain more than one spectral class, and hence more than one training sample was collected for each spectral class. For example, the cropland class contains four spectral classes linking with various growth stages, and four training sample sets were selected to account for the spectral variations within the cropland class. A similar training procedure was used for other classes. To ensure the quality of training samples, an initial training sample set was determined by examining the spectral profiles at different locations within the same land cover type. Then, the final training set was selected by visually inspecting the classification performance in combination with accuracy assessment by iteratively adding more samples from the locations that have not been well distinguished in the classification. Moreover, a significant spectral confusion between vegetated wetland and mixed forest was observed. To address this issue, the vegetated wetland class was first recoded to mixed forest and then buffers were created around water bodies. And all mixed forest pixels falling within a 100m buffer around water bodies were recoded into the vegetated wetland. The 100m buffer was decided through visual inspection with respect to the high resolution orthoimagery.

For the extracted vegetation covers in the urban subset, the three shade-normalized fraction images produced from MESMA, along with the six spectral bands, were included as input for image classification. A Normalized Difference Vegetation Index (NDVI) image was also included as additional layer for maximum likelihood classification of each subset. The inclusion of NDVI, vegetation fraction, and near infrared channels has increased the separability between the selected training classes and therefore was used to better characterize different types of vegetation. Therefore, 10 “bands” were used for the urban vegetation subset (i.e., 6 reflectance bands, γ fraction bands, and NDVI), and 7 “bands” (i.e., 6 reflectance bands and NDVI) were included for the rural subset. The urban vegetation and rural (mostly agriculture) subsets contain only positive NDVI values. Therefore the values of all bands (i.e., reflectance, shade-normalized fraction, and NDVI) used in image classification range from 0 to 1, and thus it was not necessary to normalize the data. The classified spectral classes were combined into related thematic classes for each subset. Then, a final land cover map was produced by unionizing the results from the urban vegetation image and the rural subset, along with the two urban classes identified using unmixed impervious surface abundance. A copy of the final land cover map is illustrated in Figure 2.5. For comparison purpose, two related land cover products were derived using maximum likelihood classifier with the combined training samples for the two subsets on the entire Atlanta 29-county area directly or each of the rural subset and the urban subset that has not been processed through MESMA. The same method for reclassifying vegetated wetland was applied to these two products too.

2.2.6 Thematic accuracy assessment

Both qualitative and quantitative accuracy assessment were conducted on the final classification maps. The qualitative assessment was based on a visual approach that compared the classified maps with the high resolution orthoimagery, with special attention on the isolated vegetation covers and urban land patches. The quantitative assessment was based on the error matrix method (Congalton 1991). The stratified random sampling scheme was used to select validation samples. An average of 50 reference points was generated for each thematic class with the classified map (Table 2.1). The reference class for validation was identified based on the high resolution imagery and field surveys. Overall accuracy, producer‟s accuracy, user‟s accuracy, and kappa coefficient were calculated for each map with the use of the identical validation data. 22

Figure 2.5. Land cover map produced from Landsat Thematic Mapper imagery by using stratified classification and multiple endmember spectral mixture analysis (MESMA) techniques.

2.3 Results and discussion

An error matrix for each map is summarized in Table 2.2-2.4. A comparison of the accuracy assessment results is included in Table 2.5. And the classification statistics for each map are summarized in Table 2.6. Based on the quantitative accuracy assessment (Table 2.2-

2.4), it is clear that the proposed method using the stratified classification and MESMA substantially outperforms the other two related methods that applied the maximum likelihood classifier on the entire image directly or on each of the rural subset and the urban subset without being MESMA processed. A close look at the conditional kappa coefficient for each class (Table 2.5) reveals that the major improvements of the proposed method are with the classes of mixed forest, grass, fallow/barren land, low density urban, and high density urban. The first two are vegetated classes that include many isolated, fragmented patches, and the third class is not a vegetated one but does contain sparse vegetation covers (Table 2.1). These classes are not mapped well when applying the MLC to the entire image directly (Table 2.5). On the other hand, the result does not show a meaningful difference for the two other forest land classes and the vegetated wetland class when comparing to the two related methods. This was expected since the proposed method has been developed to reduce spectral confusion and resolve mixed pixels. And these two forest classes can be effectively separated with a per-pixel classifier, and vegetated wetland class was derived by GIS-based reclassification of mixed forest class within a certain distance from water bodies. Now, let us take a look at how each of the two major strategies adopted has helped improve urban land use/cover mapping. The first major strategy used was to partition the entire landscape into rural and urban subsets to reduce the spectral confusion between agricultural land and urban built-up land (Myint and Okin 2009). Figure 2.6c shows the misclassification between residential vegetation covers and agricultural land when this strategy is not used (i.e., using MLC on the entire image). With the use of the landscape partition strategy, such confusion has been substantially reduced, as shown in Figures 2.6a and 2.6b. And the quantitative accuracy assessment also confirms the robustness of using landscape partition because the two partition- based methods have led to much higher classification accuracies for the two agricultural classes (i.e., pasture and cropland) when comparing with the method without using this stratified classification strategy (Table 2.5). Table 2.4 shows that when using MLC on the entire image, the major source of errors is due to the confusion of these two agricultural classes in the rural subset with grass mainly found in golf courses and parks in the study area and mostly located in the urban subset. And this misclassification has led to a substantial overestimation of grass because of the inclusion of numerous pasture or cropland patches when using the MLC on the entire image (Table 2.6).

Table 2.2. The error matrix for the land cover map produced using stratified classification and multiple endmember spectral mixture analysis (MESMA).

Reference

Classification Fallow/ Low High Evergreen Deciduous Mixed Vegetated barren density density Row forest forest forest wetland Grass Pasture Cropland land urban urban Water total Evergreen 46 1 7 4 1 1 60 forest Deciduous 1 38 1 6 1 47 forest

Mixed forest 2 7 44 2 4 2 1 1 63

Vegetated 1 4 5 32 1 43 wetland

Grass 3 1 42 8 1 55

Pasture 2 2 1 40 4 49

Cropland 1 3 5 38 5 1 1 54

Fallow/ 1 6 2 40 49 barren land Low density 1 42 1 1 45 urban High density 1 3 42 46 urban

Water 39 39

Column total 50 53 65 46 48 63 45 49 46 44 41 550

Table 2.3. The error matrix for the land cover map produced using maximum likelihood classifier (MLC) on each of the urban and rural subsets.

Reference

Mixed forest 2 8 40 3 5 1 1 60

Vegetated 1 4 5 31 1 42 wetland

Grass 3 1 20 3 27

Pasture 2 2 1 42 4 51

Cropland 1 3 4 38 5 1 1 53

Fallow/ 1 5 1 21 1 29 barren land Low density 6 22 7 1 19 45 9 1 110 urban High density 1 4 1 32 38 urban

Water 39 39

Column total 50 53 65 46 48 63 45 49 46 44 41 550

Table 2.4. The error matrix for the land cover map produced using maximum likelihood classifier (MLC) on the entire image.

Reference

Mixed forest 8 6 33 4 3 4 1 3 62

Vegetated 1 2 2 22 1 28 wetland

Grass 1 1 7 2 25 18 3 3 3 63

Pasture 3 1 1 6 26 5 5 2 49

Cropland 2 3 4 4 33 4 1 51

Fallow/ 5 3 23 2 5 38 barren land Low density 2 4 1 7 5 4 34 15 2 74 urban High density 1 1 1 1 7 4 24 39 urban

Water 1 3 6 39 49

Column total 50 53 65 46 48 63 45 49 46 44 41 550

Table 2.5. Comparison of the classification accuracies by different approaches.

Stratified classification MLC on urban and rural MLC on the entire image Land cover and MESMA subsets PA% UA% Kappa PA% UA% Kappa PA% UA% Kappa Evergreen forest 92.00 76.67 0.74 92.00 85.19 0.84 78.00 72.22 0.69 Deciduous forest 71.70 80.85 0.79 71.70 80.85 0.79 66.04 81.40 0.79 Mixed forest 67.69 69.84 0.66 61.54 66.67 0.62 50.77 53.23 0.50 Vegetated wetland 69.57 74.42 0.72 67.39 73.81 0.71 47.83 78.57 0.77 Grass 87.50 76.36 0.74 41.67 74.07 0.72 52.08 39.68 0.34 Pasture 63.49 81.63 0.79 66.67 82.35 0.80 41.27 53.06 0.47 Cropland 84.44 70.37 0.68 84.44 71.70 0.69 73.33 64.71 0.62 Fallow/barren land 81.63 81.63 0.80 42.86 72.41 0.70 46.94 60.53 0.57 Low density urban 91.30 93.33 0.93 97.83 40.91 0.35 73.91 45.95 0.41 High density urban 95.45 91.30 0.90 72.72 84.21 0.83 54.55 61.54 0.58 Water 95.12 100.00 1.00 95.12 100.00 1.00 95.12 79.59 0.78 Overall Accuracy 80.55 71.27 60.55 (%) Overall Kappa 0.79 0.68 0.57 Note: UA = user‟s accuracy; PA = producer‟s accuracy; MESMA = multiple endmember spectral mixture analysis; MLC = maximum likelihood classifier.

Table 2.6. Comparison of the classification results by different approaches.

Stratified classification and MLC on urban and rural MLC on entire image Land cover MESMA subsets ha % ha % ha % Evergreen forest 272595 11.71 251551 10.81 274288 11.78 Deciduous forest 214825 9.23 213361 9.17 250345 10.75 Mixed forest 898726 38.61 804200 34.55 656357 28.19 Vegetated 42080 1.81 38502 1.65 34980 1.50 wetland Grass 162443 6.98 85301 3.66 339605 14.59 Pasture 312087 13.41 312087 13.41 238367 10.24 Cropland 48645 2.09 48643 2.09 68115 2.93 Fallow/barren 57145 2.46 38993 1.68 49799 2.14 land Low density 131418 5.65 333819 14.34 288573 12.40 urban High density 152880 6.57 166365 7.15 83100 3.57 urban Water 34694 1.49 34878 1.50 44605 1.92 Note: MESMA = multiple endmember spectral mixture analysis; MLC = maximum likelihood classifier.

Figure 2.6. Comparison of the land cover maps produced by different approaches. (a) Stratified classification and multiple endmember spectral mixture analysis (MESMA). (b) Maximum likelihood classifier (MLC) on each of the rural subset and the urban subset without being MESMA processed. (c) Maximum likelihood classifier (MLC) on the entire image. (d) An orthoimage with 1m resolution. Note that each image covers the same geographic area with a dimension of approximately 3.5 km by 3.5 km.

The second major strategy was to apply the multiple endmember spectral mixture analysis (MESMA) to the urban subset in order to account for small, fragmented land patches scattering in urban areas. The MESMA-derived land cover fractions were used to separate urban

29 vegetated and non-vegetated areas as well as the high density and low density urban areas before applying the MLC on the urban subset. Figures 2.7a and 2.7b illustrated the sub-pixel unmixed

Figure 2.7. Comparison between the per-pixel analysis and the sub-pixel analysis. (a) An unmixed vegetation fraction image displayed in grayscale. (b) An NDVI (Normalized Difference Vegetation Index) image displayed in grayscale. (c) The original image in false color display with the near infrared, red, and green bands as the red, green, and blue guns, respectively. (d) A fraction image in false color display with the vegetation fraction, the impervious surface fraction, and the soil fraction (all shade normalized) as the red, green and blue guns, respectively. Note that each image covers the same geographic area with a dimension of approximately 10 km by 10 km. vegetation fraction and the per-pixel NDVI layer derived from the TM image for the same area. The major differences between NDVI and the vegetation fraction is that NDVI is calculated from two channels and vegetation fractions are derived from the spectral information contained in all

30 channels (Adams and Gillespie 2006). The NDVI image tends to underestimate the abundance of evergreen forest canopy cover than other vegetation covers due to the lower reflectance at near- infrared band in the May 2007 image. The NDVI image does not allow to differentiate between the densely vegetated grass (e.g., golf courses) and the sparsely vegetated area (e.g., urban residential). In this regard, the findings here are in line with what Small (2001) reported in his vegetation abundance estimation for New York City. Figure 2.6a shows the pattern of vegetation patches interspersed in the residential area that were extracted from moderate resolution imagery by assigning threshold values on the three land cover fractions. When using the high-resolution image (Figure 2.6d) as the reference, these fragmented patches have been delineated quite well through the sub-pixel MESMA processing. In contrast, most of those isolated vegetation patches have been misclassified as built-up land with the pixel-based methods (Figure 2.6b and 2.6c). In addition to the vegetation classes, the sub-pixel analysis also performs much better in mapping the two urban categories, i.e., high- density urban and low-density urban, which are classified using the impervious surface fractions derived from MESMA. The use of MESMA has significantly improved the user‟s accuracy of low density urban, while per-pixel classification tended to misclassify a significant amount of isolated vegetation patches into low-density urban that has led to a substantial overestimation of this urban class (Table 2.6). Although the proposed method has shown some major merits, there are several potential limitations. First, it does not accurately distinguish specific vegetation types scattered in the residential area. Table 2.2 shows the confusion between grass and forest types, which can be visually confirmed from Figures 2.6a in which most residential vegetation patches have been classified as mixed forest. Another possible limitation is that a simple buffering and GIS-based recoding method were used to classify vegetated wetland that may introduce both user‟s and producer‟s errors. In order to accurately map wetland vegetation, more factors should be taken into account, such as soil moisture and water salinity. In addition, the classification accuracy of mixed forest is relatively low, which may be due to the spectral confusion between this class and the two other forest types, particularly as shown in medium resolution imagery. Mixed forest is essentially the mixture of evergreen and deciduous forest by definition, whose reflectance is influenced by the abundance and structure of the two other forest types. The error in classifying mixed forest can be attributed to the use of medium resolution data and the difficulty in

31 identifying forest species in training sample collection and accuracy assessment even with the use of high spatial resolution data. Finally, there are some other factors which may influence urban land use/cover mapping accuracy. For example, data availability can be a major constraint. When using Landsat TM imagery for large areas extending more than one path, there is no adjacent scene that could be acquired on the same date due to the satellite system design. The temporal difference among the multiple satellite scenes may introduce errors due to atmospheric and phenological effects. The temporal difference could be more complicated when a large amount of cloud cover exists. In this specific project, a small area of the August image from Path 18 was mosaicked with the May images from Path 19 which could lead to the difficulty in image preprocessing and classification.

2.4 Conclusions

Mapping land use and land cover in urban areas has been challenged by the spatial and spectral heterogeneity in the urban environment. A method was developed to help improve land use/cover mapping in urban areas from medium resolution imagery through the combined use of stratified classification with multiple endmember spectral mixture analysis (MESMA) techniques. This method was applied to produce a land use/cover map for a large metropolitan area, and the performance of this method was compared with two related methods that do not adopt the stratified classification or the sub-pixel information extraction strategy. The results indicate that the use of stratified classification strategy can help suppress the spectral confusion between agricultural land and some urban land features through spatially separating the entire landscape into rural and urban subsets and then processing each subset independently. The landscape partition was accomplished by using road intersection density and weighted road buffers. Caution should be paid on the construction of the road intersection density surface with an appropriate search radius and the determination of a threshold value to separate the two subsets. Needless to say, the quality of road network data in terms of completeness and temporal accuracy is critical for the success of landscape partition that is considered as the first major step in the proposed method. The urban subset has been further processed with the MESMA technique to generate the sub-pixel V-I-S fractions. Then the urban subset image has been separated into a vegetation layer and urban built-up layer by applying threshold values to each fraction image. The two urban

32 classes, i.e., high density and low density urban, have been derived from the urban built-up layer by using a thresholding technique. To classify various vegetation types within the extracted vegetation area, the sub-pixel V-I-S fraction images were included as additional layers to be used in the supervised classification protocol. Finally, the maximum likelihood classifier has been applied to each of the urban and rural subset. Training samples have been carefully selected for each subset independently that can account for the spectral variability of each land cover class in urban and rural regions. Contrasting with many existing studies that have mostly restricted on using the sub-pixel mapping technique to derive the land cover abundance information, the method developed in this study has extended into the area of mapping spatial distribution of various land use/cover types in a complex urban environment through stratified classification combined with a sub-pixel level information extraction strategy.

CHAPTER THREE

MONITORING LAND CHANGES IN AN URBAN AREA USING SATELLITE IMAGERY, GIS AND LANDSCAPE METRICS

This chapter presents a method for land change mapping and analysis through the combined use of satellite imagery, geographic information systems (GIS), and landscape metrics. The method consisted of two major components: remote sensing-based land classification and GIS-based land change analysis. Specifically, a stratified image classification strategy was combined with a GIS-based spatial reclassification procedure to map land classes from Landsat Thematic Mapper (TM) scenes acquired in two different years. Then, a post-classification change detection strategy was employed to analyze land changes through a variety of GIS-based operations. Landscape metrics was further used to examine the size, pattern, and nature of land changes. The results reveal a far-reaching suburbanization process, which has rampantly altered the forest ecosystems, as indicated by the change in forest landscape structure and patterns in the study site. This study has demonstrated the usefulness of integrating remote sensing with GIS and landscape metrics that allows land change analysis to go beyond simple statistic description and into the characterization of the spatial characteristics and nature of land changes in a complex urban environment.

3.1 Introduction

Land changes in urban areas, especially the conversion of cropland and forest land to urban uses, is one of the most important forms of global environmental changes (Briassoulis 2000). Since the mid-20th century, many American metropolises have experienced dramatic growth, which was dominated by a suburbanization process with new development spreading outward from the urban core towards suburbs and exurbs. While urban development has always been viewed as a sign of the regional economic prosperity, the emerged low density and leapfrog built-up land patterns in suburban and exurban areas have begun to undermine environmental sustainability. Monitoring land changes in urban areas can support decision making in urban planning and resource management (Lambin et al. 2001; Turner, Lambin, and Reenberg 2007). The advances in remote sensing and geospatial information techniques offer a promising

34 framework to monitor land changes in urban areas (Elvidge et al. 2004; Rindfuss et al. 2004; Southworth and Gibbes 2010). Remote sensing provides a cost-effective alternative to the ground-based survey for land use/cover mapping and change analysis. Time-series of remotely sensed data allow examining the temporal dynamics of urban attributes or processes. And post-classification comparison methods produce “from-to” change information between land classes that can help capture the nature of land changes (Jensen 2004). Given the wide availability and long-time archive, Landsat data have been widely used in land use/cover classification and change detection at regional scales (e.g., Vogelmann et al. 2001; Yang and Lo 2002; Yang and Liu 2005; Yuan et al. 2005; Liu and Weng 2013). However, land use/cover classification in urban areas using medium resolution remotely sensed data (e.g., Landsat Thematic Mapper) can be challenging due to the presence of heterogeneous urban features and the spectral similarity between different urban land cover types (Welch 1982; Forster 1983; Gao and Skillcorn 1998; Small 2001; Herold, Gardner, and Roberts 2003; Guindon, Zhang, and Dillabaugh 2004). Over the past years, a sizable number of research has demonstrated the usefulness of sub-pixel analysis to deal with the "mixed" pixel problem associated with using medium resolution remotely sensed data in urban land mapping (e.g., Roberts et al. 1998; Small 2001; Rashed et al. 2003; Small 2003; Song 2005; Small and Lu 2006; Powell et al. 2007; Franke et al. 2009; Myint and Okin 2009; Liu and Yang 2013). For change analysis, the sub-pixel analysis has mostly been applied to the detection of land cover fraction change, such as percent imperviousness change (e.g., Yang et al. 2003), vegetation fraction change (e.g., Small and Miller 1999, 2000; Taramelli et al. 2014). Some research has shown the potential of incorporating sub-pixel fraction in thematic land use/cover classification (e.g., Lu and Weng 2006; Liu and Yang 2013). However, using the spectral response from remote sensing alone may not be sufficient to differentiate specific land types in urban areas, which can be valuable for various applications such as driving force analyses, urban morphological studies, and land use modeling. Previous studies have identified the importance of incorporating ancillary data in image classification (e.g., Guindon, Zhang, and Dillabaugh 2004; Bock et al. 2005; Lu and Weng 2006). While the post-classification change detection provides insight into the nature of land change, the integration of remote sensing and geographic information system (GIS) can be quite useful for characterizing the spatial patterns of urban land change (e.g., Chen, Zeng, and Xie

2000; Aspinall and Hill 2008). GIS provides a flexible environment for entering, analyzing, and displaying digital data from various sources necessary for land type identification, change detection, and database development. On the one hand, integrating GIS with remote sensing can help improve land mapping in urban areas. Examples include on-screen digitizing urban features from high resolution satellite imagery (e.g., Lathrop 2004) and combining GIS ancillary data to improve image classification (e.g., Boteva, Griffiths, and Dimopoulos 2004; Guindon, Zhang, and Dillabaugh 2004; Mundia and Aniya 2005; Yang and Liu 2005; Shalaby and Tateishi 2007; Liu and Yang 2013). On the other hand, the integration of remote sensing and GIS can help generate useful information on how much, where, and what types of land changes have occurred. For example, various GIS spatial analysis functions ranging from overlay to cluster analysis has been applied tothe spatial pattern analysis of land changes (e.g., Weng 2002; Yang and Lo 2002; Li and Yeh 2004; Mundia and Aniya 2005; Wu et al. 2006; Xiao et al. 2006). Moreover, landscape metrics can be used to quantify the spatial structure of urban areas and thus add insights to the remote sensing-based change detection. For examples, some research has demonstrated the importance of using landscape metrics to study urban growth patterns through interpreting the trends of selected landscape metrics (e.g., Herold, Scepan, and Clarke 2002; Herold, Goldstein, and Clarke 2003; Carrion-Flores and Irwin 2004; Seto and Fragkias 2005; Ji et al. 2006). Integrating spatial metrics with remote sensing and GIS facilitates the examination of the spatial structures of land changes in urban areas, including location, distribution, size, shape, and arrangement, which are important variables in measuring urban sprawl. This study presents a method for land change mapping and analysis through the combined use of satellite imagery, geographic information systems (GIS), and landscape metrics. The study site, Atlanta metropolitan area, is one of the fast-growing large metropolises in the United States, which contains a mosaic of complex landscape types. This method consisted of two major components: remote sensing-based image classification and GIS-based land change analysis. Specifically, a stratified image classification strategy combined with a GIS-based spatial reclassification procedure were adopted to map land classes from Landsat Thematic Mapper (TM) scenes acquired in two different years. Then, a post-classification change detection strategy was employed to analyze land changes through a variety of GIS-based operations. Landscape metrics was further used to examine the size, pattern, and nature of land changes. The

36 following sections will detail the research methodology and discuss the spatial characteristics and nature of land changes.

3.2 Research methods

The land change mapping and analysis used here was based on the post-classification change detection strategy (Jensen 2004). The specific working procedural route is illustrated in Figure 3.1. The sections to follow will detail each component, including data collection and preprocessing, image processing, and land change analysis.

Figure 3.1. Flowchart of the working procedural route.

3.2.1 Data collection and preprocessing

A variety of data were collected in support of the research activities. Specifically, six cloud-free TM scenes acquired during the late spring/early summer in 2000 and 2010 were collected from the USGS EROS Data Center (Table 3.1). Note that three TM scenes (i.e.,

Path/Row 19/36, 19/37, and 18/37of the Worldwide Reference System 2 (WRS-2)) are needed in order to cover the entire study area. The path 18 scene only covers less than 2% of the study site, and therefore the 7-day interval in acquisition dates between adjacent paths of the Landsat system can be ignored. Only the six reflective bands with 30 m resolution were used for further data analysis while the thermal infrared band with coarse spatial resolution of 120 m was excluded. In addition to the TM images used for land classification, an accurately rectified 1997 Landsat TM image supplied by Space Imaging EOSAT was collected as the reference image in geometric correction that will be discussed later.

Table 3.1. List of the Landsat Thematic Mapper (TM) scenes used.

RMSE Acquisition date Path/Row Sun elevation(°) Sun azimuth(°) Atmospheric correction (GCP No.) 2000/05/08 18/37 61.19 116.88 Yes 2000/05/15 19/36 61.90 116.82 0.091 (15) Yes 2000/05/15 19/37 62.29 114.01 Yes 2010/04/18 18/37 58.93 130.23 Yes 2010/04/09 19/36 55.25 135.18 0.097 (16) Yes 2010/04/09 19/37 56.07 133.26 Yes 1997/07/29 center-shifted* 61.00 106.00 reference No * The center of the 19/36 scene has been shifted downward 50%. The scene size is approximately 185 by 185 km

In addition to the satellite imagery, several ancillary datasets were assembled to assist the land change mapping. They included a street centerline dataset for use in creating an urban mask to partition the entire landscape. The street centerline dataset consisted of all paved roads across local, state, and interstate levels, which was produced in 2007 through a joint effort by Georgia Department of Transportation (GDOT), Carl Vinson Institute's Office of Information Technology Outreach Services (ITOS), ARC, and the ARC county governments. The digitized land use databases (LandPro) were collected for use in identifying detailed land use types. The databases provided specified land use/cover classes that can be useful for this study. The ARC's LandPro GIS databases were created by on-screen photo-interpretation and digitizing of orthorectified aerial photography with 1 m pixel resolution for 2001 and 1.64 foot resolution for 2010. In addition, the National Wetland Inventory (NWI) 1:24,000 wetland datasets for 2012 were also obtained from the U.S. Fish and Wildlife Service to facilitate wetland mapping. Finally, field data and aerial photographs were collected to assist land use/cover classification and thematic accuracy assessment. Two extensive field surveys were conducted

38 across the entire study area in December 2010 and May 2011. The field itineraries were designed according to the author‟s knowledge about the study area and a preliminary examination of the satellite images. During the field trips, geographic positions and land use/cover types at each predefined sample location were recorded with a Trimble GPS receiver. In addition, two sets of color infrared aerial photographs were collected and used as the reference data: (1) USGS aerial photographs at a scale of 1:40,000 derived from National Aerial Photography Program (NAPP) taken in January 2000; and (2) USDA aerial photographs with 1 m spatial resolution derived from National Agriculture Imagery Program (NAIP) acquired in August and September of 2010. These datasets were mainly used to assist thematic accuracy assessment. The Landsat TM imagery data were preprocessed through several procedures, namely, atmospheric correction, image mosaicking, geometric correction, and image subsetting. The Chavez (1996) COST model was employed to convert the digital number of the six TM scenes into surface reflectance. The COST model uses the cosine of the solar zenith angle to approximate the atmospheric transmittance for the dates and sites which can help maximize the accuracy. Then, the three reflectance images of the same year were mosaicked to produce a larger image that covers the entire study area. To address the spectral disparities between scenes at adjacent WRS-2 paths (i.e., Path18/Row37 and Path19/Row37) due to varying acquisition dates, the overlaying area of the two TM scenes that covers less than 2% of the entire study site were spectrally matched. Then the mosaicked images were geometrically rectified with reference to the 1997 Landsat TM image, which has already been accurately rectified and georeferenced to the UTM (Universal Transverse Mercator) map projection (Zone16), NAD83 datum and GRS1980 ellipsoid. The number of ground control points and resultant average root mean square error (RMSE) are shown in Table 3.1. The first-degree polynomial fit was used in image transformation given the relatively flat terrain relief in this region. The nearest-neighbor resampling method was applied to avoid changing the images‟ original pixel values. Finally, the boundary of the 29-county Atlanta region was used to subset the image to fit the study area. Note that in order to account for the possible georeferenced error, a 200 m buffer was created immediately surrounding the 29-county polygon that was actually used for the image subsetting. To ensure the spatial accuracy of the land use/cover classification with ancillary data, both the street centerline data and the LandPro data were geometrically rectified with reference to the mosaicked image.

3.2.2 Image processing

3.2.2.1. Land classification scheme. Based on the research objectives, data resolution, and field surveys, a land use/cover classification scheme was designed based on a modification from the Anderson scheme (Anderson et al. 1976):  Residential land uses range from high density multiple-unit structures in urban cores to low density large lots areas found in peri-urban areas.  Commercial/Industrial land uses include commercial areas used predominantly for the sale of products and services and industrial areas used for manufacturing.  Other urban/built-up land consists of areas mainly used for transportation, communication, utilities, and some rural constructions.  Agricultural land consists of areas used for crop production and herbaceous covers planted for livestock grazing or the production of seed or hay crops.  Forest land consists of areas dominated by trees greater than 5 meters tall, including all deciduous, evergreen, and mixed tree species.  Grass/Shrub land encompasses areas dominated by herbaceous vegetation and/or shrubs, young trees, and stunted trees. This category also contains small parks, cemeteries, and golf courses.  Wetland is characterized by vegetation covers where the soil or substrate is periodically saturated with or covered with water, including both forested and non-forested wetland.  Fallow/barren land is characterized by areas with sparse vegetation covers (less than 20%), including quarries, exposed rocks, clear-cuts, fallowed agricultural land, unpaved rural roads, and urban transitional land.  Water consists of all areas of open waters, including lakes, streams, rivers, and reservoirs. 3.2.2.2. Stratified classification. The first strategy in the stratified classification was to spatially partition the entire landscape into urban and rural subsets to be processed separately. The goal here was to suppress the spectral confusion between urban impervious surfaces and agricultural land that was fallowed or at a certain growth stage. The landscape partition was done by using an urban mask created through the use of road density. Specifically, the road

40 intersection density surfaces were firstly generated using road intersections extracted from the street centerline data. Note that the road data for 2000 and 2010 were derived from the ARC 2007 street centerline data which were adjusted with the reference of the satellite images acquired at the two years. Secondly, the threshold values were determined iteratively to single out the urban built-up area from the rural part for each road intersection density surface. Any area with density values greater than the thresholds was then grouped into the urban portion. Finally, the generated urban portion was combined with street centerline buffers with varying distance values by road levels to create the complete urban masks. The area outside the urban mask boundary was defined as the rural subset. To deal with the mixed pixel problem with using medium resolution data, the second strategy adopted here was to implement sub-pixel mixture analysis (i.e., MESMA) to the urban subset to help extract the isolated urban and vegetation patches. More details on the rationale and methodology of MESMA can be found in (Roberts et al. 1998; Dennison et al. 2003; Dennison and Roberts 2003a, 2003b; Dennison, Halligan, and Roberts 2004; Powell et al. 2007). The endmember library for each year was built according to a four-endmember model with shade endmember included into the vegetation-impervious surface-soil (V-I-S) model (Ridd 1995). Three methods were used to facilitate the identification of spectrally “pure” pixels: pixel purity index (PPI) (Boardman, Kruse, and Green 1995), count-based endmember selection (CoB) (Roberts et al. 2003), and endmember average root mean square error (EAR) (Dennison and Roberts 2003a). The final set of endmembers for the 2000 image consists of 4 spectra in the vegetation category, 15 distinct endmember spectra in the impervious surface category, and 2 spectra in the soil category. The endmember library for the 2010 image consists of 4, 17, and 2 for the V-I-S classes respectively. With the endmember library available, each image pixel can be modeled with all the possible combinations of endmember spectra of the three non-shade endmembers. For model implementation, the non-shade fractions were constrained between - 0.05 and 1.05; the maximum RMS error allowed was set to 0.025; and the maximum allowable shade fraction was set to 0.5 so that the water bodies can be excluded by making them remain unmodeled (Powell 2011). Then the model that met all the constraints and had the lowest RMS error was selected to estimate the V-I-S and shade fractions for each pixel within the urban subset. The output was an image with six layers, including the fractions of impervious surface, vegetation, soil, and shade, model number, and RMS error. The fraction values at each pixel

41 were normalized by dividing each V-I-S fraction by the sum of non-shade fraction because the shade fraction represents a variation of brightness but not certain types of land cover (Adams and Gillespie 2006). Note that the shade-normalized fractions were further used as input for thematic land use/cover classification. The shade-normalized V-I-S fraction images were used to extract all urban built-up area in the urban subset by defining a threshold value on the impervious surface fractions with a decision tree classifier. The thresholds were determined by testing a range of fractions and then choosing the one that best delineates the boundary between urban and non-urban pixels. The detailed land use classes within the urban built-up category were further identified through a GIS-based spatial reclassification with the assistance of LandPro data, which will be discussed in Section 2.4.3. The non-urban pixels within the urban subset were processed through supervised classification that will be discussed in the next section. The maximum likelihood classifier (MLC) was adopted to classify each of the rural subset and the non-urban pixels within the urban subset. A separate training procedure was conducted for each subset. The training samples were carefully selected for each class described in Section 2.3 through an iterative procedure based on an examination of their representativeness and separability between training classes, as well as results in thematic accuracy assessment. Due to the significant spectral confusion observed between wetland and forest classes, the wetland class was grouped into the forest category for the supervised classification. A Normalized Difference Vegetation Index (NDVI) image was included as an additional layer for maximum likelihood classification of each subset. Also the three shade-normalized fraction images were included as input layers in the supervise classification for the urban subset. Therefore, 10 “bands” were actually used for the urban subset (i.e., 6 reflectance bands, 3 fraction bands, and NDVI), and 7 “bands” (i.e., 6 reflectance bands and NDVI) were included for the rural subset. Finally, the classified outcomes from the two subsets and the previously extracted urban classes were combined to produce a complete map. 3.2.2.3. Spatial reclassification. A series of GIS-based spatial reclassification procedures were conducted to refine the outputs from the stratified image classification. Firstly, the wetland layer generated from the NWI datasets for each year was overlaid with each of the classified land use/cover map. This can help address the significant spectral confusion between the wetland and forest classes during the supervised classification. The wetland polygon data for each year were

42 manually modified from the NWI 2012 datasets with reference to the satellite imagery and aerial photos. The wetland polygons of each year were then converted into binary layers that were then overlaid with the classified map to produce the wetland classes. Therefore, all the classified forest pixels that overlapped with the wetland layer were reclassified as wetland in the output maps for both years. Secondly, the urban built-up category for each year was further classified into three different land use classes (i.e., residential, commercial/industrial, and other urban) by using the ARC‟s LandPro GIS databases. Specifically, the polygons of each urban land use class were extracted and converted into binary layers that were then overlaid with the classified map to produce the detailed land use/cover maps. Because the ARC‟s LandPro data were not available for the year of 2000, the LandPro 2001 layer was used as a substitute and was modified with reference to the Landsat TM 2000 images. In addition, the spatial extents of LandPro data for both years did not cover the entire 29-county Atlanta region, therefore the outside urban area was manually digitized based on visual image interpretation as described by Campbell (2002). Note that the un-covered region for each year only contains a relative smaller percent of the total urban area. The GIS overlay procedure used the classified output as the base map, while only the urban pixels within the classified map were further reclassified into the three corresponding land use categories. The primary reasons for this procedure include: (1) the mapping units of the LandPro data for both years were relatively coarse compared to the satellite imagery‟s pixel resolution; and (β) the mapping units of the two different years‟ LandPro datasets were different which may be problematic in land change analysis. Therefore the classified map from remotely sensed data was referred to provide both spatial accuracy and temporal consistency in the change analysis. 3.2.2.4. Thematic accuracy assessment. Both qualitative and quantitative accuracy assessment were conducted for the final classified maps of both years. The qualitative assessment was based on a visual approach that inspected the classification results with reference to the aerial photos. The quantitative assessment was conducted through a standard error matrix approach (Congalton 1991). A stratified random sampling scheme was used to select validation samples, which stratified the number of random points to the distribution of thematic layer classes. At least 50 reference points were then generated for each thematic class within the classified maps. The reference classes for validation were identified with reference to the aerial

43 imagery and field surveys. Overall accuracy, producer's accuracy, user's accuracy, and kappa coefficient were calculated for each year‟s classified map (Table 3.2 and 3.3). Both maps met the minimum 85% overall accuracy specified by the Anderson classification scheme (Anderson et al. 1976), which is an indication of the effectiveness of the image processing techniques adopted here. The final land use/cover classification maps for the two years are shown in Figure 3.2. The statistics of land use/cover categories and the changes between 2000 and 2010 for the 29-county Atlanta region are summarized in Table 3.4.

3.2.3 Urban land change analysis

Land change analysis conducted here was mainly based on post-classification change detection through the use of several GIS-based operations and landscape metrics measured from the land use/cover maps produced by remote sensing. Among many possible trajectories of land changes, this study focused on the growth of the three urban land classes during the period of 2000 and 2010. And several qualitative and quantitative techniques were adopted to examine the spatial patterns and the nature of land changes. In addition, the observed patterns of residential and commercial land use were also compared with past findings documented by Yang (2002). Specifically, the net growth of urban area between 2000 and 2010 for the three urban land use classes was visualized, which is shown in Figures 3.3. This visual approach can help reveal the spatial patterns (e.g., location, size, shape, fragmentation) of urban growth in the study area. Quantitatively, the land change statistics were compared for the ARC's10-county, 20- county, and 29-county region with a GIS overlay function. The geographic extent of the three regions represents the spatial expansion of the Atlanta metropolitan area. Therefore, the comparison of land change statistics across the different regions can help reveal the regional disparity in urban development (Figure 3.4). With the relative percentage values, it was able to identify the degree to which each land use class is spread out through the metropolitan area over the ten years. Landscape pattern metrics quantify the size, shape and spatial arrangements of land use/cover patterns, thus providing information on the spatial characteristics of changes. A set of landscape metrics were calculated for the three urban classes to measure the three types of urban growth patterns: fragmentation, dispersion, and compactness (Carrion-Flores and Irwin 2004). Firstly, the number of patches, the mean patch size, and the total edge metrics (i.e., the sum of

Table 3.2. Thematic accuracy assessment for the 2000 land use/cover map produced from Landsat Thematic Mapper (TM) data.

Reference data Classified Residential Commercial/ Other Agricultural Forest Grass/ Wetland Fallow/ Water Row User's Kappa data industrial urban shrub barren total accuracy

Residential 57 2 1 60 95.00% 0.943

Commercial/ 52 1 53 98.11% 9.979 industrial

Other urban 1 46 1 2 50 92.00% 0.910

Agricultural 1 43 2 6 51 84.31% 0.822

Forest 1 1 4 3 52 1 3 1 63 82.54% 0.802

Grass/shrub 8 2 41 3 58 70.69% 0.673

Wetland 1 50 51 98.04% 0.978

Fallow/ 2 2 46 50 92.00% 0.910 barren

Water 51 51 100% 1.000

Column total 59 53 54 57 58 50 53 52 51 487

Producer‟s 96.61% 98.11% 85.19% 75.44% 89.66% 82.00% 94.34% 88.46% 100% accuracy Overall 89.94% accuracy Overall 0.887 Kappa Note: The bold face in the table shows the number of correctly classified samples for each thematic class.

Table 3.3. Thematic accuracy assessment for the 2010 land use/cover map produced from Landsat TM data.

Reference data Classified Residential Commercial/ Other Agricultural Forest Grass/ Wetland Fallow/ Water Row User's Kappa data industrial urban shrub barren total accuracy

Residential 50 1 3 54 92.59% 0.916

Commercial/ 52 1 53 98.11% 0.979 industrial

Other urban 1 50 2 53 94.34% 0.936

Agricultural 1 50 1 5 57 87.72% 0.860

Forest 1 50 1 2 1 55 90.91% 0.898

Grass/shrub 3 4 41 8 56 73.21% 0.700

Wetland 51 51 100% 1.000

Fallow/ 2 3 5 42 52 80.77% 0.784 barren

Water 1 50 51 98.04% 0.978

Column total 55 53 56 58 51 52 54 53 50 482

Producer‟s 90.91% 98.11% 89.29% 86.21% 98.04% 78.85% 94.44% 79.25% 100% accuracy Overall 90.46% accuracy Overall 0.893 Kappa Note: The bold face in the table shows the number of correctly classified samples for each thematic class.

Figure 3.2. Land use/cover maps for 2000 and 2010, which were derived from Landsat Thematic Mapper (TM) imagery.

Table 3.4. Land use/cover changes between 2000 and 2010 for the 29 counties under the Atlanta Regional Commission (ARC).

2000 2010 2000-2010* Land use/cover Area (ha) % Area (ha) % Area (ha) % Residential 124303 5.34 137702 5.92 13400 10.78 Commercial/industrial 68166 2.93 76016 3.27 7850 11.52 Other urban 45512 1.96 54401 2.34 8889 19.53 Agricultural 258015 11.08 265120 11.39 7105 2.75 Forest 1401184 60.20 1523467 65.45 122282 8.73 Grass/shrub 279130 11.99 137084 5.89 -142047 -50.89 Wetland 60546 2.60 63556 2.73 3010 4.97 Fallow/barren 45403 1.95 26592 1.14 -18811 -41.43 Water 45369 1.95 43693 1.88 -1676 -3.69 *The percent changes of each class are calculated by dividing the net area changes between 2000 and 2010 by the class area in 2000.

Figure 3.3. Spatial growth of the three urban land classes during the period of 2000 and 2010: (a) Residential land; (b) Commercial/industrial land; and (c) Other urban land. Note that the location of expressways, including all interstate highways and Georgia Highway 400, is also shown.

49 the perimeters of all patches within the same class) were calculated for each year to represent the size and degree of landscape fragmentation. Secondly, the mean nearest neighbor metric (i.e., the average of the shortest distances between patches of the same class) was calculated to measure the dispersion of land use patches. Thirdly, the mean perimeter/area ratio (i.e., the ratio of the mean perimeter length to the mean patch area for certain class) was used to measure the compactness of the landscape pattern. The selected landscape metrics for each urban land use classes are calculated with the original classified thematic map at 30 m pixel size and the results are summarized in Table 3.5.

Table 3.5. Selected landscape metrics for the three urban land classes in 2000 and 2010.

Residential Commercial/industrial Other urban Landscape Metrics 2000 2010 2000 2010 2000 2010 Number of patches 17987 22488 14055 18480 15174 13917 Mean patch size (ha) 0.6829 0.5949 4.8949 4.1490 0.2912 0.3793 Total edge (km) 59589.15 68086.61 12564.19 14886.80 30238.22 30767.39 Mean nearest neighbor (m) 81.78 81.74 184.57 163.63 116.38 123.13 Mean perimeter/area ratio (m/m2) 1031.88 1053.89 758.50 784.38 1174.10 1154.31

Land use/cover conversion between classes can be useful to indicate the nature of land changes. A two-way cross-tabulation or matrix analysis was adopted to characterize the conversions from different land use/cover categories to the three urban classes. The matrix analysis produces a thematic layer that contains a separate class for every possible conversion between classes in two layers. As mentioned before, this study focused on the changes of the three urban land use classes, and therefore, the total possible combinations are grouped into 15 conversion types as shown in Table 3.6. The C1-C12 categories in Table 3.6 consist of the conversion from forest, agricultural, grass/shrub, and fallow/barren land to the three urban classes. The C13 and C14 categories denote the changes of a few vegetation classes especially forest land. Note that the C0 category includes all the unchanged pixels (approximately 74% of the landscape) and other combinations that are not considered here. The spatial characteristics of land conversions are illustrated in Figure 3.5.

Table 3.6. Land use/cover conversion between 2000 and 2010.

Nature of change Land Conversion Code From To Area (ha) % C1 Forest Residential 41699 1.79 C2 Agricultural Residential 2715 0.12 C3 Grass/shrub Residential 18836 0.81 C4 Fallow/barren Residential 1995 0.09 C5 Forest Commercial/industrial 8432 0.36 C6 Agricultural Commercial/industrial 919 0.04 C7 Grass/shrub Commercial/industrial 4336 0.19 C8 Fallow/barren Commercial/industrial 855 0.04 C9 Forest Other urban 16739 0.72 C10 Agricultural Other urban 1970 0.08 C11 Grass/shrub Other urban 7713 0.33 C12 Fallow/barren Other urban 2838 0.12 C13 Grass/shrub and fallow/barren Forest 188127 8.08 C14 Forest Grass/shrub and Fallow/barren 36776 1.58 C0 All other combinations 1993640 85.65

3.3 Results and discussion

The yellow pixels in Figure 3.3 represent the smallest amount of urban land uses in 2000 which were produced through the GIS minimum dominate overlay function on the two years‟ urban uses that was used by Yang (2002). And the red pixels are the net growth of each urban land use between 2000 and 2010. This visual approach gives some general indication on where the changes have occurred. Based on Figure 3.3a, the spatial distribution of residential land expansion can be clearly visible. In 2000, the residential land (in yellow) occupied about 124,303 hectares, or 5.34% of the total area for the 29 counties (Table 3.4). Overall, the spatial pattern of residential land in 2000 shows some degree of fragmentation and dispersal. Based on Figure 3.4a, it is clear that the majority of residential area was located within the ARC's 20-county region in 2000. Note that the values at y-axis in the charts represent the percent area of land use/cover classes to the total area of the study area. Continuing residential growth was observed during the period of 2000-2010, with a net addition of 13,400 hectares or 10.78% increment. In terms of the spatial distribution, Figure 3.3a reveals that residential land growth between 2000

Figure 3.5. Land use/cover conversion for the 29 counties in the Atlanta metropolitan area during the period 2000-2010. Note that some classes of conversion are combined. and 2010 was found both around existing development in central cities and suburbs and at the peri-urban fringes. Additionally, most of the residential land growth was still found within the 10-county region (Figure 3.4c), particularly in the counties of Gwinnett, Fulton, Cobb, Cherokee, Forsyth, Henry, and Douglas. And some new residential constructions were also found in the external counties within the 20-county Atlanta region, especially in the counties of Coweta and Paulding. As a result, most of the residential land in 2010 was located within the 20-county

52 region with a limited extension into the outer-ring suburban and exurban areas. The evolution of Atlanta's residential land patterns between 2000 and 2010 is further quantified by using several selected landscape metrics (Table 3.5). From 2000 to 2010, the mean patch size of residential land decreased while the number of patches and total edges increased rapidly, indicating that residential land had become more fragmented over time. The mean nearest neighbor metric of residential land was fairly constant between the two years, implying that new residential development largely maintained a dispersed pattern. The increase of mean perimeter/area ratio over time reveals that a less compact pattern had emerged. Overall, the landscape metrics analysis suggests that new residential development has evolved in a sprawl pattern during the period of 2000-2010. Compared with the growth of low-density urban use (mainly residential) from 1973 to 1999 reported by Yang (2002), the rate of residential growth in the Atlanta metropolis between 2000 and 2010 was well below its historic pace since the 1970s. This residential development slowdown was also in line with the trend of decreasing population growth since 2000 according to the latest estimates from U.S. Bureau of Census. The current research has also revealed some evolving growth patterns in residential growth directions after 2000. In addition to the northern counties, new development was also quite visible in several southern counties, such as Henry and southern Fulton. The spatial growth of commercial/industrial land is shown in Figure 3.3b. In 2000, the commercial/industrial land occupied 68,166 hectares or 2.93% of the total land area for the 29- county Atlanta region (Table 3.4). Compared with the spatial distribution of residential land, commercial/industrial land shows a highly concentrated pattern, especially along the major interstate and state highways and suburban centers, as well as some other transportation routes in linear forms as shown in Figure 3.3b. From 2000 to 2010, commercial/industrial land increased to 76,016 hectares, or 3.27% of the 29-county area (11.52% increment). The new commercial/industrial development clearly followed the multiple transportation routes and the existing suburban centers. Based on the landscape metric measurements (Table 3.5), the increasing number of patches and total edges and the decreasing mean patch size indicate that commercial/industrial land had become more fragmented from 2000 to 2010. The decreasing mean nearest neighbor metric represents a less dispersed pattern, indicating that new development tended to infill around the existing developed area during the 10-year period. The mean perimeter/area ratio had increased between 2000 and 2010, suggesting that

53 commercial/industrial land patches had become less compact. Overall, a similar sprawling pattern has also been found for commercial/industrial land, developed in a more fragmented and less compact manner but mainly infilling around existing developed areas. The rate of commercial/industrial land growth between 2000 and 2010 also shows a slowdown trend compared with the past three decades that was reported by Yang (2002). The new development has followed the multinucleated pattern evolved in the past around suburban centers, mostly found within the 10-county region, such as the counties of Fulton, Henry, Gwinnett, Forsyth, Cobb, and Douglas. In 2010, commercial/industrial land was mainly concentrated within the 20- county region, showing no sign of new centers emerging in the outer-ring suburbs and exurbs. Figure 3.3c illustrates the spatial evolution of all other urban land including the area used for transportation, communication, utilities, and some rural constructions. In 2000, this category occupied 45,512 hectares or 1.96% of the total land area for the 29-county Atlanta region (Table 3.4). The spatial pattern of this category is characterized by linear features along transportation routes combined with some regional clusters around large transportation and utility facilities, such as the Hartsfield-Jackson Atlanta International Airport, several other regional airports and stations of the regional rail system. From 2000 to 2010, this category showed an increase of 8,889 hectares or 19.53% increment. In terms of the spatial distribution, the growth of all other urban land spread outward across the entire 29 counties, especially within the 20-county region, including the counties of Fulton, Henry, Paulding, Cherokee, and Bartow. The growth was largely contributed by the construction of several transportation facilities (e.g., the Paulding Northwest Atlanta Airport opened in Fall 2008) and urban facilities under construction. Landscape metric measurements reveal some differences between this category of urban use and the residential and commercial/industrial uses. Firstly, the increase in the mean patch size and the total edges indicates that the other urban land patches grew larger but also became more fragmented. The increase in the mean nearest neighbor indicates that a more dispersed pattern had evolved because the new development was dominated by many regional facility clusters scattered across the metropolis. While the mean perimeter/area ratio decreased between 2000 and 2010, implying that the construction of transportation and other facilities had become more spatially compact as the share of clustered features (e.g., facility centers) increased and linear features decreased (e.g., transportation routes) according to visual interpretation.

Based on Table 3.6 and Figure 3.5, the nature of land changes can be clearly revealed by the land use/cover conversion between classes occurred in the metropolis from 2000 to 2010. The growth of all three types of urban land consumed a large area of forest land in the metropolis. Among the total net addition of residential land, 63.7% came from the loss of forest land (C1) and 28.83% came from the loss of grass/shrub land (C3). The loss of agricultural land (C2) and fallow/barren land (C4) only contributed to 4.27 percent and 3.20 percent growth of residential land, respectively. For the growth of commercial/industrial land, 57.14% came from the loss of forest land (C5) and 30.16% came from the loss of grass/shrub land (C7). While the loss of agricultural land (C6) and fallow/barren land (C8) each contributed to 6.35% in the growth of commercial/industrial land. The majority net addition of other urban land category still came from the loss of forest land (C9) and grass/shrub land (C13), about 57.6% and 26.4%, respectively. The loss of agricultural land (C10) and fallow/barren land (C12) contributed to 6.4% and 9.6% in the growth of other urban land, respectively. Besides the urban land growth, the dynamic changes between forest and a few other vegetative classes are evident from the C13 and C14 category in Table 3.6. Between 2000 and 2010, the conversion from grass/shrub land and fallow/barren land to forest land (C13) accounted for 188,127 hectares or 8.08% of the total study area. And the conversion from forest land to grass/shrub land and fallow/barren land (C14) was about 36,776 hectares or 1.58% of the entire metropolis. These conversions represent the ecological and human-induced succession and disturbance within the forest ecosystem, such as clear-cut harvesting and forest regrowth. The spatial distribution of these conversions is clearly represented in Figure 3.4c and Figure 3.5. Overall, most of the conversion to urban classes (C1- C12) took place within the 20-county Atlanta region. Specifically, the loss of forest to urban uses (C1, C5, and C9) was widely spread in the counties of Cherokee, Forsyth, Gwinnett, Fulton, Cobb, Douglas, and Paulding. Although not dominant, the loss of agricultural land to urban land (C2, C6, and C10) and grass/shrub land to urban land (C3, C7, and C11) showed a concentrated pattern in several counties within the 20-county Atlanta region, as revealed by the decrease of agricultural land within the 20-county region as shown in Figure 3.4c. While the forest dynamics from 2000 to 2010 mainly occurred in a spread pattern scattered in the metropolis with larger clusters in the exterior suburban and remote rural areas, as revealed by the significant changes across the three regions illustrated in Figure 3.4c.

3.4 Conclusions

Rapid urban growth has been observed worldwide in the past several decades, which prompt concerns over the accompanied environmental issues and the degradation of economical sustainability. Monitoring land changes therefore can help provide valuable information for regional management and planning, as well as for a better understanding of the underlying socioeconomic and biophysical processes shaping the observed land changes in urban areas. Satellite remote sensing allows a retrospective, synoptic viewing of large regions. When integrated with other geospatial technologies, it can effectively assess land changes in urban areas in both the spatial and the temporal dimensions. This study has demonstrated the usefulness of integrating satellite imagery with GIS and landscape metrics for land change mapping and analysis in a complex urban environment. The method developed in this study was based on a thorough understanding of landscape features, sensor characteristics and information extraction techniques. A series of techniques were adopted in order to ensure accurate land use/cover classification from the multi-temporal data. The Landsat TM data were radiometrically normalized to minimize the influences of atmospheric conditions on spectral response between the time series. A stratified classification approach was adopted to derive the urban class and several other land cover classes in the study area. This approach was implemented with several procedures including landscape partition, multiple endmember spectral mixture analysis (MESMA), urban cover extraction, and supervised classification. Then, a GIS-based spatial reclassification procedure was employed to refine the urban class into specific urban land uses through the use of ancillary data. The accuracy assessment results confirmed that the proposed image processing procedure was effective. In addition, the combined use of post-classification comparison, GIS operations, and landscape metrics allows land change analysis to go beyond simple statistic description and into the characterization of the spatial characteristics and nature of land changes in a complex urban area. This study has extended the earlier work conducted by Yang (2002) who examined urban growth and landscape changes in the 13 counties within the Atlanta metropolitan area over the period of 1973-1999. This current work examines the changes in the entire metropolis (28 counties) after 2000. Until 2010, the majority of the developed area was found within the ARC's 20-county region with a limited extension into the outer-ring suburban and exurban areas. The landscape metrics measurements have revealed that a more fragmented and dispersed pattern had

56 emerged for the residential and commercial/industrial land from 2000 to 2010, suggesting that a sprawling development pattern still dominated the regional growth. In terms of the nature of change, the growth of all three types of urban land was achieved at the cost of forest and grass/shrub land. In addition to urban growth, significant changes were observed for the forest ecosystem, as indicated by the changes in forest landscape structure and patterns in the entire metropolis. While the method and techniques identified in this study have been quite efficient in analyzing the spatial characteristics and nature of land changes in the urban environment, there may be some limitations. For example, statistical analysis may have a limited power to explain the causality behind the observed changes. Correlation may imply causation but in some cases, such a relation may not be true. In this research, the integrated remote sensing techniques, GIS operations, and landscape metrics can help reveal the land change patterns in a quantitative form. Ancillary data, in situ observations and qualitative methods need to be incorporated in order to fully capture the underlying processes behind the observed land changes. These potential limitations open some room for further research.

CHAPTER FOUR

A SCALE-DEPENDENT ANALYSIS OF THE FACTORS DRIVING URBAN LAND USE CHANGES

This chapter examined the factors driving urban land use changes through a scale-dependent statistical analysis. Specifically, two types of urban land uses, namely, residential and commercial/industrial land use were taken as the dependent variables, which were to be associated with a set of potential explanatory factors through statistical analysis. To explore the effects of scale on the statistical analysis, this study was performed at three nested aggregation levels, i.e., counties, census tracts and block groups, and over three different spatial extents, i.e., the entire metropolis, the urban core area, and the outer suburban area. The results suggested that population density and location measures were among the most significant factors driving urban land use changes. Both aggregation levels and spatial extents influenced the statistical analysis results. The effects of aggregation have reduced variability, which led to stronger correlation and better fitted models at the coarse level (e.g., county). The different spatial extents also caused the detected driving factors to alter, which was an indication of spatial non-stationarity within the data set. Some limitations when using the exploratory statistical analysis to study land use change drivers have also been discussed, along with some areas for suggestions for future land use driving force research.

4.1 Introduction

Understanding the factors driving land use/cover change has been recognized as a key component for land change science and environmental sustainability research (Turner et al. 2007). While mapping and monitoring land use/cover changes can capture the spatial pattern of the changes, driving factor analysis can help further explain the processes behind the observed land changes, which can support land use decision-making for environmental sustainability. Turner et al. (1995) proposed a framework to examine drivers of land changes at three dimensions: socio-economic, biophysical, and proximate causes (e.g., land management). And the three dimensions need to be put into specific social contexts at various scales. Various factors have been documented as causes of land change in different contexts (Turner, Moss, and Skole 1993; Lambin et al. 2001; Geist and Lambin 2002; Waggoner and Ausubel 2002; Seto and Kaufmann 2003; Carrion-Flores and Irwin 2004). The candidate driving forces can be identified 58 into six categories: population, level of affluence, technology, political economy, political structure, attitude and values (Turner, Moss, and Skole 1993). The first three categories have also been linked to environmental change in the I = PAT framework which considers environmental impact (I) to be a function of population (P), affluence (A), and technology (T) (Commoner 1972). Natural factors, such as soil characteristics, climate, vegetation, and topography, may constrain land change. Therefore, a three-dimensional structure is modified from Turner et al. (1995) to represent the factors driving urban land use change (Figure 4.1).

Figure 4.1. Three dimensions of forces driving urban land use change (modified from Turner et al. 1995).

Analyzing driving forces of land change is challenged by several theoretical and methodological issues. One of the major issues relates to the choice of appropriate surrogate variables. Although the major categories and frameworks of candidate driving forces have been identified, it is still not certain which specific variables would be best represent the factors in each category (Turner, Moss, and Skole 1993). Especially the incorporation of human driving 59 forces has been hampered by data availability and collection levels. Another issue lies in the need to integrate biophysical and socioeconomic data that collected at different spatial units. The relevant spatial units for biophysical processes are essentially different from those for human decision-making (Martin 1996). The above two issues are further complicated by the issue of scale. The proximal, social and biophysical factors are multi-scalar in nature and their interactions over multiple scales together determine the emergent land system changes at different scales (Turner et al. 1995). Scale is one of the most fundamental issues in land change research (Verburg and Chen 2000; McConnell and Moran 2001; Veldkamp et al. 2001; Walsh et al. 2001). All scales consist of extent and resolution: extent refers to the size of a dimension, e.g. the size of the study area or the duration of time under consideration, whereas resolution refers to the precision used in measurement, i.e. grain size (Turner et al. 1989). The choices over scale, extent, and resolution can critically affect the observed patterns and processes. For example, patterns that occur at one level of resolution may be lost at lower or higher levels; patterns that appear over one extent of a dimension may be lost if the extent is increased or decreased (Allen and Hoekstra 1991; O‟Neill et al. 1991; Bergkamp 1995; Gibson, Ostrom, and Ahn 2000). Moreover, land change is essentially a multi-scale process, where the driving forces at multiple scales interacting with each other in driving pattern changes. As a result, no single scale can be regarded as “optimal” for studying the land use systems. It is one of the early steps to identify an appropriate scale (e.g., extent and resolution) for the analysis of spatial phenomena, such as land changes. Although there may be an optimal scale of analysis that can generate the highest predictability for a specific data set, this is usually not consistent through analysis with different locations or data sets (Briassoulis 2000). Given the wide recognition of the effects of scale on the observed patterns and processes, research has long been dedicated to examining the scale dependency in land change analysis. The factors driving land changes were found to vary across spatial scales. For example, Irwin (2006) has identified ten determinants of land use changes by the global, regional/metropolitan and neighborhood/parcel scales. Different studies have also shown that the grain size or resolution of analysis can influence the statistical analysis results between land use and its driving factors (e.g., Veldkamp and Fresco 1997; Walsh et al. 2001; Lo and Yang 2002). However, relatively few studies have paid attention on the effects of extent on land change driving forces (Verburg and Chen 2000), especially in the urban area. The effects of spatial

60 extent on the observed landscape patterns cannot be ignored given the heterogeneity of the urban environment (see Chapter 3). Therefore, examination of the effects of aggregation level (i.e., resolution) and spatial extent and their interactions may bring insights into the understanding of the complex urban land use change processes. The objective of this study was to test the hypothesis that the driving factors of land use changes are scale dependent at different aggregation levels and over different spatial extents in the study area. At the current stage, exploratory statistical analysis was performed to help examine the scale dependencies of the detected driving factors of land use changes. Correlation and multivariate regression analysis were both used to account for the potential drivers that may be omitted otherwise. Important factors driving residential and commercial land use change were identified based on the statistical analysis results. These analysis results were further examined to test their scale dependencies. The multi-scale analysis approach at different aggregation levels and over different spatial extents helped review the processes underlying urbanization and suburbanization which may be lost when using single-scale analysis. The following sections will describe the research methods and discuss the scale-dependent driving forces of urban land uses and the effects of scale on the analysis results.

4.2 Research methods

The research methods consisted of two major procedures: data preparation within GIS and statistical analysis of the driving factors for urban land use changes. The detailed procedures were illustrated in Figure 4.2.

4.2.1 Data preparation

The study reported here targeted the 20 counties in the Atlanta metropolitan area (Figure 1.2). And the land changes between 2000 and 2010 were observed with remote sensor imagery (see Chapter 3 for the details). Various biophysical and socioeconomic factors may be responsible for the land changes in urban areas. The candidate explanatory factors considered in this study were selected based on the three-dimension conceptual framework (see Section 4.1) and a review of the literature dealing with the driving forces of land change (Turner, Moss, and Skole 1993; Verburg and Chen 2000; Lo and Yang 2002; Carrion-Flores and Irwin 2004; Xie et al. 2005; Iwrin 2006; Zhu et al. 2010). A list of the candidate explanatory variables was presented in Table 4.1. The data used in this research consisted of land use data extracted from

Figure 4.2. A flowchart of the working procedure route. satellite images, socioeconomic variables assembled from the decennial census data, topographic variables derived from elevation data, and location variables measured from ancillary GIS data. A zonal approach was adopted to organize the spatial data from different sources. A zonal file for each aggregation level and spatial extent was first produced in vector format. This zonal file was then used to extract each variable in Table 4.1 from these data layers zone by zone by means of cross-tabulation. Therefore, a zone-based table with multiple data attributes for each aggregation level and spatial extent was generated for further statistical analysis. Table 4.2-4.4 displayed three sample tables prepared for residential land use proportions in 2000, 2010, and the changes at the county level, 20-county region. A total of 54 tables were generated at the different spatial extents and aggregation levels for the two types of land use proportions in 2000, 2010 and their changes.

Table 4.1. Description of land use types and candidate explanatory factors used in the analysis.

Variable Name Description Source Land use types Residential Percentage of total area used for residential Landsat Thematic uses Mapper (TM), LandPro Commercial Percentage of total area used for Landsat TM, LandPro commercial/industrial uses Demography and socioeconomics Population density Total population density (per hectare) Census Summary File (SF) 1 Income Per capita income (1000 US dollar) Census 2000 SF3, American Community Survey Topographic measures Terrain elevation Mean elevation (100 meters) National Elevation Dataset (NED) Elevation range Standard deviation of elevation (meter) National Elevation Dataset (NED) Slope Mean slope (percent) National Elevation Dataset (NED) Location measures Distance to urban centers Mean Euclidean distance to urban centers Atlanta Region (km) Information System (ARIS) Distance to highways Mean Euclidean distance to highways (km) Atlanta Region Information System (ARIS) Distance to node points Mean Euclidean distance to node (km) Atlanta Region Information System (ARIS) Distance to water Mean Euclidean distance to major water Atlanta Region bodies (km) Information System (ARIS) Distance to parks Mean Euclidean distance to parks (km) Atlanta Region Information System (ARIS)

4.2.1.1. Administrative/statistical boundaries. The boundaries for the 20 counties were extracted from the 2003 U.S. Census Bureau Tiger/Line files. The boundaries of 676 census tracts and 1894 block groups for 2000 were extracted from the 2000 Tiger/Line files. The boundaries of 948 census tracts and 2565 block groups for 2010 were extracted from the 2010 Tiger/Line files. All these boundary data have a nominal scale of 1:100,000. Note the boundaries and total number of census tracts and block groups have changed between 2000 and 2010. This is common for decennial census data as physical changes in street patterns and population

Table 4.2. A sample table prepared for residential land use in 2000 at the county level, 20-county region.

Population Distance Distance Land use density Per capita Terrain to urban Distance to to node Distance Distance proportion (per income elevation Elevation Slope centers highways points to water to parks County (%) hectare) (1000 US$) (100 m) range (m) (percent) (km) (km) (km) (km) (km) Fayette 7.701 0.297 2.950 2.665 15.308 5.211 3.296 1.164 2.155 5.183 2.071 Carroll 2.406 0.111 1.763 3.360 25.941 5.132 4.649 1.005 2.134 4.341 9.907 Rockdale 9.231 0.342 2.229 2.449 18.349 5.546 5.245 0.931 1.208 3.981 3.197 Cobb 17.105 1.132 2.788 3.067 21.423 6.460 3.154 0.898 1.864 6.165 1.351 Forsyth 5.777 0.278 2.913 3.469 22.076 8.050 5.443 1.352 2.499 6.352 2.589 Clayton 18.312 1.057 1.811 2.694 14.545 5.564 2.181 0.724 1.216 6.877 1.719 Henry 7.999 0.234 2.294 2.457 19.746 5.279 4.121 1.146 1.528 6.396 2.497 Bartow 2.305 0.105 1.903 2.462 27.806 5.408 4.430 1.627 2.415 5.768 5.503 Newton 5.355 0.144 1.932 2.241 17.749 4.997 5.693 1.050 2.362 3.945 5.959 Gwinnett 17.900 0.864 2.503 3.039 22.719 6.702 2.634 0.974 1.582 6.969 2.161 Fulton 11.062 0.968 3.035 3.006 26.514 6.639 2.144 0.681 1.205 6.061 0.973 Hall 3.116 0.224 1.971 3.530 31.232 8.485 5.643 0.910 1.720 11.623 2.389 Spalding 4.164 0.188 1.681 2.651 18.284 4.142 4.551 1.379 2.341 9.192 2.781 Douglas 6.104 0.291 2.117 3.152 25.074 5.815 4.000 1.178 1.576 4.167 1.963 Coweta 3.090 0.128 2.195 2.720 15.436 4.948 3.463 1.335 2.549 9.615 4.051 DeKalb 17.218 1.567 2.404 2.804 25.274 5.960 2.686 0.657 1.194 4.433 0.882 Cherokee 3.410 0.213 2.489 3.086 26.819 8.587 2.955 1.752 2.400 4.282 2.012 Walton 4.158 0.118 1.947 2.663 18.249 5.391 3.480 1.218 2.525 13.349 8.431 Barrow 5.027 0.182 1.838 2.812 21.032 5.493 2.477 0.963 1.464 6.940 2.235 Paulding 4.238 0.166 1.996 3.169 30.854 6.218 5.992 1.646 2.408 6.402 4.060

Table 4.3. A sample table prepared for residential land use in 2010 at the county level, 20-county region.

Population Distance Distance Land use density Per capita Terrain to urban Distance to to node Distance Distance proportion (per income elevation Elevation Slope centers highways points to water to parks County (%) hectare) (1000 US$) (100 m) range (m) (percent) (km) (km) (km) (km) (km) Fayette 11.444 0.347 3.508 2.665 15.359 5.308 3.288 1.162 2.138 5.206 2.114 Carroll 3.804 0.140 2.052 3.365 26.007 5.422 4.685 1.065 2.167 4.365 9.764 Rockdale 12.869 0.415 2.437 2.436 18.218 5.829 5.312 0.913 1.200 3.884 3.145 Cobb 27.501 1.281 3.311 3.058 22.178 6.693 3.238 0.946 1.902 6.232 1.363 Forsyth 11.108 0.495 3.539 3.460 22.538 8.656 5.534 1.361 2.531 6.121 2.510 Clayton 25.389 1.159 1.896 2.693 14.667 5.617 2.188 0.712 1.225 6.771 1.743 Henry 12.366 0.400 2.577 2.460 19.904 5.339 4.186 1.150 1.536 6.508 2.603 Bartow 3.549 0.138 2.224 2.500 29.908 6.097 4.532 1.568 2.391 5.868 5.410 Newton 7.370 0.232 2.158 2.228 17.870 5.287 5.795 1.081 2.359 3.801 5.906 Gwinnett 26.792 1.183 2.690 3.044 23.314 6.925 2.693 1.018 1.622 7.138 2.189 Fulton 18.520 1.105 3.721 2.969 27.231 6.848 2.273 0.705 1.247 6.239 1.073 Hall 4.919 0.289 2.368 3.494 32.697 9.141 5.835 0.963 1.764 11.169 2.441 Spalding 5.244 0.206 1.961 2.644 18.466 4.274 4.564 1.416 2.361 9.222 2.744 Douglas 11.952 0.418 2.452 3.123 26.528 6.327 4.164 1.254 1.611 4.010 2.197 Coweta 5.274 0.183 2.616 2.706 15.725 5.347 3.504 1.424 2.600 9.830 4.160 DeKalb 25.165 1.635 2.841 2.802 25.635 6.071 2.643 0.649 1.193 4.307 0.895 Cherokee 6.718 0.321 3.022 3.109 28.285 9.341 3.286 1.608 2.401 4.227 2.040 Walton 5.243 0.163 2.252 2.659 18.378 5.578 3.447 1.235 2.560 13.486 8.409 Barrow 6.825 0.274 2.088 2.787 21.595 5.885 2.745 0.965 1.486 6.899 2.437 Paulding 8.885 0.288 2.345 3.148 32.942 7.027 6.222 1.688 2.392 6.215 3.916

Table 4.4. A sample table prepared for residential land use changes between 2000 and 2010 at the county level, 20-county region.

Population Distance Distance Land use density Per capita Terrain to urban Distance to to node Distance Distance proportion (per income elevation Elevation Slope centers highways points to water to parks County (%) hectare) (1000 US$) (100 m) range (m) (percent) (km) (km) (km) (km) (km) Fayette 3.743 0.050 0.558 0.000 0.050 0.097 -0.008 -0.002 -0.017 0.023 0.043 Carroll 1.398 0.029 0.289 0.005 0.066 0.290 0.036 0.060 0.033 0.024 -0.143 Rockdale 3.638 0.073 0.208 -0.013 -0.131 0.283 0.067 -0.018 -0.008 -0.097 -0.052 Cobb 10.396 0.149 0.523 -0.009 0.755 0.233 0.084 0.048 0.038 0.067 0.012 Forsyth 5.331 0.217 0.626 -0.009 0.462 0.606 0.091 0.009 0.032 -0.231 -0.079 Clayton 7.077 0.102 0.085 -0.001 0.122 0.053 0.007 -0.012 0.009 -0.106 0.024 Henry 4.367 0.166 0.283 0.003 0.158 0.060 0.065 0.004 0.008 0.112 0.106 Bartow 1.244 0.033 0.321 0.038 2.102 0.689 0.102 -0.059 -0.024 0.100 -0.093 Newton 2.015 0.088 0.226 -0.013 0.121 0.290 0.102 0.031 -0.003 -0.144 -0.053 Gwinnett 8.892 0.319 0.187 0.005 0.595 0.223 0.059 0.044 0.040 0.169 0.028 Fulton 7.458 0.137 0.686 -0.037 0.717 0.209 0.129 0.024 0.042 0.178 0.100 Hall 1.803 0.065 0.397 -0.036 1.465 0.656 0.192 0.053 0.044 -0.454 0.052 Spalding 1.080 0.018 0.280 -0.007 0.182 0.132 0.013 0.037 0.020 0.030 -0.037 Douglas 5.848 0.127 0.335 -0.029 1.454 0.512 0.164 0.076 0.035 -0.157 0.234 Coweta 2.184 0.055 0.421 -0.014 0.289 0.399 0.040 0.089 0.051 0.215 0.109 DeKalb 7.947 0.068 0.437 -0.002 0.361 0.111 -0.043 -0.008 -0.001 -0.126 0.013 Cherokee 3.308 0.108 0.533 0.023 1.466 0.754 0.331 -0.144 0.001 -0.055 0.028 Walton 1.085 0.045 0.305 -0.004 0.129 0.187 -0.033 0.017 0.035 0.137 -0.022 Barrow 1.798 0.092 0.250 -0.025 0.563 0.392 0.268 0.002 0.022 -0.041 0.202 Paulding 4.647 0.122 0.349 -0.021 2.088 0.809 0.230 0.042 -0.016 -0.187 -0.144

66 growth/decline may require boundary revisions. To deal with the spatial inconsistency of areal units for the two years, it involved an areal interpolation procedure to redistribute the Census 2000 demography and socioeconomic data with respect to the Census 2010 areal units (see Section 4.2.1.3). The land use data and other explanatory variables for both years were also extracted using the geographic boundaries of the 2010 census tracts and block groups. 4.2.1.2. Land use data. For the purpose of this study, two types of urban land uses, namely residential land uses and commercial/industrial land uses, were analyzed and taken as the dependent variables to be associated with the candidate explanatory variables. Land use/cover data for 2000 and 2010 were produced from Landsat Thematic Mapper (TM) imagery with 30 m spatial resolution using a stratified classification procedure described by in Chapter 2. The ARC's LandPro GIS databases were used to further derive specific urban land uses as described in Chapter 3. The ARC's LandPro GIS databases were created by on-screen photo-interpretation and digitizing of orthorectified aerial photography with 1 m pixel resolution for 2001 and 1.64 foot resolution for 2010. Each land use class was converted to a binary layer with value of 1 for the land use and value of 0 for the background. Then the proportion of each urban use was calculated for each areal unit at the level of counties, census tracts and block groups. The distribution of land uses was also used to extract the respective topographic and location measures that will be discussed in Section 4.2.1.4 and Section 4.2.1.5. 4.2.1.3. Population and income data. Population and income were derived from the 2000 and 2010 census data. The population data for both years were obtained from the summary file 1 (SF1) of Census 2000 and 2010 that presents 100 percent population and housing figures. Total population was normalized by dividing the number of population by the area of each spatial unit. The income data for 2000 can be retrieved from the summary file 3 (SF3) of Census 2000, which is also known as the long form for Census that contains detailed tables of socioeconomic characteristics. However, the decennial SF3 in 2010 has been replaced by the American Community Survey (ASC). The income data for 2010 were then retrieved from the ASC 5-year estimates, which is available at census-tract and block-group levels and comparable to the corresponding Census 2000 SF3 table. A dasymetric mapping procedure was employed to the Census 2000 data in order to deal with the spatial inconsistency between the two years‟ census boundaries (Langford and Unwin 1994). As an areal interpolation technique, dasymetric mapping basically involves transforming

67 spatial data collected in one set of areal units to another set of areal units. Specifically, it represents the process of disaggregating spatial data collected at some aggregation level (usually arbitrary) into a map that more accurately depicts the distribution of the data. Then the data can be re-aggregated to a set of desired areal units to enable areal interpolation (Holt and Lu 2011). The dasymetric mapping technique was adopted here to measure the population density from census population data and residential land use data in 2000. It was assumed that the distribution of residential land use can better represent the extent of populated areas than the census aggregation unit. Firstly, the population density layer was produced by dividing the population data by the area of residential land use within each block group. This procedure was based on the assumption of homogeneous distribution of population within each areal unit (i.e., block group). A portion of the derived dasymetric map of population density for 2000 by block group is illustrated in Figure 4.3.

Figure 4.3. Dasymetric map of population density by block groups 2000. Only a portion (approximately 31 × 31 km) of the entire map is shown here for illustration purpose.

With the population density layer, it was possible to re-aggregate the population data in 2000 within each set of the Census 2010 aggregation areal units. To transform the income data, the total income for the Census 2010 areal units can be firstly calculated by associating the per capita income with the population. The total income was then divided by the total population within each Census 2010 areal unit to calculate the per capita income. 4.2.1.4. Topographic measures. Three topographic measures, namely, terrain elevation, elevation range, and slope, were also included as the candidate driving factors. Topographic measures can be related to land use suitability which were found to be important factors of land use distributions (Verburg and Chen 2000; Lo and Yang 2002; Zhu et al. 2010). These topographic measures were derived from the National Elevation Dataset (NED) with a 30 m pixel size (Figure 4.4a). The NED is a raster product assembled by the U.S. Geological Survey (USGS) that provides national elevation data in a seamless form. The initial data source of this dataset is the USGS 7.5-minute digital elevation model (DEM) data, which were edge matched and mosaicked by the USGS. The value at each pixel represents the elevation in meters. The three topographic measures were extracted with respect to the distribution of the two urban land use types in 2000 and 2010, respectively. The mean and standard deviation of elevation values were summarized for each areal unit to represent the mean elevation and elevation range, respectively. Slope (in percentage) was calculated from the NED data using the Spatial Analyst Tools built within ArcGIS (ESRI 2010) (Figure 4.4b). Mean slope was summarized for each zone as mean slope gradient.

Figure 4.4. The elevation and slope layers to be associated with the distribution of land uses for the extraction of topographic measures at each set of areal units. 69

4.2.1.5. Location measures. Distance to urban centers, highways, node points, parks and waters were generated from various sources to represent the location measures. While the first three measures can represent the accessibility of land uses to major urban services and facilities, the last two may represent the aesthetic values of each location. A Euclidean distance surface was calculated for each data layer and used as location measures. The mean distance value was then extracted with respect to the distribution of the two urban land use types in 2000 and 2010, respectively. Figure 4.5 displayed the five layers of distance surface in 2000 to be associated with the land use distribution for further extraction at each set of areal units.

Figure 4.5. The layers of distance surface to be associated with the distribution of land uses for the extraction of location measures at each set of areal units.

The data used for deriving location measures were from the Atlanta Region Information System (ARIS) GIS data sets in vector formats published by the ARC. And they were further adjusted and modified with reference to the 2000 and 2010 satellite imagery to add additional features and correct spatial inconsistencies. Urban centers were digitized based on orthophotography of traditional municipal downtowns and significant regional centers that was

70 developed by ARC's Land Use Planning Division. There are more than 100 such regional and town centers across the 20-county region. The highway layer is a subset of the Georgia Department of Transportation‟s (GDOT) DLGF street centerline database, which consists of all interstate and state highways, as well as a number of additional roads that were identified by ARC‟s Transportation Planning Division (TPD) as major roads. Node points represent highway exits and junctions where major highways run across. These node point data were extracted from the above-mentioned street centerline data. Park and water proximity can be favored by residential uses as major aesthetic attractiveness. The park data layer was created by a joint effort of the ARC's Land Use Planning Division in coordination with various planning partners, including Georgia DNR, local GIS inventory of parks and local governments. The water layer is a subset of GDOT‟s statewide Georgia DLG-F Linear Hydrographic Features dataset which contains hydrographic features including lakes, ponds, reservoirs, swamps, and islands.

4.2.2 Statistical analysis

4.2.2.1. Spatial analysis scheme. To explore the scale dependencies of the driving factors for land use changes, the analysis was performed at multiple scales, including various spatial extents and aggregation levels. The effect of aggregation levels was assessed at three nested administrative/statistical levels: counties, census tracts and block groups. These are the aggregation levels where population and income data are mostly available. The effects of spatial extent was also analyzed at the three ARC-designated regions, including the ARC 20-county region as a whole, and the subdivision with the ARC 10-county region and the outer 10-county region (Figure 1.2). The regional subdivision was used to represent the outward expansion of the Atlanta metropolitan area. Statistical analysis was employed to explore how the driving factors of land use changes would change at varying aggregation levels and spatial extents. 4.2.2.2. Statistical methods. Statistical techniques can be used to explore the scale dependency of the factors driving urban land use change. Both correlation analysis and multivariate regression analysis were performed for interpretation purposes because: (1) it is possible for several independent variables to be individually correlated with a dependent variable, but not all of them are statistically significant in the same multivariate regression model; and (2) it is also possible for a variable to become significant only when other variables are accounted for. The statistical analysis was performed with land use proportions and their changes as the dependent variable and a group of explanatory factors as independent variables for 2000 and 71

2010. The data have been logarithmically transformed before the statistical analysis to allow for non-linearity. Firstly, simple correlation analysis was applied to determine whether an association between land use patterns and all the individual explanatory variables exists. Pearson correlation has been mostly often used to discover the associations between driving factors and land use (e.g., Verburg and Chen 2000; Lo and Yang 2002), which tells both the strength and direction (i.e., positive or negative) of the significant associations between pairs of variables. Pearson correlation is a parametric test built upon the assumption of linearity, normality, and homoscedasticity. To avoid bias, a visual inspection of the histograms and scatter plots has been performed for these variables to check the above assumptions. The correlation coefficients were considered statistically significant at the 0.01 level (2-tailed) or 0.05 level (2-tailed). If the observed significance level is less than 0.05 or 0.01, the null hypothesis of no association is rejected. Multivariate regression was then used to further derive comprehensive models that can describe the patterns of land uses and their changes as a collective function of the explanatory variables. When using real world spatial data, multicollinearity is a common problem that some of the explanatory variables can be highly correlated. Therefore, the stepwise regression method was used to select variables that yield a significant contribution to explain the land use patterns and their changes. The stepwise procedure selects variables that can pass the tolerance criterion, specified as the probability of F (the square of t value) less than 0.050. A variable was also excluded if it would cause the tolerance of another variable already in the regression model to drop below the tolerance criterion, specified as the probability of F larger than 0.100. All entered variables with tolerance larger than the specified level were removed from the model. For the regression output, two types of measures were recorded for each regression model and the selected independent variables in the models. The adjusted coefficient of determination (adjusted R2) was recorded as a measure of goodness of fit that represents the variation of dependent variable explained by the regression model. The standardized coefficients (beta) indicate the number of standard deviation changes in the dependent variable associated with a standard deviation change in the independent variable. Therefore, the standardized coefficients were used as a measure of the relative contribution of a variable in a regression model.

A common issue when performing statistical analysis using spatial data is the phenomena of spatial autocorrelation. Conventional statistical methods, including linear regression, assume the observations to be statistically independent and random (Cliff and Ord 1981). The presence of positive autocorrelation may overestimate the significance of statistical test, and vice versa (Legendre and Legendre1998). To account for the potential influences of spatial autocorrelation on the regression analysis, a structured sampling method was conducted by fitting the regression model with randomly sampled observations. And the experiments did not show significant alteration of the analysis results for these datasets.

4.3 Results of driving force analysis

This section discussed the major findings from the statistical analysis of the driving factors for residential and commercial land uses and their changes. The potential implications in the contexts of urban growth were also examined. Note only the most relevant results were presented here for illustration purposes.

4.3.1 Residential land use

Population density was found to be the most important factor being positively correlated with residential land use proportions and the changes. The correlation coefficients ranged from 0.457 at the census tract and block group levels to 0.953 at the county level for all of the three spatial extents. For per capita income, it was only significantly correlated with residential land use proportion at the two subdivided extents, but with opposite signs. Specifically, negative correlation coefficients were detected around -0.1 for the inner 10-county region, while positive correlation coefficients ranging from 0.272 to 0.692 were found for the outer 10-county region. The opposite patterns found in the two regions reflected the different patterns in housing prices between the central urban area and the surrounding suburban area. The association between topographic measures and residential land use was stronger at the finer levels (i.e., census tract and block group levels) where more topographic variations can be revealed. At finer scales, the elevation range was found to be negatively correlated with residential land use proportions for both years across the three spatial extents, with correlation coefficients ranging from -0.420 to -0.276. Slope was also found to be negatively correlated with the residential land use proportion, but with lower strength. Mean elevation was positively

73 correlated with the residential land use proportion in the inner 10-county region with correlation coefficients around 0.1. As for the location measures, distance to highways, node points and urban centers were found to be negatively correlated with the residential land use proportion for both years. Stronger associations between distance to highway and residential use proportion were observed at the county level in the inner 10-county and the entire 20-county region (-0.630 to -0.765) for both years. This observation suggested the role of highway proximity in residential development over these years. Distance to parks was also found to be negatively correlated with residential land use proportions at both county and census levels, which may be considered as a primary attractiveness for housing location. However, the association was not significant for the inner-10 county region. The association between distance to water and residential use proportion was weak and displayed opposite signs for the 20-county (negative) and inner 10-county (positive) region at the finer aggregation levels. It can be difficult to interpret the results from multivariate regression since even very little correlation among independent variables may change the relative importance of the parameters. However, the importance of population density to residential land use can be further confirmed by the regression analysis, which has been included as an independent variable in almost every regression model with the largest standardized coefficient. In addition, regression equations that included population and income, and at least one topographic measure and one location measure as the independent variables generally can explain more variance of the residential use proportions, as indicated by higher value of the adjusted R2.

4.3.2 Commercial/industrial land use

Population density was also found to be an important factor for commercial land use. However, the association strengths declined more sharply at the finer tract and block group level (from 0.9 to 0.2 approximately). This observation can be related to the clustered patterns of commercial use, which caused more variations in land use proportion at finer level. Per capita income showed a slightly negative correlation (around -0.2) with the commercial use proportion at the census tract and block group levels across the three extents. However, a strong positive correlation (0.716) was detected at the county level for the outer 10-county region (primarily suburban area) in 2010.

Mean elevation displayed a slightly positive correlation (around 0.15) with the commercial land use at the finer aggregation levels, indicating that the commercial land use tended to develop at locations with higher elevations. The elevation range showed negative correlation (around -0.12) with the commercial land use at the census tract level and positive correlation (around 0.1) at the block group level. This indicated that the elevation range for commercial land use was a function of distance. Slope also showed a positive correlation with the commercial land use (0.632) at the county level for the outer 10-county region in 2010. Location measures, especially the accessibility to urban centers, highways and node points, were found to be significant at all aggregation levels and across all extents for commercial land use. All the three location measures were negatively correlated with the proportion of commercial use. Proximity to urban centers can be explained by the agglomeration patterns for developing commercial land use. Highways were the major transportation corridors that support the functioning of commercial/industrial uses. Node points represented highway exits and junctions where major highways run across, which are favorable sites for commercial and industrial activities to locate. When looking at the results from regression analysis, the inclusion of location measures of proximity to urban centers, highways and node points has contributed to explaining more variance for the commercial land use proportion for both years.

4.4 Discussion

4.4.1 Effects of aggregation levels

The effects of spatial scale were firstly assessed for the nested aggregation levels. Firstly, it was clear that the detected correlation coefficients for the driving factors change at different aggregation levels. Generally, correlation coefficients increase at the coarser aggregation level (e.g., counties). Figure 4.6 illustrated the correlation coefficients (in absolute value) between the commercial land use proportion and the significant factors for the 20-county region in 2010. The increase in correlation coefficients may be caused by the fact that aggregation tends to suppress the spatial variability. However, the increasing slope across the three aggregation levels was quite different for each variable. For example, the coefficient for population density increased rapidly with increasing aggregation level, while it only showed a slight increase for the distance to node points and to roads. This was due to the differences in the spatial variability of these variables and the distances over which these variables affects land use. While similar patterns

75 were observed for residential land use, the change in correlation strength was larger for the commercial land uses. This can be explained by the spatial agglomeration of commercial land use patterns that led to high spatial variation at finer levels. With the increasing size of the aggregation units, more commercial land uses falls within the same areal unit as the population, leading to stronger correlation.

Figure 4.6. Correlation coefficient (in absolute value) between commercial land use proportions and five explanatory variables at three aggregation levels for the 20-county region for 2010.

Similar patterns were also found in the regression analysis. The goodness of fit for multiple regression models (judged by adjusted R2) was higher at more aggregated levels. Table 4.5 displayed the results of the stepwise regression models that explain the residential land use proportion at different aggregation levels for the 20-county region for 2010. The overall explanation of the residential use proportion was very high at the county level (adjusted R2 = 0.902). In fact, several location measures were also negatively correlated with the residential use proportion at the county level, but population density was the only variable entered in the stepwise regression model. Increasing numbers of factors collectively contributed to the explanation of residential land use proportions at finer census tract and block group levels. However, the overall explanation of residential use proportion at finer levels remained low compared to the county level. This can be explained by the increasing spatial variability and

76 heterogeneity at the finer levels which goes beyond the explanatory power of the statistical regression analysis (Kolasa and Rollo 1991).

Table 4.5. Results of the stepwise regression models explaining the proportion of residential land use at different aggregation levels for the 20-county region in 2010.

Independent variable Standardized beta* Block group level (n = 2565) adj-R2 = 0.372 Population density 0.466 Elevation range -0.210 Distance to park -0.134 Distance to highway 0.049 Slope 0.047 Tract level (n=948) adj-R2 = 0.397 Population density 0.457 Elevation range -0.189 Distance to park -0.178 Distance to highway 0.089 County level (n=20) adj-R2 = 0.902 Population density 0.953 *significant at p<0.0001

4.4.2 Effects of spatial extents

The effects of spatial scale on the analysis results were further examined over different spatial extents. The ARC‟s 20-county region under study was further subdivided into an inner 10-county region (i.e., ARC‟s 10-county region) and an outer 10-county region. The subdivision can be taken as a simplified representation of the inner urban and suburban region for the entire metropolitan area, although the inner 10-county region tends to be more spatially heterogeneous. Based on the correlation analysis, the factors that were significantly correlated with the urban land use proportions tended to change across different spatial extents. Even for the same factor, the strength and direction of their correlation with land use can change across different spatial extents. For example, per capita income tended to be negatively correlated (-0.1) with the residential land use proportion over the inner 10-county region. Their correlation became

77 positive ranging from 0.15 to 0.69 for the outer 10-county region. The compact land use pattern and lower home price found in the inner urban area may collectively contribute to the observed negative correlation between the residential land use and income. And the large lot size and higher price of suburban housing may help explain the positive correlation observed in the outer region. This varying correlation results between the urban land use proportions and the explanatory variables was also an indication of the spatial non-stationarity in the dataset. In this example, the opposite correlation relationships found in the subdivided regions caused the correlation less significant for the entire region. The effect of spatial extents was further confirmed by the regression analysis. Table 4.6 showed the results of the stepwise regression models explaining the proportion of residential land use over different spatial extents at the tract level for 2010.

Table 4.6. Results of the stepwise regression models explaining the proportion of residential land use over different spatial extents at tract level (n=948) in 2010.

Independent variable Standardized beta* 20-county region adj-R2 = 0.397 Population density 0.457 Elevation range -0.189 Distance to parks -0.178 Distance to highways 0.089 Inner 10-county region adj-R2 = 0.271 Population density 0.429 Elevation range -0.227 Distance to nodes 0.168 Elevation 0.085 Distance to parks 0.071 Outer 10-county region adj-R2 = 0.763 Population density 0.847 Income 0.281 Elevation -0.192 Distance to water -0.099 *significant at p<0.0001

Firstly, the variables entered in the models vary across different extents at the same aggregation level (i.e., census tract). For example, per capita income only appeared to be important over the outer 20-county region, which is probably due to the higher housing price in the suburban area as discussed earlier. In addition, the common factors for all the models (e.g. population density) also varied in their relative importance in explaining the residential land use proportion, as judged by the standardized beta. Moreover, the model for the outer 10-county region had the highest adjusted R2 value, suggesting that various variables can collectively explain more variance of the residential land use in the outer suburban area. This may be explained by the fact that the outer region consisted of mainly low density suburban area which are relatively homogeneous compared with the inner region with a mixture of high density and low density development.

4.5 Conclusions

This chapter explored the factors driving urban land use changes using a scale-dependent statistical analysis. Specifically, a set of socioeconomic and biophysical factors were associated with two types of urban land uses, namely, residential and commercial land use using statistical analysis. Overall, population growth and location measures were found to be among the most significant factors for both residential and commercial land use change at varying levels and extents. This study focused on some generic factors driving urban land use changes. However, there may be other factors important for residential and commercial land uses in the Atlanta metropolitan area. For example, race and ethnicity are factors driving social segregation which will affect the urban and land use patterns. School district and quality may influence the location choices of residents and therefore need to be considered for analyzing residential land uses in future studies. Some other socioeconomic variables, such as household characteristics (e.g., age, composition), home values also need to be modeled to derive a comprehensive understanding of the driving factors of land use changes in the study area. To test the effects of scale on the analysis results, this study was performed at three nested aggregation levels, namely, counties, census tracts and block groups, and over three different spatial extents that subdivide the 20-county region into an inner 10-county and an outer 10-county region. Both aggregation level and spatial extent influenced the statistical analysis results. The effects of aggregation have reduced variability, which has reaffirmed previous findings in the literature (Kolasa and Rollo 1991; Verburg and Chen 2000; Lo and Yang 2002). 79

The different spatial extents also led to changing patterns, which was an indication of spatial non-stationarity in the dataset. Further analysis is needed to better understand the spatially varying patterns using a local form model, such as the Geographically Weighted Regression (GWR). The scale-dependent analysis has suggested that characteristics of larger units are not simple combinations of attributes of smaller units, which was an indication of complexity (i.e., emergent properties). This study explored the driving factors of urban land use change with statistical exploratory methods at the current stage. However, there are several limitations when using conventional statistical methods for spatial land use analysis. Firstly, statistical exploratory analysis, such as the stepwise regression, can be used in the early stages in view of data mining for the purposes of theory development. However, such exploratory analysis can be limited when we are more concerned with theory testing. Further research is needed towards a more comprehensive understanding of the factors driving urban land use changes. Moreover, the selected candidate explanatory variables can explain the majority of variations of land use change, especially at the coarser level and in more homogeneous regions. However, there may be other variables that can account for the land use changes at the finer scales and in more heterogeneous environments which are not generally available. In addition, statistical analysis does not always reveal causal relationships. Although some novel approaches has been developed to study the causality in complex systems using time-series data, they are not well suited for analyzing spatial data (Sugihara et al. 2012). In addition to the quantitative methods, more location specific factors need to be derived and accounted for, which can be based on field surveys, interviews, and other qualitative approaches. Finally, land change is essentially a multi- scale process, where the driving forces at multiple scales interacting with each other in driving pattern changes. In addition to the study on scale dependency, the multi-scale interactions and feedbacks of socioeconomic and biophysical processes in the land use systems can be incorporated using more advanced techniques, such as agent-based models, as factors driving human decision-making.

CHAPTER FIVE

LAND CHANGE MODELING: STATUS AND PROSPECT

Over the past several decades, various land change modeling approaches have been developed. They provided insights into the functioning of land changes at aggregated and individual levels, across various spatiotemporal scales, as well as in human, natural, or the coupled systems. This chapter surveyed recent research publications and examined several frequently used land change modeling approaches including statistical regression models, artificial neural networks, Markov chain models, cellular automata, economic models, and agent-based models. For each modeling approach, the theoretical and methodological basics and their relative strengths and weaknesses were examined. Moreover, several important theoretical and methodological issues in land change modeling were discussed: (i) coupling human-environment systems, (ii) scale dependency and multilevel interactions, and (iii) temporal dynamics and complexity. Finally, a review of research that integrates land change modeling with other environmental analysis and modeling techniques for studying global environmental change was provided.

5.1 Introduction

Among the four focus areas of land change science (Turner et al. 2007), the development of land change modeling is closely related to the other research dimensions for global environmental change and sustainability science. Firstly, the advances in land observation technologies, especially in remote sensing, provide a valuable data source for analyzing the dynamics of land cover/use changes that can enrich the theories and data input for model calibration and validation. Over the past decade, more remotely sensed data became available at finer spatial, spectral, radiometric, and temporal resolutions, or with novel instruments (e.g., LiDAR). Improved image processing technologies (e.g., sub-pixel analysis, object-based analysis) also offer the potential to advance land change modeling. Meanwhile, the simulated land change information using modeling techniques can in turn facilitate the spectral-based monitoring of land changes from satellite imagery (e.g., Liu and Cai 2012; Jin and Mountrakis 2013). Furthermore, land change modeling also supports the research on the coupled human- environment systems in terms of the causes and impacts of land changes. Various modeling techniques have been employed to determine the potential drivers of land changes either at the

81 scale of human-environment systems as a whole based on statistical modeling or at finer scale using agent-based models to represent individual decision making (e.g., Hu and Lo 2007; Ligmann-Zielinska 2009). The development of land change models for simulating multilevel interactions and feedbacks may improve our understanding of the processes underlying land change. In order to better understand the impacts and feedbacks of land change to other components of earth systems, it is essential to integrate land change modeling with other environmental modeling or to produce outputs that are useful for land change impact assessment. Finally, the use and integration of models will lead to a comprehensive understanding of the complexity of the coupled human-environment systems (i.e., synthesis and assessment issues). In the context of global environment change and sustainability science, increasing concerns have been given to research on sustainability that can inform practice and decision making in planning and management domains. Over the past several decades, various modeling approaches have been developed, which provide insights into the functioning of land changes at aggregated and individual levels, across various spatiotemporal scales, as well as in human, natural, and the coupled systems. Models allow us to link human behaviors with landscape patterns for simulating the processes of land changes in the past (e.g., Clarke, Hoppen, and Gaydos1997; Xie, Batty, and Zhao 2007), for forecasting future landscape dynamics under different scenarios (e.g., Yang and Lo 2003), and for informing decision-making towards sustainable land and resource management (see Section 5.4). Meanwhile, there are numerous theoretical and technological challenges for the modeling of land change given the complexity of the coupled human and environmental systems (Rindfuss et al. 2004). Advances in geospatial theories, technology and data provide great opportunities for addressing various challenges and for developing the next generation of land change models. With increasing demand for better assessing the impacts of land-use/cover change to the earth system, or simply for gaining insights into the land systems, it is timely to evaluate the progress and prospect of these modeling approaches and their contributions to land change science for global environmental change studies. This chapter examined several frequently used land change modeling approaches including statistical regression models, artificial neural networks, Markov chain models, cellular automata, economic models, and agent-based models. For each modeling approach, the theoretical and methodological basics and the relative strengths and weaknesses were discussed with selected examples. Some outstanding issues in land change modeling as for

82 the study of coupled human and natural systems were further identified and discussed. Research that couples land change modeling with other environmental modeling to study were also reviewed. For the purpose of this study, a collection of articles was assembled through two steps. The first step was a search on Web of Science using the Keywords: (Topic = “land change” or “land use change” or “land cover change” or “land use and land cover change” or “urbanization” or “urban growth” or “urbanization” or “deforestation” or “farmland”) AND (Topic = model or simulation). These keywords were included in the search criteria to encompass the different terms found in the literature. The online search led to 696 results which were further refined to exclude articles beyond our scope on spatially-explicit models of land change patterns and processes. In other words, articles focusing on non-spatial models or integrated environmental models were excluded from the publication library. As a result, 246 articles between 1994 and 2014 were sorted out at this first step. The second step was complementary to the online search through assembling publications from the author‟s personal archive since 2000 which includes articles that are not in the database on Web of Science. At this step, 23 publications were added to complete the publication library. The relatively small number of newly added publications also indicated the efficiency of the online search criteria. Therefore, a total of 269 publications between 1994 and 2014 were finally included in the publication library for further analysis. An overview of these publications clearly reveals the interdisciplinary nature of land change modeling research, with contributions from geospatial information science, remote sensing, computer science, economics, environmental science, and planning communities. The top journals (with number of related papers in parenthesis) were: International Journal of Geographical Information Science (21), Ecological Modelling (13), Environment and Planning B-Planning & Design (13), Environmental Modelling & Software (12), Agriculture Ecosystems & Environment (10), Landscape and Urban Planning (9), Computers Environment and Urban Systems (8), Photogrammetric Engineering and Remote Sensing (8), Journal of Environmental Management (8). On the temporal dimension, the publications in land change modeling have increased dramatically since the new century, with an explosion in the recent few years (Figure 5.1). Note that the publications in 2014 (12 publications by the end of May) was not included in this chart as more papers are incoming.

Figure 5.1. Number of publications on land change modeling between 1994 and 2013 based on the search criteria. An exponential trend line is shown in gray.

5.2 Land change modeling approaches

5.2.1 Statistical regression model

The basic structure of statistical regression model is based upon empirical analyses that link between land use and land cover changes (i.e., dependent variable) and a set of environmental and socio-economic explanatory variables. The derived relationships are usually used to generate maps of land transitional probability to predict potential land changes in the future. Some frequently used statistical methods for land change modeling include logistic regression (Hu and Lo 2007), generalized linear models (Aspinall 2004), generalized additive models (Brown et al. 2002), and Bayesian statistics (Agarwal et al. 2005). A popular example is the CLUE-S (the Conversion of Land Use and its Effects at Small regional extent) model developed by Verburg et al. (2002). The CLUE-S model consists of a non-spatial demand module and a spatially explicit allocation module. The non-spatial module estimates the

84 aggregate demand of land changes, and the spatial module allocates the land demands at various locations on a raster space based on stepwise logistic regression. Logistic regression is a form of multivariate models where the dependent variable has a categorical output, e.g., change or no- change of land use. Logistic regression can be binomial or multinomial. It takes the logit transformation of the categorical dependent variable to ensure that the dependent variable of the regression is continuous. Given the less demand of computational resources and easy operability, statistical modeling has become one of the most popular approaches for land change research communities. Statistical methods provide valuable information on key factors of land changes but are relatively deterministic compared to more advanced forms of model. It can also contribute to theory building and testing (Lesschen, Verburg, and Staal 2005). However, it has very limited capability to represent the complex interactions and the temporal dynamics within the coupled human- environmental systems.

5.2.2 Artificial neural networks

Artificial neural networks (ANN) are developed based on machine learning algorithms (e.g., Li and Yeh 2002; Liu and Seto 2008). The functioning of ANN is relating to regression models in that they both seek to associate land change and its potential drivers. ANN is characterized by its „learning‟ ability which can be used to detect non-linear relationships through the incorporation of a hidden layer. The algorithms of ANN calculate weights for input layers, hidden layers, and output layers by introducing the input in a feed-forward manner. For example, Liu and Seto (2008) presented an ART-MMAP neural network model for urban growth prediction from historical data. A set of proximity, neighborhood, and physical factors were included. This paper also applied a multi-resolution analysis to test the model‟s performance. In general, spatial aggregation results in higher accuracies. By comparing with a null model, two random models and a naive model, neural network outperforms other models at finer resolution. The strength of neural networks lies in their flexibility and non-linearity (Lesschen, Verburg, and Staal 2005) in predicting future change. However, it provides little interpretability because the relationships between variables remain invisible, criticized as a “black box”. ANN is commonly used for predicting future land cover/use based on the „knowledge‟ learned from the patterns and behaviors observed from historical data. The assumption here is that past and

85 present trend will continue to the future (i.e., stationarity), which tends to oversimplify the temporal complexity of land change processes.

5.2.3 Markov chain modeling

The Markov chain modeling approach employs a discrete stochastic process to determine the transition probability of land conversion. There is a set of discrete states in the modeling structure. In the context of land change modeling, each state usually represents different types of land cover/use. The model moves from one state (e.g., land cover/use type) to the other with some transition probability depending on the current state but not the previous ones (often called a process without memory). Transition probabilities are computed based on the observed land change data which represent the probability that the land cover/use type within a cell (i.e. spatial unit) will convert (or move) to another land type within the same period of time in the future. For example, Muller and Middleton (1994) applied Markovian analysis to time series data to quantify the land use changes over a human-dominated landscape. Markovian analysis can represent all the multi-directional land use changes between land use categories. Sequent time series data were used to simulate land use change over a longer time period. Markov models usually do not account for specific drivers of land change, which assumes that collective forces that functioned to produce the observed pat-terns in the past will continue to do so into the future. In other words, Markov models are used to project future land change based on the assumption of stationarity. Markov model can be dynamic by changing transition probabilities in some sort of regular patterns over time (Howard et al. 1995). Given the capability of automatically computing land transition probability with temporal data, Markov chain models are often integrated with more complex forms of models such as cellular automata and agent-based model that will be discussed shortly.

5.2.4 Cellular automata

A conventional modeling framework describes systems in equilibrium or as moving between equilibriums. However, the evolution of urban areas usually does not reach a stable equilibrium but exhibits features of complexity (e.g., edge of chaos, emergence, non-linearity). The concept of complexity emphasizes on the interdependence among constituent parts. Therefore, complex adaptive system (CAS) is a system composed of interconnected parts that as a whole exhibit one or more properties not obvious from the individual parts. Cellular automata

(CA) models are built upon static cell-based environment where each cell has a state and can transfer to others based on the current state and the interactions with its neighbor-hoods using a set of transition rules (Batty and Xie 1994; Clarke, Hoppen, and Gaydos 1997; Miller and Page 2007). The four major components of CA therefore are state, landscape/space, neighborhoods and transition rules. For each of the four components, their structures vary from simple to more complex forms according to existing literatures (e.g., Stevens and Dragicevic 2007). The transition rules are usually set to represent spatial and temporal constraints (Sante et al. 2010). One of the well tested CA model is the SLEUTH (Slope, Land use, Exclusion, Urban extent, Transportation, Hillshade) model developed by Clarke, Hoppen, and Gaydos (1997) for urbanization simulation. This model defines complex rules representing control parameters that allow the model to self-modify under the circumstances it generates. More applications of this model are found in Clarke and Gaydos (1998), Yang and Lo (2003), Jantz, Goetz, and Shelley (2004), Mahiny and Clarke (2012), and Akin, Clarke, and Berberoglu (2014). As a dynamic modeling tool, CA model has gained great popularity among all modeling approaches. Although offering a framework for studying complex systems, CA modeling does not explicitly incorporate drivers of change except for the neighborhood interactions and transition rules. In addition, CA does not explicitly account for human decision makings in their modeling structures as the cells cannot move and their transition in states mainly represent the physical processes of land conversion.

5.2.5 Economic models

Economic models generate land use patterns as aggregate outcomes from the underlying microeconomic behavior that determines demand and supply relation-ships. These models explicitly involve human choices and economic behaviors and thus address the human dimension of land change, mainly focused on land uses. The basic idea of economic models of land changes is based on market equilibrium (e.g., market clear with zero excess demand and zero excess supply). Economic models can operate at aggregate scale (e.g., sector-based models) and disaggregate scale (e.g., spatially disaggregate models). Sector-based models represent the global economy and the interactions between different sectors (i.e., general equilibrium models) or only some specific sectors as a closed system (i.e., partial equilibrium models). Therefore, sector- based models describe the amount of land allocated to different uses by demand-supply structures (Sohngen et al. 1999). Spatially disaggregate models simulate the optimal land use 87 decision based on profits or utility maximization or cost minimization (Bockstael 1996; Wu et al. 2004). These models explicitly represent individual decision-making at the micro level that will lead to land change outcomes at the aggregate level. Economic models explicitly represent human land use decisions based on market and price mechanism compared with most statistical, machine learning and cellular models. The spatially disaggregate models are promising in accounting for the market feedbacks and dynamics within the land change systems. These models are often used in the agent-based framework to simulate the decision-making processes of human agents. Economic models are useful for non-marginal land change simulation and prediction. However, given the complexity of human choices and data scarcity, it is quite challenging for economic models to build the underlying assumptions.

5.2.6 Agent-based models

Agent-based models (ABM), or the multi-agent system models (MAS), are developed based upon the assumption that “agent” is the major driver of a system (e.g., Parker et al. 2003; Batty 2005; Torrens and Benenson 2005; Xie, Batty, and Zhao 2007). ABMs are similar to CA models which are both spatial transition models built on a bottom-up perspective for the simulation of emergent properties of complex adaptive systems (Couclelis 2001). The three primary components of an ABM are the agent, landscape and their interactions. Within the modeling structure, the agents can interact with each other as well as the environment across multiple scales. Agents could employ high degree of rationality and information-processing ability in decision making which will influence the behavior of the systems (Miller and Page β007). A number of ABMs apply the utility function as drivers for agents‟ decision-making on location choices (e.g., Brown and Robinson 2006; Xie, Batty, and Zhao 2007; Ligmann- Zielinska 2009). Usually, an agent will select a location that can maximize utility or profit. Although the traditional ABM is built on the bottom-up perspective, researchers in geographic and ecological modeling have proposed that ABM should not be restricted to the bottom-up simulation (Xie, Batty, and Zhao 2007). In the paper by Xie, Batty, and Zhao (2007), the author considers both macro level and micro level spatiotemporal urban dynamics. The macro level model is based on a stepwise regression model to project the aggregated rate of change. The micro level model is to allocate the changes at the cellular level. The interaction among the two levels is also modeled through incorporating township competition in the utility function. 88

The structure of ABM is promising for land change research in that it explicitly represents human-nature interactions and feedbacks which are essential components for land change as coupled human and natural systems. However, given its complexity in model design and implementation, much effort needs to be done to examine its operability for simulating real world processes and to fully realize the potential of ABM. Moreover, the advancement of ABM is challenged by the lack of detailed data to represent and validate complex human decision- making processes and interactions between actors at the micro level.

5.3 Major issues in land change modeling

The usefulness and complexity of land change models lie in the necessity to treat land changes as a coupled human-environmental system with complex interactions and feedbacks at multiple spatiotemporal scales (Turner et al. 2007). This section discusses several important theoretical and methodological issues in land change modeling: (i) coupling of human decision- making and environmental conditions, (ii) scale dependency and multilevel interactions, and (iii) temporal dynamics and complexity. These proposed issues are important for developing a comprehensive understanding of land change in an integrative framework for global environmental change.

5.3.1 Coupling human-environment systems

Land changes are both causes and consequences of earth system changes, including the biophysical and the socioeconomic processes. Models taking specific drivers into considerations have tried to include factors from both subsystems. One major challenge arises from the integration of data and processes representing biophysical conditions and human decision making. Difficulty lies in the different level of aggregation and spatial unit of observation (Rindfuss et al. 2004). In social-demographic analysis, data are usually collected at some levels of aggregation, whereas direct measurements and remote sensing techniques have been more commonly used in extracting biophysical variables (Jensen 1983). As a result, research of the coupled human-environmental system has to deal with the problem of (i) integrating different types of data (e.g., raster and vector), (ii) integrating spatial data at different scales, (iii) integrating spatial data from different dimensions (e.g., point, line, polygon), and (iv) integrating data acquired at different locations (Gotway and Young 2002). These four types of spatial data integration problems are often intertwined, which leads to even more challenges.

The issue of coupling human and environmental systems is also related to the scale issues in that statistical modeling and machine learning are designed at the scale of human- environmental system as a whole while cell-based models can represent multilevel dynamics in both dimensions. Moreover, it is quite challenging to fully represent the processes in the human subsystems due to the lack of specific data on human decision-making and a high level of uncertainty. Towards a comprehensive understanding of the coupled system, the potential interactions and feedbacks within the land change system need to be incorporated in the models. In this sense, the structures of agent-based models and integrated models seem promising for integrating biophysical feedbacks and human behaviors. Its capability in representing temporal dynamics further facilitates the realization of simulating system feedbacks in land change processes.

5.3.2 Scale dependency and multilevel interactions

Research on the coupled human-environmental systems is further complicated by the issue of scale dependency and the multilevel interactions within the system. One of the early steps in spatially explicit modeling is to identify an appropriate scale (e.g., extent and resolution) for analyzing the spatial phenomena, such as land changes. This is known as the modifiable areal unit problem (MAUP) in geospatial science, that is, the correlation between variables may change with scales (Openshaw and Taylor 1979). The common approach to deal with the MAUP problem is to apply a multi-scale analysis to examine how the relationships among variables change with varying levels of aggregation and different ways of zoning (e.g., Veldkamp and Fresco 1997; Walsh et al. 2001; Evans and Kelley 2004; Hu and Lo 2007). Multilevel statistical modeling has also been used for analyzing land change driving factors at nested hierarchical levels (Hoshino 2001). Land change modeling is further complicated by the potential interactions and feedbacks among different levels of processes (Verburg 2006). In simulating the multilevel interactions, the modeling frameworks of cellular automata and agent-based model allow for the representation and incorporation of processes at multiple levels. The current land use models focus on two types of cross-scale dynamics: top-down and bottom-up simulation. The top-down control is represented by the government policies and global interactions affecting land demand and growth suitability. From the bottom-up perspective, human makes decisions on land allocation which produces the aggregate land use patterns. Further exploration on their capacity is needed 90 given the challenges in theoretical development and data availability, as well as the high computing demand of agent-based modeling.

5.3.3 Temporal dynamics and complexity

Simulating temporal dynamics is another critical issue for land change modeling, which brings about the need to handle time lags and feedback responses in the temporal dimension of land change processes (Agarwal et al. 2002). Under the assumption of stationarity, statistical modeling and machine learning have very limited capability to represent temporal dynamics and complexity of the land change processes. They often assume the factors leading to the observed patterns and processes in the past will continue to do so into the future. This assumption is problematic as it is very likely the factors will alter their future behaviors given changes in the landscape or some exogenous conditions. To the contrary, the framework of cellular automata and agent-based models allows for the temporal dynamics to be considered as the behaviors at individual level may alter in response to landscape changes or incorporated external variables at each simulation time step. The ecological and socioeconomic responses within the coupled human-environmental systems may not be immediately observable or predictable because the existence of time lags between the human-nature interactions and the appearance of ecological and socioeconomic consequences. To address this issue, a temporally lagged variable can usually be included in some models such as the statistical regression. More complex models have the flexibility to represent time lags in land use decisions. For example, Irwin and Bockstael (2002) treat the interactions among neighboring agents making a residential conversion decisions as a temporally lagged process to better represent the real world decision-making.

5.4 Integrated land change modeling for global environmental changes

The use and integration of models will lead to a comprehensive understanding of the complexity of the coupled human-environmental systems (i.e., synthesis and assessment issues). In the context of global environmental change and sustainability science, increasing concerns are given to research on sustainability that can inform practice and decision making in planning and management domains. The development of the next generation of land change models needs to take these concerns into consideration towards an integrated research framework for land change and earth system studies. In this section, four research articles were reviewed to illustrate the

91 progress of coupling land change modeling with other environmental analysis and modeling techniques for studying the interactions between land change and other components of global environmental changes, such as climate change, hydrological processes, soil degradation, and biodiversity loss. Kerr et al. (2003) described an integrated process-based modeling approach that couples ecological modeling of Carbon dynamics with economic modeling of land use for the prediction of land use and Carbon storage. This integrated model contains three components to simulate the interactions and feedbacks between ecosystems and human land-use activities. The ecological model and economic model were coupled through the land manager‟s choice of land use at each time step. The complex interactions were then realized through the exchange of individual model outputs as endogenous variables that will affect the next step of simulation. For example, the ecological model provides inputs to the land use choice model through estimates of biomass productivity. The key outputs from the integrated model include both land use and Carbon stocks. Lin et al. (2007) developed an approach for modeling the impacts of future land use and climate changes on hydrological process through integrating the CLUE-S model (Verburg et al. 2002) and the generalized watershed loading functions model (Haith and Shoemaker 1987). The structure of the CLUE-S model was described earlier in Section 5.2.1. The hydrological model is a combined distributed/lumped parameter watershed model that simulates runoff, sediment, and nutrient loadings in watershed using variable sized source areas of different land use/cover types. The simulated land use and cover types have different coefficient values that are used to determine the evapotranspiration in the hydrological model. Moreover, climate change scenarios generated from general circulation models (GCM) simulations have also been included to examine the impact of climate change on the hydrological cycle. Van Rompaey et al. (2002) loosely coupled land use change model with soil erosion model to predict future soil degradation and its on-site and off-site consequences. They firstly applied stochastic simulations to simulate future land changes based on the calculated afforestation and deforestation probabilities from historical land use maps. Then a spatially distributed soil erosion/sediment delivery model, SEDEM, was used to quantify the effect of afforestation or deforestation on soil erosion and sediment delivery. Land use classes are not directly involved in calculating the soil erosion component of SEDEM. But the probability of

92 land conversion and soil erosion rate are both affected by the same factor of slope gradient. The simulated future land use patterns were used as input for the sediment transport component in SEDEM, with a transporting capacity coefficient estimated for each land use class. Reidsma et al. (2006) assessed the relationship between land use intensity and related biodiversity in agricultural landscapes. For land use simulation, an integrated model was applied to quantify the area changes in agricultural land use and the CLUE model was used for land use allocation. Biodiversity in this study was measured using the ecosystem quality, which is expressed as the mean abundance of species originally present in the natural ecosystems relative to their abundance in undisturbed situations. Following the land use scenarios, the ecosystem quality of agricultural landscapes can be calculated as conditioned by land use. Then the impact of agricultural land use changes on overall biodiversity was assessed by comparing the relative size of nature area and the average ecosystem quality of natural ecosystems.

5.5 Conclusions

Land changes are processes in which human and natural systems interacting over space and time to reshape the earth‟s surface. They are both causes and consequences of global change that interacts with other components of the earth system. Land change science has recently emerged as a fundamental component of global environmental change and sustainability science. However, the complexity of land systems leads to many challenges for the research communities. Among the research components in land change science, land change modeling appears to be promising in improving our understanding of land use and land cover change as a coupled human-environmental system. A wide variety of modeling approaches have been developed to simulate the processes of land change. This chapter reviewed some commonly used approaches, including statistical regression models, artificial neural networks (ANN), Markov chain modeling, cellular automata, economic models, and agent-based models (ABM). These different approaches are built upon various theoretical and methodological foundations. The order of these approaches generally represents the theoretical transition of land change modeling from aggregate to individual modeling frameworks. The best model to use depends on specific applications given their unique strengths and weaknesses. The complexity for land change modeling is owing to their need to represent the spatiotemporal dynamics of the coupled human-environmental systems. For coupling the factors 93 from human and environmental systems, development of data integration techniques can help address the differences in spatial data. However, more comprehensive understanding and representation of the integrated processes within the coupled system is one of the major challenges for land change modeling. To deal with the influences of spatial dependency, multi- scale analysis is necessary to address the modifiable areal unit problem (MAUP). Another important issue is to model the interactions and feedbacks among multiple scales in the land change processes. New models need to take into consideration of the multilevel processes and to integrate alternative perspectives into the existing modeling framework. In modeling land change processes, a temporally dynamic modeling framework is critical to capture the necessary behavior changes in during the modeling time periods. Moreover, the factor of time lags needs to be considered to avoid biased simulation. The advances in land change modeling offer great opportunities to study global environmental change in an integrated framework. The examples reviewed in Section 5.4 shed light on the progress of coupling land change modeling with other ecological modeling and analysis techniques for analyzing the interactions between land change and other components of global environmental change. Many of the integrated frameworks are based on the use of simulated land use patterns or other land use/cover derived variables as input to the ecological models. More complex examples make use of the process-based models that integrate land change models and ecological models through individual decision-making using outputs from each model.

CHAPTER SIX

SIMULATING RESIDENTIAL LAND USE CHANGE THROUGH AN AGENT-BASED MODELING APPROACH

With the multi-decadal rapid urban population growth and the accompanying socio-ecological changes, there have been growing interests in developing spatially-explicit models that can help understand the underlying processes of land changes to support decision-making. This chapter described an agent-based modeling approach that has been developed for residential land use simulation. Three agent groups and their behaviors were simulated, namely, resident agents, developer agents, and government agents. Their different roles and objectives in residential location choices were represented as land utility maximization for residents, investment profitability maximization for developers, and zoning regulation by government, respectively. The land use patterns were simulated from the bottom up as a collective result from the interactions of agents‟ different objectives. The model was implemented to simulate residential land changes in a fast-growing suburban area. The initial model implementation revealed the importance of representing multiple decision makers in land change models although some further refinements are needed. Future work to improve the proposed model was discussed for both model implementation and conceptualization.

6.1 Introduction

Rapid urban population growth has been observed since the mid 20th century. As a result, human settlement patterns in the urban areas spread outward from the urban core towards the suburbs and exurbs, which was largely driven by the development of high-speed transportation and communication systems, dispersal of economic activities and a preference for rural lifestyles (Kaplan, Wheeler, and Holloway 2008). While land development can be viewed as a sign of the regional economic prosperity, the emerged low density and leapfrog built-up patterns in the suburban and exurban areas have begun to undermine the environmental and socioeconomic sustainability (Steffen et al. 2004; Camill 2010). There have been increasing concerns over the potential problems of sprawl, congestions, housing affordability and loss of open space (Waddell 2002). Understanding the linkages between human behaviors and the emergent landscape

95 patterns therefore becomes critical to support decision-making in land use planning and resource management. Recently, geospatial modeling has become increasingly popular in land change studies, which helps improve the understanding of the real world land change processes and also contributes to the development of urban theories (Batty, Couclelis, and Eichen 1997; Torrens 2002; Parker et al. 2003; Lo 2004; Batty 2005; National Research Council 2013). Over the past few decades, various geospatial modeling approaches have been developed, which are based upon different theories and methods. Examples include logistic regression (Verburg et al. 2002; Hu and Lo 2007), generalized linear regression (Aspinall 2004), artificial neural networks (Liu and Seto 2008; Wang and Mountrakis 2011), Markov chains (Muller and Middleton 1994; Tang, Wang, and Yao 2007), cellular automata (Clarke, Hoppen, and Gaydos 1997; Jenerette and Wu 2001; Sui and Zeng 2001; Sante et al. 2010), and agent-based models (Evans and Kelley 2004; An et al. 2005; Li and Liu 2007; Xie, Batty, and Zhao 2007; Arsanjani, Helbich, and Vaz 2013). Land change models have been employed to study the driving forces of urban growth (e.g., Seto and Kaufmann 2003; Hu and Lo 2007), to explore the effects of specific factors on land use patterns (e.g., Parker and Meretsky 2004; Brown and Robinson 2006; Ligmann-Zielinska 2009), to simulate and predict the evolution of urban forms (e.g., Clarke and Gaydos 1998; Xie, Batty, and Zhao 2007), and to generate alternative scenarios to inform land use planning (e.g. Yang and Lo 2003; Alcamo et al. 2011). The major challenge for land change modeling lies in the need for understanding land change as a coupled human and environmental system, which exhibits properties of complexity, such as edge of chaos, emergence, and nonlinearity (Liu et al. 2007; Turner et al. 2007). Both cellular automata (CA) and agent-based models (ABM) were developed for simulating the emergent properties of complex adaptive systems (Torrens 2006; Miller and Page 2007). CA models are usually built in a static cell-based environment where each cell has a state and can transfer to others based on a set of transition rules that take into consideration the current state and the interactions with its neighbors (Batty and Xie1994; Clarke, Hoppen, and Gaydos 1997; Sante et al. 2010). While CA models focus on the local interactions of physical factors, ABMs explicitly incorporate human behaviors in their modeling framework by including the autonomous and interacting “agents” (Parker et al. 2003; Batty 2005; Torrens and Benenson 2005; Xie, Batty, and Zhao 2007). Modeling human behaviors and decision-making is

96 particularly useful for understanding the complexity in the coupled systems (Liu et al. 2007; An 2012). Therefore, ABM has emerged as a promising approach for understanding the complex land change processes (Parker et al. 2003; Batty 2005). Agent-based land change models have been developed for various applications, from theoretical models that explain land use patterns and theory tests, to more realistic predictive models in connection to policy analysis and planning practices. The flexibility of ABMs offers great opportunities for representing detailed, dynamic processes in land use decision-making. However, fully realizing the capability of ABM in land change studies is challenging given the lack of detailed empirical data inputs that are needed to support model design and implementation. Many existing models therefore focused on explaining the effects of particular factors on land use patterns. For example, Parker and Meretsky (2004) examined the effects of edge-effect externalities on land use patterns using an agent-based model. Brown et al. (2004) evaluated how the location and width of a greenbelt affect the urban sprawl patterns. However, the feasibility and applicability of using ABM as an operational decision support tool for the real world decision-makers are still in question (Matthews et al. 2007; National Research Council 2013). Some examples of the comprehensive predictive models have been very case specific and require substantial local knowledge to be incorporated in the model (e.g., An et al. 2005; Manson and Evans 2007; Evans and Kelley 2008). It is therefore necessary to further explore the applicability of using ABM as a decision support tool in land change planning and management. This study explored an agent-based model for simulating the residential development decision-making processes and the emergent land use patterns. Three agent groups were identified, namely, resident agents, developer agents, and government agents. The three-agent structure allowed the exploration of the interactions and negotiation processes underlying residential land use development in a suburban area. The proposed modeling framework therefore focused on the interactions among the different perspectives of residents, developers and governments in the residential development processes. The proposed model was then implemented to simulate residential land change patterns in Gwinnett County, Georgia (Figure 1.2). The following sections will describe the methods for model conceptualization and model implementation and discuss some future research directions to refine and extend the proposed modeling framework.

6.2 Research methods

6.2.1 Model conceptualization

The process of residential land development is complicated and involves multiple groups of agents‟ decision-making and negotiation. The proposed model focused on the development of an operational model that is easy-to-implement and can be used to support realistic simulation. Three interacting groups of agents and their different objectives were considered in the conceptual model: residents, developers, and government. The emergent land use patterns were therefore determined by the location choices and interactions of these three agent groups. The following subsections will specify the objectives and decision-making for each group of agent and the spatial allocation strategy of residential land use that couples the decision-making processes of the three agent groups. The procedure of the proposed conceptual model was illustrated in Figure 6.1.

Figure 6.1. A conceptual model to simulate residential development proposed in this dissertation research.

6.2.1.1. Resident agents. Residents are the prospective home buyers whose behaviors are among the primary determinants of the residential development growth demand (Brown and Robinson 2006). While residents are not directly involved in the land development process, their housing location preferences will affect the investment decisions of land developers. The developers need to consider the local housing demand and preferences of the potential home buyers in order to make profits. In search for potential home locations, residents‟ location preference is simulated based on the trade-off between factors representing the affordability and livability (e.g., easy access to urban activities, close to natural attraction) of a certain location. Thus, resident agents‟ behavior can be defined by the following additive utility function: ( , ) = ( ) + ( ) + ( ) (6.1) where ( , ) represents� the land utility of a location (�) for agent at time , () represents the factor of home price, ( ) represents the factor of accessibility� to urban activities which can be calculated with the composite proximity to major roads and urban centers, and ( ) represents the factor of natural attractiveness which can be measured by the� composite proximity to parks and major water bodies. And the values of , , and represent the relative importance (weights) of each factor in resident agents‟ location choices and their sum equals to 1. To make these factors comparable in decision-making, it is necessary to normalize their initial values. For example, home price is considered as a negative factor in resident agents‟ site selection, that is, lower price is preferable. Therefore, the initial values for the home price factor are normalized to the range of 0-1 using the following function so that the higher score of the normalized factor is considered preferable: Max x (6.2) x = Max Min ′ − The raster layers used to produce the composite accessibility and attractiveness are initially − calculated using Euclidean distance to the locations of interests, which are also negative factors (shorter distance is preferable). The distance values are also normalized using the same function in Equation 6.2. And these normalized distance layers were combined to produce the composite spatial criteria representing accessibility and attractiveness using a weighted summation aggregation function within GIS (Randolph 2004). If no detailed information available, the weights can be determined based on iterative experiments and comparing the combined results.

Residents may be further characterized into subtypes by their demographic and socioeconomic characteristics, such as income, age, and household size or structure (e.g., with or without children). These differences will lead to their unique combinations of preferences in housing location choice, which are reflected by the different weights ( , , and ) assigned in

Equation 6.1. And the interactions between different groups of resident agents are responsible for the evolution of the physical and social structures of residential land use in an urban area. However, the heterogeneity in residents‟ behaviors is not considered in the current model in order to keep the model simple and stylish. 6.2.1.2. Developer agents. Among the multiple decision-makers, land developers take an active and dominant role who leads the planning, financing, and construction processes in residential development (Peiser 1990; Coiacetto 2000; Gillen and Fisher 2002). Land developers make investment decisions based on their estimations of the local housing demand according to the analysis of demographic and socioeconomic conditions. Meanwhile, developer agents can guide residents in location choice based on economic factors for selecting locations that favor both parties. Their development decisions are also subject to the land use regulation by local governments. Realistic representation of developers is difficult given the complicated decision-making and negotiation processes. The primary goal of developers identified in this model is to maximize investment profit by taking into consideration the preferences of prospective home buyers and land use regulation by local governments. A simplified representation is therefore adopted here to simulate the profit maximization behaviors of property developers through the land residue method (Li and Liu 2007). The following equation is used to assess the spatial criteria of the investment profit for developer agents: = (6.3) where represents the investment� profit,� − � represents− � the home sales price, represents� the land acquisition price, and represents� the development cost, which �involves both direct and indirect costs, including engineering� costs, construction costs, marketing costs, and professional fees for market feasibility analysis and appraisal, legal and accounting fees, and financing costs. And the development probability for the developer agents can be computed as follows (Li and Liu 2007):

100

(6.4) ( , ) = � − � � where ( , ) represents the development� probability− � of a location ( ) for agent at time ,� represents the investment profit, is the maximum profit, is a threshold � value identified to represent the minimum� anticipated profit from the property� investment. The threshold value can be determined by expert knowledge if available. In the model implementation, developer agents are assumed to have the same attributes as more detailed information is unavailable. 6.2.1.3. Government agents. The objectives of government agents can be very different from those of land developers and therefore merit separate representation in the simulation (Fisher 2005). Their major objectives are to promote efficient distribution of urban land use over feasible sites in order to meet the local housing need and land and resource management plans. Government will examine the development suitability of a potential location according to the existing land use/cover, surrounding environment, transportation, and facilities. Two sets of spatial criteria are considered at this stage: zoning districts and existing land use/cover. Rather than imposing a Boolean image with restricted and non-restricted area for development, a suitability surface is produced to represent the land regulation by government agents. The zoning districts map and land use/cover map were both reclassified so that higher value representing higher suitability for residential development. The two spatial criteria are then combined using a multiplicative function as specified below: ( ) = ( ) × ( ) (6.5) where ( ) represents the development� suitability of a location , ( ) represents the factors of the initial land use/cover types, and ( ) represents � the factor of zoning plan with permitted uses. For each land use/cover type, their suitability for conversion to residential land use can be very different. For instance, wetland area will have lower suitability for urban land use conversion. In fact, the different types of land cover can be ranked between 0- 1 based on their suitability of conversion to residential land use. The order can be determined based on expert knowledge. The second criterion introduced in the simulation is the zoning district with permitted land use types as designated by the local government. In general, zoning districts designation represents the most comprehensive land use plan that separate one set of uses from another. While the detailed regulation specify the density, height and many other 101 spatial characteristics, only the most general categories of land uses (e.g., residential, commercial, industrial, etc.) are considered at this initial stage of model development. For the simulation of residential land use change, the permitted zones for residential use will be given substantial priority by assigning high suitability values. Developer agents may be eligible to apply for re- zoning of a piece of land with lower suitability. And the interactions between developers and governments are represented in the model by the trade-offs among land utility, profitability and suitability in the land allocation processes. 6.2.1.4. Spatial allocation strategies. As described above, the residential development decision-making processes are represented by a probability surface for each of the three agent groups using different spatial criteria measurements. Given the different objectives identified for each agent group, it is necessary to couple these different objectives in the land allocation process to generate the residential land use patterns using agent-based approach. The model therefore focused on the interactions among land utility for resident agents, investment profitability for developer agents, and land regulation for government agents. It is defined as a function to combine the three probability or suitability layers for the agents to select potential development site. A linear form is adopted here for easy implementation and interpretation (Xie, Batty, and Zhao 2007). The combined allocation probability is defined as: ( , ) = ( , ) + ( , ) + ( ) (6.6) where �( , ) representsμ the allocation �� probability of a location� ( ) for agent at time , �( , ) represents the land utility for residents agents of a location ( ) based on the factors of affordability, accessibility and attractiveness calculated using Equation 6.1, ( , ) represents the development probability of a location ( ) for developer agents based� on investment profitability calculated using Equation 6.4, and ( ) represents the land suitability of a location ( ) designated by the government agents based on initial land µ use/cover types and zoning constraints using Equation 6.5. And the values of , λ, and ψ represent the relative importance (weights) of each probability criteria for the three agent groups and their sum equals to 1. In this sense, the actual agents entering the landscape represent the combined entities of the residents, developers, and governments. In principle, what each agent is doing is to develop residential sites by maximizing the combined allocation probability ( , ). � 102

For the spatial allocation of residential development, it is assumed here that 1 year to be an appropriate time step for the simulation (Ligmann-Zielinska and Jankowski 2010). The total land demand is estimated for each time step. Agents enter the landscape at the beginning of the simulation. Each agent draws a sample of developable locations to represent the bounded rationality of human decision-making due to incomplete information about the local real estate market (Arthur 1994; Brown et al. 2004; Brown and Robinson 2006). With the allocation probability calculated for each location, the agents order the candidate sites from best to worst. Then the agents will allocate land development at the selected sites around its 3×3 Von Neumann neighborhood. This neighborhood allocation strategy is employed to simulate the agglomerated urban development patterns. Note the neighborhood growth is subject to land suitability defined by the government agents. Therefore, only the most suitable neighbors of the selected site will be developed. The allocation process will repeat until meets the residential land demand. The site searching procedure by the agents can be explained as a procedure of the interactions between resident agents choosing “optimal” sites to maximize utility and developer agents determine the development site to maximize profitability, which are subject to the government agents‟ decision on whether to approve the development project by evaluating the land suitability.

6.2.2 Model implementation

The conceptual model was experimented with Gwinnett County, Georgia as a case study area (Figure 1.2). The model was implemented using the open source Recursive Porous Agent Simulation Toolkit for Java (Repast J) version 3 (North, Collier, and Vos 2006). Repast J was chosen given its support for writing advanced models and importing GIS data in both vector and raster format. The GIS support is important for constructing realistic simulation models, such as land change models. Therefore, the spatial criteria were firstly processed within GIS and then converted and imported into Repast J as model input layers. 6.2.2.1. Data preparation. The analysis was performed using a two-dimensional raster data format with a cell size of 120 m and an extent of 394 columns by 394 rows. The cell size was selected for computational efficiency, which was the smallest cell size with acceptable running time. The specific procedures for preparing each input data layer was specified below. Residential land use data for model calibration and validation were derived from satellite imagery given the lack of consistent land use/cover data for the study site. Specifically, Landsat Thematic Mapper (TM) imagery with 30m resolution in years of 2000 and 2010 were acquired 103 from the USGS EROS Data Center for extracting urban land uses through a sub-pixel classification method developed in Chapter 2 and 3. The spatial pattern of residential land use growth between 2000 and 2010 was then produced through a GIS minimum dominate overlay function on the two years of land use map (Figure 6.2). Therefore, the yellow pixels in Figure 6.2 represent the minimum amount of residential land use in 2000 and the red pixels represent the net residential growth between 2000 and 2010.

Figure 6.2. Spatial patterns of residential land use growth in Gwinnett County, 2000-2010. The location of expressways, including all interstate highways and Georgia Highway 400, is also shown.

The spatial criteria for resident agents included the factors of home price, accessibility and attractiveness. Detailed home price data were not generally available. In the model implementation, the average home sales price data at zip code level were acquired from Zillow,

104 an online real estate database that includes data on sales prices at local levels. The data have also been used by the Atlanta Regional Commission (ARC) for studying and reporting housing- related statistics. The average home sales price data in 2000 were attributed to each zip code polygon within GIS and were further normalized using Equation 6.2 and converted to raster format (Figure 6.3a). Spatial data for calculating accessibility and attractiveness were collected from the Atlanta Region Information System (ARIS) GIS data sets that was adjusted with respect to the year 2000 satellite image. The factor of accessibility was calculated by combining the raster layers of distance to major roads and existing urban centers based on a weighted summation aggregation function. Three weight combinations were experimented in the aggregation: {0.3, 0.7}, {0.7, 0.3}, and {0.5, 0.5}. The resultant layers showed some degree of similarity in their spatial patterns which was further justified by a correlation analysis with a minimum correlation coefficient of r = 0.908 (p<0.01). Since none of the two criteria seem more important than the other, the equal weight case (i.e., {0.5, 0.5}) was finally used to produce the accessibility layer (Figure 6.3b). The rationale for calculating the factor of attractiveness was similar, which was a composite measure of the distance to parks and distance to water bodies. Positive correlation (minimum correlation r = 0.845, p<0.01) was also found between pairs of resultant aggregated layers and therefore the equal weight case was selected to produce the spatial criteria of attractiveness (Figure 6.3c). Due to the lack of detailed information on the relative importance of the three factors for the residents in the study area, the possible situations of their preferences were tested. Seven combinations of the relative importance of the three spatial criteria (α, , and ) were experimented in the simulations under the conditions of three factors being equally important, one factor being more important and one factor being less important (Table 6.1). To calculate the investment profit for land developers (Equation 6.3), the average home sales price data at the zip code level in 2000 from Zillow were used again. However, the land price and development cost data for measuring the investment profits of land developers were not available in the study area. Some assumptions were thus adopted to deal with the data issue. In general, land price is closely related to home sales price. Therefore, it is assumed the land price is in proportion to the home sales price. In addition, it is assumed the development costs are the same across the study area, which can then be omitted in Equation 6.3. In order to calculate the development probability for land developers (Equation 6.4),

105

Table 6.1. Agent preferences (weights) experimented for calculating land utility for resident agents using Equation 6.1.

Experiment number Home price Accessibility Attractiveness (α) () () 1 Equal importance 0.3333 0.3334 0.3333 2 One more important factor 0.6 0.2 0.2 3 0.2 0.6 0.2 4 0.2 0.2 0.6 5 One less important factor 0.4 0.4 0.2 6 0.4 0.2 0.4 7 0.2 0.4 0.4 the threshold value representing the minimum acceptable profit needs to be determined based on local knowledge. In the model implementation, this value was determined as the minimum home sales price in the entire Atlanta metropolitan area to take the regional interactions into consideration (ARC 2013). The produced development probability surface for developer agents was shown in Figure 6.4.

106

Figure 6.4. Development probability surface for developer agents calculated based on Equation 6.4.

As formulated earlier in the model conceptualization, the land suitability determined by the government agents was produced based on a combination of the existing land use/cover map and the zoning district map. Firstly, the initial land use/cover types in 2000 (see Chapter 3) were ordered to represent the land conversion suitability to residential land use. The order was determined based on the land use/cover change analysis discussed in Chapter 3 according to the percent conversion from each land use/cover class to residential land use between 2000 and 2010. Specifically, the agents will try to occupy barren land (1) first, then forest (0.8), grass/shrub land (0.6), agricultural land (0.4), and finally water and wetlands (0.1). The initial developed land areas have been excluded for the land allocation by assigning them a suitability value of 0. Then, the 2000 zoning map acquired from the Gwinnett County GIS Digital Data Sets was used to generate the spatial criteria representing zoning regulation. The zoning areas were ordered based on the types of permitted uses to give the residential zones higher priority, and other land use zones low priority. The specific suitability assigned to each type of permitted use is: residential (1), mixed use (0.8), commercial (0.5), industrial (0.3), and conservation (0.1). The suitability value was ranked to represent the possibility for residential land development and only the relative values are important for the earlier stage model experiments. The two generated suitability surfaces were then combined to produce a land suitability layer using the multiplicative product of the two raster layers within GIS (Figure 6.5).

107

Figure 6.5. The land suitability surface representing the land regulation for government agents produced based on Equation 6.5.

6.2.2.2. Model calibration and validation. The land demand at each simulation time step (i.e., 1-year) was estimated using linear interpolation of the residential land changes between 2000 and 2010. Model calibration was conducted to determine the values for the three parameters µ, λ, and ψ in Equation 6.6 which represent the relative importance of the three spatial criteria for resident, developer, and government agents, respectively. A heuristic approach was used here to explore the possible combinations of these parameters that can generate the best simulation results compared to the actual observation from 2000 to 2010. This procedure involved running the model hundreds of iterations in a batch mode with changing parameter values and observing the model performance in Repast. At the current stage, the goodness of fit of the simulation results was measured with the Kappa coefficient (Pontius and Schneider 2001; Pontius, Huffaker, and Denman 2004) from a confusion matrix for the developed and non-developed pixels. Note the confusion matrix was produced only for the developable cells (i.e., non-developed cells) in 2000. The experiments have also been conducted with the seven possible weight combinations for determining the land utility for resident agents listed in Table 6.1.

108

6.3 Results and discussion

The input parameter sets that can create the best-fit model were finally identified as µ ~0.4, λ ~0.4, and ψ ~0.2, when the accessibility factor being more important for resident agents (Experiment 3 in Table 6.1). The accuracy assessment results for the simulation results and actual observation in 2010 were summarized in Table 6.2. The simulation results for residential development with the best combination of parameters were shown in Figure 6.6. Note this may not be the optimal results as only a limited number of the input parameter combinations was experimented due to the computational constraints. However, it may reveal some general patterns of parameter distribution which is acceptable at this initial stage of model implementation. During the model experiments using varying combinations of the input parameters, different patterns were generated based on the visual observation. This is an indication of the importance of incorporating the different objectives for the three different agent groups in generating the emergent land use patterns. However, this three-tier model was constructed based on a set of simplifications given the constraints in data availability and computational limitations. Future work is needed in order to better explain the realistic interactions among the different types of decision-makers. The specific future research directions for improving the current model were identified and discussed below from the perspectives of model implementation and conceptualization. A few improvements in the model implementation need to be done for future studies. The first issue is relating to data availability. This model relies upon the use of GIS analysis to produce various spatial criteria, and the lack of detailed data has posed great challenges in the model implementation. Some assumptions and simplifications were adopted which may weaken

Table 6.2. Model validation using Kappa coefficient based on cell-by-cell comparison of the simulated and the actual resident development in 2010.

Simulated developed Simulated non-developed

Actual developed 4279 3216

Actual non-developed 3221 46298

Kappa coefficient 0.506

109

Figure 6.6. Simulated residential land use growth patterns in Gwinnett County, Georgia for 2010 using the proposed agent-based model. Note the blue pixels represent the initial built-up area in 2000, and the red pixels represent the simulated residential land change between 2000 and 2010. the capability of the model in explaining the interactions. For example, it is clear that the coarse pattern of home sales data at zip code level has affected the simulation results in an aggregate manner when comparing the simulated residential land change patterns with the actual observation. To develop predictive models, there has been an increasing need to acquire detailed empirical data to facilitate model design and construction (National Research Council 2013). Therefore, the immediate need for refining this model is to obtain data from various sources, such as using qualitative methods, to support realistic simulation. In the model validation, a cell- based comparison using Kappa coefficient was applied to evaluate the simulation results from the proposed agent-based model at this initial stage. While it can reveal the general goodness of fit of the model performance, more advanced validation approach needs to be applied for the future refined model in order to capture the patterns or morphologies of urban development. Some studies have recommended the use of spatial metrics in the process of validating land change models (Clarke and Gaydos 1998; Herold, Couclelis, and Clarke 2005). A sensitivity

110 analysis can also be helpful to explore the effects of the input parameters on the simulation results (Ligmann-Zielinska and Jankowski 2010). A number of extensions of the model conceptualization have also been identified based on the simulation experiments. Firstly, the three agent groups in the current simulation were assumed to be homogeneous (i.e. uniform behavior) for simplification purposes. However, heterogeneous preferences of agents can lead to very different land use patterns (Brown and Robinson 2006). To further refine the model structure, heterogeneous groups of resident and developer agents and their decision-making need to be characterized. Take the resident agents as an example, smaller households and households without children are likely to seek smaller dwelling units such as apartments, condos, and townhouses, rather than the detached single- family houses. Similarly, residents with high income can afford good quality home at location with higher price, while lower income residents can only afford homes with cheaper price. Their unique preference as interacted with the environment will change the housing demand as well as the location and patterns of land use, which merit specific representation in simulating land use change. Similar to the resident agents, property developers may differ in their investment strategies which ultimately determine the spatial location, intensity and timing of development projects. And the competition among different groups of developers may lead to different emergent land use patterns from the bottom up. However, to characterize the heterogeneity within each agent group, it is necessary to rely on qualitative methods such as sample surveys, interviews to appropriately characterize the agents and their unique preferences in residential land development. Another important factor that has been omitted in the current conceptual model is the factor of market forces. For example, if the demand from residents for a particular site increases, property price will increase accordingly. Such an increase in home price may attract developer agents‟ investment. If the price rises extremely high that exceeds the affordability of the prospective home buyers, they may seek location to reside elsewhere. It will inevitably lead to overdevelopment. It is therefore desirable to incorporate such dynamic supply-demand interactions in simulating their influences on the aggregate land use patterns. From the planning perspective, understanding these dynamics in a spatially-explicit way can further support the decision-making in land use regulation. The mismatch between housing need and supply found in Gwinnett County and many other counties in the Atlanta metropolitan area suggests the need

111 to consider the demographic trends and market forces in future land use planning (ARC 2013). One possible refinement is to incorporate a feedback mechanism in the simulation to represent the housing price dynamics so that incremental demand from the resident agents for a particular site will affect the home price factor for both residents and developer agents.

6.4 Conclusions

This chapter described an agent-based model for residential land change simulation in a fast-growing suburban area. The decision-making roles of three agent groups have been considered, namely, resident agents, developer agents, and government agents. They held different objectives in the residential development process. Resident agents prefer sites with lower price, easy access to urban activities, and close to attractive areas. Property developers take the housing needs of prospective residents into consideration and meanwhile try to maximize the investment profit. Governments make policies on land regulation that optimize the regional land use configurations. Their different objectives were coupled in the agent-based simulation that collectively generates the emergent residential land use patterns. The simulation experiments indicate the importance of representing decision-making and interactions among different types of decision-makers in land change modeling. Future refinement is needed toward the development of a more realistic representation to support decision-making. From the perspective of model implementation, it is necessary to acquire detailed data in developing predictive model of real world land changes. The model validation approach is also a key component which not only evaluates the model performance but also provides insights into explaining the spatial patterns. From the perspective of model conceptualization, the factors of agent heterogeneity and housing market dynamics need further exploration which may influence the simulated land use patterns.

112

CHAPTER SEVEN

SUMMARY AND CONCLUSIONS

This dissertation research has explored the feasibility and applicability of integrating geographic information technologies for analyzing and modeling land change in an urban area. The focus has been given to the three specific dimensions in land change science: observation and monitoring of land changes (Chapter 2 and 3), driving force analysis (Chapter 4), and spatially- explicit land change modeling (Chapter 5 and 6). Overall, this dissertation research has demonstrated the important role of integrating various geographic information technologies, such as remote sensing, GIS, spatial analysis and modeling, in land change science for global environmental change research. The technological integration also provides the opportunities for coupling the theories and methods from human and environment sciences towards more comprehensive understanding of land changes as a coupled human-environment system.

7.1 Land use and land cover mapping

Land use and land cover mapping has been largely relying upon the advance in remote sensing acquisition and processing techniques. However, challenges exist when mapping land use and land cover in the heterogeneous urban environment from medium resolution satellite imagery. In Chapter 2, a stratified classification method combined with multiple endmember spectral mixture analysis (MESMA) has been developed. The combined use of these techniques has been designed to address two important issues in urban land use and land cover mapping, namely, the spectral confusion between urban and rural features and the “mixed pixel” problem in the urban area. A GIS-based landscape partition was first performed based on road network analysis to spatially separate those spectrally confused pixels in the urban and rural areas. The MESMA technique was then applied to deal with the mixed pixel problem in the heterogeneous urban area to help extract isolated land patches. The method distinguishes from previous work by extending sub-pixel analysis into the area of mapping thematic land use/cover types. At the technology level, this study has demonstrated the usefulness of integrating various ancillary data, spatial analysis, and image processing techniques in land use and land cover mapping from remotely sensed data. There is also an indication to develop novel approaches making better use of the existing tools towards effective land use and land cover mapping in different contexts.

113

7.2 Land change analysis

Land use and land cover mapping from time-series of remotely sensed data provides the basis to explore the temporal dynamics of land change. A comprehensive land change analysis needs to characterize different aspects of these temporal dynamics. In Chapter 3, a change analysis was conducted through the combined use of satellite imagery, GIS, and landscape metrics. The remote sensing-based land classification was first applied on time-series of satellite imagery to produce the detailed land use and land cover maps. Post-classification change detection, GIS-based operations, and landscape metrics were then employed to conduct the land change analysis in the study area. Specifically, the size, location, distribution, spatial configuration, and nature of land change were characterized and examined with the combined methods to help reveal the undergoing processes. At the technology level, this study has demonstrated the usefulness of integrating remote sensing with GIS and landscape metrics that allows land change analysis to go beyond simple statistic description and into the characterization of the spatial characteristics and nature of land changes in a complex urban environment. At the application level, the results revealed a far-reaching suburbanization process in the study area with significant alterations in the forest ecosystems.

7.3 Driving forces of land change

The dynamics of land change is driven by the complex interactions among various biophysical and socioeconomic factors. Understanding the causes of land change can help explain the mechanisms of land change which implicates decision making. Scale has been recognized as one of the fundamental issues for land change research. The choices over scale, extent, and resolution can critically affect the observed patterns and processes. Chapter 4 has examined the driving factors of urban land use change in the study area using a multi-scale statistical analysis. Specifically, correlation analysis and multivariate regression have been employed to detect the important biophysical and socioeconomic factors driving urban land use change at varying scales. The analysis has been performed at different nested aggregation levels and over different spatial extents. At the conceptual level, this analysis confirmed the scale dependency of factors driving urban land use change. On the other hand, the effects of spatial extents on the analysis results suggested further studies to examine the spatial non-stationary patterns of these driving factors. At the application level, some important factors of urban land

114 use change in the study area have been identified, such as population density and location measures.

7.4 Land change modeling

As one of the major dimensions of land change science, land change models have been developed for a wide range of applications. Given the large number of land change models, Chapter 5 has reviewed some frequently used modeling techniques, including statistical regression models, artificial neural networks, Markov chain models, cellular automata, economic models, and agent-based models. The theoretical and methodological fundamentals of each modeling approach were discussed. Some outstanding issues for land change modeling were also identified. In the context of global environmental change research, some examples that integrate land change modeling with environmental models were also reviewed. Among the different modeling approaches, agent-based model appears to be promising for simulating land change as a coupled human-environmental system. In Chapter 6, an agent- based model was developed to simulate the residential land development processes in a fast- growing suburban area. Specifically, the decision-making and interactions among three agent groups were represented, namely, resident agents, developer agents, and government agents. The proposed model focused on how their different objectives collectively generate the aggregate land use patterns. At the initial stage of model implementation, the simulation results have revealed the importance of representing multiple decision makers in land change modeling. Further refinement is needed in order to apply the model to support real world decision-making.

7.5 Future studies

Future studies will extend this dissertation research at two broad directions. The first direction is identified at the technology and methodology level to further explore and improve the capability and applicability of integrating geographic information technologies for analyzing and modeling land changes. Firstly, recent improvements in remotely sensed data offer more options for land use and land cover mapping using remote sensing techniques. While this dissertation research focused on the use of medium resolution imagery, the potential of higher resolution data (e.g., very-high-spatial-resolution, hyperspectral) and active remote sensor data (e.g., radar, lidar) can be explored in different circumstances. Secondly, more advanced methods need to be explored for the driving force analysis of land change. The capability of empirical

115 statistical analysis is challenged by its data mining nature that is usually used for theory development at the earlier stage. Further analysis is needed to analyze the driving factors of land change in the context of the coupled human-environmental system with complex and dynamic features. Finally, more work needs to be done for the spatially-explicit modeling of land changes. In this dissertation, an agent-based modeling framework has been employed to test the simple interactions among multiple decision makers. Alternative modeling approaches may offer complementary perspectives that can help refine the representations of the processes underlying land changes. Another direction is identified at the conceptual level towards a more comprehensive understanding of the land change system. This dissertation focused on three specific dimensions in land change science. Another important and comprehensive dimension in land change science focuses on the synthesis and assessment issue. To a large degree, the issue of synthesis and assessment is based on a thorough understanding of the three components explored in this dissertation. It is therefore important to integrate land change mapping, driving factor analysis, and spatial modeling to support the development of land change science for global environmental change research. Among the various geographic information technologies, modeling offers great flexibility for the purpose of integrating these components of land changes as well as the dynamics in the coupled system.

116

REFERENCES

Adams, John B., and Alan R. Gillespie. 2006. Remote Sensing of Landscapes with Spectral Images: A Physical Modeling Approach. Cambridge ; New York: Cambridge University Press.

Agarwal, D. K., J. A. Silander, A. E. Gelfand, R. E. Dewar, and J. G. Mickelson. 2005. "Tropical deforestation in Madagascar: analysis using hierarchical spatially explicit, Bayesian regression models." Ecological Modelling 185 (1):105-31. doi: 10.1016/j.ecolmodel.2004.11.023.

Akin, A., K. C. Clarke, and S. Berberoglu. 2014. "The impact of historical exclusion on the calibration of the SLEUTH urban growth model." International Journal of Applied Earth Observation and Geoinformation 27:156-68. doi: 10.1016/j.jag.2013.10.002.

Alcamo, J., R. Schaldach, J. Koch, C. Kolking, D. Lapola, and J. Priess. 2011. "Evaluation of an integrated land use change model including a scenario analysis of land use change for continental Africa." Environmental Modelling & Software 26 (8):1017-27. doi: 10.1016/j.envsoft.2011.03.002.

Allen, T. F. H., and T. W. Hoekstra. 1991. "Role Of Heterogeneity In Scaling Of Ecological Systems Under Analysis." Kolasa, J. and S. T. a. Pickett (Ed.). Ecological Studies Analysis and Synthesis, Vol. 86. Ecological Heterogeneity. Xi+332p. Springer-Verlag New York, Inc.: Secaucus, New Jersey, USA; Berlin, Germany. Illus. Maps:47-68.

An, L. 2012. "Modeling human decisions in coupled human and natural systems: Review of agent-based models." Ecological Modelling 229:25-36. doi: 10.1016/j.ecolmodel.2011.07.010.

An, L., M. Linderman, J. Qi, A. Shortridge, and J. Liu. 2005. "Exploring complexity in a human- environment system: An agent-based spatial model for multidisciplinary and multiscale integration." Annals of the Association of American Geographers 95 (1):54-79.

Anderson, James Richard, Ernest E. Hardy, John T. Roach, and Richard E. Witmer. 1976. A Land Use and Land Cover Classification System for Use with Remote Sensor Data: USGS Professional Paper 964, Sioux Falls, S.D., USA.

Arsanjani, J. J., M. Helbich, and E. D. Vaz. 2013. "Spatiotemporal simulation of urban growth patterns using agent-based modeling: The case of Tehran." Cities 32:33-42. doi: 10.1016/j.cities.2013.01.005.

Arthur, W. B. 1994. "Inductive Reasoning And Bounded Rationality." American Economic Review 84 (2):406-11.

Aspinall, R. 2004. "Modelling land use change with generalized linear models - a multi-model analysis of change between 1860 and 2000 in Gallatin Valley, Montana." Journal of Environmental Management 72 (1-2):91-103. doi: 10.1016/j.jenvman.2004.02.009.

117

Aspinall, Richard J., and Michael J. Hill, eds. 2008. Land Use Change: Science, Policy and Management. Boca Raton: CRC Press.

Atlanta Regional Commission (ARC). 2013. The Atlanta Region. Available from: http://www.atlantaregional.com.

Banzhaf, E., V. Grescho, and A. Kindler. 2009. "Monitoring urban to peri-urban development with integrated remote sensing and GIS information: a Leipzig, Germany case study." International Journal of Remote Sensing 30 (7):1675-96. doi: 10.1080/01431160802642297.

Batty, M. 2005. "Agents, cells, and cities: new representational models for simulating multiscale urban dynamics." Environment and Planning A 37 (8):1373-94. doi: 10.1068/a3784.

Batty, M., H. Couclelis, and M. Eichen. 1997. "Urban systems as cellular automata." Environment and Planning B-Planning & Design 24 (2):159-64.

Batty, M., and Y. Xie. 1994. "From Cells To Cities." Environment and Planning B-Planning & Design 21:S31-S8.

Bergkamp, G. 1995. "A hierarchical approach for desertification assessment." Environmental Monitoring and Assessment 37 (1-3):59-78. doi: 10.1007/bf00546880.

Boardman, Joseph. 1995. Mapping Target Signatures Via Partial Unmixing of Aviris Data.

Bock, Michael, Panteleimon Xofis, Jonathan Mitchley, Godela Rossner, and Michael Wissen. 2005. "Object-oriented methods for habitat mapping at multiple scales - Case studies from Northern Germany and Wye Downs, UK." Journal for Nature Conservation (Jena) 13 (2-3):75-89. doi: 10.1016/j.jnc.2004.12.002.

Bockstael, N. E. 1996. "Modeling economics and ecology: The importance of a spatial perspective." American Journal of Agricultural Economics 78 (5):1168-80. doi: 10.2307/1243487.

Boteva, Dimitrina, Geoffrey Griffiths, and Panayotis Dimopoulos. 2004. "Evaluation and mapping of the conservation significance of habitats using GIS: an example from Crete, Greece." Journal for Nature Conservation (Jena) 12 (4):237-50. doi: 10.1016/j.jnc.2004.09.002.

Briassoulis, H. 2000. Land Use Change: Theoretical and Modeling Approaches. Web Book of Regional Science, S. Loveridge, ed. Regional Research Institute, West Virginia University. http://www.rri.wvu.edu/WebBook/Briassoulis/contents.htm.

Brown, D. G., S. E. Page, R. Riolo, and W. Rand. 2004. "Agent-based and analytical modeling to evaluate the effectiveness of greenbelts." Environmental Modelling & Software 19 (12):1097-109. doi: 10.1016/j.envsoft.2003.11.012.

118

Brown, D. G., and D. T. Robinson. 2006. "Effects of heterogeneity in residential preferences on an agent-based model of urban sprawl." Ecology and Society 11 (1). doi: 46.

Brown, D. G., D. T. Robinson, L. An, J. I. Nassauer, M. Zellner, W. Rand, R. Riolo, S. E. Page, B. Low, and Z. F. Wang. 2008. "Exurbia from the bottom-up: Confronting empirical challenges to characterizing a complex system." Geoforum 39 (2):805-18. doi: 10.1016/j.geoforum.2007.02.010.

Bullard, R., G. Johnson, and A. Torres. β000. “The Costs and Consequences of Suburban Sprawl: The Case of Metro Atlanta.” Georgia State University Law Review 17(4): 9γ6- 989. Available at: http://scholarworks.gsu.edu/gsulr/vol17/iss4/11

Camill, P. β010. “Global Change.” Nature Education Knowledge γ(10): 49.

Campbell, James B. 2002. Introduction to Remote Sensing, Third Edition. New York: The Guilford Press.

Carlson, T. N., and S. T. Arthur. 2000. "The impact of land use - land cover changes due to urbanization on surface microclimate and hydrology: a satellite perspective." Global and Planetary Change 25 (1-2):49-65. doi: 10.1016/s0921-8181(00)00021-7.

Carrion-Flores, C., and E. G. Irwin. 2004. "Determinants of residential land-use conversion and sprawl at the rural-urban fringe." American Journal of Agricultural Economics 86 (4):889-904. doi: 10.1111/j.0002-9092.2004.00641.x.

Chavez, P. S. 1996. "Image-based atmospheric corrections revisited and improved." Photogrammetric Engineering and Remote Sensing 62 (9):1025-36.

Chen, S. P., S. Zeng, and C. G. Xie. 2000. "Remote sensing and GIS for urban growth analysis in China." Photogrammetric Engineering and Remote Sensing 66 (5):593-8.

Clarke, K. C., and L. J. Gaydos. 1998. "Loose-coupling a cellular automaton model and GIS: long-term urban growth prediction for San Francisco and Washington/Baltimore." International Journal of Geographical Information Science 12 (7):699-714.

Clarke, K. C., S. Hoppen, and L. Gaydos. 1997. "A self-modifying cellular automaton model of historical urbanization in the San Francisco Bay area." Environment and Planning B- Planning & Design 24 (2):247-61.

Cliff, A. D., and J. K. Ord. 1981. Spatial Processes Models and Applications. London: Pion Ltd.

Coiacetto, E. J. β000. “Places Shape Place Shapers? Real Estate Developers' Outlooks Concerning Community, Planning and Development Differ between Places.” Planning Practice and Research 15: 353-37.

Commoner, Barry. The Closing Circle. New York: Alfred A. Knopf.

119

Congalton, R. G. 1991. "A Review Of Assessing The Accuracy Of Classifications Of Remotely Sensed Data." Remote Sensing of Environment 37 (1):35-46. doi: 10.1016/0034- 4257(91)90048-b.

Couclelis, H. β001. “Modeling Frameworks, Paradigms, and Approaches.” In Geographic Information Systems and Environmental Modeling, edited by K.C. Clarke, B.O. Parks, M.P. Crane, 36-50. Upper Saddle River, New Jersey: Prentice Hall.

Dear, M., and S. Flusty. 1998. "Postmodern urbanism." Annals of the Association of American Geographers 88 (1):50-72.

Dear, Michael J. 2000. The Postmodern Urban Condition. Oxford, OX., England: Wiley- Blackwell.

Dennison, P. E., K. Q. Halligan, and D. A. Roberts. 2004. "A comparison of error metrics and constraints for multiple endmember spectral mixture analysis and spectral angle mapper." Remote Sensing of Environment 93 (3):359-67. doi: 10.1016/j.rse.2004.07.013.

Dennison, P. E., and D. A. Roberts. 2003. "The effects of vegetation phenology on endmember selection and species mapping in southern California chaparral." Remote Sensing of Environment 87 (2-3):295-309. doi: 10.1016/j.rse.2003.07.001.

Dennison, P. E., and D. A. Roberts. 2003. "Endmember selection for multiple endmember spectral mixture analysis using endmember average RMSE." Remote Sensing of Environment 87 (2-3):123-35. doi: 10.1016/s0034-4257(03)00135-4.

Dennison, P. E., D. A. Roberts, S. R. Thorgusen, J. C. Regelbrugge, D. Weise, and C. Lee. 2003. "Modeling seasonal changes in live fuel moisture and equivalent water thickness using a cumulative water balance index." Remote Sensing of Environment 88 (4):442-52. doi: 10.1016/j.rse.2003.08.015.

Eastman, Ronald J. 1997. Idrisi for windows: user's guide: Clark University, Graduate School of Geography.

Elvidge, C.D., P.C. Sutton, T.W. Wagner, R. Ryznar, S. J. Goetz, A. J. Smith, C. Jantz, K.C. Seto, M. L. Imhoff, Y. Q.Yang, C.Milesi, and R.Nemani. β004. “Urbanization” In Land Change Science: Observing, Monitoring, and Understanding Trajectories of Change on the Earth's Surface, edited by G. Gutman, A. Janetos, C. Justice, E. F. Moran, J.F. Mustard, R. R. Rindfuss, D. Skole, B. L. Turner II, and M. A. Cochrane, 315-328. New York: Springer.

ESRI. 2010. ArcGIS Spatial Analyst: Advanced Raster Spatial Analysis.

Evans, T. P., and H. Kelley. 2004. "Multi-scale analysis of a household level agent-based model of landcover change." Journal of Environmental Management 72 (1-2):57-72. doi: 10.1016/j.jenvman.2004.02.008.

120

Evans, T. P., and H. Kelley. 2008. "Assessing the transition from deforestation to forest regrowth with an agent-based model of land cover change for south-central Indiana (USA)." Geoforum 39 (2):819-32. doi: 10.1016/j.geoforum.2007.03.010.

FAO Global Forest Resources Assessment. 2010.

Fisher, P. β005. “The Property Development Process: Case Studies from Grainger Town.” Property Management 23: 158–175.

Foody, G. M. 2000. "Estimation of sub-pixel land cover composition in the presence of untrained classes." Computers & Geosciences 26 (4):469-78. doi: 10.1016/s0098-3004(99)00125- 9.

Foody, G. M., N. A. Campbell, N. M. Trodd, and T. F. Wood. 1992. "Derivation And Applications Of Probabilistic Measures Of Class Membership From The Maximum- Likelihood Classification." Photogrammetric Engineering and Remote Sensing 58 (9):1335-41.

Forster, B. 1983. "Some Urban Measurements From Landsat Data." Photogrammetric Engineering and Remote Sensing 49 (12):1693-707.

Franke, J., D. A. Roberts, K. Halligan, and G. Menz. 2009. "Hierarchical Multiple Endmember Spectral Mixture Analysis (MESMA) of hyperspectral imagery for urban environments." Remote Sensing of Environment 113 (8):1712-23. doi: 10.1016/j.rse.2009.03.018.

Gao, J., and D. Skillcorn. 1998. "Capability of SPOT XS data in producing detailed land cover maps at the urban-rural periphery." International Journal of Remote Sensing 19 (15):2877-91. doi: 10.1080/014311698214325.

Garreau, Joel. 1991. Edge City: Life on the New Frontier. New York: Anchor.

Geist, H. J., and E. F. Lambin. 2002. "Proximate causes and underlying driving forces of tropical deforestation." Bioscience 52 (2):143-50. doi: 10.1641/0006- 3568(2002)052[0143:pcaudf]2.0.co;2.

Gibson, C. C., E. Ostrom, and T. K. Ahn. 2000. "The concept of scale and the human dimensions of global change: a survey." Ecological Economics 32 (2):217-39. doi: 10.1016/s0921- 8009(99)00092-0.

Gillen, M. and Fisher, P. β00β. “Residential Developer Behaviour in Land Price Determination.” Journal of Property Research 19: 39-59.

Gotway, C. A., and L. J. Young. 2002. "Combining incompatible spatial data." Journal of the American Statistical Association 97 (458):632-48. doi: 10.1198/016214502760047140.

Guindon, B., Y. Zhang, and C. Dillabaugh. 2004. "Landsat urban mapping based on a combined spectral-spatial methodology." Remote Sensing of Environment 92 (2):218-32. doi: 10.1016/j.rse.2004.06.015.

121

Gutman, Garik, Anthony C. Janetos, Christopher O. Justice, Emilio F. Moran, John F. Mustard, Ronald R. Rindfuss, David Skole, Billy Lee Turner Ii, and Mark A. Cochrane. 2004. Land Change Science: Observing, Monitoring and Understanding Trajectories of Change on the Earth's Surface. Dordrecht; New York: Springer.

Haberl, H., K. H. Erb, F. Krausmann, V. Gaube, A. Bondeau, C. Plutzar, S. Gingrich, W. Lucht, and M. Fischer-Kowalski. 2007. "Quantifying and mapping the human appropriation of net primary production in earth's terrestrial ecosystems." Proceedings of the National Academy of Sciences of the United States of America 104 (31):12942-5. doi: 10.1073/pnas.0704243104.

Haith, D. A., and L. L. Shoemaker. 1987. "Generalized Watershed Loading Functions For Stream-Flow Nutrients." Water Resources Bulletin 23 (3):471-8.

Hall, Tim. 2000. Urban Geography: Routledge.

Herold, M., H. Couclelis, and K. C. Clarke. 2005. "The role of spatial metrics in the analysis and modeling of urban land use change." Computers, Environment and Urban Systems 29 (4):369–99.

Herold, M., M. E. Gardner, and D. A. Roberts. 2003. "Spectral resolution requirements for mapping urban areas." Ieee Transactions on Geoscience and Remote Sensing 41 (9):1907-19. doi: 10.1109/tgrs.2003.815238.

Herold, M., N. C. Goldstein, and K. C. Clarke. 2003. "The spatiotemporal form of urban growth: measurement, analysis and modeling." Remote Sensing of Environment 86 (3):286-302. doi: 10.1016/s0034-4257(03)00075-0.

Herold, M., J. Scepan, and K. C. Clarke. 2002. "The use of remote sensing and landscape metrics to describe structures and changes in urban land uses." Environment and Planning A 34 (8):1443-58. doi: 10.1068/a3496.

Holt, James B., and Hua Lu. 2011. "Dasymetric mapping for population and sociodemographic data redistribution." In Urban remote sensing: monitoring, synthesis and modelling in the urban environment edited by X. Yang, 195-210. Chichester, UK: Wiley-Blackwell.

Hoshino, S. 2001. "Multilevel modeling on farmland distribution in Japan." Land Use Policy 18 (1):75-90. doi: 10.1016/s0264-8377(00)00048-x.

Howard, D. M., P. J. A. Howard, and D. C. Howard. 1995. "A Markov Model Projection Of Soil Organic-Carbon Stores Following Land-Use Changes." Journal of Environmental Management 45 (3):287-302. doi: 10.1006/jema.1995.0076.

Hu, Z. Y., and C. P. Lo. 2007. "Modeling urban growth in Atlanta using logistic regression." Computers Environment and Urban Systems 31 (6):667-88. doi: 10.1016/j.compenvurbsys.2006.11.001.

122

Irwin, E. G. β006. “Determinants of Land-Cover and Land-Use Change at Different Spatial Scales.” Paper presented at the Urban Ecology Workshop, Duluth, MN, May 18-20.

Irwin, E. G., and N. E. Bockstael. 2002. "Interacting agents, spatial externalities and the evolution of residential land use patterns." Journal of Economic Geography 2 (1):31-54. doi: 10.1093/jeg/2.1.31.

Jantz, C. A., S. J. Goetz, and M. K. Shelley. 2004. "Using the SLEUTH urban growth model to simulate the impacts of future policy scenarios on urban land use in the Baltimore- Washington metropolitan area." Environment and Planning B-Planning & Design 31 (2):251-71. doi: 10.1069/b2983.

Jenerette, G. D., and J. G. Wu. 2001. "Analysis and simulation of land-use change in the central Arizona-Phoenix region, USA." Landscape Ecology 16 (7):611-26. doi: 10.1023/a:1013170528551.

Jensen, J.R. 198γ. “Biophysical Remote-Sensing.” Annals of the Association of American Geographers 73(1): 111-132.

Jensen, J. R. 2004. Introductory Digital Image Processing. Upper Saddle River, N.J: Prentice Hall.

Ji, Wei, Jia Ma, Rima Wahab Twibell, and Karen Underhill. 2006. "Characterizing urban sprawl using multi-stage remote sensing images and landscape metrics." Computers Environment and Urban Systems 30 (6):861-79. doi: 10.1016/j.compenvurbsys.2005.09.002.

Jin, H. R., and G. Mountrakis. 2013. "Integration of urban growth modelling products with image-based urban change analysis." International Journal of Remote Sensing 34 (15):5468-86. doi: 10.1080/01431161.2013.791760.

Kaplan, David, James O. Wheeler, and Steven Holloway. 2008. Urban Geography: Wiley.

Kerr, S., S. G. Liu, A. S. P. Pfaff, and R. F. Hughes. 2003. "Carbon dynamics and land-use choices: building a regional-scale multidisciplinary model." Journal of Environmental Management 69 (1):25-37. doi: 10.1016/s0301-4797(03)00106-3.

Kolasa, J. and C.D. Rollo 1991. “Introduction: The Heterogeneity of Heterogeneity: A Glossary.” In Ecological Heterogeneity, Edited by J. Kolasa and S.T.A. Pickett, 1-23. New York: Springer-Verlag.

Lambin, Eric F., and Helmut J. Geist. 2006. Land-Use and Land-Cover Change: Local Processes and Global Impacts: Springer.

Lambin, E. F., B. L. Turner, H. J. Geist, S. B. Agbola, A. Angelsen, J. W. Bruce, O. T. Coomes, et al. 2001. "The causes of land-use and land-cover change: moving beyond the myths." Global Environmental Change-Human and Policy Dimensions 11 (4):261-9. doi: 10.1016/s0959-3780(01)00007-3.

123

Langford, M., and D. J. Unwin. 1994. "Generating And Mapping Population-Density Surfaces Within A Geographical Information-System." Cartographic Journal 31 (1):21-6.

Legendre, P. and L. Legendre. 1998. Numerical Ecology: Developments in Environmental Modelling 20. Amsterdam, the Netherlands: Elsevier.

Lesschen, J.P., P.H. Verburg, and S.J. Staal. 2005. Statistical Methods for Analysing the Spatial Dimension of Changes in Land Use and Farming Systems. LUCC report series No. 7. International Livestock Research Institute/LUCC Focus 3 Office, Nairobi/Wageningen.

Li, X., and X. P. Liu. 2007. "Defining agents' behaviors to simulate complex residential development using multicriteria evaluation." Journal of Environmental Management 85 (4):1063-75. doi: 10.1016/j.jenvman.2006.11.006.

Li, X., and A. G. Yeh. 2002. "Neural-network-based cellular automata for simulating multiple land use changes using GIS." International Journal of Geographical Information Science 16 (4):323-43. doi: 10.1080/13658810210137004.

Li, X., and A. G. O. Yeh. 2004. "Analyzing spatial restructuring of land use patterns in a fast growing region using remote sensing and GIS." Landscape and Urban Planning 69 (4):335-54. doi: 10.1016/j.landurbplan.2003.10.033.

Ligmann-Zielinska, Arika. 2009. "The impact of risk-taking attitudes on a land use pattern: an agent-based model of residential development." Journal of Land Use Science 4 (4):215- 32.

Ligmann-Zielinska, A., and P. Jankowski. 2010. "Exploring normative scenarios of land use development decisions with an agent-based simulation laboratory." Computers Environment and Urban Systems 34 (5):409-23. doi: 10.1016/j.compenvurbsys.2010.05.005.

Lin, Y. P., N. M. Hong, P. J. Wu, and C. J. Lin. 2007. "Modeling and assessing land-use and hydrological processes to future land-use and climate change scenarios in watershed land-use planning." Environmental Geology 53 (3):623-34. doi: 10.1007/s00254-007- 0677-y.

Liu, D. S., and S. S. Cai. 2012. "A Spatial-Temporal Modeling Approach to Reconstructing Land-Cover Change Trajectories from Multi-temporal Satellite Imagery." Annals of the Association of American Geographers 102 (6):1329-47. doi: 10.1080/00045608.2011.596357.

Liu, Hua, and Qihao Weng. 2013. "Landscape metrics for analysing urbanization-induced land use and land cover changes." Geocarto International 0 (0):1-12.

Liu, J. G., T. Dietz, S. R. Carpenter, M. Alberti, C. Folke, E. Moran, A. N. Pell, et al. 2007. "Complexity of coupled human and natural systems." Science 317 (5844):1513-6. doi: 10.1126/science.1144004.

124

Liu, T., and X. J. Yang. 2013. "Mapping vegetation in an urban area with stratified classification and multiple endmember spectral mixture analysis." Remote Sensing of Environment 133:251-64. doi: 10.1016/j.rse.2013.02.020.

Liu, W. G., and K. C. Seto. 2008. "Using the ART-MMAP neural network to model and predict urban growth: a spatiotemporal data mining approach." Environment and Planning B- Planning & Design 35 (2):296-317. doi: 10.1068/b3312.

Lo, C. P. 2004. "Testing urban theories using remote sensing." GIScience & Remote Sensing 41 (2):95–115.

Lo, C. P., and X. J. Yang. 2002. "Drivers of land-use/land-cover changes and dynamic modeling for the Atlanta, Georgia Metropolitan Area." Photogrammetric Engineering and Remote Sensing 68 (10):1073-82.

Lu, D. S., E. Moran, and M. Batistella. 2003. "Linear mixture model applied to Amazonian vegetation classification." Remote Sensing of Environment 87 (4):456-69. doi: 10.1016/j.rse.2002.06.001.

Lu, D. S., and Q. H. Weng. 2004. "Spectral mixture analysis of the urban landscape in Indianapolis with landsat ETM plus imagery." Photogrammetric Engineering and Remote Sensing 70 (9):1053-62.

Lu, D. S., and Q. H. Weng. 2006. "Use of impervious surface in urban land-use classification." Remote Sensing of Environment 102 (1-2):146-60. doi: 10.1016/j.rse.2006.02.010.

Mahiny, A. S., and K. C. Clarke. 2012. "Guiding SLEUTH land-use/land-cover change modeling using multicriteria evaluation: towards dynamic sustainable land-use planning." Environment and Planning B-Planning & Design 39 (5):925-44. doi: 10.1068/b37092.

Manson, Steven M., and Tom Evans. 2007. "Agent-based modeling of deforestation in southern Yucatan, Mexico, and reforestation in the Midwest United States." Proceedings of the National Academy of Sciences of the United States of America 104 (52):20678-83. doi: 10.1073/pnas.0705802104.

Martin, David. 1996. Geographic Information Systems: Socioeconomic Applications. London ; New York: Routledge.

Matthews, R. B., N. G. Gilbert, A. Roach, J. G. Polhill, and N. M. Gotts. 2007. "Agent-based land-use models: a review of applications." Landscape Ecology 22 (10):1447-59. doi: 10.1007/s10980-007-9135-1.

McConnell, W. 2002. "Meeting in the middle: the challenge of meso-level integration." Land Use Policy 19 (1):99-101. doi: 10.1016/s0264-8377(01)00042-4.

Miller, John H., and Scott E. Page. 2007. Complex Adaptive Systems: An Introduction to Computational Models of Social Life: Princeton University Press.

125

Miller, Matthew D. 2012. "The impacts of Atlanta's urban sprawl on forest cover and fragmentation." Applied Geography 34:171-9. doi: 10.1016/j.apgeog.2011.11.010.

Muller, M. R., and J. Middleton. 1994. "A Markov Model Of Land-Use Change Dynamics In The Niagara Region, Ontario, Canada." Landscape Ecology 9 (2):151-7.

Mumby, P. J., and A. J. Edwards. 2002. "Mapping marine environments with IKONOS imagery: enhanced spatial resolution can deliver greater thematic accuracy." Remote Sensing of Environment 82 (2-3):248-57. doi: 10.1016/s0034-4257(02)00041-x.

Mundia, C. N., and M. Aniya. 2005. "Analysis of land use/cover changes and urban expansion of Nairobi city using remote sensing and GIS." International Journal of Remote Sensing 26 (13):2831-49. doi: 10.1080/01431160500117865.

Myint, S. W. 2006. "Urban vegetation mapping using sub-pixel analysis and expert system rules: A critical approach." International Journal of Remote Sensing 27 (13):2645-65. doi: 10.1080/01431160500534630.

Myint, S. W., and G. S. Okin. 2009. "Modelling land-cover types using multiple endmember spectral mixture analysis in a desert city." International Journal of Remote Sensing 30 (9):2237-57. doi: 10.1080/01431160802549328.

Myint, S.W., and D. Stow. 2011. "An object-oriented pattern recognition approach for urban classification." In Urban remote sensing: monitoring, synthesis and modelling in the urban environment, edited by X. Yang, 129-40. Chichester, UK: Wiley-Blackwell.

National Research Council. 2013. Advancing Land Change Modeling: Opportunities and Research Requirements. Washington, DC: The National Academies Press.

Nichol, J., and C. M. Lee. 2005. "Urban vegetation monitoring in Hong Kong using high resolution multispectral images." International Journal of Remote Sensing 26 (5):903-18. doi: 10.1080/01431160412331291198.

North, Michael J., Nicholson T. Collier, and Jerry R. Vos. 2006. "Experiences creating three implementations of the repast agent modeling toolkit." Acm Transactions on Modeling and Computer Simulation 16 (1):1-25. doi: 10.1145/1122012.1122013.

Nowak, D. J., R. A. Rowntree, E. G. McPherson, S. M. Sisinni, E. R. Kerkmann, and J. C. Stevens. 1996. "Measuring and analyzing urban tree cover." Landscape and Urban Planning 36 (1):49-57. doi: 10.1016/s0169-2046(96)00324-6.

Oneill, R. V., S. J. Turner, V. I. Cullinan, D. P. Coffin, T. Cook, W. Conley, J. Brunt, J. M. Thomas, M. R. Conley, and J. Gosz. 1991. "Multiple Landscape Scales - An Intersite Comparison." Landscape Ecology 5 (3):137-44. doi: 10.1007/bf00158061.

Openshaw, S. and P.J. Taylor. 1979. “A Million or so Correlation Coefficients: Three Experiments on The Modifiable Areal Unit Problem.” In Statistical Applications in The Spatial Sciences, edited by N. Wrigley, 127–144. London: Pion.

126

Parker, D. C., S. M. Manson, M. A. Janssen, M. J. Hoffmann, and P. Deadman. 2003. "Multi- agent systems for the simulation of land-use and land-cover change: A review." Annals of the Association of American Geographers 93 (2):314-37.

Parker, D. C., and V. Meretsky. 2004. "Measuring pattern outcomes in an agent-based model of edge-effect externalities using spatial metrics." Agriculture Ecosystems & Environment 101 (2-3):233-50. doi: 10.1016/j.agee.2003.09.007.

Pauleit, S., and F. Duhme. 2000. "Assessing the environmental performance of land cover types for urban planning." Landscape and Urban Planning 52 (1):1-20. doi: 10.1016/s0169- 2046(00)00109-2.

Pauleit, S., R. Ennos, and Y. Golding. 2005. "Modeling the environmental impacts of urban land use and land cover change - a study in Merseyside, UK." Landscape and Urban Planning 71 (2-4):295-310. doi: 10.1016/j.landurbplan.2004.03.009.

Peiser, R. 1990. "Charting The 1990s - Who Plans America - Planners Or Developers." Journal of the American Planning Association 56 (4):496-503. doi: 10.1080/01944369008975453.

Pontius, R. G., D. Huffaker, and K. Denman. 2004. "Useful techniques of validation for spatially explicit land-change models." Ecological Modelling 179 (4):445-61. doi: 10.1016/j.ecolmodel.2004.05.010.

Pontius, R. G., and L. C. Schneider. 2001. "Land-cover change model validation by an ROC method for the Ipswich watershed, Massachusetts, USA." Agriculture Ecosystems & Environment 85 (1-3):239-48. doi: 10.1016/s0167-8809(01)00187-6.

Powell, Rebecca. 2011. "Characterizing urban subpixel composition using spectral mixture analysis." In Urban remote sensing: monitoring, synthesis and modelling in the urban environment, edited by Xiaojun Yang, 112-28. Chichester, UK: Wiley-Blackwell.

Powell, Rebecca L., and Dar A. Roberts. 2010. "Characterizing Urban Land-Cover Change in Rondônia, Brazil: 1985 to 2000." Journal of Latin American Geography 9 (3):183-211.

Powell, R. L., D. A. Roberts, P. E. Dennison, and L. L. Hess. 2007. "Sub-pixel mapping of urban land cover using multiple endmember spectral mixture analysis: Manaus, Brazil." Remote Sensing of Environment 106 (2):253-67. doi: 10.1016/j.rse.2006.09.005.

Rafiee, R., A. S. Mahiny, N. Khorasani, A. A. Darvishsefat, and A. Danekar. 2009. "Simulating urban growth in Mashad City, Iran through the SLEUTH model (UGM)." Cities 26 (1):19-26. doi: 10.1016/j.cities.2008.11.005.

Randolph, John. 2011. Environmental Land Use Planning and Management: Second Edition. Washington, DC: Island Press.

Rao, Nirmala. 2007. Cities in Transition: Growth, Change and Governance in Six Metropolitan Areas. London: Routledge.

127

Rashed, T., J. R. Weeks, D. Roberts, J. Rogan, and R. Powell. 2003. "Measuring the physical composition of urban morphology using multiple endmember spectral mixture models." Photogrammetric Engineering and Remote Sensing 69 (9):1011-20.

Reidsma, P., T. Tekelenburg, M. van den Berg, and R. Alkemade. 2006. "Impacts of land-use change on biodiversity: An assessment of agricultural biodiversity in the European Union." Agriculture Ecosystems & Environment 114 (1):86-102. doi: 10.1016/j.agee.2005.11.026.

Ridd, M. K. 1995. "Exploring A V-I-S (Vegetation-Impervious Surface-Soil) Model For Urban Ecosystem Analysis Through Remote-Sensing - Comparative Anatomy For Cities." International Journal of Remote Sensing 16 (12):2165-85.

Rindfuss, R. R., S. J. Walsh, B. L. Turner, J. Fox, and V. Mishra. 2004. "Developing a science of land change: Challenges and methodological issues." Proceedings of the National Academy of Sciences of the United States of America 101 (39):13976-81. doi: 10.1073/pnas.0401545101.

Roberts, D. A., P. E. Dennison, M. E. Gardner, Y. Hetzel, S. L. Ustin, and C. T. Lee. 2003. "Evaluation of the potential of Hyperion for fire danger assessment by comparison to the Airborne Visible/Infrared Imaging Spectrometer." Ieee Transactions on Geoscience and Remote Sensing 41 (6):1297-310. doi: 10.1109/tgrs.2003.812904.

Roberts, D. A., M. Gardner, R. Church, S. Ustin, G. Scheer, and R. O. Green. 1998. "Mapping chaparral in the Santa Monica Mountains using multiple endmember spectral mixture models." Remote Sensing of Environment 65 (3):267-79.

Roberts, D. A., I. Numata, K. Holmes, G. Batista, T. Krug, A. Monteiro, B. Powell, and O. A. Chadwick. 2002. "Large area mapping of land-cover change in Rondonia using multitemporal spectral mixture analysis and decision tree classifiers." Journal of Geophysical Research-Atmospheres 107 (D20). doi: 8073 10.1029/2001jd000374.

Sanders, R. A. 1986. "Urban Vegetation Impacts On The Hydrology Of Dayton, Ohio." Urban Ecology 9 (3-4):361-76. doi: 10.1016/0304-4009(86)90009-4.

Sante, I., A. M. Garcia, D. Miranda, and R. Crecente. 2010. "Cellular automata models for the simulation of real-world urban processes: A review and analysis." Landscape and Urban Planning 96 (2):108-22. doi: 10.1016/j.landurbplan.2010.03.001.

Seto, K. C., and M. Fragkias. 2005. "Quantifying spatiotemporal patterns of urban land-use change in four cities of China with time series landscape metrics." Landscape Ecology 20 (7):871-88. doi: 10.1007/s10980-005-5238-8.

Seto, K. C., and R. K. Kaufmann. 2003. "Modeling the drivers of urban land use change in the Pearl River Delta, China: Integrating remote sensing with socioeconomic data." Land Economics 79 (1):106-21. doi: 10.2307/3147108.

128

Seto, K. C., C. E. Woodcock, C. Song, X. Huang, J. Lu, and R. K. Kaufmann. 2002. "Monitoring land-use change in the Pearl River Delta using Landsat TM." International Journal of Remote Sensing 23 (10):1985-2004. doi: 10.1080/01431160110075532.

Settle, J. J., and N. A. Drake. 1993. "Linear Mixing And The Estimation Of Ground Cover Proportions." International Journal of Remote Sensing 14 (6):1159-77.

Shalaby, Adel, and Ryutaro Tateishi. 2007. "Remote sensing and GIS for mapping and monitoring land cover and land-use changes in the Northwestern coastal zone of Egypt." Applied Geography 27 (1):28-41. doi: 10.1016/j.apgeog.2006.09.004.

Small, C. 2001. "Estimation of urban vegetation abundance by spectral mixture analysis." International Journal of Remote Sensing 22 (7):1305-34.

Small, C. 2001. "Multiresolution analysis of urban reflectance." Ieee/Isprs Joint Workshop on Remote Sensing and Data Fusion over Urban Areas:15-9. doi: 10.1109/dfua.2001.985717.

Small, C. 2003. "High spatial resolution spectral mixture analysis of urban reflectance." Remote Sensing of Environment 88 (1-2):170-86. doi: 10.1016/j.rse.2003.04.008.

Small, C., and J. W. T. Lu. 2006. "Estimation and vicarious validation of urban vegetation abundance by spectral mixture analysis." Remote Sensing of Environment 100 (4):441- 56. doi: 10.1016/j.rse.2005.10.023.

Small, Christopher, and Roberta Balstad Miller. 1999. "Monitoring the Urban Environment from Space." In The International Symposium on Digital Earth. Beijing, China.

Small, Christopher, and Roberta Balstad Miller. 2000. "Spatiotemporal Monitoring of Urban Vegetation." In International Symposium on Remote Sensing of Environment. Capetown, South Africa.

Smith, M. O., S. L. Ustin, J. B. Adams, and A. R. Gillespie. 1990. "Vegetation In Deserts .1. A Regional Measure Of Abundance From Multispectral Images." Remote Sensing of Environment 31 (1):1-26. doi: 10.1016/0034-4257(90)90074-v.

Sohngen, B., R. Mendelsohn, and R. Sedjo. 1999. "Forest management, conservation, and global timber markets." American Journal of Agricultural Economics 81 (1):1-13. doi: 10.2307/1244446.

Song, C. H. 2005. "Spectral mixture analysis for subpixel vegetation fractions in the urban environment: How to incorporate endmember variability?" Remote Sensing of Environment 95 (2):248-63. doi: 10.1016/j.rse.2005.01.002.

Southworth, Jane, and Cerian Gibbes. 2010. "Digital Remote Sensing within the Field of Land Change Science: Past, Present and Future Directions." Geography Compass 4 (12):1695– 712.

129

Steffen, W., A. Sanderson, P. Tyson, J. Jager, P. Matson, B. Moore III, F. Oldfield, K. Richardson, Schellnhuber H-J, B.L. Turner II, R. Wasson. 2004. Global Change and The Earth System: a Planet Under Pressure. Berlin: Springer.

Stevens, D., and S. Dragicevic. 2007. "A GIS-based irregular cellular automata model of land- use change." Environment and Planning B-Planning & Design 34 (4):708-24. doi: 10.1068/b32098.

Sugihara, G., R. May, H. Ye, C. H. Hsieh, E. Deyle, M. Fogarty, and S. Munch. 2012. "Detecting Causality in Complex Ecosystems." Science 338 (6106):496-500. doi: 10.1126/science.1227079.

Sui, D. Z., and H. Zeng. 2001. "Modeling the dynamics of landscape structure in Asia's emerging desakota regions: a case study in Shenzhen." Landscape and Urban Planning 53 (1- 4):37-52.

Sukopp, H., M. Numata, and A. Huber. 1995. Urban Ecology as the Basis of Urban Planning. Edited by H. Sukopp, M. Numata and A. Huber: Backhuys Publishers.

Tang, J., L. Wang, and Z. Yao. 2007. "Spatio-temporal urban landscape change analysis using the Markov chain model and a modified genetic algorithm." International Journal of Remote Sensing 28 (15):3255-71.

Taramelli, A., E. Valentini, L. Cornacchia, S. Mandrone, J. Monbaliu, S. P. G. Hoggart, R. C. Thompson, and B. Zanuttigh. 2014. "Modeling uncertainty in estuarine system by means of combined approach of optical and radar remote sensing." Coastal Engineering 87:77- 96.

Tooke, T. R., N. C. Coops, N. R. Goodwin, and J. A. Voogt. 2009. "Extracting urban vegetation characteristics using spectral mixture analysis and decision tree classifications." Remote Sensing of Environment 113 (2):398-407. doi: 10.1016/j.rse.2008.10.005.

Torrens, P.M. β00β. “Cellular Automata and Multi-Agent Systems as Planning Support Tools.” In Planning Support Systems in Practice, edited by S. S. Geertman and J. Stillwell, 205– 222. London: Springer-Verla.

Torrens, P. M. 2006. "Simulating sprawl." Annals of the Association of American Geographers 96 (2):248-75.

Torrens, P. M., and I. Benenson. 2005. "Geographic automata systems." International Journal of Geographical Information Science 19 (4):385-412. doi: 10.1080/13658810512331325139.

Turner, B. L., E. F. Lambin, and A. Reenberg. 2007. "The emergence of land change science for global environmental change and sustainability." Proceedings of the National Academy of Sciences of the United States of America 104 (52):20666-71. doi: 10.1073/pnas.0704119104.

130

Turner II, B.L., R.H. Moss, and D.L. Skole. 1993. Relating Land Use and Global Land-Cover Change: a Proposal for an IGBP-HDP Core Project. IGBP-HDP Report no 5, IGBP, Stockholm.

Turner II, B.L., D.L. Skole, S. Sanderson, G. Fischer, L. Fresco, and R. Leemans. 1995. Land Use and Land-Cover Change Science/Research Plan. IGBP Report no 35, IGBP, Stockholm.

Turner, M. G. 1989. "Landscape Ecology - The Effect Of Pattern On Process." Annual Review of Ecology and Systematics 20:171-97.

U.S. Global Change Research Program. 2014. Available from: http://globalchange.gov.

Van de Voorde, T., J. Vlaeminck, and F. Canters. 2008. "Comparing different approaches for mapping urban vegetation cover from Landsat ETM+ data: A case study on Brussels." Sensors 8 (6):3880-902. doi: 10.3390/s8063880.

Van Delm, A., and H. Gulinck. 2011. "Classification and quantification of green in the expanding urban and semi-urban complex: Application of detailed field data and IKONOS-imagery." Ecological Indicators 11 (1):52-60. doi: 10.1016/j.ecolind.2009.06.004.

Van Rompaey, A. J. J., G. Govers, and C. Puttemans. 2002. "Modelling land use changes and their impact on soil erosion and sediment supply to rivers." Earth Surface Processes and Landforms 27 (5):481-94. doi: 10.1002/esp.335.

Veldkamp, A., and L. O. Fresco. 1997. "Reconstructing land use drivers and their spatial scale dependence for Costa Rica (1973 and 1984)." Agricultural Systems 55 (1):19-43. doi: 10.1016/s0308-521x(95)00080-o.

Veldkamp, A., P. H. Verburg, K. Kok, G. H. J. de Koning, J. Priess, and A. R. Bergsma. 2001. "The need for scale sensitive approaches in spatially explicit land use change modeling." Environmental Modeling & Assessment 6 (2):111-21. doi: 10.1023/a:1011572301150.

Verburg, Peter H. 2004. "Land use change modelling: current practice and research priorities." GeoJournal (61):309-24.

Verburg, Peter H. 2006. "Simulating feedbacks in land use and land cover change models." Landscape Ecology 21 (8):1171-83. doi: 10.1007/s10980-006-0029-4.

Verburg, P. H., and Y. Q. Chen. 2000. "Multiscale characterization of land-use patterns in China." Ecosystems 3 (4):369-85. doi: 10.1007/s100210000033.

Verburg, P. H., W. Soepboer, A. Veldkamp, R. Limpiada, V. Espaldon, and S. S. A. Mastura. 2002. "Modeling the spatial dynamics of regional land use: The CLUE-S model." Environmental Management 30 (3):391-405. doi: 10.1007/s00267-002-2630-x.

131

Verburg, P. H., J. R. R. van Eck, T. C. M. de Nijs, M. J. Dijst, and P. Schot. 2004. "Determinants of land-use change patterns in the Netherlands." Environment and Planning B-Planning & Design 31 (1):125-50. doi: 10.1068/b307.

Vitousek, P. M. 1994. "Beyond Global Warming - Ecology And Global Change." Ecology 75 (7):1861-76. doi: 10.2307/1941591.

Vogelmann, J. E., S. M. Howard, L. M. Yang, C. R. Larson, B. K. Wylie, and N. Van Driel. 2001. "Completion of the 1990s National Land Cover Data set for the conterminous United States from Landsat Thematic Mapper data and Ancillary data sources." Photogrammetric Engineering and Remote Sensing 67 (6):650-62.

Waddell, P. 2002. "UrbanSim - Modeling urban development for land use, transportation, and environmental planning." Journal of the American Planning Association 68 (3):297-314. doi: 10.1080/01944360208976274.

Waggoner, P. E., and J. H. Ausubel. 2002. "A framework for sustainability science: A renovated IPAT identity." Proceedings of the National Academy of Sciences of the United States of America 99 (12):7860-5. doi: 10.1073/pnas.122235999.

Walsh, S. J., T. W. Crawford, W. F. Welsh, and K. A. Crews-Meyer. 2001. "A multiscale analysis of LULC and NDVI variation in Nang Rong district, northeast Thailand." Agriculture Ecosystems & Environment 85 (1-3):47-64. doi: 10.1016/s0167- 8809(01)00202-x.

Wang, F. 1990. "Fuzzy Supervised Classification Of Remote-Sensing Images." Ieee Transactions on Geoscience and Remote Sensing 28 (2):194-201. doi: 10.1109/36.46698.

Wang, J. D., and G. Mountrakis. 2011. "Developing a multi-network urbanization model: A case study of urban growth in Denver, Colorado." International Journal of Geographical Information Science 25 (2):229-53. doi: 10.1080/13658810903473213.

Welch, R. 1982. "Spatial resolution requirements for urban studies." International Journal of Remote Sensing 3 (2):139-46.

Weng, Q. H. 2002. "Land use change analysis in the Zhujiang Delta of China using satellite remote sensing, GIS and stochastic modelling." Journal of Environmental Management 64 (3):273-84. doi: 10.1006/jema.2001.0509.

Weng, Q. H., and D. S. Lu. 2009. "Landscape as a continuum: an examination of the urban landscape structures and dynamics of Indianapolis City, 1991-2000, by using satellite images." International Journal of Remote Sensing 30 (10):2547-77. doi: 10.1080/01431160802552777.

Wu, C. S. 2004. "Normalized spectral mixture analysis for monitoring urban composition using ETM plus imagery." Remote Sensing of Environment 93 (4):480-92. doi: 10.1016/j.rse.2004.08.003.

132

Wu, C. S., and A. T. Murray. 2003. "Estimating impervious surface distribution by spectral mixture analysis." Remote Sensing of Environment 84 (4):493-505. doi: Pii s0034- 4257(02)00136-0 10.1016/s0034-4257(02)00136-0.

Wu, J. J., R. M. Adams, C. L. Kling, and K. Tanaka. 2004. "From microlevel decisions to landscape changes: An assessment of agricultural conservation policies." American Journal of Agricultural Economics 86 (1):26-41. doi: 10.1111/j.0092-5853.2004.00560.x.

Wu, Qiong, Hong-qing Li, Ru-song Wang, Juergen Paulussen, Yong He, Min Wang, Bi-hui Wang, and Zhen Wang. 2006. "Monitoring and predicting land use change in Beijing using remote sensing and GIS." Landscape and Urban Planning 78 (4):322-33. doi: 10.1016/j.landurbplan.2005.10.002.

Xiao, J. Y., Y. J. Shen, J. F. Ge, R. Tateishi, C. Y. Tang, Y. Q. Liang, and Z. Y. Huang. 2006. "Evaluating urban expansion and land use change in Shijiazhuang, China, by using GIS and remote sensing." Landscape and Urban Planning 75 (1-2):69-80. doi: 10.1016/j.landurbanplan.2004.12.005.

Xie, Y. C., M. Batty, and K. Zhao. 2007. "Simulating emergent urban form using agent-based modeling: Desakota in the suzhou-wuxian region in china." Annals of the Association of American Geographers 97 (3):477-95.

Xie, Y. C., Y. Mei, G. J. Tian, and X. R. Xing. 2005. "Socio-econornic driving forces of arable land conversion: A case study of Wuxian City, China." Global Environmental Change- Human and Policy Dimensions 15 (3):238-52. doi: 10.1016/j.gloenvcha.2005.03.002.

Xie, Y. C., Z. Y. Sha, and M. Yu. 2008. "Remote sensing imagery in vegetation mapping: a review." Journal of Plant Ecology-UK 1 (1):9-23. doi: 10.1093/jpe/rtm005.

Yang, L. M., G. Xian, J. M. Klaver, and B. Deal. 2003. "Urban land-cover change detection through sub-pixel imperviousness mapping using remotely sensed data." Photogrammetric Engineering and Remote Sensing 69 (9):1003-10.

Yang, X., and C. P. Lo. 2002. "Using a time series of satellite imagery to detect land use and land cover changes in the Atlanta, Georgia metropolitan area." International Journal of Remote Sensing 23 (9):1775-98.

Yang, X. J. 2002. "Satellite monitoring of urban spatial growth in the Atlanta metropolitan area." Photogrammetric Engineering and Remote Sensing 68 (7):725-34.

Yang, X. J., and Z. Liu. 2005. "Using satellite imagery and GIS for land-use and land-cover change mapping in an estuarine watershed." International Journal of Remote Sensing 26 (23):5275-96.

Yang, X. J., and C. P. Lo. 2003. "Modelling urban growth and landscape changes in the Atlanta metropolitan area." International Journal of Geographical Information Science 17 (5):463-88. doi: 10.1080/1365881031000086965.

133

Yuan, F., K. E. Sawaya, B. C. Loeffelholz, and M. E. Bauer. 2005. "Land cover classification and change analysis of the Twin Cities (Minnesota) Metropolitan Area by multitemporal Landsat remote sensing." Remote Sensing of Environment 98 (2-3):317-28. doi: 10.1016/j.rse.2005.08.006.

Zhang, X. Y., X. Z. Feng, and H. Jiang. 2010. "Object-oriented method for urban vegetation mapping using IKONOS imagery." International Journal of Remote Sensing 31 (1):177- 96. doi: 10.1080/01431160902882603.

Zhou, X. L., and Y. C. Wang. 2011. "Spatial-temporal dynamics of urban green space in response to rapid urbanization and greening policies." Landscape and Urban Planning 100 (3):268-77. doi: 10.1016/j.landurbplan.2010.12.013.

Zhu, Z. Q., L. M. Liu, Z. T. Chen, J. L. Zhang, and P. H. Verburg. 2010. "Land-use change simulation and assessment of driving factors in the loess hilly region-a case study as Pengyang County." Environmental Monitoring and Assessment 164 (1-4):133-42. doi: 10.1007/s10661-009-0880-2.

134

BIOGRAPHICAL SKETCH

Professional Preparation

 Peking University (China), Urban and Rural Planning, B.S., 2006  Florida State University, Geographic Information Science, M.S., 2009  Florida State University, Geography, Ph.D., 2014

Appointments

 Instructor/Teaching Assistant (Since 2009), Department of Geography, Florida State University  DeVoe L. Moore Dissertation Research Fellow (2012-2013), DeVoe Moore Center, Florida State University  Research Assistant (2010-2011), Department of Geography, Florida State University  Web Developer (Intern; 2009), Florida Resources and Environmental Analysis Center (FREAC)

Publications

 Liu, T. and Yang, X. Operationalizing GIS-based land change modeling (to be submitted to International Journal of Geographical Information Science; in process)  Liu, T. and Yang, X. A scale dependent analysis of the factors driving urban land use change (to be submitted to Computers, Environment and Urban Systems; in process)  Yang, X. and Liu, T. Quantifying land patterns and estuarine nitrogen loading relationship at four different aggregation units (to be submitted to Journal of Coastal Research; in process)  Liu, T. and Yang, X. Monitoring land changes in an urban area using satellite imagery, GIS and landscape metrics (under review; submitted to Applied Geography)  Liu, T. and Yang, X. 2013. Mapping vegetation in an urban area with stratified classification and multiple endmember spectral mixture analysis. Remote Sensing of Environment, 130:251-264  Liu, T. and Yang, X. 2012. Geospatial modeling of urban landscape changes through an agent-based approach. Proceedings of the 2012 AutoCarto International Symposium on Automated Cartography, Columbus, Ohio, September 16-18, 2012  Zhao, T., Brown, D., Fang, H., Theobald, D., Liu, T., and Zhang, T. 2012. Vegetation productivity consequences of human settlement growth in the eastern United States. Landscape Ecology, 27(8):1149-1165

135