<%BANNER%>

Interdisciplinary Contributions to Spatial and Temporal Analyses for Land Cover Change

Permanent Link: http://ufdc.ufl.edu/UFE0022765/00001

Material Information

Title: Interdisciplinary Contributions to Spatial and Temporal Analyses for Land Cover Change
Physical Description: 1 online resource (158 p.)
Language: english
Creator: Marsik, Matthew
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2008

Subjects

Subjects / Keywords: bolivia, change, climate, cover, florida, fragmentation, gis, hydrology, interdisciplinary, land, landscape, modeling, pando, pattern, process, remote, sensing, spatial, tarcoles, temporal, variability, virilla
Geography -- Dissertations, Academic -- UF
Genre: Geography thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Land cover change drives global change through interactions with climate, ecosystem processes, biogeochemical cycles, biodiversity, and human activities. By understanding the linkages between the patterns and causes of land cover change, we can predict possible outcomes and offer alternatives to ameliorate potential negative effects. Methods that link and analyze spatial and temporal patterns of land cover improve the understanding of the complexity and dynamics of land cover change in various geographic locations. I present intedisciplinary methods and accompanying theory from landscape ecology, remote sensing and hydrologic modeling for analyzing impacts and characteristics of the spatial and temporal changes of land cover. Development of new multiscalar analytical methods to analyze observed patterns captured by remotely sensed data can yield appropriate temporal and spatial scale domains important for the applied study of linking landscape and ecological patterns and processes. The incorporation and scaling of spatial, temporal and spectral information into land cover change analyses and greatly improves the amount of information obtained facilitating linkages between landscape pattern and process. Combining techniques from remote sensing and landscape ecology allows the observation of land cover and the analysis of changing patterns and rates of the spatial structure in northern Bolivia. The combined simulation modeling and statistical analysis to investigate changing discharge contributions to the Rio Grande de Tarcoles an integrated approach to assess the impacts of land cover change within a coupled hydro-climatic system. Despite confounding results of determining whether climate variability or anthropogenic causes created the observed change in river discharge, the combined approach highlights the complexity of the hydro-climatic system and investigates, simultaneously, the various aspects of the spatial and temporal heterogeneity of a complex watershed. The goal is to combine the strengths of each discipline to form intedisciplinary tenants for spatial-temporal analyses for land cover change. A intedisciplinary approach to land cover change would comprise pattern-process linkages across multiple scales with the spatial analysis of historic regional and global remotely sensed data with the detailed description of ecosystem sub-processes.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Matthew Marsik.
Thesis: Thesis (Ph.D.)--University of Florida, 2008.
Local: Adviser: Waylen, Peter R.
Local: Co-adviser: Southworth, Jane.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2010-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2008
System ID: UFE0022765:00001

Permanent Link: http://ufdc.ufl.edu/UFE0022765/00001

Material Information

Title: Interdisciplinary Contributions to Spatial and Temporal Analyses for Land Cover Change
Physical Description: 1 online resource (158 p.)
Language: english
Creator: Marsik, Matthew
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2008

Subjects

Subjects / Keywords: bolivia, change, climate, cover, florida, fragmentation, gis, hydrology, interdisciplinary, land, landscape, modeling, pando, pattern, process, remote, sensing, spatial, tarcoles, temporal, variability, virilla
Geography -- Dissertations, Academic -- UF
Genre: Geography thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Land cover change drives global change through interactions with climate, ecosystem processes, biogeochemical cycles, biodiversity, and human activities. By understanding the linkages between the patterns and causes of land cover change, we can predict possible outcomes and offer alternatives to ameliorate potential negative effects. Methods that link and analyze spatial and temporal patterns of land cover improve the understanding of the complexity and dynamics of land cover change in various geographic locations. I present intedisciplinary methods and accompanying theory from landscape ecology, remote sensing and hydrologic modeling for analyzing impacts and characteristics of the spatial and temporal changes of land cover. Development of new multiscalar analytical methods to analyze observed patterns captured by remotely sensed data can yield appropriate temporal and spatial scale domains important for the applied study of linking landscape and ecological patterns and processes. The incorporation and scaling of spatial, temporal and spectral information into land cover change analyses and greatly improves the amount of information obtained facilitating linkages between landscape pattern and process. Combining techniques from remote sensing and landscape ecology allows the observation of land cover and the analysis of changing patterns and rates of the spatial structure in northern Bolivia. The combined simulation modeling and statistical analysis to investigate changing discharge contributions to the Rio Grande de Tarcoles an integrated approach to assess the impacts of land cover change within a coupled hydro-climatic system. Despite confounding results of determining whether climate variability or anthropogenic causes created the observed change in river discharge, the combined approach highlights the complexity of the hydro-climatic system and investigates, simultaneously, the various aspects of the spatial and temporal heterogeneity of a complex watershed. The goal is to combine the strengths of each discipline to form intedisciplinary tenants for spatial-temporal analyses for land cover change. A intedisciplinary approach to land cover change would comprise pattern-process linkages across multiple scales with the spatial analysis of historic regional and global remotely sensed data with the detailed description of ecosystem sub-processes.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Matthew Marsik.
Thesis: Thesis (Ph.D.)--University of Florida, 2008.
Local: Adviser: Waylen, Peter R.
Local: Co-adviser: Southworth, Jane.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2010-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2008
System ID: UFE0022765:00001


This item has the following downloads:


Full Text

PAGE 1

1 INTERDISCIPLINARY CONTRIBUTIONS TO SPATIAL AND TEMPORAL ANALYSES FOR LAND COVER CHANGE By MATT MARSIK A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2008

PAGE 2

2 2008 Matt Marsik

PAGE 3

3 To the memory of Carolyn Sue Marsik and Fredrick Henry Studenka

PAGE 4

4 ACKNOWLEDGMENTS First, I would like to thank my committee. I am grateful to Peter Waylen for giving me the opportunity to pursue my masters and doctorate degrees with his guidance and supervision. I thank Jane Southworth for her encouragement a nd guidance to incorporate remote sensing and land cover change into my research repertoi re. Tim Fik provided stimulating and challenging discussions about advanced statis tics, most of which I generally understood at the time. I thank Graeme Cumming for always challenging me to expand my conceptual and theoretical bounds, even though my head hurt afterwards. As for those people who directly and indirect ly supported me, I first thank Jim Sloan for introducing me to GIS and making me do it the r ight way. I thank David Coley and Forrest Stevens for being the people I could actually ask for research and techni cal help and advice. Folks in LUECI include Lin Cassidy, Andrea Ch avez, Amy Daniels, Andrea Gaughan, Daniel Godwin, Jackie Hall, Joel Hartter, Luke Rost ant, Gaby Stocks, Tracy Van Holt, Jaime Waggnor, and Miriam Wyman. I also thank the myriad prof essors that supported me during my university education. I thank those student s who endured my GIS classes over the four years I taught. I profusely thank Julia Williams a nd Desiree Price for all their ad ministrative help in making my graduate experience easier. A la rge debt of gratitude goes to Dr. Corene Matyas for allowing access to her super-computer, the Weather Machine. I especially thank Dr Cesar Caviedes for his friendship, guidance, advice, support, and wa tchful eye during my graduate career. I am grateful for general, and at times, speci fic, social interfacing w ith Jose Torres, David Coley, Jim Penn, Jim Sloan, Matt Langholtz, Br ian Becker, Brian Condon, Norman Breuer, Adam and Melania Szylagi, Valerio Gomes, Am y Duchelle, Jon Colburn, and Alfredo Rios. From the Department of Geography I am gratef ul to know Andrea Wolf, Anna Szyniszewska, Carlos Caas, Kofi Abu-Brempong, Keith Year wood, Jim Rasmussen, Kwadwo Owusu, Phillip

PAGE 5

5 Morris, and Cerian Gibbes. Along th ese same lines, I am grateful to have fallen in with and had the chance to play Frisbee with Greg Parent, Lisa Seales, Maria DiGia no, Jon Engels, Miriam Wyman, Forrest Stevens, Alex Cheeseman, Joel Ha rtter, Kelly Biedenweg, Rachel Hartter, JG Collomb, and the random people who showed up to join in the game. I also thank the Latino community I had a chance to know over the past couple of years. I thank my colleague Marvin Quesada fo r collaborating on Costa Rican hydrology and climatology over eight years, providing a place to stay during my travels, taking me on some pretty great tours of Costa Rica, and being a good friend. During my time in Costa Rica Trino Barrantes and Ana Cecilia JimenezBarrantes and their family were always gracious to take me in and treat me as one of their own; for that no amount of words and gr atitude are sufficient. Funding for research conducted in Bolivia was provided by the Working Forests in the Tropics Program at the University of Florid a supported by the Nati onal Science Foundation (DGE-0221599) for field work during summer 2007, a nd by a two-year research assistantship from the National Science Foundation Human and Social Dynamics program (#0527511) awarded to the University of Florida through the Department of Sociology. Angelica Almeyda, Chris Baraloto, Grenville Barnes, Kelly Bieden weg, Eben Broadbent, I. Foster Brown, Liliana Cabrerra, Lin Cassidy, Andrea Chavez, Jaime Ch avez, Graeme Cumming, Frank Paul de la Barra, Monica de Los Rios, Amy Duchelle, Hugo Dueas Linares, Florida Ferreira, Carlos Varelio Gomes, Dean Kenji Vaca, Erika Llanos, El sa Mendoza, Stephen Perz, Karla Rocha, Cara Rockwell, Daniel Rojas, Julio Rojas, Allie Sh enkin, and Jackie Vadjunec helped support and encourage me during my time in Bolivia.

PAGE 6

6 I thank my family for fully supporting me dur ing my graduate career even though, at times, they didnt know where I was at during my field studies. I am just glad I finished my PhD before my kid-sister finished her undergraduate degree. The final thanks, gratitude and support go to Kelly Biedenweg, for being my sunshine during the past two years. Also, earne st earns a spot for being a good cat.

PAGE 7

7 TABLE OF CONTENTS page ACKNOWLEDGMENTS...............................................................................................................4 LIST OF TABLES................................................................................................................. ........10 LIST OF FIGURES................................................................................................................ .......11 ABSTRACT....................................................................................................................... ............13 CHAPTER 1 INTRODUCTION..................................................................................................................15 Land Cover Change.............................................................................................................. ..15 Linking Pattern and Process to Understand Land Cover Change...........................................17 Scale.......................................................................................................................... ..............18 Intedisciplinary Approaches to Spatial-Temporal Analysis..................................................19 Importance of Study............................................................................................................ ...21 2 LINKING SPATIAL AND TEMPORAL VARI ATION AT MULTIPLE SCALES IN A HETEROGENEOUS LANDSCAPE.................................................................................23 Introduction................................................................................................................... ..........23 Study Area..................................................................................................................... .........27 Methods........................................................................................................................ ..........29 Local Variance Analysis..................................................................................................30 Data Analysis.................................................................................................................. .31 Results and Discussion......................................................................................................... ..34 Conclusions and Recommendations.......................................................................................37 3 RATES AND PATTERNS OF LAND COVE R CHANGE and fragmentation IN PANDO, NORTHERN BOLIVIA FROM 1986 TO 2005.....................................................49 Introduction................................................................................................................... ..........49 Remote Sensing for Change Detection............................................................................51 Previous Work in the Amazon........................................................................................52 Study Area..................................................................................................................... .........55 Methods........................................................................................................................ ..........57 Image Selection and Preprocessing.................................................................................57 Image Classification........................................................................................................58 Change Detection............................................................................................................60 Classification Accuracy...................................................................................................61 Fragmentation Analysis...................................................................................................62 Results........................................................................................................................ .............63 Decision Tree Model Accuracies....................................................................................63

PAGE 8

8 Classification and Trajectory Accuracies........................................................................64 Forest and Non-Forest Extents and Rates.......................................................................64 Fragmentation Metrics: Pando and Buffer Extents.........................................................65 Discussion..................................................................................................................... ..........68 Methodological Considerations.......................................................................................68 Extents and Rates of Land Cover Change in Pando........................................................69 Forest Fragmentation Patterns.........................................................................................72 Conclusions.................................................................................................................... .........74 4 CHANGING DISCHARGE CONTRIBUTI ONS TO THE RO GRANDE DE TRCOLES....................................................................................................................... .....89 Introduction................................................................................................................... ..........89 Tropical Hillslope Hydrology: Undisturbed Conditions.................................................89 Tropical Hillslope Hydrology: Altered Conditions.........................................................90 Mesoscale Watershed Studies.........................................................................................91 Research Justification......................................................................................................92 Study Site..................................................................................................................... ...........93 Methods........................................................................................................................ ..........94 SWAT Model..................................................................................................................94 Data........................................................................................................................... .......95 Statistical Analysis: Change detection............................................................................97 Statistical Analysis: Climate Variability.........................................................................98 SWAT Model Construction.............................................................................................98 Results........................................................................................................................ ...........101 Statistical Analysis........................................................................................................101 SWAT Model................................................................................................................102 Simulations....................................................................................................................103 Climate Va riability........................................................................................................104 Discussion..................................................................................................................... ........106 Model Performance.......................................................................................................106 Non-Linear Nature of Sub-Basins.................................................................................107 Challenges of Mesoscale Modeling Studies..................................................................108 Conclusions.................................................................................................................... .......110 5 CONCLUSIONS..................................................................................................................131 Significance of Findings.......................................................................................................131 Identification of Disciplinary Strengths................................................................................133 Recommendations................................................................................................................ .134 APPENDIX A LANDSAT IMAGES USED IN LAND COVE R CHANGE ANALYSIS FOR PANDO, BOLIVIA........................................................................................................................ ......137

PAGE 9

9 B ASTER IMAGES USED IN LAND COVER CHANGE ANALYSIS FOR PANDO, BOLIVIA........................................................................................................................ ......138 C TEST STATISTICS FOR FRAGMENTATION METRICS AT THE PANDO EXTENT......................................................................................................................... ......142 LIST OF REFERENCES.............................................................................................................143 BIOGRAPHICAL SKETCH.......................................................................................................158

PAGE 10

10 LIST OF TABLES Table page 3-1. Accuracy results for classificati on rules developed using Compumine.................................84 3-2. Classification accuracies for 2005 clas sification images, and 2000-2005 trajectory image.......................................................................................................................... ........85 3-3. Absolute values and percentages of forest and non-forest in Pando......................................86 3-4. Percentages of forest and non-forest, and rates of deforestation and reforestation for Pando.......................................................................................................................... ........87 3-5. Rates of deforestation and extents of forest and non-forest from other research conducted in Pando, Bolivia..............................................................................................88 4-1. Model parameters adjusted during manual and automatic calibrations, and sensitivity and uncertainty analyses..................................................................................................123 4-2. Measures of model fit for ma nual calibration and validation...............................................124 4-3. Measures of model fit for automatic calibration..................................................................125 4-4. Output SWAT parameter ranges and pe rcentages of value range from PARASOL automatic calibration routine...........................................................................................126 4-5. Output SWAT parameter ranges and perc entages of value range from SUNGLASSES uncertainty analysis routine.............................................................................................127 4-6. Measures of fit for the PARASOL automa tic calibration routine for each subbasin and land cover year................................................................................................................ .128 4-7. Measures of fit for the SUNGLASSES uncer tainty analysis for each subbasin and land cover year..................................................................................................................... ....129 4-8. Percentages of land cover and land cover change per subbasin...........................................130 A-1 Landsat image platform, path, row, and acquisition date information.................................137 B-1. ASTER images used for classificati on and trajectory accuracy assessment.......................139 B-1. Continued................................................................................................................. ............140 B-1. Continued................................................................................................................. ............141 C-1. Kruskall-Wallis statistics for Mean Patc h Size and Perimeter-Area Corrected metrics......142

PAGE 11

11 LIST OF FIGURES Figure page 2-1. Study area of the southeastern coastal plai n, with focus region and location of Landsat TM image footprint............................................................................................................41 2-2. Landsat TM Image NDVI products created for each image date, each filter size, and each pair of dates............................................................................................................. ..42 2-3. Illustration of the spatial a nd temporal variance calculations................................................43 2-4. Regression models used in analysis........................................................................................44 2-5. Flowchart of analysis methods and null model creation........................................................45 2-6. Adjusted R2 relationships between spatial versus temporal variance, across the different time-steps and window size and area.................................................................................46 2-7. Slope relationships between spatial versus temporal vari ance, across the different timesteps and window size and area.........................................................................................47 2-8. Intercept relationships betw een spatial versus temporal variance, across the different time-steps and window size and area.................................................................................48 3-1. Study area, major roads, population centers, and surrounding geographies..........................76 3-2. Landscape patch size metrics for traj ectory images at the Pando extent................................77 3-3. Landscape patch shape metrics for tr ajectory images at the Pando extent.............................78 3-4. Temporal changes in mean patch size metrics for multi-distance road buffer and distance from Cobija along road........................................................................................79 3-5. Temporal changes in corrected patch peri meter-area metrics for multi-distance road buffer and distance from Cobija along road......................................................................80 3-6. Temporal changes in fractal dimensi on metrics for multi-distance road buffer and distance from Cobija along road........................................................................................81 3-7. Temporal changes in aggregation inde x metrics for multi-distance road buffer and distance from Cobija along road........................................................................................82 3-8. Fragmentation and deforestation by distan ce from road and from Cobija between 2000 and 2005....................................................................................................................... ......83 41. Study area and monthly runoff regimes..............................................................................112

PAGE 12

12 42. Matrix of possible combinations under changing precipitation and land cover conditions for SWAT simulations...................................................................................113 43. Sub basin annual runoff as a percentage of confluence runoff compared to sub basin.......114 44. Deviates from median annual runoff...................................................................................115 45. Regression summaries of monthly runoff step plots...........................................................116 46. Deviates from medi an annual precipitation.........................................................................117 47. Regression summaries of mont hly precipitation step plots.................................................118 48. Mean monthly runoff differences for th e Rio Grande de San Ramon under various land cover and precipitation combinations..............................................................................119 49. Mean monthly runoff differences fo r the Virilla under various land cover and precipitation combinations...............................................................................................120 410. Seasonal standardized precipitation devi ates for the Rios Grande de San Ramon and Virilla........................................................................................................................ .......121 411. Non-linearities in subbasin responses fo r runoff, precipitation, and evapotranspiration..122

PAGE 13

13 Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy INTEDISCIPLINARY CONTRIBUTI ONS TO SPATIAL AND TEMPORAL ANALYSES FOR LAND COVER CHANGE By Matt Marsik December 2008 Chair: Peter Waylen Cochair: Jane Southworth Major: Geography Land cover change drives globa l change through interactions with climate, ecosystem processes, biogeochemical cycles, biodiversit y, and human activities. By understanding the linkages between the patterns and causes of la nd cover change, we can predict possible outcomes and offer alternatives to ameliorate potential negative effects. Methods that link and analyze spatial and temporal patterns of land cover improve the understanding of the complexity and dynamics of land cover change in various geog raphic locations. I pres ent intedisciplinary methods and accompanying theory from lands cape ecology, remote sensing and hydrologic modeling for analyzing impacts and characteristics of the spatial and tem poral changes of land cover. Development of new multiscalar analytical me thods to analyze observed patterns captured by remotely sensed data can yiel d appropriate temporal and spatial scale domains important for the applied study of linking landscape and ecological patterns and processes. The incorporation and scaling of spatial, temporal and spectral information into land cover change analyses and greatly improves the amount of information obt ained facilitating linka ges between landscape pattern and process. Combining techniques from remote sensing and landscape ecology allows

PAGE 14

14 the observation of land cover and the analysis of changing patterns and rates of the spatial structure in northern Bolivia. The combined si mulation modeling and statistical analysis to investigate changing discharge contributions to the Rio Grande de Tarcoles an integrated approach to assess the impacts of land cover cha nge within a coupled hydro-climatic system. Despite confounding results of determining whether climate variability or anthropogenic causes created the observed change in river discharge, the combined a pproach highlights the complexity of the hydro-climatic system and investigates, s imultaneously, the various aspects of the spatial and temporal heterogeneity of a complex watershed. The goal is to combine the strengths of each di scipline to form intedisciplinary tenants for spatial-temporal analyses for land cover change A intedisciplinary approach to land cover change would comprise pattern-p rocess linkages across multiple scales with the spatial analysis of historic regional and global remotely sensed da ta with the detailed description of ecosystem sub-processes.

PAGE 15

15 CHAPTER 1 INTRODUCTION I present intedisciplinary methods and accompanying theory from landscape ecology, remote sensing and hydrologic modeling to charac terize spatial and temporal changes in land cover and their consequent impacts. The resear ch is presented as thre e separate studies in publication format for submission to academic jour nals. The first study (Chapter 2) develops an approach for scaling of space and time as applied to satellite images, which form the core data source for land cover change studies. The second study (Chapter 3) inve stigates spatial and temporal rates and patterns of land cover cha nge and fragmentation in the face of population increase and road development and expansion in Pando, the northernmost department in Bolivia. The final study, (Chapter 4) focuses on the potential interpla y between land cover change and climate variability and their imp act on river discharge. These studi es are topically diverse and are unified by the investigation of me thods to determine the spatial a nd temporal characteristics of land cover change. With a background in physical geography and guiding theory from land change science these studies in corporate concepts from landscape ecology, remote sensing, and hillslope hydrology and modeling to identify spat ial and temporal methods capable of enhancing the understanding of land cover change and its advance as a science. Land Cover Change Land cover exerts a fundamental impact on, a nd links many parts of, human and physical environments and changes in land cover are rega rded as the single most important process of global change affecting ecol ogical systems (Vitousek 1994). Observed land cover patterns represent the net result of indi vidual, communal or societal d ecision-making processes regarding the relative returns to land use (Currie 1981) se t within a local, regional or national context. Human activity by way of individual or societal needs or wants dr ives present land cover change

PAGE 16

16 (Ojima et al. 1994) and these purposes define la nd use the intent and underlying manner in which the biophysical attributes of the land ar e manipulated (Lambin and Geist 2006) which purports land cover change. Land cover change can drive global change th rough interactions with climate, ecosystem processes, biogeochemical cycles, biodiversit y, and human activities (Turner et al. 1995). Important natural resources of climate, soils vegetation, water resources and biodiversity (E.F.Lambin et al. 1999) reflect human histor y and are linked with economic development, population growth, and technology. Global effect s include the conversion of potentially productive land with diminished capacity to support crops, forests and people, and the irreplaceable loss of species and the emission of chemically active and heat-trapping gases to the atmosphere (Ojima et al. 1994). Po tential deleterious lo cal and regional effect s of deforestation to pastures, for example, include erosion of soils, reduced rainfall, reduced capacity of soils to hold water, increased frequency and severity of floods, and siltation of dams (Houghton 1994). Land cover change has been linked to changes in flood and drought freque ncy (Nepstad et al. 2001) and impacts on water quality (Rogers 1994). Forest clearing affect s local and regional hydrology through reduced infiltration and evap otranspiration, and changes to river and groundwater regimes (Giambelluca 2002; Bonell 1998; Bruijnzeel 2004). These myriad effects of land cover change at the human-environment interface necessitate the quantification of and cataloging of land cover change at multiple scales. Gutman (2004) has proposed the development of a multidisciplinary land change science, which draws on the diverse disciplines of geogr aphy, remote sensing, GIScience, and various other social and ecological sciences. To furthe r cultivate the development of this young science, land change theory requires the application of a diverse array of tec hniques of analyses and

PAGE 17

17 models (Rindfuss et al. 2004). The ultimate and broad objective of land cover change is to improve understanding, and to accrue knowledge of regionally based, interactive changes between land uses and covers (Turner et al. 199 5). Thus we can gain an understanding of the nature of human decisions that alter land cover at regional scales where policy interventions are possible and effective (E.F.Lambin et al. 1999). Linking Pattern and Process to Understand Land Cover Change Development of a land change scien ce (Woodcock and Ozdogan 2004) relies on the understanding of linkages between pa tterns and processes. The hierar chical examination of levels in a system as the outcomes of a collection of results from smaller, underlying processes or behaviors (Levin 1992) comprises the main goal of patte rn-process linkages. Remote sensing and spatial statistics, along with other types of sp atial analyses (Schroder and Seppelt 2006), can describe spatial distributions, at multiple scales using multiple data sources and indicate the controls and constraints that determine their nature. Methods of spatial and temporal analysis of land cover change can help better link observed pattern with the underlyi ng processes of formation. Howeve r, there is certain inherent complexity in distinguishing processes of form ation, either anthropogenic or climate driven (Rindfuss et al. 2004). Lambin a nd Geist (2006) note that the complexity of land cover change embodies multiple scales and rates of modificati ons. We have regional change in land cover conditions caused by climate variability overlying localized human-induced land cover modifications. They note that [m]ultiple spatial a nd temporal scales of change, with interactions between climate-driven and anthropogenic changes, are a significant source of complexity in the assessment of land cover change (Lambin and Geist 2006). Analyzing both space and time together in the study of land cover change can help ameliorate this dual complexity.

PAGE 18

18 Interdisciplinary spatial-temporal analysis meth ods are necessary to tackle the complexities associated with land cover ch ange (Rindfuss et al. 2004). Scale The definitions and concepts of scale and sca ling are well known and have be reported in the literature time again (Gibson et al. 2000; Marceau 1999; Meentemeyer 1989; Turner et al. 1989). Currently no unified theory or science of scale and scaling exis ts. Possible steps to identify and remediate scale effects are: (1) to recognize scale dependency in observed phenomena, (2) to predict and control for scal e effects using spatially explicit methods, (3) explicitly state the scale or scales of analysis, and (4) to make use of a solid unified theoretical framework, such as hierarchy theory (Allen and St arr 1982; O'Neill et al. 1989) to derive and test hypotheses, and generalize analysis results. Hierarch ical structuring (O'Ne ill et al. 1989) simply means that, at a given level of resolution, a syst em is composed of interacting components (i.e., lower-level entities) and is itself a component of a larger system (i.e., higher level entity). This concept emphasizes that the behavior of a system is limited (1) by the pote ntial behaviors of its lower level components (i.e., biotic potential) and (2) by the environmental constraints (e.g. abiotic and biotic environmental limits) imposed by higher levels. Remotely sensed data can fit nicely into this concept of hierarchically organized structures as the pixel is the minimum mapping unit and can be easily aggregated upward in a nested fashion from its base resolution. Remote sens ing offers great potential for scaling (Woodcock and Strahler 1987; Quattrochi and Goodchild 199 7; Stewart et al. 1998; Marceau and Hay 1999) as it provides the required data for upscaling and downscaling physical models, provides the possibility of conducting empirical studies to understand the behavior of variables when changing scales, and to derive the appropriate rules for scali ng. Interdisciplinary methods for spatial-temporal analyses of land cover change are inherently multiscalar with the combination

PAGE 19

19 of many pattern forming processe s occurring at many scales. By considering landscapes and land cover change hierarchically organized, multilevel analysis techniques can be applied that are potentially able to resolve cri tical landscape thresholds, domains of scale, and correctly linked data between scales a nd landscape components. Intedisciplinary Approaches to Spatial-Temporal Analysis Landscape ecology has a history of space-time st udies (Turner et al. 1989; Turner 2005; Turner 1990) used primarily to deduce dominant pr ocesses of formation from observed patterns. Patterns observed from remotely sensed data provide the foundation upon which simulation modeling, ideally empirically driven, verified, an d executed at multiple sc ales, facilitates the identification of ecological pr ocesses responsible for genera ting the patterns (Levin 1992). Bellehumeur and Legendre (1998) suggest carefu l planning of sampling design coupled with pattern detection techniques (e.g. spatial correlograms, geostatis tical analyses and frequency analysis) applied at multiple scales to an ecologi cal community or landscape to help detect key scales of spatial variation. W ith a strong foundation in explanation of spatial patterns and the effects of changing scale on key pattern formi ng processes (Turner 1990), the integration of multi-temporal levels is recognized and are inco rporated into sampling designs and analyses (Southworth et al. 2006). To better understand pattern-process interdepen dencies, we need to determine important scales of landscape heterogeneity, human ac tivities, and ecosystem processes that most effectively explain heterogeneous spatial patterns. The first study (Chapter 2) found at smaller spatial scales the spatial varian ce was greater than temporal vari ations in the landscape, that temporal variance increased with increasing sp atial scales, and lands capes show inherent temporal variation without spatial changes in landscape pattern. The development of new methods, as occurs in this study, specifically of multiscalar analysis as applied to observed

PAGE 20

20 patterns captured by remotely sensed data, can yield appropriate temporal and spatial scale domains important for the applied study of linking landscape and ecological patterns and processes. The incorporation and scaling of spa tial, temporal and spectral information into land cover change analyses greatly improves the amount of information obtained. Remote sensing and associated analytical me thods provide the capacity to analyze space and time together (Mertens and Lambin 2000; Gutman 2004; Lambin and Geist 2006). Calculation of land cover traj ectories (Petit and Lambin 2001; Mertens and Lambin 2000; Southworth et al. 2002) creates temporal change classes (i.e., categories between two or more dates), and the resulting image s hows the spatial distribut ion of these changes. This approach has permitted the development of a multivariate spatial models to determine the probability of change in land cover (Mertens and Lambin 2000), enhanced model predictions of the expansion of disturbed and anthropogenic land cover at the expense of na turally vegetated land covers (Petit et al. 2001), and detected reforestation patterns in west ern Honduras (Southworth et al. 2002). The second study (Chapter 3) presents an in tegrated set of tools for the spatial and temporal rates and patterns of land cover change and landscape level fragmentation. The multiscalar analysis of land cover fragmen tation at the regional extent (65,000 km2) and with increasing distances (1-15 km) along a major access road and from a major population center provides a comprehensive spatial and temporal view of land cover change with regard to population increases and road improvements. Thes e spatial-temporal pa tterns coupled with socio-economic data in a modeling framework may permit the identification of social actors and drivers of land cover change.

PAGE 21

21 Hydrologic models often relegate land cover change to a posi tion secondary to description of watershed processes. Land cover is usually ex plicitly represented (Wu et al. 2007; Marsik and Waylen 2006; Marshall and Randhi r 2008) and derived from satellite images or existing GIS data. Specific data describing changes in surface roughness, which affect overland flow velocities, or canopy cover, which controls the partition of precipitation into interception and throughfall, are not commonly measured. Rarer still are the observations and analyses of temporal changes in important hydrologic variab les resulting from land cover change. Instead many detailed hydrologic studies (Elsenbeer et al. 1992; Elsenbeer et al 1999; Western et al. 1999; Western et al. 2001; Western et al. 2004) focus on the spatial distributions of hydrologically important variable s (i.e., soil moisture, hydrauli c conductivity, and other soil hydraulic properties), which occur in small, experi mental watersheds, usually with very little change in land cover and over relatively short time frames. The third study (Chapter 4) investigates the potential ro les of non-lin ear hydrologic responses to changes in land use and land cover, and climate variab ility within two adjacent subbasins in the Central Valley, Costa Rica. A semi-distributed h ydrologic model, SWAT, incorporating historic la nd cover changes delineated is coupl ed with statisti cal analysis of precipitation variability to better understand possible causes of a change noted in the mid-1970s of the relative contribut ions of the Rios Virilla and Gra nde de San Ramn, the principle tributaries of the Rio Grande de Tarcles. This re search illustrates that changes in land cover, contemporary to changes in climate variability may further amplify or dampen the observed response of river discharge. Importance of Study Together these studies contribu te to contemporary methods to analyze spatial and temporal rates and patterns of land cover change base d on techniques used in landscape ecology, remote

PAGE 22

22 sensing, and hillslope hydrology. Current research agendas for land cover change science are based upon methods from multiple disciplines such as geography, remote sensing, sociology, anthropology, and GIScience (Rindfuss et al. 2004; Gutman 2004), which serve to analyze the biophysical and environmental dimens ions of land cover change. The analyses presented in these studies offer one geographic perspective on an in terdisciplinary approach and draw on strengths of spatial analyses used in multiple sub-di sciplines of physical geography. Linking patternforming processes is one outcome of using these types of spatial analyses for land cover change studies.

PAGE 23

23 CHAPTER 2 LINKING SPATIAL AND TEMPORAL VARIAT ION AT MULTIPLE SCALES IN A HETEROGENEOUS LANDSCAPE Introduction The global impacts of the earths human populati on are reflected in extensive changes in the spatial patterns of land cove r and land use (O'Neill et al. 1996). Interacting anthropogenic, ecological and land-surface processes occur in la ndscapes at multiple scales. If we are to understand and manage the causes an d consequences of anthropogenic effects on landscapes, it is imperative that we develop approaches to unde rstanding spatial and temporal variation, the processes that produce the patter ns that we observe, and the wa ys in which pattern-process relationships change with scale. Remote sensing has traditionally been considered an ideal tool for providing data to describe landscape patterns and dynamics However, our understanding of the scale dependency of landscap e pattern-process inte ractions is limited (Moody and Woodcock 1995). Understanding scaling effects is critical to our ability to better understand, model and/or predict landscape dynamics, and specifically for unde rstanding the roles of spatial and temporal heterogeneity and the hierarchical arrangeme nt of landscape elements (Qi and Wu 1996). Most natural systems across the world exhibit so me form of spatial structure. As stated by Wu et al. (2000), Spatial heterogeneity is the most fundamental characteristic of all landscapes and scale multiplicity is inherent in sp atial heterogeneity. Thus, multiscale analysis is imperative for understanding the structure, func tion and dynamics of landscapes. Studies of spatial structure occur across a suite of disc iplines, and many diffe rent approaches and methodologies for the analysis of spatial pattern have been developed (Dal e et al. 2002). It is valuable in such studies to characterize heteroge neity and to assess the ways in which it changes in both space and time (O'Neill et al. 1996). Due in part to the demands of data collection, both landscape ecology and remote sens ing studies typically focus on either spatial or temporal

PAGE 24

24 variation. The central goal of this study is to conduct a geographical analysis of landscape change in North-Central Florida in a way that directly addresse s variation in both space and time by integrating landscape ecology, re mote sensing, and GIS technologies. In this way we hope to link form, both spatially and temporally, to func tion. As stated in Goodchild (2004), geographic information science (GIScience) is an emphasis on form with the inclusion of process to bolster our understanding, which is what we attempt to do in this research using an integration of GIScience and landscape ecology. Landscape studies that are intended to determine pattern-process interacti ons often occur at a single scale, which dictates the relationships that are found and th e patterns of spatiotemporal heterogeneity that are seen (Habeeb et al. 2005). In order to u nderstand pattern-process interdependencies, we need to find the scales of landscape heterogeneity, human activities, and ecosystem processes that most effec tively explain variation in spatial patterns. In other words, spatial and temporal scales of observati on must match those of ecological patterns and processes (Marceau 1999). The question of scale has been a central topic in remote sensing since the earliest days of nonmilitary research, and in geography for even long er (Gehlke and Biel 1934). Openshaw (1977; 1978; 1984) suggested an approach in which the determination of an optimal spatial resolution is a required early step in any spatial analysis. This protocol is supported by some very recent works (Wu and David 2002), although there is also an increasing realization that a multi-scale understanding is important. In the remote sensing arena, Cohen and Justice (1999) hypothesize that there is a fundamental grain size of each landscape (or biome) a bove which error rates accelerate when modeling NPP; test of hypothe sis using field data, Landsat ETM +, and geostatistical models, implying that the fundamental grain size must be determined.

PAGE 25

25 Much attention has been paid to the questions of scale in ecology (P eterson and Parker 1998), environmental science, land-cove r/land-use change (Evans et al 2001) remote sensing and GIS (Quattrochi and Goodchild 1997), hydrology (D roogers and Kite 2002), and geology and geomorphology, even to the point that a book ch apter (O'Neill and King 1998) was titled Why are there so many books on scale? De spite this intellectual activity, the role of scale in remote sensing studies is still poorly understood, in both subjective and obj ective contexts; our scale of measurement influences our perceptions of the wo rld, while landscape processes vary in magnitude and rate across a range of scales. The developmen t of scaling relationships typically involves the repeated measurement of a quantity of interest at different dimensions, whether length, space, mass, or time. For example, the loglinear body mass-meta bolic rate relationship is a classical scaling relationship that is widely cite d in ecology. Many past studies of landscapes have examined how changes in spatial scale (predominantly in grain) can impact measures of habitat configuration and composition (Wu and David 2002). Fewer studies have addressed the role of changing temporal scale in landscape analyses, alt hough there has been substantial re search within the fields of geography, landscape ecology, and remote sensing on temporal variation in ecosystems and anthropogenic influences on landscapes. Many statistical techniques ha ve been developed for the sc aling analysis of remotely sensed imagery. A seminal paper by Woodcock a nd Strahler (1987) prop osed a local variance technique to determine the appropr iate resolution (i.e., cell size) of a remotely sensed image to view the spatial structure of the landscape. They used the mean value of the standard deviation calculated with a 3 x 3 window around each pixel, excluding the edge pixels of the image. The local variance measures denote, within the 3 x 3 window, the similarity or spatial dependence of pixel values on the center pixe l under investigation. Woodcock and Strahler (1987) extended

PAGE 26

26 their method by progressively degrading the image resolution to measure lo cal variation within the image at multiple resolutions. This degradation or coarsening of resolution led to multiple representations of the same la ndscape, with fewer pixels in the coarser resolution images. Measures of local variance at multiple resolutions were then plotted as local variance versus image resolution to determine the scales at which the dominant landscape patterns were occurring. The calculation of the local variance for an image described the size of discrete landscape objects (e.g. forest stands or pastures) within the image and could be used to select a scale at which to analyze these objects. Other researchers have utilized this scaling approach in a variety of applications. For instance, Coops and Catling (1997) applied a modi fied local variance tec hnique to videographic imagery of individual tree canopies. The mean valu e of the standard deviation of each spectral channel was calculated at increasingly large pi xel windows (i.e., 3 x 3, 5 x 5.... 49 x 49), yielding local variance values for each pixel. They f ound low within-window va riance at a window size significantly smaller than individu al tree canopies in the scene. Local variance increased to a maximum as window size increased to the ca nopy size (around 20 m), and constant local variance resulted within the window as the wi ndow size increased beyond the size of the individual tree crowns. Coops and Culvenor (2000) applied their modi fied local variance or texture variance technique to determine if mini mized values of local variance (with window sizes of 5 x 5, 20 x 20, 30 x 30, and 80 x 80 pixels) resulted from regular spatial distribution of high-resolution tree canopy structures within a simulated forest e nvironment. Conversely, maximum variance would result from clumped or aggregated tree canopies As the percentage of canopy cover increased,

PAGE 27

27 the texture variance decreased, with a plateau at the maximum local variance regardless of the inclusion of additional tree canopy object s to a simulated forest scene. Our research expands on some of these ideas in the context of GISc ience, as driven by scientific and technological motivations, to de velop a multiscale approach to investigate the related patterns of spa tial and temporal scales concurren tly (Walsh and Crews-Meyer 2002). Of particular interest is the question of whether uni ed comparisons of spatial and temporal variation across a range of differe nt scales yield any general pa tterns or scaling relationships (Woodcock and Strahler 1987; Wu et al. 2000). In this sense the development of the technique presented here employs commonly used geogr aphic technologies (remote sensing, variance analysis, regression analysis, custom applic ation development for randomizing null landscape models, etc.) to facilitat e analysis of seemingly disparate issu es of temporal and spatial scaling. Our conceptual study lends itself to furthering GIScience in that the development of new methods, specifically of multiscale analysis, can yield appropriate temporal and spatial scale domains important for the applied study of the patterns and processes of landscape analyses. Study Area Our study area is in north centr al and northeast Florida, and southern Georgia on the U.S. Atlantic coastal plain. The area is defined by Landsat WRS 2 path 17, and row 39 (Figure 2-1) with bounding latitude and longitu de coordinates of approximately 31 13' N, 83 10' W (northwest corner), and 29 20' N, 81 41' W (southeast corner). The entire scene footprint covers an area of about 34,000 km2. The study area is useful for a ddressing scaling questions for several reasons. First, the Coasta l Plain of the southeastern Unite d States east of the Mississippi River (Figure 2-1 inset) is large (nearly 600,000 km2), covering a land area 2.5 times the size of the United Kingdom. Second, despite its size, it is also of rela tively low relief, reducing the importance of gross topography (but not fine-sca le topography) as a fa ctor structuring the

PAGE 28

28 landscape, and has a generally uniform and modera te subtropical climate. Elevation across the study landscape ranges from near sea level to a high point of 63 m in the north. The topography is flat to very gently rolling, with slopes > 5% rare. Third, the Coastal Plain has a large and rapidly growing human population with a long history of habita tion and rapid land-use change, and a very high diversity of natural ecosystems and endemic biodiversity, controlled mainly by the subtle variations in topography, geology and soils (Myers 1990). This diverse and heterogeneous landscape consists of a mixture of natural and plantation fo rests, urban centers, urban and rural residential areas, and commercial and small-scale agricultural operations. The dominant upland forests are pine flatwoods, with mixed hardwoods and pines at intermediate moisture levels, leading to bottomland hardw ood or mixed cypress forests near streams and rivers. Only 1.5% of the historical area of the upland Pinus palustris (longleaf pine) forest remains after clearing during the past 150 y ears, although this ecosystem once covered 25 million ha (Myers 1990; Ware et al. 1989). Wide ly spaced trees and frequent, low-intensity ground fires characterized the original forests. Mo st of the natural forests in the region today are mixed pine stands that have regenerated under a variety of fi re regimes after being used for cattle grazing. Most of the flatw oods in Florida were converted to plantations of the native slash pine ( Pinus elliottii ), starting in the 1950s. The total area in planta tions has been roughly constant for the past 20 years, while the total area of forest in the region has declined. The hydrologically isolated cypress we tlands that dot the landscap e are largely unmanaged, although most have had their larger trees harvested at some time in the past century. This di versity of land cover types is spatially heterogeneous and patc h sizes of the various vegetation classes vary across a wide range of scales. The temporal sc ales of management and land-cover change are

PAGE 29

29 also variable depending on the size of forest st ands, areas of rapid urbanization, and kinds of agricultural operations. Methods The data consisted of Landsat TM/ETM+ scenes from the US GS EROS Data Center with level-1G preprocessing. The data used here comp rise six cloud-free images acquired on January 16, 1985 (TM); February 12, 1989 (TM); January 20, 1992 (TM); January 17, 1997 (TM); January 4, 2001 (ETM+); and February 11, 2003 (ETM+). We used NDVI (Normalized Difference Vegeta tion Index) as the quantity of interest throughout this research. NDVI repr esents a continuous variable related to productivity of land cover or vegetation biomass, which varies both in space and time. The Normalized Difference Vegetation Index (NDVI) is based upon the observati on that healthy leaves reflect near-infrared light while absorbing red light that provides the bulk of energy for photosynthesis. The NDVI provides an assessment of vegetative canopy cover characteristics, and is strongly correlated with the fraction of photosynthe tically active radia tion intercepted by the canopy (Jensen 2005). NDVI (Equation 2-1) is calculated where NIRBand4 is the digital number (DN) that describes reflectance in near-infrared wave lengths, and REDBand3, reflectance in red wavelengths for Landsat TM/ETM+ imagery. ) 3 ( ) 4 ( ) 3 ( ) 4 ( Band Band Band BandRED NIR RED NIR NDVI (2-1) NDVI values were calculated for every image for the entire spatial extent of each Landsat scene (Figure 2-2). The use of a continuous variable such as NDVI contrasts with more traditional classification schemes, where each pixel (or spatial unit) is repr esented by a single categorical value (Lambin 1999). In general, th e incorporation of bot h spatial and spectral information into

PAGE 30

30 land-cover change analyses greatly improves the am ount of information that is obtained (Figure 2-2) (Southworth et al. 2004a). For example, La mbin and Strahler (1994) found that changes in the spatial extent are more likely to reveal longer-lasting and longer-term land-cover changes, while spectral differences are more sensitive to shorter-term fluctu ations e.g., inter-annual variability in climatic conditions. The image dates represent the winter phenologi cal season (January and February) within six weeks of each other to keep the biomass amo unt consistent enough not to yield extreme interannual differences in NDVI. Since NDVI values were needed for the entire study area, we worked only with cloud-free images that fell w ithin a six-week window during the winter season. The image dates selected avoid extreme climatol ogic years of drought or excess precipitation as indicated by the El Nio Southe rn Oscillation (ENSO) index according to the Center of OceanAtmospheric Prediction Studies (COAPS) (http://www.coaps.fsu.edu/ research/jma_index1.shtml ). Avoidance of extreme climate years prevents climatic variability influences in NDVI values for vegetation between each image date. Local Variance Analysis We used the modified local variance or texture variance approach (Coops and Catling 1997; Coops and Culvenor 2000). However, instead of degrading the origin al resolution of the images through averaging cell values, we used the neighboring pixel values at a given window size to calculate the va riance about a pixel value (Figures 2-2 and 2-3). Using the moving window maintains the sp atial integrity of th e landscape without changing the NDVI values, while providing a measure of neighborhood varian ce. This approach may be superior to resolution coarsening (Woodcock a nd Strahler 1987) which alters the original NDVI values and degrades variance measures.

PAGE 31

31 Data Analysis All image processing was performed in ERDAS Imagine 8.5. Image preprocessing required geometric correction to the UTM Zone 17 NAD1983 coordinate system. NDVI values were calculated for all image dates used. Im age subtraction was performed between the 2001 image date and the remaining five image dates (F igure 2-2). The spatial variance was calculated for the 2001 image date using the six window sizes The temporal variance was calculated for the image subtraction products (Figure 2-3). The spat ial and temporal variance image products were combined into one image for each image subtraction pair, in which a sub-sample of every twentieth column and row were output to an ASC II text file for regression analysis in SPSS version 10 (Figures 2-4 and 2-5). Image to GPS control point geometric co rrection was performed for a September 1997 base image, which, although not used in this an alysis, served as the base image for geometric registration for all other image dates. Thirty to forty control points, collected using a Magellan handheld GPS with an accuracy of 7-10 meters, we re used for the geometric correction of the September 1997 base image. A second order polynomial algorithm was used with nearest neighbor resampling to a cell size of 30 meters and an RSME of +/0.5 pixel or 15 meters. Image to image correction was performed for the remaining image dates. No radiometric correction/calibration was performed before NDVI calculation because of NDVI is a normalized ratio. The January 4, 2001 image date served as the base image standard for the calculation of spatial variance, as it was the latest image date with complete cloud-free coverage for the entire scene (since the research began we obtained a 2003 image which was then added to the analysis). To analyze spatial variance, a moving window wa s passed over the entire NDVI image, pixel by pixel. The variance that was calculated within the window was assigned to the focal (central)

PAGE 32

32 pixel (Figure 2-2b, Figure 2-3). In different runs, we varied the size of the moving window to include 3 x 3, 10 x 10, 25 x 25, 50 x 50, 100 x 100, or 250 x 250 pixels. These window extents represented observations of landscap e patterns across a range of s cales. The spatial analysis of variance using a moving window resulted in 30 im age products (spatial variance for 5 dates x 6 window extents). Each of the five remaining NDVI image dates in the time series was subtracted from the 2001 NDVI image to produce an image subtraction pr oduct (Figure 2-3). Five subtraction images were generated. These represented the changes in landscape patterns over time. The NDVI image subtraction products served as i nput for the calculation of temporal variance (Figure 2-2c). As for the calculation of spatial variance, a moving window was applied to the five NDVI subtraction image products and the temporal variance within the window was assigned to the focal (central) pixel. Pixels t hus contained a measure of the ch ange in spatial variance through time. The calculation of temporal variance for each of the five image date pairs at each of the six window sizes yielded 30 temporal variance images that were paired with the previously generated spatial variance images. For input into regression anal ysis, the spatial and temporal variance data values were exported to ASCII text format. We reduced the si ze of the data set because of the large number (greater than 80 million) of pixel values that re sulted from stacking each pair of temporal and spatial variance images. NDVI values at every twentieth row and column were exported from ERDAS into ASCII text format for re gression analysis in SPSS (Figure 2-4). This approach allowed for a more manageable regression analysis for each image date pair in the statistics package SPSS 10. Linear regressi on without any data transformation was run on the image products for each of the window sizes at each of five time intervals using the spatial

PAGE 33

33 variance values as the dependent variable and the temporal vari ance values as the independent variable (Figure 2-4). Regression parameters were compared across image dates with each respective parameter plotted agai nst the log (base 10) of window area. In other words, we examined how the relationship between spatial and temporal variance varies across multiple scales of analysis in the same landscape. Transf orming areas using logarithms allowed for easier visualization of the regression results (the area within the window varies from 8100 m2 for the 3 x 3 pixel window to 56.25 x 106 m2 for the 250 x 250 pixel window). The moving window approach to scaling resu lts in a varying sample size for the calculation of variance. In order to check that our results were not merely an artifact of the sampling procedure, we created a null varian ce model (Figure 2-2d, Figure 2-5). NDVI values from the 2001 and 1997 image dates were used to create a neutral landsca pe in which the NDVI values from each of the original images were randomly rearranged about the landscape. To generate the neutral landscape, we c onverted the 2001 and 1997 NDVI images from ERDAS native image format to ASCII text form at. The ASCII values were read into a VB randomizer program, which calculated a frequenc y histogram with 2000 bins, to create output NDVI values with a numerical precision (thoug h not necessarily accura cy) of three decimal places. Absolute and relative frequency histograms were calculated from the sampled frequency histogram. For each original input NDVI value, random numbers were generated and were compared to the values of the relative frequency histogram to determine the spatial location of the output NDVI values. If the random number was less than the relative fr equency at a certain bin number then the random number was placed in that bin, becomi ng the new output NDVI value. The new output NDVI values were writte n to a new ASCII text file, which was then reconstituted into the ERDAS native image fo rmat for the variance calculations. This

PAGE 34

34 randomization was performed for both the 2001 a nd 1997 image dates. The spatial and temporal variances at each moving window size were calc ulated for the 2001 and 1997 random images as described above. A linear regression was conducte d following the same format as described above for the original NDVI spatial and tempor al variance images (Fi gure 2-5). The hypothesis tested by this analysis was that if the results from the analysis of actual landscapes were an artifact of our sampling procedure, they shoul d follow the same form as the results from the neutral model. Results and Discussion The results showed interesting patterns in the relationship between spatial and temporal variation. The strength of the re lationship between spatial varian ce and temporal variance, as measured using the adjusted R2 value, was highest at a windo w size of 10 x 10 pixels or 300 x 300 m2 (Figure 2-6). Beyond this window size, the relationship between spatial and temporal variance is less similar than expected by chance; i.e., spatial variance explains less temporal variance than is expected by chance. There were al so two clusters of resp onses in the temporal dimension. The longer time scales (9-16 yr windows) were gene rally more strongly related to spatial variation than were the shorter time scales, except for the smallest window size. The R2 values for the null model are high becau se the randomization (in space) of NDVI results in more equal variances from window to wi ndow when the window size is larger than 3 x 3. The expectation for randomly distributed (in space) data is that the varian ce of a sub-sample of the entire image (e.g. 10 x 10 cells or larger) will equal the variance of the entire image. Thus the spatial and temporal variances are necessarily mu ch more strongly related to one another than those of the real la ndscape (Figure 2-6). The results for the slope and intercept relations hips (Figures 2-7 and 2-8) revealed that the magnitude of the relationship between spatia l variance and temporal variance was highest at

PAGE 35

35 a window size of 50 x 50 pixels (1500 x 1500 m2) for slope (Figure 27), and 100 x 100 pixels (3000 x 3000 m2) for the intercept (Figure 2-8). The ratio of temporal variance to spatial vari ance (indicated by the slope) increased as we increased the spatial scale of analysis, meaning that at smaller spatial scales, the level of spatial variation was greater than that of temporal variation in the landscape. This result is consistent with Lambin and Strahlers ( 1994) observation that longer-term land-cover changes tend to occur over larger spatial extents. Conve rsely, at increasing spatial scales we found increased temporal variance. Hence, at increasing scales, the diffe rence between actual and random data increased with an increasing decoupling of spatial and temporal varia tion (coupling peaks at a 10 x 10 pixel window size, Figure 2-6). So, spatial variation explained less and less temporal variation (compared to the null model) as the sp atial scale of analysis increased. A non-zero intercept of the regression line rela ting spatial to temporal variation indicates that even when the land cover does not change spa tially, there is an inhere nt amount of temporal variation, perhaps the conseque nce of climate variation (Fig ure 2-8). This suggestion is supported further because slope increases with increasing spatial window (Figure 2-7), up to a plateau at 100 x 100 and above, but is lower with higher temporal windows. The inherent temporal variation is larger in larger areas (possibly because larger areas include a higher diversity of vegetation with more variable responses to climate fl uctuation), but is damped when integrated over longer time spans. The results show that an interesting relati onship may exist between spatial and temporal variation in this landscap e. Although the mechanisms behind the patterns that we have described are unclear, and will take further research to fully understand, the finding that spatial and temporal variance are correlated at certain scales but less so at others is intriguing. We can

PAGE 36

36 imagine instances in which spatial and temporal variation could interact with one another to drive changes in landscapes. For example, freque nt fires are characteristic of many landscapes, including our study area. Differences in the ra te of accumulation of flammable plant matter through space will affect the frequency of fire s, and hence temporal variation in primary production. Areas in which fuel accumulates ev enly through space will experience hot, homogeneous and less frequent fire s that return succession to an early stage throughout the area. By contrast, areas in which fuel accumulates unevenly may experience cooler, heterogeneous and more frequent fires, result ing in a mosaic of patches of vegetation at differing successional stages. Another process is the pl anting, growth, and harvest of larg e tracts of tree plantations that dominate the study landscape. The spatial scale th at explains most of the temporal variance (105 to 106 m2) includes the range of management unit sizes used by forest-products companies, and the most influential temporal scal e (9-16 years) is the closest to the 20-25 year plantation harvest cycle. This kind of mechanism implies that there is a strong possibility that positive and negative feedbacks can occur betw een spatial and temporal vari ation. Disturbances that are facilitated by heterogeneity may create further he terogeneity, leading to further disturbance, and so on. The end consequence of such a feedback w ould be extreme anisotropy in landscapes, possibly originating from quite small initial differences. Another possibility would be dampening, in which spatial variat ion in landscapes is reduced by the slow rate of change in more heterogeneous areas. Dampening would lead to greater homogeneity in landscapes than the dominant processes themselves would imply. We would expect such dynamics to be strongly non-linear and highly scale-dependent. Unfortunatel y, demonstrating causal interactions between spatial and temporal variation will be difficult; the standard, correlative approaches that are

PAGE 37

37 typically used to analyze remote ly sensed data are generally inadequate for causal analysis. Rigorous tests of causation will require simu lation modeling and ideally, experimental manipulations (if possible). Conclusions and Recommendations An important contribution that our approach offers is a framework in which landscape variation through space and time can be considered together, rather than individually. This is useful for problems where we know that both spa tial and temporal variation are likely to be important but there has been no quantitative way of considering the two t ogether. In addition, our approach has the potential to help link pattern to process and to highli ght the spatiotemporal grain(s) at which dominant processes occur in a given landscape. Previous researchers have addressed spatial variance in some detail, and most studies in the field of remote sensing concern changes over space or time, but fe w researchers have managed to link the two in a quantitative framework. One preliminary attempt to do so inco rporated trajectories of land cover change (based on multiple dates of satellite imagery land cover classifications) into a landscape fragmentation analysis (Nagendr a et al. 2003; Southworth et al. 2002). Similarly, Mertens and Lambin (2000) attempted to model land cove r change using landscape trajectories, again incorporating time into the analysis. Other resear chers in remote sensing are also starting to incorporate such methods into th eir research (Munroe et al. 2 004). However, these studies use land cover classifications, a discre te data source, and so lose mu ch of the inherent variability within the dataset prior to analysis. While some researchers are trying to overcome this limitation by working more with continuous data (Southworth et al. 2004a) such res earch is still in its infancy. Research assessing spatial variance is much more common in the field of remote sensing than that assessing temporal variance. The scaling relationships of spatial patterns have received

PAGE 38

38 some attention (Wu and David 2002), primarily th rough studies looking at how spatial grain and extent impact landscape metrics. Wu et al. ( 2000) found that in order to adequately quantify spatial heterogeneity, response cu rves (i.e., of metrics to changing scale) were necessary, as opposed to the more widely used single-scale meas ures. ONeill et al. (1996) in a study for the southeastern U.S., developed a method using pa ttern state space and a distance metric that measured the overall differences between landsca pes or changes through time. Lambin (1997), in a study of landscape disturbance in Africa, found that spatial hetero geneity was a key variable of interest; landscapes with either very high or very low levels of disturba nce were characterized by low spatial heterogeneity, while areas of mode rate disturbance were very heterogeneous. However, as our results would suggest, this relations hip varied with the scale of the analysis. In a study using local spatial variance measures of simulated high resolution imagery to predict the spatial patterns of forest stands, Coops and Culvener (2000) found a technique using local variance and spatial pattern statis tics to be successful. Their results suggested that for their study the relationship between spatial patterns of objects a nd variance is strongest at window sizes between 20-30 m, which matches the reso lution of SPOT and Landsat TM data. Ecology has a longer history than remote sens ing of addressing both spatial and temporal patterns in plant and animal populations. In part icular, attempts to explain spatiotemporal variation in community structur e have yielded a range of stat istical approaches that could conceivably be applied to remotely sensed data. For example, a common approach in the analysis of community patterns is to partition variance in community composition between environmental (spatial) factors and temporal factors (W illiams 1982; Rundle and Jackson 1996; Ault and Johnson 1998; Legendre and Legendre 1998; Herbert and Gelwick 2003). Once the computational difficulties of working with very large matrices have been overcome, approaches

PAGE 39

39 such as Redundancy Analysis (RDA) and Pr incipal Coordinates of Neighborhood Matrices (PCNM) (Borcard and Legendre 2002; Legendr e and Gallagher 2001) have considerable potential for application to remotely sensed data. Another new but promising approach in ecology uses Markov Chain models to assess sp atio-temporal relationships (Hill et al. 2002), extending some of the trajectory analyses described above. The primary challenges in applying methods from community and landscape ecology to land cover data sets are computational rather than conceptual, and hence should be relativel y easy to overcome as computing technology continues to improve. Wu et al. (2000) found that scaling analysis of variance can be a powerful method for describing and detecting the dominant patterns with in a given landscape. They proposed that the scientific community should address three rela ted groups of questions: (1) how does changing the scale of observation or analys is impact the results, and does it do so in a predictable way; (2) are the systems hierarchically or multi-scaled structures, and how do we relate these to the resultant pattern and process in a landscape; a nd (3) what scaling laws exist in heterogeneous landscapes? Wu et al. (2000) conclu ded that while these research questions are critical to the development of landscape ecology, they are also among the most challenging to address. The approach that we have outlined offers one way in which we can start to use remotely sensed data to identify key scales in landscap es and to look for interactions between spatial and temporal patterns at different scales. While our study was lim ited in scale, with no sub-pixel or finer scale spatial variation being included below a 3x3 pixel area (Butler et al. 2004; Malanson et al. 2002) remote sensing data can be used to track variat ion at both finer and coarser scales. Many remote sensing studies lack transferability across space and time of established relationships; it remains to be seen whether our approach is more gene ralizable to other landscap es and across remote

PAGE 40

40 sensing platforms. Although the full integration of spatial and temporal data across multiple scales will ultimately require a more mechanistic approach, the results that we have presented offer a novel and interesting way in which to star t investigating some of the questions that are most critical for the further theoretical deve lopment and unification of remote sensing, GIS approaches, and landscape ecology. Finally, this study is focused on the linkage s between landscape ecology and GIScience and remote sensing applications, and used data from remote sensing coupled with development of a multiscale spatiotemporal variance analysis approach to address a fundamental question in landscape ecology. All three "disciplines" were necessary to conduct th e study, but one could argue that the substa ntive questions came fr om landscape ecology while remote sensing and GIScience provided the materials and methods. The methods developed and employed here, addressing form and function within a landscape across multiple spatial and temporal scales, ultimately heeds the call of the GIScience research agenda in terms of better representing landscape dynamics in association with a closer c oupling of analysis and the conceptualization of process (Goodchild 2004). This con uence of landscape ecology a nd GIScience is inevitable many multiscale techniques, as employed here, are commonplace for analyzing pattern and process issues relating to scale. Where else wi ll we obtain sufficient in formation about landscape dynamics, and where else will we develop methods to answer the questions?

PAGE 41

41 Figure 2-1. Study area of the southeastern coas tal plain, with focus region and location of Landsat TM image footprint.

PAGE 42

42 Figure 2-2. Landsat TM Image NDVI products created for each image date, each filter size, and each pair of dates, here using the exam ple for (a) January 1997 NDVI Image, (b) January 2001 spatial variance of NDVI Imag e with a 10 x 10 filter, (c) January 1997 temporal variance NDVI, and (d) January 1997 Null Model NDVI.

PAGE 43

43 Figure 2-3. Illustration of the spatia l and temporal variance calculations.

PAGE 44

44 Figure 2-4. Regression mode ls used in analysis.

PAGE 45

45 Figure 2-5. Flowchart of analysis methods and null model creation.

PAGE 46

46 Figure 2-6. Adjusted R2 relationships between spatial vers us temporal variance, across the different time-steps and window size and area.

PAGE 47

47 Figure 2-7. Slope relationships between spatial ve rsus temporal variance, across the different time-steps and window size and area.

PAGE 48

48 Figure 2-8. Intercept relationships between spatial versus temporal variance, across the different time-steps and window size and area.

PAGE 49

49 CHAPTER 3 RATES AND PATTERNS OF LAND CO VER CHANGE AND FRAGMENTATION IN PANDO, NORTHERN BOLIVIA FROM 1986 TO 2005 Introduction Mertens and Lambin (2000) c ontend that land use change is necessarily a complex and often multidirectional process of biophysical an d socioeconomic interactions. While land-use change is a fundamentally local process, it is nested in a structure of hierarchical decisionmaking (Moran et al. 1998). Observed land cove r patterns are the net result of individual decision-making processes regarding the relative returns to land use (Cur rie 1981) set within the regional or national context. Hence, land cover, along with pattern analysis and social science measures, can be used to indicate the changing patterns in land use (i.e ., to link land cover to land use). Patterns of land cover change in mo st tropical countries relate significantly to anthropogenic impacts and are extremely complex, with changes occurring across multiple spatial and temporal scales (Woods a nd Skole 1998; Duncan et al. 1999). Land cover change is regarded as the single most important variable of global change affecting ecological systems (Vitousek 1994) with an impact on the environment that is at least as large as that associated with climate cha nge (Skole et al. 1994; Ch en 2002). Several factors need to be addressed while monitoring land cove r change: what kinds of alterations are taking place? Where do they occur? What are the rates of change? What are the patterns of change? What other factors influence each of the above? Much attention has been paid to the issues of human-induced land cover change within the la st few decades as evidenced by such multidisciplinary, multi-national programs as NASAs Land Cover Land Use Change (LCLUC) programs (http://lcluc.gsfc.nasa.gov/ ) or the International Geos phere Biosphere Programmes (http://www.igbp.kva.se/cg i-bin/php/frameset.php ).

PAGE 50

50 According to the Land-Use and Land-Cover Change Science/Research Plan of the International Geosphere-Biosphere Programme (IGBP), to explain land-cover change, particularly tropical deforestati on, we must consider the likely de terminants of land use. Such factors include: population size or density, technology, level of af fluence, political structures, economic factors such as systems of exchange or ownership, and attitudes and values (Turner et al. 1995). Thus, land-cover change is in great part the realization of ch anging patterns in land use. Most researchers break up the determinants of land use into two broad categories: proximate and ultimate driving forces. At a regional or la ndscape level, causal mechanisms include broader economic forces (price policies, wa ge trends), and intersectoral li nkages, such as factor markets and trade (Moran et al. 1998; C oxhead et al. 2001). At this la ndscape scale, one can answer questions appropriate to this le vel of analysis, such as: Does infrastructure development cause deforestation? Do markets cause deforestation? A second level of analysis has been called the proximate causes of land-cover change (Turner et al. 1995). Other researchers have ca lled this level the enabling environment (Coxhead et al. 2001) or direct causes (Panay otou and Sungsuwan 1994). Land-use change is a fundamentally local process, but it is nested in a structure of hier archical decision-making (Moran et al. 1998). At the local level, one can consider such factors as changes in population density, technical innovation, or ch anges in agricultural producti on systems (e.g., switching from a subsistence crop to a cash crop). Incentives fo rmed at the landscape level, such as policy changes, are realized by human decision-making at the local level. For example, in Bolivia INRA and the Forest Law allow a specific amount of forest cleared per head of cattle (about 5 ha).

PAGE 51

51 Remote Sensing for Change Detection Remote sensing is an attractive source of land cover data as it provide s a representation of the Earths surface that is spatially continuous and highly consistent. Such data are available at a range of spatial and temporal scales. Remote se nsing data collected for the same season during a year over many years provides a temporal sequ ence of information as to the changes and evolution of land cover on a landscape or a regi on. Much information about land cover types and changes has been derived from satellite images, specifically the Landsat platform, in the past 30 years. Analyzing changes in spatial pattern in land cover over time may allow for the identification of biophysical pro cesses driving changes in land c over (Brown et al. 2000). These spatio-temporal patterns coupled with socio-econ omic data in a modeling framework may permit the identification of social actors and drivers of land cover change for a particular area. Thematic mapping from remotely sensed data is based typically on an image classification; and while a thematic map provides unquestioned simplif ication of reality, it is only one model or representation of the depicted theme (Woodcock and Gopal 2000). While most classification stra tegies have focused on the us e of spectral dimension, the spatial domain (as represented by th e spatial organization or pattern of the data) also contains important information that to da te has not been utilized well with in land cover and classification methodologies (Cihlar 2000). Land cover fragmentat ion analyses are used frequently to help interpret the impact of land cover changes with in a landscape, by calculating for each land cover class a range of metrics to describe fragmentati on and spatial distributi on, often from satellite based land cover classifications. Analyzing changes in spatial pattern over time will facilitate the identification of the social and biophysical proc esses driving these chan ges (Brown et al. 2000). A discrete classification is ge nerally used for this purpose, to partition a landscape into

PAGE 52

52 homogeneous units with distinct boundaries, so that interpreti ng numbers that refer to the geometry and arrangement of these discrete units is conceptually simple (Pearson 2002). It is an exciting time for the use of remo te sensing for monitoring global land cover change. A multitude of new satell ite platforms have been launched within the past decade and more recently, offering a variety of data product s with varying spatial, temporal, and spectral resolutions. A variety of derive d data products (e.g. variance a nd local spatial autocorrelation indices as measures of image texture) from satellite images are ava ilable along with improved methods for image classification accounting for th ese and additional ancillary data products. The use of images from multiple satellite platforms allows for improved identification of patterns and hotspots of land cover change. Previous Work in the Amazon The Amazon basin has seen accele rated rates of deforestation since the 1970s (Laurance et al. 2004). Business-as-usual estimates predict that forest cover will be reduced to 53% of its original area by 2050 with large agricultural and timber industrie s as the primary beneficiaries (Soares et al. 2006). Current development of the Amazon by deforestation pr edicts increased fire frequency, and modified regional climate (Laurance et al. 2004; Nepstad et al. 2001; Soares et al. 2006) The Amazon forest provides ecosystem servic es for carbon sequestration in biomass and soils from the global atmosphere, regulation of th e water balance and flow of the entire Amazon river system, the modulation of regional climat e and air patterns over much of South America and the potential prevention of the spread of vectorand water borne diseases (Foley et al. 2007). Enhanced drying of the forest floor, increased fire frequency and lo wered productivity are examples of collateral damage incurred by su rrounding forests through canopy damage due to deforestation and selective logging (Foley et al. 2007; Nepstad et al. 2001; Alencar et al. 2004).

PAGE 53

53 Much previous work on deforestation and land cover change has been highlighted by research occurring in the Brazilian Amazon. Su ch studies have been the proxy for global deforestation rates and have garnered much atten tion to the processes of deforestation (Roberts et al. 2003; Nepstad et al. 2001; Laurance et al. 2004; Moran 1993; Lu et al. 2007). Given the longevity of NASAs Landsat program coupled w ith newer satellite platforms (e.g. MODIS and ASTER) remote sensing affords a reasonably ec onomic and attractive source of data for the monitoring and assessment of deforestation in th e Amazon basin. The latest, most reliable annual rate of deforestation in the legal Br azilian Amazon was calculated at 27,379 km2 yr-1 for 2004 (INPE 2008) with unofficial annual rates of 18759 km2 yr-1, 14039 km2 yr-1, and 11224 km2 yr-1, respectively, for 2005, 2006, and 2007. Many other studies have more focused local investigations of deforestation and forest frag mentation in various parts of the Brazilian Amazon where deforestation hotspots occur. (Roberts et al. 2003) summarize early research results from the Large Scale Biosphere Atmosphere experiment in Amazonia of projects using remote sensing for monitoring of the regional ecosystem functi oning of the area. Land cover mapping along with forest degradation were two of many applications of remote sensing for monitoring the land cover/land use change occurring in the Brazi lian Amazon. Although much attention of the scientific community on defore station is directed to the Brazilian Amazon, some studies document land cover changes in the countries that comprise the Amazon Basin (Oliveira et al. 2007; Armenteras et al. 2006; Etter et al. 2006; NaughtonTreves 2004; Imbernon and Branthomme 2001). Tropical deforestation can be couched in terms of Von Thne n bid-rent theory where land is allocated to the highest value (Walker 2004; Kaimowitz et al. 1998). In the case of deforestation in frontier zones of tropical countries, ease of acce ssibility factors into the land

PAGE 54

54 rent; therefore forested land nearest to populat ion centers and roads is most likely to be deforested sooner than remote areas. In additi on to industrial agricultu re, cattle ranching and colonization projects (Steininge r et al. 2001), roads and construc tion of new roads have been reported as proximate cause for forest clearing and widespread deforest ation, especially in frontier areas of tropical countries where gove rnments encourage new colonization projects (Arima et al. 2005). New roads offer market acce ss for timber and agricultural products from previously remote areas and roads lower transp ortation costs for internal migration, land access and land clearing for subsistence farming (Cho mitz and Gray 1996). Road construction into forested areas increases incentives to log these ar eas and allows access to forest resources, which increases the probability for deforestation nelson and (Nels on and Hellerstein 1997). Forest conversion along road corridors resu lts in habitat fragmentation and exposes the forest to various forms of degradation (Chomitz and Gray 1996). Lo cations nearer to existing deforestation tend to be deforested (Mertens and Lambin 2000) and forest fragments are more accessible thus susceptible to new deforestation than larger, c ontiguous areas of forest (Kaimowitz et al. 1998; Mertens and Lambin 2000). These impacts of road s on forest integrity and health are important to consider here given the multitude of infrastr ucture projects underway in Brazil (Fearnside 2002) and Peru in the southwestern Amazon (P erz et al. 2008), whic h will directly impact northern Bolivia. The southwestern Amazon is slowly receiving increased attention from the scientific community (Oliveira et al. 2007; Asner et al 2005; Naughton-Treves 2004). Deforestation around and extending from Puerto Ma ldonado, the capital of Madre de Dios, Peru accounted for 23% of total forest damage (deforestation and fo rest disturbance) incurred in the Peruvian Amazon between 1999 and 2005 (Oliveira et al. 2007) Asner et al. (2005) found an average of

PAGE 55

55 76 km2 yr-1 of selective logging and 564 km2 yr-1 of deforestation in Acre, Brazil between 1999 and 2002. Contrastingly, defore station in the Amazonian lo wlands of Tambopata, Peru proceeded at the net annual ra te of 0.1% (Naughton-Treves 2004). These examples scratch the surface; however, more attention is needed on la nd cover change in this region given the vast area, large number of people, and settlement and road improvement projects encouraged by respective national governments. As the northern-most department in Bolivia, Pa ndo is coming into the scientific spotlight with the recent paving of the Interoceanic highway in nearby Brazil (Hamilton 2006). Pando is the only Bolivian department that is entirely with in the Amazon and holds the most intact forest. There has been limited research in quantitatively inventorying the amounts of forest loss in Pando both at the department and municipal levels. Few departme nt wide studies have been conducted with results accessible outside Bo livia; the most notable is the Zonificacion Agroecolgica y Socioeconmica y Perfil Ambi ental del Departamento de Pando (Agroecological and Socioeconomic Zoning and Environm ental Profile of the Department of Pando) conducted in 1997 (ZONISIG 1997). This current research comprises a multi-date, and multiscene high spatial resolution regi onal investigation of deforesta tion in Pando, Bolivia between 1986 and 2005 using Landsat TM and ETM+ images acq uired every five years. The objective of this study is to determine the extent, rates and patterns of land cover change and associated forest/non-forest trajectory classe s for Pando, Bolivia from 1986 to 2005. Study Area Pando is the northern most department of Bolivia, with an area of 63,872 km2 that represents 5.8% of the total area of Bolivia. Pando cont ains five provinces comprised of fifteen municipalities with the capital, C obija, located on the bor der with Acre, Brazil. With an elevation range between 90 to 289 meters Pando is considered tropica l lowlands. Approximately 98%

PAGE 56

56 forest cover that remains is comp rised of evergreen forests broadly classified into two types of upland forest and floodplain forests with minor amounts of open sava nna with tree islands (Mostacedo et al. 2006). The clim ate of Pando is classified as warm-humid tropical and tropical wet and dry climate (Kppen Aw) with a pronounced dry period from May to September and a rainy season from November to March, during wh ich the north-south migration of Intertropical Convergence Zone produces heavy rains due to low air pressure and extremely unstable atmospheric conditions. From 1944 to 1990 average monthly temperatures were reported from 23.6 to 26.4C with an annual average of 25.4C. For the same time period the annual average precipitation 1,834 mm (ZONISIG 1997). From the 1992 census the tota l population in Cobija was about 11,000 of the 38,000 persons residing in Pando. The most recent census placed the 2001 populations at slightly more than 22,000 and 52,500 for Cobija and Pando, respectivel y, with the 2005 estimated populations at almost 32,000 and 67,000 (Instituto Nacional de Es tadsta de Bolivia, www.ine.gov.bo). Rural economic activities consist of Brazil nut harves ting, timber extraction, and cattle ranching while some gold mining occurs within river alluvi al bars (ZONISIG 1997). Petroleum has been discovered in southern Pando, but is not extracted. Commercial and re tail activities, especially in Cobija, along with communications and road co nstruction are other lead ing economic activities (ZONISIG 1997). Two major road corridors dissect Pando with the central we st to east CobijaPorvenir-El Sena-Puerto Copacabana-Riberalta (i n the neighboring department of Beni) road, and the north-south Porvenir-San Silvestre-Chive road along the western edge of Pando (Figure 3-1). The main access into the in terior of Pando is by paved road from Cobija to Porvenir, and then by dirt road from Porvenir through Puerto Rico to El Sena and on to Riberalta in the neighboring department of Beni (Figure 3-1).

PAGE 57

57 In contrast to the widespread deforestati on and larger patches in Acre, Brazil, the deforestation landscape in Pando is relatively recent (i.e., in past few decades). Spatially extensive deforestation, extending up to 2 kilome ters, occurs on privately owned lands adjacent to the roads, close to the ma in population center of Cobija. Af ter about 20 kilometers along the roads to Sena and to Extrema, these large non-forest areas become less pronounced on the landscape and deforestation becomes patchy an d more linear along the road less than 0.5 kilometers into the forest. In Pando under the Ley Forestal or National Forestry Law of Bolivia (Repblica de Bolivia 1996), landholders are permitted to clear one he ctare of land per year for chacos.) Clearing occurs through traditional slash-and-burn methods The cleared land is in chaco for one year after which it is generally maintained for agricultur e for another year or so before left fallow. In a few cases, fallows are converted to pasture. The maintenance of pa sture is cyclical with active burning every year near the end of the dry season in September (later in the year than most images obtained for this study). Annual burning is used to remove stumps and felled trees over time. Methods Image Selection and Preprocessing Landsat 4 and 5 TM and 7 ETM+ images (Appendix A) were acquired for 1986, 1991, 1996, 2000 and 2005 during the dry season (May to Oc tober) totaling 40 images for the area. This seasonal time period permits image acqui sition with relatively minimal cloud cover and haze from smoke from the burning of pastures and recently cleared forests. Due to some clouds a few images were acquired outside of this time peri od. Best efforts were made to minimize intraand inter-annual precipitation differences betw een images although inco mplete and often nonoverlapping annual and monthly record s made this process challenging.

PAGE 58

58 The images were corrected for atmospheric a nd seasonal differences following (Green et al. 1999; Green et al. 2001) and were georeferenced to less than 0.5 pixels or 15 meters using the University of Maryland Global Land Cover Faci lity Geocover 2000 images, which make up part of the analysis. The images were mosaicked to generate a seamless coverage of the area for each date. Clouds, shadows and water were remove d from each individual image and from the mosaics using a PCA image differencing and th resholding method (Varl yguin et al. 2001). The masks were combined together from individual dates and applied to the mosaics in preparation for classification. Next, secondary image products were derived to assist with the image classification. The tasseled cap indices (K auth et al. 1976), a mid-infrared index (Boyd and Petitcolin 2004) and a 3 by 3 moving window calculation of the variance of each pixel for bands 4, 5 and 7 (as a measure of image texture) were generated for each mo saic. The Kauth-Thomas transformation bands, particularly the greenness band along with the mi d-infrared bands, are useful for discriminating forest structure from other types of vegetation and the brightness band is useful in identifying non-forest areas. Texture is useful for classifi cation of forest versus non-forests (Boyd and Danson 2005) and useful in aiding classification of forests where selective logging has occurred (Chan et al. 2003) resulting in small clea red patches in a complex forest matrix. Image Classification Image classification using decision trees handle non-parametric data at different spatial, spectral and temporal resolutions, generate easily interp retable and explicit classification rules, and produce an importance score for each variab le (Breiman 1984). Given the vast geographic extent of the study region and its complex ecosy stems and eco-regions, a rule-based or decision tree classification (Breiman 1984) was chosen over traditional unsuperv ised and supervised techniques. The visible and therma l bands contained striping, whic h proved to be difficult to

PAGE 59

59 remove, thus limiting the available information fo r a traditional classifi cation. The rule based classification approach provided flexibility to eliminate these bands, using only the nearand mid-infrared bands along with secondary derived products. From an initial field visit to Pando in May 2006 coupled with visual interpretation of the 2005 mosaic, known locations of forest, pasture a nd bare-built were identified and used as training samples for the creation of the decision rules. Three initial classes were created for forest, pasture, and bare/built. Pasture and bare/built were aggregated to create the non-forest class. The pixel values of these areas were extr acted from the mosaics of the infrared bands 4, 5, and 7; the Kauth-Thomas brightness, gree nness, and wetness images; and a 3x3 windowed calculation of image variance for each of the infrared bands. The data mining software Compumine (www.compumine.com ) was used to create the decision rules for each mosaic. Similar to logistic regression, Compumine predicts user specified classes based on the recursive po rtioning of continuous input data resulting in a decision tree based on variable importance in the model. The va riable importance is a relative measure of how much each of the variables used in the model c ontributes to reducing the prediction error of the model. A split sample validation was used for th e training sample points whereby 85% were used to train or create the de cision tree and 15% were used to test the tree. This process was repeated for each mosaic date. Performance assessment statistics of accura cy, precision, and recall are provided for the resulting decision rules and acco mpanying tree models. Accuracy is the number of correctly predicted test examples divided by the total number of test exam ples reported as a percentage. Precision (Equation 3-1), reported between 0 and 1, is the measure of accuracy that a specific

PAGE 60

60 class has been predicted where tp and fp are the number of true positive and false positive predictions for the considered class. fp tp tp precision (3-1) Recall (Equation 3-2) measures the ability to select instances of a certain class from a data set where tp and fn are the numbers of true positive and false nega tive predictions for the considered class. The total number of test examples of the considered class is tp + fn. fn tp tp recall (3-2) The resulting classified images were subset to the extent of Pando that coincided within the mosaic extent. That is, due to the selection of th e Landsat image path and rows for the creation of the mosaics and due to the major road corridors through Pando which omitted the northeastern municipality of Federico Roman, the spatial extent of Pando descri bed here is smaller than that of the department of Pando. Change Detection Since deforestation is categorical ch ange (Woodcock and Ozdogan 2004) change trajectories were chosen as the best means to an alyze temporal changes in forest cover and to calculate the rates of deforest ation (Southworth et al. 2004a; Pe tit et al. 2001). Image change trajectories are defined as se quences of successive changes in land cover types providing information changes between two or more time peri ods of an area or regi on. That is, the number of change categories (Equation 3-3) is defined as mt, the number of change trajectories, mc the number of land cover classes defined, and the superscript t as the number of images (Petit et al. 2001). t c tm m (3-3)

PAGE 61

61 Change trajectories were calculated between th e five dates of classified mosaics resulting in four date pairs for 1986-1991, 1991-1996, 19962000, and 2000-2005. Four trajectory classes of stable forest (F-F), deforestation (F-NF), re forestation (NF-F), and stable non-forest (NF-NF) are possible from this type of analysis. These si mple two-date pairs were selected for ease of analysis as calculating a trajectory image using a ll five dates for the classification would result in 32 change trajectory classes, whic h would be difficult to interpret. The calculation of trajectory images for pairs of images also allows for the calculation of rates of conversion for deforestation and reforestation between the two dates, thus ma king the results comparable to other studies of the region. Classification Accuracy Classification accuracy was assessed using Kappa coefficient and overall percent accuracy for each class and for the overall classification. Field observations about the land cover type at locations close to transportation corridors (i.e ., roads and rivers) ar ound the study area were collected in 2006 and 2007. The 2006 poi nts were collected around four communities in Acre, Brazil and seven communities in Pando. These points were aggregated to a FNF classification from a finer level of classification and 30 points were selected for each forest and non forest class. The data collected in 2007 were based on a stratified random sample of 300 points within 1.5 kilometers of the Cobija-Sena road corridor for each cover type derived from a preliminary FNF classification using the 2005 mosaic. Of these 300 points, at least ten percent were selected to be visited within the field based on accessibi lity. One hundred and fifty points were collected during the 2007 field season; all were used for accuracy assessment of the 2005 FNF classification. Since field data do not exist for the 2000 FNF classified mosaic, temporally and spatially corresponding ASTER images (Appendix B) were used as an independe nt data source for

PAGE 62

62 additional accuracy assessment. The classifi cation accuracy of 2000-2005 FNF trajectory image was assessed also using ASTER images from 2000 and 2005. Since ASTER has a higher spatial resolution than Landsat, making it easier to discern forest from non forest, it is appropriate for use as an accuracy assessment data set. The ASTER images were acquired from the USGS GLOVIS archive and not all of the study extent was covered by one year; therefore 2001 and 2006 images were included with the 2000 and 2005 ASTER images, respectively resulting in 31 and 35 images used for each date. A stratif ied random sampling design of 100,000 points was generated within the overlapping extent of the ASTER images and classification mosaics. Great care was taken to ensure the accuracy sampling poi nts were well within the FNF and trajectory patches by buffering each patch edge by 60 meters to compensate for edge effects between forest and non-forest patches and differences in reso lution between ASTER a nd Landsat. From these mass points 150 for each class (300 for the FNF cl assification and 600 point s for the four class 2000-2005 trajectory image) were ra ndomly selected as testing poi nts. The points were assigned a class of forest or non-forest based on visual comparison of the point s to their corresponding location on the ASTER images. In this way, the test points derived from the ASTER data along with sampling points from the 2006 and 2007 field seasons provide suffici ent information to test the 2000 and 2005 classifications along with the 2000-2005 trajectory im age. No historical data exist to test the remaining FNF classifications or the trajectory images. Fragmentation Analysis A fragmentation analysis was performed on the FNF trajectory images at the Pando extent and at road buffer distances of 1, 3, 5, 8, 10, and 15 km using APACK (Mladenoff and DeZonia 2000). The buffer distances correspond to influences of roads on the forest structure for near: with 1 and 3 km corresponding to along-the-road de velopment, intermediate distances (5 and 8

PAGE 63

63 km) and remote distances at which 10 and 15 km represent influences fr om secondary roads or no road influence. Given the duplication of information using multiple landscape metrics measuring essentially the same aspect of a la ndscape (Gustafson 1998; Riitters et al. 1995) basic measures of mean patch size (MPS) and corr ected perimeter-area ratio (PARC), along with fractal dimension (FRAC) (Loehl e and Li 1996), and an aggrega tion index (AI) (He et al. 2002) were used to quantify the spatial configurations of the forest/non-forest in the land cover change trajectory images. These metrics capture the esse ntial information about the landscape structure to assist in describing the spatia l patterns of land cover change in the study area at all spatial extents. A Kruskall-Wallis test (Appendix C) wa s applied to MPS and PARC metrics calculated for the four FNF trajectory image dates (1986-1991, 1991-1996, 1996-2000, and 2000-2005) to determine if any of the date pairs were st atistically significant from one another. Results Decision Tree Model Accuracies The overall accuracy for the classification rule s (Table 3-1) were quite high ranging from 98% (1996) to 99.8% (2005), and these were corrobor ated by the high valu es of total area under the curve. The number of rules used in each classification ranged from six for the 1986 classification to sixteen for the 1996 classificatio n. At the class level th e precision, comparison of the true and false positives, ranged from 0.923 (the lowest in the 1996 classification for barebuilt) to 1.000. Recall, comparing true positives against false negatives, also measured high ranging from 0.93 (also in 1996) to 1.0. Overall, the classification rules distinguished the classes well. The precision and recall for the 1996 mosa ic showed the most confusion between predicting true positives and fals e negatives for the pasture and bare-built classes. This is reflected in the overall accuracy being the lowest in the sets of classification rules and is also addressed by grouping these classes t ogether to represent non-forest.

PAGE 64

64 Classification and Trajectory Accuracies An overall accuracy of 87.85% and an overa ll Kappa of 0.6762 (Table 3-2) was reported for the 2005 FNF image based on the field samples. Accuracy assessment based on samples from the 2005-2006 ASTER images showed 97.96% overall accuracy and an overall Kappa of 0.9592. Although the producers accuracy for forest base d on the field samples was 100%, the users accuracy for forest was low compared to the AS TER assessment of the same classification. The overall accuracy of the 2000 FNF classificati on compared to the 2000-2001 ASTER images was 96.96% with an overall Kappa of 0.9592. Based on the ASTER comparison with the 2000 and 2005 FNF classifications at the cl ass level the producers and us ers accuracies and Kappa are comparable in magnitude. The overall accur acy and Kappa for the 2000-2005 FNF trajectory image are 82.01% and 0.7592, respectively. The NF-F transition class has the lowest users accuracy compared to the remaining three change classes, with the F-F class having the highest users accuracy. Stable non-forest has the lowest producers accuracy and the NF-F class has the highest producers accuracy, with the lowest Kappa at the class level. Forest and Non-Forest Extents and Rates The area of interest of Pando analyzed in th is study is 5,631,414 hect ares or 88.2% of the 6,384,526 hectares actually in department of Pando. The masked area makes up 3.8% or 206,346 ha of the study area. From 1986 to 2005 the extent of forest (Table 3-3) decreased very slightly from 95.96% (slightly less than 5.5 million ha ) in 1986 to 94.53% (about 5.3 million ha) in 2005. The amount of non-forest quintupled from about 21,000 ha in 1986 to a little more than 100,000 ha in 2005; a change of only 1.25% increase from less than 0.5% to almost 2% in 20 years. Although the amount of non-forest has increased between 1986 and 2005, Pando still has very little non-forest.

PAGE 65

65 From 1986 to 2005 stable forest (Table 34) decreased from 5,381,361.27 ha (99.16% of the total area) to 5,319,193.59 ha (97.91%) while stable non-fore st increased slightly from 14,684.94 ha (0.27%) to 50,515.74 ha (0.93%) in 2005. Th e amount of deforestation doubled from about 24,500 ha (0.45%) in the first decade (1986-1996), when it was constant, to over 51,000 ha (0.95%) in the second decade (1996-2005). De forestation rates were stable in the first decade at about 4,100 ha (0.075%) per year while they increased to almost 10,000 ha (about 0.2%) per year between 1996 and 2000 and dropped to about 8,600 ha (0.16%) per year between 2000 and 20005. Reforestation increased also, from just fewer than 7,000 ha (0.12%) in 1986 to over 22,000 ha (0.41%) in 2005 with 1,100 ha (0. 02%) per year between 1986 and 1991 and a little more than 3,000 ha (between 0.06% and 0.07%) per year between 1991 and 2005. Fragmentation Metrics: Pando and Buffer Extents Fragmentation metrics are first discussed at the extent of Pando (Figures 3-2 and 3-3). Comparison among classes for MPS and aggregati on values show a general decrease in value from stable forest to stable nonforest to patches of deforested and reforested areas except for patches of deforestation for the 1991-1996 date pair Stable forest has the highest values for fractal dimension, but th e differentiation among linearity of th e other three classes is more ambiguous. The aforementioned trends break down for the PARC values, although the PARC values in conjunction with the other two shape me trics can help differentiate which "non-forest" classes have more two dimensional shapes and those that have a more linear shape. Metrics for the road buffers are shown in Figures 3-4 through 3-7. These plots represent the four landscape metrics as a function of lateral di stance (i.e., buffer distance) away from the road and in relation to increasi ng distance from Cobija in 50 km sections. That is, as one goes further from Cobija along the C obija-Sena road, one would expect the various metrics to change as a function of distance, hence the sectioning of the road buffers. The Cobija-Sena road is 250

PAGE 66

66 km in length and thus choosing 50 km sections was convenient for this part of the metric analysis. Lateral distance is represented by the X axis, the distance from Cobija on the Y axis and the metric value on the Z value. The plots sh ould be read as the observer standing in the lowermost corner at a distance of one kilometer fr om the road and within the first 50 km section (0-50km) from Cobija, looking to the right along the Y axis. Each intersection on the data mesh represents a node for lateral dist ance and distance from Cobija. Fo r example, reading the plot of MPS for stable forest in Figure 3-4a, one would observe, in genera l, an increase in MPS with lateral distance from the road and with increasin g distance from Cobija. In section one, the first 50 km segment from Cobija, MPS values incr ease by an order of magnitude from a buffer distance of 1 km to 15 km from the road. Tracking MPS through time (Figure 3-4) requires one to view the data plot vertically where MPS in 1986 is larger gene rally, regardless of buffer dist ance and distance from Cobija, than MPS in 2005. In the 200-250 km section from Cobija, MPS values appear similar for all time periods and increase with distance from the road. The plots of AI and FRAC also exhibit similar trends as the MPS plot for stable forest. Stable forest PAC values appear higher at close and far lateral distance extremes to the road and Cobija. No pronounced temporal trend is obvious. The spatial and temporal trends for the ot her three classes are not as neat and cannot be as easily described. In general MPS and AI values (Figures 3-4 and 3-7) for deforested patches follows a declining trend with distance from Cobija and seem indifferent to buffer distance regardless of time; however, in the 150-200km segment from Cobija in 1991-1996 MPS values decline noticeably. FRAC values decrease in a curvilinear fashion with increase d distance from Cobija and increase with lateral distance from the road (Figure 3-6). PA C values seem insensitive to

PAGE 67

67 lateral distances from the road and at distance fro m Cobija. No distinct temporal trend is present for any of the metrics (Figure 3-5). MPS and AI (Figures 3-4 and 3-7) for refore sted patches increase with distance from Cobija except for a downturn in 150-200 km secti on while lateral distance seems negligible in influence. Reforested patches increase thei r MPS slightly through time from 1986 to 2005 except for the 1991-1996 time period that ex hibits larger MPS value rega rdless of spatial position. At close distances to the road AI values increase considerably in the 200-250 km section following a pronounced decrease in the 150-200 km section. FRAC values generally decrease with distance from Cobija and appear to increase slightly with increasing lateral distance from the road. FRAC values less than one indicate that not enough reforestation patches were available for the metric calculation. PAC values for refore station patches decrease from Cobija towards the 150-200 km section and then increase in the last section. La teral distance influences cause higher PAC values closest to the road great er than 150 km from Cobija and then decrease with increased distance from the road in these sections. Again no distinct temporal trend is obvious. For the stable non-forest patches there is a steady decline in MPS and AI values for all time periods for all 50 km sections from Cob ija except for the MPS values in 200-250 km section, which show an order of magnitude increas e. This section shows the lowest AI values, especially close to the road. The influence of lateral distance from the road appears negligible on MPS and AI values regardless of time and space. FRAC values decrease with distance from Cobija except at 3 and 5 km buffer distance from the road, where they remain invariant to lateral distance within 150 km of Cobija and show a clear increase through time from 1986 to 2005. PAC values increase slightly through time and decrease with distance from Cobija with a

PAGE 68

68 minimum in the 150-200 km section and then incr ease slightly in the last section. PAC values appear to decrease with lateral distance from the road at all distances from Cobija. Discussion Methodological Considerations There were many opportunities for innovation in the classification of multiple-scene Landsat mosaics for multiple time steps. This is the first time a mosaic at a large extent has created from Landsat images to cover the tri-na tional area of Brazil, Bolivia, and Peru. The novel combination of near and mid infrared bands, tassel ed cap indices, and vari ance texture in a rulebased classification scheme was used as the vi sible bands showed heavy striping, which were difficult to resolve. Lu et al. (2004) noted that inclusion of te xture data improved classification results of complex landscapes of rural settlem ent areas in Rhondonia. The selection of mid infrared band during data mining for the creation of the classification rule s shows that the near and mid-infrared wavelengths, particularly the latt er, are useful for the discrimination of forest (Boyd and Danson 2005). Although only three land cove r types were classifi ed, the decision tree classification approach has been endorsed to cla ssify complex landscape and environments in the Amazon (Lu et al. 2004). The overall classification accuracies were qui te high despite some challenges with the mosaicking procedure, as well as spatial, temporal, a nd spectral resolution differences between Landsat and ASTER. Additional challenges includ ed the potential spectral confusion between burned areas and flooded forests and non-forest areas. This was especially true in Landsat images taken in the middle of the dry season (Jul y and August) when slash and burn begins; thus what was once vegetated (potentia lly classified as forest) is now burn scars (classified as nonforest). Given the considerable spectral variabil ity of this landscape on a region scale, the rule

PAGE 69

69 based classification using the infrared bands and derived image products proved robust to discriminate between the desired sp ectral classes for classification. Extents and Rates of Land Cover Change in Pando Increased attention in last 10 to 15 years has been focused on reporting the trends and rates of deforestation in Bolivian lowland forests (Tab le 3-5). Within this time period these forests have been listed under the World Heritage sites of most diverse remaining Amazonian forests (Killeen et al. 2007; Steininger et al. 2001) and have received a ttention from the Kyoto protocol for carbon storage in attempts to reduce carbon em issions and combat global warming (Killeen et al. 2007). Additionally, the department of Santa Cr uz in the past two decades has experienced increased rates of deforestation due to corporat e agriculture for soybean production for export to Asia (Steininger et al. 2001; Ka imowitz et al. 1999) bringing Bolivia into the deforestation spotlight. In previous studies (Steininge r et al. 2001; Killeen et al. 2007 ), only temporal trends of deforestation rates in Bolivia ar e highlighted, without much atten tion to the spatia l patterns of land cover change, especially in Pando. Previous work on deforestation rates and extents in Bolivia was focused on the more populated areas in the lowlands of Santa Cruz, La Paz, and Cochabamba (Steininger et al 2001; Killeen et al. 2007). Alt hough deforestation rates for the northern departments were reporte d, little is known about the spatial extents and configurations of deforestation, particularly with respect to roads and population centers. There is the need of focal studies to better unde rstand the nature and extent s of deforestation in Pando. As shown in Figure 3-5 results for the tem poral rates and amounts of deforestation in Pando generally match those previously reported for Pando (Killeen et al. 2007; Wachholtz et al. 2006; Rojas et al. 2003; Steini nger et al. 2001; CUMAT 1992) although there are some discrepancies in deforestation rates for certain periods. Notabl e exceptions include the following.

PAGE 70

70 Deforestation rates from Killeen et al. (2007) for the periods 1987-1991 and 1992-2000, overand under-estimated, respectively, ours. The annual deforestation rate re ported by the Wachholtz et al. (2006) are about half of the annual rate for the 2000s. Both Rojas et al. (2003) and Steininger et al. (2001) report slightly higher (about double) rate percentages for 1993-2000, and 1984-2000. In looking at the overall landscape, the amount of forest clearing from 1986 to 2005 in Pando is minimal, from 96% to 94.5% (an area of 886.43 km2), and has only increased recently. The spatial locations and patterns, however, ar e very interesting. Anal ysis of deforestation extents and patterns in 50 km sections of the ro ad corridor from Cobija to Sena (Figure 3-8) showed larger clearings for stable non-forest (i.e., pasture) al ong the Cobija-Porvenir road, east and west of Porvenir as evidenced by larger m ean patch sizes in this section. As we go further from Porvenir, MPS for stable non-forest decreases as the lateral extent (i.e., buffer distance) decreases creating long, narrow patc hes along the road about one-qua rter to one-third kilometer in width. The more isolated patches of transitiona l deforestation and reforestation tend to be the dominant pattern on the landscap e in the 50 to 150 km sectio ns, although relatively speaking they do not comprise a great area. These sections contain a number of co lonist communities, in which they may have isolated patches of chacos, or land cleared for personal agriculture. In the last section (200 to 250 km) we see a large cluster of clearings persistent through time around the community of Conquista, which co ntain large clearings used for pasture thus inflating the MPS for this area, particularly clos e to the road (within one kilome ter). Interestingly, this area has experienced considerable regrowth observed in the more recent images from 2000 and 2005. Rapid reforestation detected with the trajectory images occurs on chacos after the first year of clearing before the burning season in September. Regrowth of vegetation occurs very rapidly,

PAGE 71

71 even in the dry season when images were collec ted, and can appear as re forestation due to high reflectance in the near and mid infrared bands. This vegetation regrowth, called barbecho, provides a challenge to the classi fication process as the amount of regrowth seen in five years, the time between image acquisition, is substant ial enough to alter the re flectance response to spectrally appear as forest but is structurally distinct from matu re forest. In field verification proves difficult as the structure of the barbecho is complex with rapid growth of emergent tree species and canopy species, but the density is not sufficient to be classified as secondary forest. Compared to other Amazonian lowland areas, the percentage of rema ining forest in Pando is high and there are low rates of deforestation. Pando has considerably much higher forest cover (about 95% in 2005, Table 3-3) co mpared to results reported by Imbernon and Branthomme (2001) for similar areas in western Amazon 61% and 74% remaining forest cover for Yurimaguas and Pulcallpa, Peru in 1996, and 50% and 65% forest cover remained in 1996 in Theobroma, Rhodonia and Pedro Peixoto, Acre, Brazi l. Along similar trends, Etter et al. (2006) calculated 30% loss of lowland forests by 1998 for ar eas in the Caribbean re gions of Colombia. More on par with the rates and trends of defo restation we see in Pando are other studies in the Amazon fringe outside of the Brazilia n Amazon. In the Colombian Amazon between 1985 and 2001 Armenteras et al. (2006) reported annual deforestati on rates were between 0.97 to 3.73% in high population areas and less than 0.3 1 to 0.01% in relatively unpopulated areas of indigenous communities. Elsewher e in the southwestern Amazon, deforestation rates of 2.4% yr1 (1986-91) and 1.4% yr-1 (1991-97) were measured within eight kilometers of the trans-Peruvian Interoceanic highway (Naughton-Treves 2004). These rates and trends are higher, except for the Armenteras et al. (2006) in unpopulated areas, than those fo r Pando during the similar time periods. The low deforestation rates and high perc entage remaining forest are due to relative

PAGE 72

72 remoteness of the area, low histor ic population density a nd growth, and restricted access into the interior of Pando. Forest Fragmentation Patterns Our results show that dominant patterns of land cover change in Pando follow road networks. The patterns of stable forest and nonforest metrics exhibit di stinct scaling patterns along the Cobija-Sena road in pe rpendicular fashion along with a distinct longitudinal gradient leaving Cobija traveling towards Sena (Figure 3-8) Not surprisingly forest patches increase with size away from the road in longitudinal and perp endicular directions. Furthest along the road from Cobija forest patches become long and linear in shape with the major ity of landscape being dominated by large forest patches constituting a ma trix. Non forest patches or cleared areas are highly aggregated along the road between Cobija and Porvenir, and west and east of Porvenir. These pasture areas comprise large parcels of land devoid of any intact forest cover. Similarly, large non-forest (i.e., agriculture or crops) occurred along roads w ith the core forest remaining intact at farther from roads (Imbernon and Branth omme 2001). This is para llel to the situation seen in Pando. As distance increases along the ro ad from Porvenir towards Puerto Rico, these consolidated patches diminish in number a nd size becoming long and narrow pastures of no more than one-half a kilometer in width. Furt her, perpendicularly from the road non-forest patches become more isolated in nature acce ssible only by paths and trails (Figure 3-8). The patterns of fragmentation resulting from infrastructure development in frontier areas are linked directly to scale and th e type of road buildi ng activates in the la ndscape (Arima et al. 2005). The fragmentation of Pandos forests along road networks and near urban areas follows similar patterns as those found in other parts of the Amazon. Oliveira Filho and Metzger (2006) reported in Mato Grosso, Brasil that independent settlements e xpanded in fishbone patterns and caused greater landscape fragmentation than large properties. Etter et al. (2006) identified large

PAGE 73

73 proportions of agricultural and grazing lands within ten kilomete rs of roads and agricultural clearings in colonist areas in the Amazon a nd Pacific regions of Co lombia. Cumming (2003) found the strongest effects of roads on mean patch sizes below 150m and forest patches decreased just 13 km from the average urban cen ter. By comparison, within 65 km of Cobija most road effects are observed within 3 km while at greater distances, road effects decrease to less than 1 km. Additionally, spatially intact fore st with large MPS is seen well under the 13 km threshold Cumming (2003) reports. In the rural areas of Pando deforestation is constrained along the road, with minimal fragmentation, and in isol ated patches in a predominately forest matrix. This is similar to Imbernon and Branthomme (2001) who found relativ ely homogenous patches of deforestation along towns in Yuramangus, Peru and along roads in Pucullpa, Peru with patchy clearings at di stance from both. Compared to the widespread deforestation and fragmentation associated with fishbone patterns in Brazil (Lambin et al. 2003) Pando may be in the early stages of the evolution of deforestation and fragmentation co nditions. In the rural areas of Pando outside of the Cobija and Porvenir corridor (Figure 3-8) deforestation occurs parallel to the road, but the classic fishbone pattern may not have evolved yet due to the lack of majo r secondary roads and population centers. However, the spine of defo restation along the road already exists near Puerto Rico and other communities. As advert ised by the departmental government of Pando, paving of the Porvenir-Sena road is planned, but the timing is unknown (Fi gure 3-8). Other road improvement projects are underway towards Extr ema and, possibly, Chive on the Madre de Dios River, and towards Bolpebra near the tri-nati onal border with Peru and Brazil (Figure 3-1). Bridge construction was completed in May 2008 to span the Rio Orthon in Puerto Rico, about 180 km from Cobija, facilitating ea sier access to Sena and to Ribe ralta. These projects will allow

PAGE 74

74 more rapid accessibility to the more remote pa rts of Pando. After the plan ned paving of the road from Porvenir to El Sena, deforestation is likely to increase given the increased access to more distant areas and communities towards El Sena Isolated population centers of Puerto Rico, Conquista, and El Sena (Figure 3-8) may gain population servi ng as nodes in th e transportation network. The improvement of existing secondary roads and construction of new secondary roads will increase ease of access to the interior of the large forest patches. This access potentially will allow for increased forest clearing. Conclusions This work highlights the pattern s of land cover change in Pando, northern Bolivia as a last frontier of relatively intact forest for the lowla nd tropics in Bolivia. Deforestation is clustered for all years around the Cobija-Porveni r corridor with large tracts of deforested areas for pasture here. As of 2005 94% of Pando's land cover was clas sified as forest. The rates of deforestation from 1986 to 2005 are less than one-tenth of one pe rcent per year, which is very small compared to Santa Cruz, one of the most de forested departments in Bolivia. From an applied standpoint of remote sensing, this work highlights th e extents and rates of land cover change in one of the important non-Brazilian Amazonian fringe areas. With much of the scientific attention focused on deforestati on in the Brazilian Amaz on, it is easy to overlook the surrounding countries like Bolivia that cont ain important diversity of flora and fauna. As seen in previous works in the Amazon (Etter et al. 2006; Arima et al. 2005; Imbernon and Branthomme 2001), deforestation increase s with increased accessi bility and improved infrastructure conditions. From these result s one could extrapolate that in Pando, after competition of major infrastructure improvement projects, we would expect to see increased deforestation along these transportation corridors. As byproducts of primary road improvements,

PAGE 75

75 secondary roads would increases access in the fore st, allowing for improved access to the forest interior facilitating more deforestation. Future work could look at the increase in deforestation along the Cobija-Sena road as the paving of this road progresses. An agent-base d modeling application including key social and economic variables would be able better explain de forestation rates patterns due to infrastructure improvements and population growth.

PAGE 76

76 Figure 3-1. Study area, major roads, populat ion centers, and surrounding geographies

PAGE 77

77 A B Figure 3-2. Landscape patch size metrics for traject ory images at the Pando extent. A) Mean patch size. B). Aggregation in dex. Classes codes are as foll ows: (F-F) stable forest, (F-NF) deforestation, (NF-F) reforest ation, and (NF-NF) stable non-forest.

PAGE 78

78 A B Figure 3-3. Landscape patch shape me trics for trajectory images at the Pando extent. A) Fractal dimension. B) aggregation index. Classes codes are as follows: (F-F) stable forest, (FNF) deforestation, (NF-F) reforesta tion, and (NF-NF) stable non-forest.

PAGE 79

79 Figure 3-4. Temporal changes in mean patch size metrics for multi-distance road buffer and distance from Cobija along road. Letters refer to the following classes: a) stable forest (F-F), b) deforestation (F-N F), c) reforestation (NF-F), and d) stable non-forest (NFNF).

PAGE 80

80 Figure 3-5. Temporal changes in corrected patch perimeter-area metrics for multi-distance road buffer and distance from Cobija along road. Letters refer to the following classes: a) stable forest (F-F), b) deforestation (F-NF) c) reforestation (NFF), and d) stable nonforest (NF-NF).

PAGE 81

81 Figure 3-6. Temporal changes in fractal dimens ion metrics for multi-distance road buffer and distance from Cobija along road. Letters refer to the following classes: a) stable forest (F-F), b) deforestation (F-N F), c) reforestation (NF-F), and d) stable non-forest (NFNF).

PAGE 82

82 Figure 3-7. Temporal changes in aggregation index metrics for multi-distance road buffer and distance from Cobija along road. Letters refer to the following classes: a) stable forest (F-F), b) deforestation (F-N F), c) reforestation (NF-F), and d) stable non-forest (NFNF).

PAGE 83

83 Figure 3-8. Fragmentation and deforestati on by distance from road and from Cobija between 2000 and 2005. The numbers correspond to the number of 50km sections from Cob ija heading towards El Sena. As represen ted on this map, Cobija and El Sena are exactly 250km apart, yielding five sections of 50km each.

PAGE 84

84Table 3-1. Accuracy results for classifi cation rules developed using Compumine Date Number of rules Accuracy (%) Total AUC Actual class Predicted precision Predicted recall Predicted AUC Predicted class 1 Predicted class 2 Predicted class 3 1 1.000 1.000 1.000 363 0 0 2 1.000 0.958 1.000 0 69 3 1986 6 99.423 1.000 3 0.966 1.000 1.000 0 0 85 1 1.000 1.000 1.000 256 0 0 2 0.925 0.980 0.999 0 49 1 1991 11 98.783 1.000 3 0.990 0.962 1.000 0 4 101 1 0.974 0.933 0.998 112 0 8 2 1.000 1.000 1.000 0 350 0 1996 16 98.067 0.999 3 0.923 0.970 0.998 3 0 96 1 0.998 0.993 1.000 419 3 0 2 0.966 0.966 0.999 1 84 2 2000 12 99.015 1.000 3 0.980 1.000 1.000 0 0 100 1 1.000 1.000 1.000 358 0 0 2 0.990 1.000 0.999 0 100 0 2005 8 99.832 0.999 3 1.000 0.993 0.999 0 1 135

PAGE 85

85Table 3-2. Classification accuracies for 2005 classification images and 2000-2005 trajector y image. Classes codes are as follow s: (F-F) stable forest, (F-NF) deforestation, (NF-F) reforestation, and (NF-NF) stable non-forest. Year Overall accuracy (%) Overall kappa Class name Reference totals Classified totals Number correct Producers accuracy (%) Users accuracy (%) Kappa F 33 55 33 100.00 60.00 0.511 2005 87.85 0.6762 NF 148 126 126 85.14 100.00 1.000 F 145 148 142 97.93 95.95 0.921 2000 96.96 0.9392 NF 151 148 145 96.03 97.97 0.959 F 144 148 143 99.31 96.62 0.934 2005 97.96 0.9592 NF 150 146 145 96.67 99.32 0.986 F-F 179 144 144 80.45 100.00 1.000 F-NF 136 144 118 86.76 81.94 0.763 NF-F 74 132 73 98.65 55.30 0.486 20002005 82.01 0.7592 NF-NF 178 147 130 73.03 88.44 0.831

PAGE 86

86 Table 3-3. Absolute values and percenta ges of forest and non-forest in Pando Hectares Area (Sq. km.) Area (%) 1986 Forest 5,405,403 54,054.029 95.96 Non-Forest 21,338 213.380 0.38 1991 Forest 5,387,568 53,875.679 95.64 Non-Forest 39,167 391.667 0.70 1996 Forest 5,384,200 53,842.001 95.58 Non-Forest 42,912 429.116 0.76 2000 Forest 5,353,909 53,539.088 95.04 Non-Forest 72,832 728.321 1.29 2005 Forest 5,325,250 53,252.504 94.53 Non-Forest 101,861 1,018.613 1.81

PAGE 87

87 Table 3-4. Percentages of forest and non-forest, a nd rates of deforestation and reforestation for Pando. Classes codes are as follows: (F-F) stab le forest, (F-NF) deforestation, (NF-F) reforestation, and (NF-NF) stable non-forest. Date pair F-F(%) F-NF(%) NF-F(%) NF-NF(%) F-NF rate(%) NF-F rate(%) 1986-1991 95.52 0.43 0.12 0.26 0.07 0.02 1991-1996 95.21 0.44 0.37 0.33 0.07 0.06 1996-2000 94.71 0.87 0.34 0.42 0.17 0.06 2000-2005 94.13 0.91 0.40 0.90 0.15 0.07 1986-2005 94.42 1.54 0.11 0.27 0.08 0.02

PAGE 88

88Table 3-5. Rates of deforestation and ex tents of forest and non-forest from ot her research conducted in Pando, Bolivia Source Data source Date of forest cover estimate Forest cover (ha) Non-forest cover (ha) Total area (ha) Total forest cleared (ha) Deforestation rate (ha yr-1) Deforestation rate (% yr-1) pre-1985 160,000 CUMUT (1992) Unknown 1985-1990 6,311,000 6,471,000 4,900 0.08 1984-1987 5,707,000 191,800 6,133,100 61,500 Steininger et al. (2001) Landsat MSS and TM 1992-1994 5,509,800 389,000 6,133,100 135,700 6,745 0.11 Rojas et al. (2003) Landsat ETM+ 1993-2000 5,370,825 49,352 6,169 0.11 2004 27,056 Superintendencia Forestal (2006) MODIS MOD13Q1 2005 29,420 0.9 <1976 700 1976-1986 1,200 1987-1991 9,600 1992-2000 3,000 Killeen et al. (2007) Landsat TM and ETM+ 2001-2004 5,877,000 8,800 0.15 1986-1991 5,381,361 14,685 5,396,046 24,475 4,079 0.07 1991-1996 5,363,451 18,336 5,381,786 24,567 4,095 0.07 1996-2000 5,335,268 23,826 5,359,094 49,006 9,801 0.17 Current study Landsat TM and ETM+ 2000-2005 5,303,011 50,516 5,353,527 51,334 8,556 0.15

PAGE 89

89 CHAPTER 4 CHANGING DISCHARGE C ONTRIBUTIONS TO THE R O GRANDE DE TRCOLES Introduction Water quantity issues are of increasing global concern. Many countries in Central America rely on hydroelectric power generation and on su rface water withdrawals for agricultural and water supply. Costa Rica generates more than 80 % of its national power from hydroelectricity and constructed more than 30 dams in the 1990s (Anderson et al. 2006). Decreased flow from major tributaries would imperil reliable power generation, diminish surf icial and groundwater storage accumulated during the rainy season and cr eate water quality issues for municipal and agricultural sectors. In the Ro Grande de T rcoles groundwater comprises 6% of total water used for industrial and urban purposes, and provid es 60% of all water for agricultural irrigation (Blomquist et al. 2005). As such reduced wate r supply coupled with an thropogenic alteration of the surface hydrologic cycle has cr itical implications for water availability. The conceptual framework of this research is that altered vegetated and urban land cover conditions modify hillslope hydrology and produce the non-linear wa tershed responses, impacting water supply. Tropical Hillslope Hydrol ogy: Undisturbed Conditions A watersheds hydrological response to rainfa ll depends on the interplay between climatic, geological and land use/land cover. Additional im portant characteristics include the hydraulic conductivity of the soil at different depths, ra infall intensity and durat ion, and slope morphology (Dunne 1978). Generally speaking, inf iltration capacities of undisturbe d forest soils are sufficient to accommodate most rainfall inte nsities (Bruijnzeel 1990). Overla nd flow in the tropics seems spatially restricted to areas of less permeable soils, on steeper slopes, and at or near soil saturation (Bonell 1998). The thr ee flow pathways of runoff ge neration; infiltration excess

PAGE 90

90 Hortonian overland flow (HOF), saturation exces s overland flow (SOF), and subsurface flow (SSF), have all been observed in undisturbed forests. HOF (Horton 1933) is critical in humid areas with disturbed soils and vegetation, but is restricted temporally and spatially within a watershed. Overland flow is unlikely on midto upper-slopes unless a shallow impeding layer is present (Bruijnzeel 1990), occurring only on lower floodplain areas as SOF (Nortcliff and Thornes 1981). Regardless of position and land cover, during high intensity precipitation rapid runoff (HOF) likely dominates the hydrological response of a watershed. SOF generally prevails in flat bottom valleys with gentle, thinly soiled, slopes (Dunne 1978) and occurs with no obvious topographic or landscape control (Elsenbeer and Cassel 1991). SSF can contribute significantly to storm flow thr ough unsaturated ground on deeply incised convex hillslopes with well dr ained, deep, and highly permeable soils (Dunne 1978). It is propagated in undist urbed forests by macro-pores a nd soil pipes on slopes (Bruijnzeel 2004) with high field saturated hyd raulic conductivities. SSF may al so prevail mainly at depth (Nortcliff and Thornes 1981) and through hollo ws below perched water tables (Dykes and Thornes 2000). Tropical Hillslope Hydrology: Altered Conditions Conversion of tropical forest may produce pe rmanent changes in th e hydrologic response of a stream (Bruijnzeel 2004) resulting from soil compacti on, soil crusting and removal of organic forest litter, which, when coupled with high rainfall intensities, favors HOF in the long term (Bonell and Balek 1993). Several review s (Bruijnzeel 1990; Bru ijnzeel 2004; Bonell 1998; Bosch and Hewlett 1982) exist of the hydrologic responses to land cover change in small watersheds. Bosch and Hewlett (1982) examining experiments in 94 catchment ranging from 1 to 2500 hectares, concluded that increased water yiel d resulted from reduced vegetation. Similar conclusions are drawn from a review of pair ed catchment studies (B onell 1998) and indicated

PAGE 91

91 that more than a 20% loss of fore st cover is required to apprecia bly increase total annual water yield. Increases in water yield are conditioned upon the spatial and temporal variability of rainfall and degree of surface disturbance. Low saturated hydraulic conductivities associated with consolidated surfaces contribute disproportionately to storm flow responses via HOF during events of small rainfall totals and low intensities. Reduced surficial saturated hydrau lic conductivity produces a greater likelihood of HOF, while reductions in subsurface saturate d hydraulic conductivity may increase the frequency of non-Hortonian overlan d flow mechanisms (Ziegler et al. 2004). Elsenbeer (2001) posits that human disturbance diminishes s ubsurface saturated hydraulic conductivity, thereby disrupting subsurface hydrological pathways an d enhancing the generation of non-Hortonian (i.e., return flow) overland flow on fragmented hillslopes (Ziegler et al. 2004). Land cover disturbances can further alter runoff in a non-line ar manner through processes such as infiltration and evaporation. Mesoscale Watershed Studies Fewer land cover change studies have been conducted in mesoscale watersheds (Bonell 1999; Bruijnzeel 2004). To disentangle the effect s of changing land cover and climate forcing, statistical and hydrologic modeli ng methods are required (Refsgaar d et al. 1989). Lorup et al. (1998), employing this combination, detected no change in annual runoff in six semi-arid watersheds (200-1000 km2) in Zimbabwe, even though population density and urban areas increased. A 78% increase in runoff obser ved in the Comet River watershed (16,440 km2) Central Queensland, Australia, (Sir iwardena et al. 2006) following forest clearing and conversion to grasses and cropland was attribut ed partially to an 8.4% increase in rainfall. Applications of an annual water balance model and a simple conceptual daily rainfa ll-runoff model suggested that showed forest clearing increased runoff by 58% and 40%, respectively. Despite a decrease in

PAGE 92

92 forest cover of 80% to 30% from 1957 to 1995 in the 12,100km2 Nam Pong watershed in northeast Thailand, Wilk et al. (2001) observed no statistically signi ficant differences in precipitation, discharge and ev apotranspiration. However, desp ite no significant changes in precipitation, Costa et al. (2003) linked significant changes in mean annual and high-flow season discharge on the Tocantins River to a 20% incr ease in agricultural lands from 1960 to 1995. Only two dynamic simulation hydrologic m odeling studies (Colby 2001) and (van Loon and Troch 2002) have been carried out in Co sta Rica. and neither us ed a physically-based, spatially explicit hydrologic model to assess th e impacts of changing land cover. Krishnaswamy et al. (2001) used dynamic linear regression mode ling investigated land co ver and hydro-climatic effects on stream flow in the Terraba watershed (4767 km2), primarily focused on sediment production. In this region of intense human alteration and high climate variability further research is needed on climate and land cover impacts on surface hydrology. Research Justification The balance of contributions of the two major tributaries, Rios Gra nde de San Ramn and Virilla, to the Ro Grande de Trcoles, Costa Rica appears to have changed around 1975. The change could result from the non-linearity of basin runoff responses to land cover change, a change of climate forcing, or both. An applie d, multifaceted approach of simulation modeling and statistical analysis is empl oyed to investigate these potential causes for the observed changes in discharge. Objective 1: Determine the significance of changes in discharges and precipitation. Objective 2: Construct a hydrologic model for the two sub basins for long term simulations of potential non-linear responses of discharge contribution conditioned on dominant land cover conditions. Objective 3: Assess the potential of climate variab ility as cause of discharge changes.

PAGE 93

93 Hypothesis 1: Precipitation input to the Rio Virilla sub basin has increased after 1975 forcing a greater contribu tion to discharge in the Rio Grande de Tarcles. Hypothesis 2: Differences in the proportions of ve getated versus urban land covers caused the sub basins to act as non-lin ear amplifiers of precipitatio n, through differences in soil moisture and evapotranspiration fluxes. Hypothesis 3: Differential spatial responses to precipitation influenced by El Nio Southern Oscillation (ENSO) and Atlantic sea surface temperatures result in distinct discharge responses even though the basins are adjacent. Study Site The Rio Grande de Trcoles watershed (1745 km2) (Figure 4-1) lies in Costa Ricas central tectonic depression flanked to the north by the northwest-southeast tren ding Cordillera Central and to the south by the Cordillera Talamanca. The watershed encompasses much of the valley and contains most of the metropolitan population of Costa Rica. Two major sub-basins, the Rio Grande de San Ramn (916 km2) and the Rio Virilla (829 km2), comprise the Trcoles watershed. The area is topographically varied an d rugged, with steep slopes geomorphologically dominated by ancient volcanic pyr oclastic flows and lacustrine depositions. Alfisols, entisols, inceptisoles, ultisols, and vertisols are the primar y soil orders in the watershed (Centro Cientfico Tropical (CCT) 1989) constituted primarily of volcanically-derived clays. Land cover differences between the Grande de San Ramon and Virilla are marked. The former is dominantly forested, pastoral and agricult ural lands (approximately 98%) with less than 2% urban areas (Ministerio de Agricultura y Ga nadera (MAG) 1992). By contrast the Gran Metropolitan urban area of Costa Rica occupies about 15% of the Rio Virilla sub-basin a nd agriculture is the dominate vegetated land cover class (about 33%). Due to Costa Ricas location between the Caribbean and the Pacific, precipitation generating mechanisms are complex (Waylen et al 1998). The northeast trade winds bring ample precipitation to the Caribbean coast, leaving the Pacific side in rain shadow during the dry season

PAGE 94

94 (November-April) except for the southeastern co rner of the Rio Virilla sub-basin, which is exposed to a wind gap in the Cordillera Cent ral. As the Inter-Tropical Convergence Zone (ITCZ) of the Eastern Pacific (Hastenrath 2002) migrates northward through the boreal summer, convective precipitation falls over much of the Pacific coast and montane areas from May to November. Low flows during the dry season are su stained by depletion of groundwater and soil water storage from the previous rainy season. Preci pitation at the onset of wet season begins to replenish the stores. July and August witness the veranillos, a temporary dry period caused by intensification of Northeast trade winds This s easonal reduction is a prominent feature of the precipitation climatology thr oughout Central America (Magana et al. 1999). During the postveranillos months of September and October soil mo isture stores are saturated or near saturated producing high runoff regardless of precipitation totals and intensity. The southward migration of the ITCZ heralds the dry season in November, initiating the release of groundwater stores and Methods SWAT Model The Soil and Water Assessment Tool (SWAT) (Arnold et al. 1998; Arnold and Fohrer 2005) is a semi-distributed, process based, comp uter hydrologic model that simulates the land based hydrologic cycle using a water balance. Complex, mesoscale watersheds are partitioned into hydrologic response units (HRUs) compri sed of unique land cover and soil combinations. Hydro-climatic inputs drive the relative impor tance of each hydrologic components in SWAT. After canopy storage and interception are simulate d excess precipitation is (1) passed to the land surface to be infiltrated with the potential for redistribution within the soil, (2) moved out of a HRU by subsurface lateral flow, or (3) moved over the land surface as surface runoff. Flood water accumulations are calculated for main cha nnels and routed, preserving channel mass flow, through the stream network and re servoirs, ponds or lakes to th e watershed outlet. SWAT has

PAGE 95

95 been primarily applied in the United States a nd Europe with few applications in tropical countries (e.g. Kenya and Indi a) (Arnold and Fohrer 2005). Data Daily discharge data were obtained from the Co sta Rican Institute of Electricity for five gauging locations within the watershed. The length of record for the gauges on the Rio Poas at Tacares, the Rio Virilla at San Miguel and th e Rio Tarcoles at Balsa spanned 1964 to 1986. Daily discharge data for the Tarcoles and Virilla were aggregated to monthly and annual time scales, and estimates of monthly discharge for th e Rio Grande de San Ramn, which is currently ungauged, obtained by subtracting the discharge of the Virilla (immediately above the confluence) from that of the Tarcoles (below c onfluence). Data from the remaining gauges were aggregated similarly and employed as inputs for the sensitivity analysis and autocalibration in the SWAT model. Precipitation data were obtained from the Costa Rican Institute of Electricity and the Ministry of Environment and Ener gy (1988). Sixty seven stations had periods of monthly totals from 1960 to 1986, of varying lengths, of these only three report complete records. Nine stations had daily records of varying le ngths (1964 1986). Missing daily and monthly precipitation data were filled used inverse distance weighted sp atial interpolation to produce a temporally continuous record. From the interpolated monthly records, seasonal and annual time series were generated for each station. For input into the SWAT model, two random sampling designs, one for the input precipitation and th e other for the stochastic weat her generation, were generated based on the daily interpolations Two sets of 30 random points, 15 in each subbasin to ensure adequate spatial representation of precipitation, were generated and the daily precipitation data extracted at each point location.

PAGE 96

96 The NCEP/NCAR reanalysis program (Kalnay et al. 1996) provided surface atmospheric data at both 2 x 2 degree and 2.5 x 2.5 degree re solution and daily time steps from 1964 to 1986. Relative humidity, minimum and maximum temper atures, and insolation were extracted from global datasets and potential evapotranspiration estimates we re conditioned upon mean surface temperatures and average wind speeds calcu lated from the U and V vector wind data. Expanded soil properties of percentages of silt and clay, profile depth, texture, and organic content from Vasquez (1980) were related to the a GIS soils layer at 1:200,000 based on the great group classification (Centr o Cientfico Tropical (CCT) 1989) while median bulk densities were based on soil order (Alvarado and Forsythe 2005). Additional hydraulic soil parameters of available water content, field capacity and sa turated hydraulic conductivity were calculated for each great group using the pedotransfer pr ogram Rosetta (Schaap et al. 2001) Land cover data were generated from La ndsat images from 1975 (MSS 3) and 1986 (TM 4). Image preprocessing included georeferencing the MSS image to the TM image and radiometrically calibrating both images. Both im ages were classified using the RuleGen (Loh and Shih 1997) decision tree clas sifier creating five classes fo r forest, developed, grassland, cropland, and shrubland. Derived products included in the deci sion tree classification included tasseled cap transformation and loca l Morans I at a spatial lag of 3 pixels. Elevation, slope, and cloud and shadow masks were also included in the classification. Classified images were resampled to 57x57 meters resolution and Manning s n values assigned to each land cover class following Bedient and Huber (1988). HydroShed (Lehner et al. 2006) furnished elevat ion data at 15 arc seconds (approximately 92 x 92 meters) and yielded slope data for use in ArcSWAT Data were processed further using the ArcSWAT watershed delineation module to en force the observed stream drainage network,

PAGE 97

97 digitized from 1:50,000 scale t opographic maps, onto the elevation data. In this way the delineated streams closely followed actual stream courses, resulting in better accuracy of flow length and slope. Stream channel dimensions were calculated using ArcSWAT and Mannings n roughness values were assigned following Bedient and Huber (1988). Statistical Analysis: Change detection The significance ( = 0.05) of changes in annual runo ff contributions from each sub-basin before and after 1975 were determined using a hypergeometric dist ribution (Equation 4-1), which defines the probability of drawing a sample, n, (without replacement) containing x successes, from a finite population, N, consisting of R successes. n N x n R N x R x f) ( (4-1) In this case x is the number of above (below) long-r un median annual runoff (or precipitation) values before and after 1975. Monthly flow volumes from the Rio Virilla are expressed as percentages of the combined volume observed in the Rio Grande de Trcoles each year. Simple linear regression fit to the time series for each monthly plot indicate the presence of statistically significant trends( = 0.05) in percentage contribution of the Rio Virilla. Comparable monthly precip itation inputs to each sub-basin are estimated via four methods; Arith metic mean, Theissen polygons, Inverse distance weighted and Spline interpol ation (Dingman 2002), employi ng a custom application in ArcGIS. Double mass curves of sub-basin input and runoff derived from each method were compared for each sub-basin and time scal e using the Kolmogorov-Smirnov goodness of fit

PAGE 98

98 statistic. No significant differe nces (alpha = 0.20) were detected between the methods at the annual and seasonal time scales, while the Arit hmetic mean/Theissen polygon combination and the Inverse distance weighted/Spline combination curves tested similar at monthly scales. A representative precipitat ion input to each sub-basin was theref ore derived as the average of the estimates from each method. Deviations of annual and monthly inputs (sub-basin volumes) from their respective long-run medians for the Virilla and Grande de San Ramn were analyzed in a similar fashion to flows. Statistical Analysis: Climate Variability Monthly precipitation totals from 1959-1986 were aggregated into five seasons (JFMA, MJ, JA, SO, and ND) at all stations and expressed as standard normal deviates. These were then categorized according to the joint states (above/below median) of sea surface temperatures (SSTs) in the equatorial Pacific (ITCZ) and tropical At lantic (Northeast trades). Composite maps of seasonal standard deviates in each of the four possible combinat ions of oceanic states indicate the seasonally varying importance of the Pacifi c and Atlantic SSTs, a nd their interaction, on precipitation. Spatial interpolation of composite deviates across the basins reveal any geographic patterns of such influences and any changes (Waylen and Quesada 2002). SWAT Model Construction Following data preprocessing, SWAT model c onstruction proceeded using the ArcSWAT (Winchell et al. 2007) interface for ArcGIS. Th e interface provides for watershed delineation with outlet and stream definition; creation of hyd rologic response units (HRUs); input of climatic variables; writing of input f iles and running of the model. Hy drologic response units define unique combinations of land cover, soils and sl ope, which act as the spatial control in the conversion of effective precipita tion into runoff. For each land cover condition in 1975 and 1986,

PAGE 99

99 61 HRUs were delineated. In a ddition to model construction a nd execution, ArcSWAT provides tools for calibration, sensitivity and uncertainty analyses. Daily precipitation and discharge records we re split into three periods for calibration, validation and simulation dependent upon the land cover conditions under which the model was executed. Under the 1975 land cover, calib ration was from 19641970, validation, 1971-1975, and simulation, 1976-1986. Under the 1986 land cove r, SWAT was calibrated from 1976-1981, validated 1982-1986, and the simulation was from 1964-1975. The first year of calibration was omitted during the evaluation of the model perfor mance. During each calibration and validation period simulated monthly runoff was compared to observed and linear regression used to assess initial model fit with an R2 threshold of 0.60-0.80 (Neitsch et al. 2002). More quantitative measures of OLS bisector slope (Isobe et al. 199 0) and Nash-Sutcliffe efficiency metric (Nash and Sutcliffe 1970) were calculated subsequently. The SWAT model possesses a bui lt-in sensitivity analysis routine (Green and van Griensven 2008) employing a Latin-hypercube co mbined with a one-factor-at-a-time sampling. Each parameter selected following Neitsch et al (2002), the range of parameter values and the method by which the parameter is adjusted during th e sensitivity analysis are listed in Table 4-1. Given the large number of model parameters pot entially adjusted during calibration, sensitivity analysis was used to identify t hose parameters most likely to improve model performance during automatic calibration. One thousand combinations of values of the 11 parameters were run for the respective calibration periods under each la nd cover condition and the model parameters ranked by sensitivity. The three least sensitive we re omitted from the subsequent autocalibration. SWAT employs the PARASOL (Parameter Solution method) autocalibration method (Green and van Griensven 2008; van Griensven and Meixner 2004) coupled with uncertainty

PAGE 100

100 analysis of model parameters to improve ove rall model fit. PARASOL is based on the Shu ed Complex Evolution Algorithm (SCE), a global se arch algorithm that minimizes the sum of squares residuals between observed and simulated runoff, based on multiple model parameters and is run and evaluated following the manual calibration and sensitiv ity analyses. Initial autocalibration produced unreasonable results as parameters controlling runoff were preferentially selected for adjustment and unlikely ranges, consequently, only parameters controlling subsurface and groundwat er flow were selected in the autocalibration step. The Virilla seemed particularly se nsitive during manual calibration; therefore autocalibration was dominantly focused on improving the modeled r unoffs with a more f easible set of model parameters. SWAT also contains SUNGLASSES (Sources of Uncertainty Global Assessment using Split Samples) (van Griensven and Meixner 2004) as a means of assessing uncertainty in the model output attributable to model formulation ra ther than parameter uncertainty. In order to better evaluate model predictive power and prediction errors, SUNGLASSES examines the validation parameter set separate from the ca libration set (van Grie nsven and Meixner 2004). Global optimization criterion (Chi-squared, = 0.05) assesses the fit between the observed and simulated runoffs. The procedure was replicated 2000 times for the simulation periods for both subbasins and land cover conditions. Once acceptable calibration and validation resu lts were obtained, SWAT was rerun under both land cover conditions for the simulation peri ods in the sub-basins thereby creating four scenarios based on combinations of prea nd post-1975 precipitation, c oupled with the two different land cover conditions (Figure 4-2). Scenario s 1 and 4 actually occu rred, while 2 and 3

PAGE 101

101 assist in identifying the relative dominance of pr ecipitation regime and land cover over runoff. Differences in runoff from each scenario provide the basis for the e following hypotheses. H0: Precipitation and land cover do not influenc e runoff (scenarios 1 and 4 not different) HA1: Land cover changed so runoff changed (s cenarios 1 and 3, and 2 and 4, different) HA2: Precipitation changed so runoff changed (scenarios 1 and 2, and 3 and 4, different) The greatest differences in simulated runoffs might be expected between scenarios 1 and 4 due to combined changes in precipitation regi me and hydrologic processes affected by land cover. Large differences under the other hypothese s would implicate one part icular variable as the cause of the change in runoff. Further qu antitative determination of the relative importance of these controls is provided by the application of the Mann-Whitney U test of medians to the mean monthly runoff values under each scenario. Results Statistical Analysis Time series (Figure 4-3) of s ub basin annual runoff as a percen tage of total runoff from the Trcoles compared to sub basin area shows that prior to 1975, neither sub basin consistently contributes greater proportional runoff than the other. In all years, except tw o, after 1975 the Virilla contributes a greater propo rtion. Test reveal statistically significant changes in the counts above/below median annual flows preand pos t-1975 in both sub basins (Figure 4-4), although the pattern is less marked in the Virilla. Regre ssion parameters (Figure 4-5) from the monthly time series indicate a sta tistically significant increase in percentage contribution from the Virilla from January to August. Annual precipitation (Figure 4-6) displa y similar temporal patterns; however, no significant change in annual prec ipitation to the Grande de San Ramn was detected at all, the opposite to the significances observed in flows. Regression applied to th e monthly precipitation time series (Figure 4-7) only yields significant linear trends in March and June. The former is

PAGE 102

102 one of the driest months, and the trend in the latt er in the opposite to that observed in flows Small changes in sub basin inputs appear to ha ve produce disproportionate changes in flows. SWAT Model The principle purpose of the SWAT in this application is to determine the causes for observed changes in runoff rather than to focu s on model construction performance; however its credibility in performing the former is predicated on the latter. Nash-Sutcliffe (NS) efficiency measures and bisector slope values (Table 42) suggest acceptable model fits after manual calibration. Measures of fit from validation trials should be equal to, or slightly less than, those from manual calibration as witnessed in for both sub-basins under 1986 land cover. However, validation of the Rio Virilla unde r 1975 land cover yields an extremely low NS and all validation measures for the Rio Grande de San Ramn under earlier land cover are higher than the respective values for the manual calibration. The sensitivity analyses identified model para meters that affected the simulation response, particularly those that could be adjusted further to improve model performance. Runoff appears most sensitive (Table 4-1) to groundwater para meters, with the runoff curve number (RCN) and Mannings n for channel flow also rated highly. Initial experimentation with autocalibrati on produced unsatisfactory NS values and bisector slopes. Although the variance about the bisector slope decrease d considerably for the Grande de San Ramn under the 1975 land cover th e NS value decreased substantially compared to the manual calibration (Table 4-2). Specifica lly, the autocalibration fo r the Grande de San Ramn under 1975 land cover performed worse than the manual calibration (Table 4-2) and there was a general lack of consistency in res ponse of Grande de San Ramn to autocalibration. Further support for rejection of the autocalibrated models incl uded increase in intercept (all

PAGE 103

103 scenarios) and the bisector sl ope (Rio Virilla, 1975 land cove r) although the variance about the bisector slope was reduced cons istently for each simulation. Computational demands of the PARASO L and SUNGLASSES processes permitted 2,000 trial simulations and led to an incomplete anal ysis of parameter and model uncertainties, as evidenced by the lack of sampling over the parame ters range, especially for the baseflow alpha coefficient (Table 4-4). However, the results do provide insight in to the ranges of uncertainty for more sensitive groundwater flow parameters. Othe rs challenges of prem ature termination of optimizations and convergence of parameter popul ations may be corrected by increasing the number of optimization simulati ons ten-fold. This was not a f easible option dur ing this study. Simulations Figures 4-8 and 4-9 show absolute differe nces between simulated runoffs under the different precipitation and land cover scenarios and can best be viewed as paired scenarios; that is, runoff differences (red and green bars) resulting from the land cover scenarios be paired together, and those from precip itation (blue and yellow) be view ed separately. Regardless of subbasin, land cover or precipitation regime, th e null hypothesis of no change in runoff is rejected. Sensitivities of runoff (Figure 4-8) within the Grande de San Ramn follow the hydroclimatology regime of the region (Figure 4-1). La nd cover dominates runoff in the drier months and precipitation in the wetter ones. Comparing the preand pos t-75 influences of precipitation and land cover on runoff differences, equal influe nce is implied during January-April; however land cover prevails at the beginning of the rai ny season (May and June), when soil moisture storage is at a minimum. Reduced precipitation in July and August, the Veranillos, renders little difference between simulations, however, sensitiviti es switch to precipit ation during the height of the rainy season (September-November). Thes e observations are consis tent with those of Bruijnzeel (2004) and Bonnell (2004).

PAGE 104

104 The reversals of runoff sensitivities of ch anges between precipita tion periods for the changing land cover and vice vers a are perplexing. For example, the scenario combination 1/3, of pre-75 precipitation under changing land cover conditions from 1975 to 1986, produces sensitivity in runoff during the dry season (Janua ry to March), the Vera nillos (July and August) and October and November, then reverses to scenario combination 2/4 for April to June, September, and December.. Over all the Virilla displays a grea ter sensitivity to changes in precipitation (scenarios 1/2 and 3/ 4), as expected given the more extensive urban cover which proportionally converts more rainfa ll to runoff than a vegetated surface. The trend of runoff sensitivities to changing precip itation persists throughout the year, except during June and July, when land cover exerts control over runoff. The change in runoff sensitivities during the dry season and the Veranillos is not apparent. Howe ver, similar to the Grande de San Ramon, a marked reversal is seen of runoff sensitivities to changing precipitation under the two land cover conditions, from scenario 1/2 in the drier months to 3/4 in the wetter months. Again the runoff sensitivity to scenario 3/4 is pronounced from June to September. Climate Variability Climate variability, specifically the combined effects of ENSO and the Atlantic sea surface temperature anomalies (SSTA), may explain precip itation trends highlight ed by the statistical analyses. A global climate shift ha s been identified in the mid-1970s. In the time series available here the Southern Oscillation Index (SOI) indicates more warm phase events of ENSO (El Nios: droughts in the basins) than cold even ts (La Nias: excess ra infall) after 1975. The Atlantic also shows a change about 1975, w ith predominately positive SSTAs earlier and negative SST anomalies post-1975. Coupling of SSTs in the two ocean basin have been shown to heavily influence precipitation input s to the area (Poveda et al. 2006).

PAGE 105

105 Standardized seasonal precip itation deviates from 1959-1986 (F igure 4-10) indicates low precipitation (negative deviate) in season 1 (JFMA) for the Rio Gr ande de San Ramon and slight positive deviate for the Virilla. The discharge reco rd reflects the release of the groundwater and soil moisture stores during this period, and variab ility in seasonal precipitation has little impact on discharge. Season 2 (MJ) also experiences more negative deviat es in precipitation due to the cold Atlantic-warm Pacific combination. Much of the precipitation normally replenishes subsurface water stores depleted in the previ ous season. The diminished precipitation in the Grande de San Ramon appears to have b een amplified by basin characteristics (e.g. topographically-induced rain-shadow) producing even lower flows. Season 3 (JA) encompassing the veranillos, has been shown to be pa rticularly sensitive to the ENSO signal. Droughts associated with warm phases and cooler Atlantic SSTAs, are particularly apparent in the Grande de San Ramon. However, regardless of the diminished precipitation subsurface water stores are stil l filling in both seasons 2 and 3. Drought-like conditions persist during season 4 (SO) with an even greater deviations Although the Grande de San Ramon shows a greater response than the Vi rilla, discharges do not reflect this as the subsurface water stores were full and both subba sins respond similarly (as evidenced by the nonlinearity of basins responses). In season 5 (ND) the signal flips from the previous season and both subbasins show a positive response to the EN SO-Atlantic influence. In the Grande de San Ramon this was relatively unimportant since the rainy season was ending and discharge enters the recession stage. The Virilla has a higher deviate explainable by the gap over the Cerro Carpinteral near Cartago, which allows increased precipitation influenced by nortes (Waylen et al. 1998) coming down from the North American c ontinent and interactin g with the Northeast trades.

PAGE 106

106 Discussion Model Performance Although this work expounds on mesoscale watershed modeling simulations in the tropics there are some shortcomings. The inability of SWAT to assign more than one rain gauge for each watershed diminishes the realistic spatial simulation of daily rainfall distribution over the watershed. The model fit during calibration may ha ve been improved by the use of the complete set of the generated random rain gauge locations based on the interpolatio n of the actual daily gauges. This, combined with the low number, and intermittent nature, of available daily gauges, inevitably leads to an over-s implification in the represen tation of rainfall inputs. The coarse scale of the atmospheric parameters available to simulate evapotranspiration is a potential limitation. The complex topography of th e Tarcoles watershed (F igure 4-1) gives rise to great spatial variability in wind speeds, te mperatures and humidity depending on spatial and topographic position, none of which was captu red by the single value permitted for the calculation of evapotranspiration in the model. Values of important soil hydraulic parameters were assigned using literature values related to soil group without field valida tion and verification. Although show n to have little significance during the sensitivity analysis important model parameters such as saturated hydraulic conductivity and available water co ntent can have a significant effect on model performance, and impact the resulting calibrations and simulations. These modeling challenges and limitations are not unique to this st udy. Other researchers (Bonell 1999; Bruijnzeel 1990; Bruijnzeel 2004) have echoed the difficulties of conducting mesoscale watershed modeling efforts in the trop ics due to the lack of field data, and model parameterization using literature or derived parameters. Surroga te parameters from secondary sources are inadequate to capture the complex, spatial heterogeneity of mesoscale watersheds.

PAGE 107

107 Currently, however, there is a dearth of sufficien t data for mesoscale studies. Methods need to be practical and operational to determine the appropriate scale of model parameters and to validate these data for real-w orld conditions, particularly in the tropics (Bonell 1999). Non-Linear Nature of Sub-Basins The observed changes in the Virilla (Figure 49) from rainfall dominated responses under 1975 land cover in seasons one and two to sens itivities under the 1986 land cover are physically reasonable as the "more" naturally vegetate d conditions of 1975 pr oduce a stronger response during drier conditions. Under wetter conditions more disturbed land cover (e.g. urban and pasture) produces a greater proport ional response in runoff. This sensitivity is most apparent after the soil and groundwater stores have b een depleted during the dry season. In June sensitivities are almost equal, particularly in the simulation under pre 1975 rainfall and 1975 land cover. Although there appear to be no significant increases in rainfa ll, it does appear to dominate the differences in runoff in the simulation resu lts. The Virilla sub-basin has complex rainfall climatology caused by the interaction of the Pacific influences and an incursion of Caribbean air to the southeast. Observed annual precipitation and runoff in bot h basins are plotted against one another (Figure 4-11). The 1:1 line represents the specia l case of no evapotranspiration losses. An upper enveloping line represents the maximum expected runoff for a gi ven annual precipitation. The difference between the enveloping line and the 1:1 line represents minimum possible evapotranspiration. As precip itation increases, the absolute value of evapotranspiration decreases and evapotranspiration as a percentage of preci pitation declines even more markedly in both subbasins. In years of diminished rain a lower proportion of the inpu t goes to runoff; while in rainy years the opposite is true.

PAGE 108

108 With respect to this non-linear amplification of rainfall signal in runoff, there are two other notable aspects of land cover a nd the 1975 switch in runoff responses The non-linearity is not as apparent in the Virilla. Being a more urban ba sin, more runoff goes dir ectly rivers, by-passing potential sources for evapotranspiration in so il and groundwater stores. Observations of the Grande de San Ramon post-1975 clearly plot be low those of pre-1975. So although rainfall has declined little since 1975 (Figure 4-6) evapotra nspiration appears to have increased; an effect that is not as marked in the Virilla where the groundwater stores are by-passed. Challenges of Mesoscale Modeling Studies Some contemporary mesoscale watershed studie s have shortcomings, including a lack of statistical rigor in testing and others which only incorporate implic itly land use/land cover. Costa et al. (2003) used parametric sta tistics without testing for normality of their hydro-climatic data while agricultural census data (from 1960 and 1995) and satellite data (from 1995) provided empirical observations of changes in agricultu ral lands. Neither land c over nor agricultural census data were included explicit ly in the statistical analysis of the hydro-cl imatic data. Correct use of statistical methods and the explicit incl usion of land cover ar e needed to improve mesoscale watershed studies. A preferred approach is to use complementary statistical methods coupled with dynamic hydrologic modeling si mulations such as Lorup et al. (1998). Contradictory results as to the effects of land cover change on climatic and hydrologic responses are common in the litera ture. Wilk et al. (2001) reported that despite a greater than 50% reduction in forest from 1957 to 1995 no change s in seasonal or annual rainfall totals were detected. Lorup et al. (1998) found no indication of increased runoff or changes in land use although population increased signifi cantly. In contrast to previ ous research (Bonell and Balek 1993; Bruijnzeel 1990) Costa et al. (2003) propose th at alterations in hydrol ogic regime arising from land use changes would be more evident in the rainy season and th at higher discharges

PAGE 109

109 would be expected from more intense land use. Reduced infiltration, although insufficient to affect dry season flow, increased surface runo ff during the rainy season, while associated reduced evapotranspiration increased discharg e throughout the year (C osta et al. 2003). As watershed size increases, traditional paired catchment studies become infeasible due to excessive cost of instrumentation and difficulty in controlling land cove r treatments (Bosch and Hewlett 1982). Results from small watershed stud ies cannot be upscaled to larger watersheds (Bruijnzeel 2004; Wilk et al. 2001; Bonell and Balek 1993), in which it is difficult to separate the effects on river discharge from climate vari ability and anthropogenic land cover change. Additionally, large withdrawal of water for municipal, agricultural and industrial uses complicate hydrologic land cover analyses (Bruijnzeel 2004). A variety of biophysical and hydro-climatic factors complicate the analysis of hydrologic responses to land cove r change in large watersheds. Unlike well-instrumented experimental hillsl opes and small catchments, the heterogeneity resulting from combinations of soils, geology, and spatial and tem poral distribution of rainfall, makes detection of the effects of land cover ch ange considerably more difficult. Hydrologic impacts may be muted by the heterogeneous watershed characteristics and mosaic of multiple land covers and varied land use practices, unde rlain by variability in soils, geology, and topography (Bruijnzeel 2004). Land use and land c over changes are spatially and temporally heterogeneous, and impede the detection of change s in discharge and other variables in the water balance (Wilk et al. 2001). Increased surface runo ff may not be detectable in river discharge due to the spatial variability of rainfall. The effect of land cover fragmenta tion on altered hydrologic regimes is not well understood especially in larg e watersheds nor is th e hydrologic connectivity between land cover fragments. As intensive a nd extensive hydro-clima tic instrumentation in

PAGE 110

110 large watersheds is unfeasible we must continue to refine similar analytical studies with improved hydrologic models and statistical methods. Conclusions Both sub-basins show a statis tically significant d ecline in discharge at the annual time scale after 1975 the Rio Virilla less so than the Rio Grande de San Ramon. Precipitation inputs however have not changed significan tly. Basins characteristics, part icularly land cover types, act to amplify and modify the convers ion of precipitation to discharg e. The non-linear ity of response in the Rio Grande de San Ramon becomes evid ent through the analysis of simulation model results. The post-1975 decline ma y also result from to occu rrence of warm phase ENSO (drought) conditions, whose effects are amplified by a persistently cooler Atlantic. Although the basins are contiguous, the historic data indicate slightly differe nt responses to ENSO forcing, while varying land use land cover further enhance this, yielding th e noted degrees of diminished discharges over the time period. Changes in runoff stem from both land cover change and climate variability and are particularly pronounced at the monthly and seas onal scales. Following earlier literature, dry season flows are more affected by changes in land cover while precipitation exerts a greater control in the wet season. In climatically and hydrologically complex areas such as Costa Rica, it is very difficult to disentangle the effects of land cover change from climate variability on stream flow in larger scale (mesoscale) wate rsheds. The Rio Tarcoles watershed is complex, experiencing interplay between the Atlantic and Caribbean, a nd the Pacific and various time scales (5-7 years ENSO, severa l decades for Atlantic). Changes in the resultant climate variability need to be accounted for when s eeking hydrologic consequences to changing land cover conditions. Further explor ation of climate variability mi ght include the division of the simulation periods conditioned upon ENSO phase a nd investigation of differences in the mean

PAGE 111

111 monthly runoff. In addition to the standard prec ipitation deviates, the in fluences of ENSO and Atlantic SSTs can be controlled providing a desc ription of the runoff responses most expected under different ENSO phases. The application of the processbased simulation model SWAT in Costa Rica (a first) aims to resolve mesoscale watershed heterogeneity by identifying non-linear hydro-climatic responses in two sub-basins with differing dominant land cover characteristics. The impacts of long term climate variability on a 26 year historic hydro-clim atic record are identifie d using the integrative approach of hydrologic modeli ng and statistical analyses focused on addressing changing discharge contributions from two tropical sub-basins. This research augments existing mesoscale watershed studies combining simulation modeling and statistical analysis to distinguish the hydrologic impacts from land cover from those due to climate variability on discharge contributions within a tropical watershed. This is currently an understudied area with potential impacts on water withdrawals for municipal and ag ricultural use not only in Costa Ricas central valley, but for other global metropolitan areas.

PAGE 112

112 Figure 41. Study area and monthly runoff regimes

PAGE 113

113 Figure 42. Matrix of possible combinations under changing precipitation and land cover conditions for SWAT simulations

PAGE 114

114 Figure 43. Sub basin annual runoff as a percenta ge of confluence runoff compared to sub basin

PAGE 115

115 Figure 44. Deviates fr om median annual runoff

PAGE 116

116 Figure 45. Regression summaries of monthly runoff step plots

PAGE 117

117 Figure 46. Deviates from median annual precipitation

PAGE 118

118 Figure 47. Regression summaries of monthly precipitation step plots

PAGE 119

119 Figure 48. Mean monthly runoff differen ces for the Rio Grande de San Ramon unde r various land cover and precipitation combinations Mean Monthly Runoff Differences for Rio Grande de San Ramon 0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 123456789101112 MonthsAbsolute Differences (mm) H0:Scen.1/4 Ha1:Scen.1/3 Ha1:Scen.2/4 Ha2: Scen.1/2 Ha2:Scen.3/4

PAGE 120

120 Figure 49. Mean monthly runoff differen ces for the Virilla under various land cover and prec ipitation combinations Mean Monthly Runoff Differences for Virilla 0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 123456789101112 MonthsAbsolute Differences (mm) H0:Scen.1/4 Ha1:Scen.1/3 Ha1:Scen.2/4 Ha2: Scen.1/2 Ha2:Scen.3/4

PAGE 121

121 Figure 410. Seasonal standardi zed precipitation deviates for th e Rios Grande de San Ramon and Virilla

PAGE 122

122 Ro Grande Annual Precipitation (mm) 1000150020002500 Annual Runoff(mm) 1000 1500 2000 2500 1960-1975 1976-1986 1:1 R = P Virilla 1000150020002500 1000 1500 2000 2500 Annual Precipitation (mm) Input = Output +/S P = { R + E} +/0 1800 = 1000 + 800 -: E = 800 (44% of P) 2300 = 1850 + 450 -: E = 450 (20% of P) Input = Output +/S P = { R + E} +/0 1800 = 1150 + 650 -: E = 650 (36% of P) 2300 = 1800 + 550 -: E = 550 (24% of P) E E E E R R R R Figure 411. Non-linearities in subbasin responses for runoff, precipitation, and evapotranspiration

PAGE 123

123 Table 4-1. Model parameters adjusted during ma nual and automatic calibrations, and sensitivity and uncertainty analyses Rio Grande de San Ramon Rio Virilla Parameter Variation method Lower bound Upper bound 1975 1986 1975 1986 Baseflow alpha coefficient (days) 1 0 1 8 6 6 8 Manning's n for channel 1 0 1 7 8 7 7 Initial SCS CN II value 3 -25 25 3 3 3 3 Soil evaporation compensation factor 1 0 1 1 2 1 2 Groundwater delay time (days) 2 -10 10 6 7 8 6 Groundwater "revap" coefficient 2 -0.036 0.0364 4 4 4 Threshold water depth in shallow aquifer for flow 2 -1000 1000 2 1 2 1 Threshold water depth in shallow aquifer for "revap" 2 -100 100 5 5 5 5 Average slope steepness 3 -25 25 10 10 10 11 Average slope length (m) 3 -25 25 11 11 11 10 Surface runoff lag 1 0 10 9 9 9 9

PAGE 124

124 Table 4-2. Measures of model fit for manual calibration and validation Year Calibration Slope Slope theoretical Variance of slope Intercept NashSutcliffe efficiency 1975 Rio Grande de San Ramon 1.171 1.321 0.270 -27.041 0.5234 Rio Virilla 1.121 1.229 0.092 11.492 0.4018 1986 Rio Grande de San Ramon 1.017 1.034 0.037 14.125 0.7055 Rio Virilla 1.211 1.397 0.434 -29.584 0.5486 Year Validation Slope Slope theoretical Variance of slope Intercept NashSutcliffe efficiency 1975 Rio Grande de San Ramon 1.035 1.068 0.084 -28.690 0.6477 Rio Virilla 1.107 1.200 0.304 32.169 0.0768 1986 Rio Grande de San Ramon 0.984 0.968 0.131 35.335 0.6032 Rio Virilla 1.259 1.482 0.815 -17.856 0.4357

PAGE 125

125 Table 4-3. Measures of model fit for automatic calibration Year Calibration Slope Slope Theoretical Variance of Slope Intercept NashSutcliffe Efficiency 1975 Rio Grande de San Ramon 0.861 0.729 0.047 -28.495 0.0696 Rio Virilla 1.120 1.227 0.091 11.648 0.4023 1986 Rio Grande de San Ramon 0.913 0.816 0.012 20.381 0.7207 Rio Virilla 1.098 1.482 0.175 -22.010 0.6866

PAGE 126

126Table 4-4. Output SWAT parameter ranges and percentages of valu e range from PARASOL automatic calibration routine. Sub-basins are coded where RGSR is the Rio Grande de San Ramon and RV is the Rio Virilla. Parameter Variation method Minimum value (1975) Maximum value (1975) Percentage of range (1975) Minimum value (1986) Maximum value (1986) Percentage of range (1986) Baseflow alpha coefficient (days) 1 0 2.594E-05 0.00 0.0081878 0.014919 0.67 Soil evaporation compensation factor 1 0 0.44921 44.92 0 0.25236 25.24 Groundwater delay time (days) 2 -10 10 100.00 -10 -1.8251 40.87 Groundwater "revap" coefficient 2 -0.036 0.036 100.00 -0.036 -0.000518 49.28 Threshold water depth in shallow aquifer for flow 2 -1000 1000 100.00 -1000 -110.13 44.49 RGSR Threshold water depth in shallow aquifer for "revap" 2 -100 100 100.00 -39.632 100 69.82 Baseflow alpha coefficient (days) 1 0.0037725 0.013649 0.99 0.0043699 0.013839 0.95 Soil evaporation compensation factor 1 0 0.31192 31.19 0 0.42711 42.71 Groundwater delay time (days) 2 -4.9694 10 74.85 -10 10 100.00 Groundwater "revap" coefficient 2 -0.036 0.02832 89.33 -0.036 0.036 100.00 Threshold water depth in shallow aquifer for flow 2 -1000 333.28 66.66 -1000 93.768 54.69 RV Threshold water depth in shallow aquifer for "revap" 2 -100 100 100.00 -100 100 100.00

PAGE 127

127Table 4-5. Output SWAT parameter ranges and percentages of valu e range from SUNGLASSES uncert ainty analysis routine. Subbasins are coded where RGSR is the Rio Grande de San Ramon and RV is the Rio Virilla. Parameter Variation method Minimum value (1975) Maximum value (1975) Percentage of range (1975) Minimum value (1986) Maximum value (1986) Percentage of range (1986) Baseflow alpha coefficient (days) 1 0 0 0.00 0 0 0.00 Soil evaporation compensation factor 1 0 0.40532 40.53 0 0.35506 35.51 Groundwater delay time (days) 2 -10 10 100.00 -10 10 100.00 Groundwater "revap" coefficient 2 -0.036 0.036 100.00 -0.036 0.036 100.00 Threshold water depth in shallow aquifer for flow 2 701.18 1000 14.94 -1000 1000 100.00 RGSR Threshold water depth in shallow aquifer for "revap" 2 -100 100 100.00 -100 100 100.00 Baseflow alpha coefficient (days) 1 0.0028127 0.0073213 0.45 0 0.038523 3.85 Soil evaporation compensation factor 1 0 0.032286 3.23 0 0.93862 93.86 Groundwater delay time (days) 2 5.728 10 21.36 -10 10 100.00 Groundwater "revap" coefficient 2 -0.036 0.036 100.00 -0.036 0.036 100.00 Threshold water depth in shallow aquifer for flow 2 -1000 354.51 67.73 -1000 1000 100.00 RV Threshold water depth in shallow aquifer for "revap" 2 -100 100 100.00 -100 100 100.00

PAGE 128

128Table 4-6. Measures of fit for the PARASOL automatic calibration routine for each subbasin and land cover year Results using Xi squared statistic Rio Grande de San Ramon (1975) Rio Grande de San Ramon (1986) Rio Virilla (1975) Rio Virilla (1986) Minimum global objective function 410.0 877.5 1278.5 1096.0 Total number of runs 787 1001 1003 1000 Minimum objective functions 1826300 514320 2259100 1480300 Maximum objective functions 4262900 1046900 4322900 2860100 Minimum output values 33.944 27.856 32.485 25.749 Maximum output values 49.560 43.169 40.745 33.688 90% probability uncertainty analysis Rio Grande de San Ramon (1975) Rio Grande de San Ramon (1986) Rio Virilla (1975) Rio Virilla (1986) Total number of observations 820.0 1755.0 2557.0 2192.0 Number of free parameters 6 6 6 6 Limit on global objective function 415.361732 882.840759 1283.83502 1101.33711 Number of selections 540 171 504 472 Minimum objective functions 1826300 514320 2259100 1480300 Maximum objective functions 1868200 522470 2290700 1502300 Minimum output values 33.944 39.837 37.540 31.092 Maximum output values 36.366 40.630 39.135 32.257 Percentage of range 15.510 5.179 19.310 14.674

PAGE 129

129Table 4-7. Measures of fit for the SUNGLASSES uncertain ty analysis for each subbasin and land cover year Results using Xi squared statistic Rio Grande de San Ramon (1975) Rio Grande de San Ramon (1986) Rio Virilla (1975) Rio Virilla (1986) Minimum global objective function 2338.52368 1323.01172 2921.99438 2191.5 Total number of runs 710 703 2002 2000 Minimum objective functions 1928200 4792200 2875900 4885200 Maximum objective functions 7560100 6646900 7371500 8765300 Minimum output values 36.622 39 35.129 35.078 Maximum output values 43.803 48.439 39.305 40.413 90% probability uncertainty analysis Rio Grande de San Ramon (1975) Rio Grande de San Ramon (1986) Rio Virilla (1975) Rio Virilla (1986) Total number of observations 4677 2646 5844 4383 Number of free parameters 6 6 6 6 Limit on global objective function 2343.85307 1328.34636 2927.32234 2416.9 Number of selections 349 400 332 1720 Minimum objective functions 1928200 4792200 2875900 4885200 Maximum objective functions 1949400 4859400 2909900 5427100 Minimum output values 36.622 39 37.04 35.078 Maximum output values 36.990 40.475 37.812 39.808 Percentage of range 5.12463 15.62666 18.48659 88.65979

PAGE 130

130 Table 4-8. Percentages of land cover and land cover change per sub-basin Sub-basin Land cover class 1975 Area (%) 1986 Area (%) Change from 1975 to 1986 Agricultural Land-Row Crops 4.6 1.1 -3.5 Forest-Deciduous 10.1 7.0 -3.1 Pasture 14.2 27.3 13.1 Range-Brush 19.6 13.2 -6.4 Rio Grande de San Ramon Residential-Med/Low Density 0.9 0.9 0.0 Agricultural Land-Row Crops 4.2 1.4 -2.9 Forest-Deciduous 10.5 8.1 -2.4 Pasture 15.4 24.1 8.7 Range-Brush 14.3 10.1 -4.1 Residential-Med/Low Density 6.2 6.9 0.8 Rio Virilla Totals 100.0 100.0

PAGE 131

131 CHAPTER 5 CONCLUSIONS I present interdisciplinary pers pectives on spatial-temporal methods for land cover change from three interrelated sub-disciplines of physical geography landscape ecology, remote sensing, and hillslope hydrology. By linking a nd analyzing space and time together, these methods can improve the understanding of the comp lexity and dynamics of land cover change in various geographic locations (Rindf uss et al. 2004). In this way, these methods can contribute to detecting and understanding linkages between patte rns and processes of land cover change. Land change science (Gutman 2004), though currently in its infancy of development, could use generally accepted, mechanistic descriptions of the biophysical processes associated with land cover change as it develops into a science. The challenge lies in the va st number of ecosystem processes, each with distinct disciplinary representations, occurring at a range of scales. Significance of Findings The first study (Chapter 2) applies methods from remote sensing and GIScience with guidance of pattern-process lin kages from landscape ecology. To better understand patternprocess interdependencies, we need to determ ine important scales of landscape heterogeneity, human activities, and ecosystem pr ocesses that most effectively explain heterogeneous spatial patterns (Southworth et al. 2006). The developmen t of new methods, speci fically of multiscalar analysis as applied to observed patterns captured by remotely sens ed data, can yield appropriate temporal and spatial scale domains important for the applied study of linking landscape and ecological patterns and processes. The incorporation and scaling of spatial, temporal and spectral information into land cover change analyses greatly improves the amount of information obtained.

PAGE 132

132 The second study (Chapter 3) dr aws on classic remote sensin g change detection through trajectory analysis coupled with landscape leve l fragmentation analysis which is common in landscape ecology. This approach, combined with remote sensi ng, categorizes land cover types and identifies land cover changes by linking temporal rates and spatial patterns of these changes with patch-level fragmentation indices at mu ltiple scales. In this study, these changes are conditioned upon distance from popu lation centers and a major road corridor. The trajectory analysis provides a baseline for tracking change s over time of an indi vidual landscape patch between two dates, which with intermediate time steps can help identify the timings of important pattern forming processe s (Crews-Meyer 2006). The combined simulation modeli ng and statistical analysis conducted in the third study (Chapter 4) highlights the complexity of th e climate and land cover change on hillslope hydrology and investigates aspects of spatial and temporal heterogeneity of a complex watershed. The SWAT hydrologic m odel is process driven, repres enting the major components in watershed hydrology. The continuous simula tion capabilities under different land cover conditions enable the observation of the imp acts on the surface hydrolog ic cycle due to land cover change. Given the difficulties of empirical monitoring for land cover change in large, regional scale (> 1000km2) watersheds (Bonell 1999; Bruijnzeel 2004), watershed models do permit investigations into the sp atial and temporal complexities of impacted hillslope hydrology. The identification of disciplinary strengths from hillslope hydrology, remote sensing, and landscape ecology can lead to improved methods of spatial-temporal analyses of land cover change, and could improve the understanding of the complexity and potential outcomes on biophysical systems resulting from land cover change.

PAGE 133

133 Identification of Disciplinary Strengths Landscape ecology leads the way in spatial-tempor al analysis methods (Turner et al. 1989) with many approaches for detecting landscape patte rns (Dale et al. 2002) and does well at linking ecological pattern and process (Turner 1989). Remo te sensing and field mapping provide the raw materials for discerning pertinent ecological pa tterns (Aplin 2005). Curre nt challenges include observing long-term temporal chan ges, and linking satellite images and empirical field data due to scale and location dependence. Complexities of scale (Levin 1992; Turner et al. 1989) are dealt with practical implementation in the cons truction of sampling designs (Bellehumeur and Legendre 1998). With a strong foundation in the ecol ogical tradition of empi rical data collection, models analyze and predict obs erved patterns (Levin 1992) with a primary focus on how those linkages change with scale (Levin 1992; Wiens 1989). Compared to ecology and landscape ecology, remo te sensing is data and method driven and an application based scien ce (Aplin 2005). However, remote sensing has provided valuable information about the surface of the Earth (E.F .Lambin et al. 1999; Turner et al. 1995) with greater understanding of land c over patterns and changes through time (Lambin and Geist 2006), has amassed an incredible amount of data abou t the Earths surface from many diverse platforms for quite an extensive period of time (Woodcoc k and Ozdogan 2004), and has been integrated easily into many disciplines. Remotely sensed data provide unbiased observ ations of land covers, and, in this way, comprise primary data sources in many types of inte rdisciplinary simulation models, especially those with an emphasis on la nd cover change. Since image analysis focuses on the detection of landscape patterns, it is reas onable that remote sensing researchers have adopted principles from landscape ecology (Crews -Meyer 2006) to analyze patterns of land cover change. Remote sensing, tied in with both landscape ecology and geography, has made

PAGE 134

134 much progress in the past decade with developing spatial-temporal anal yses (Petit and Lambin 2001; Mertens and Lambin 2000; Southworth et al. 2004; Southwor th et al. 2006). Hillslope hydrology has a lengthy history of collecting high-resolution temporal data (Bonell 1999) and uses process-sp ecific simulation models (Beven et al. 1980; Beven et al. 1984; Beven et al. 1987) able to simulate str eamflow production. In the 1960s and 1970s hydrology made great advances in understa nding the spatial dist ributions of runoff ge neration on a hillslope (Dunne 1978) and throughout a watershed (Freez e 1972; Hewlett and Hibbert 1967). Since then, especially after the developmen t of Freeze and Harlans blueprint for a hydrologic simulation model (Freeze and Harlan 1969) more attention ha s been directed to the development of better and more detailed hydrologic models and to da ta collection at higher spatial and temporal resolutions (Schulz et al. 2006). Over the past decade, in response to the popularity and availability of GIS and remote sensing technologies, an intere st has emerged in the spatial configurations of hydrologic patterns and thei r processes of formation (Grayson et al. 2002). Simulation and field hydrology has advanced primarily due to emerging new technologies capable of generating high reso lution spatial and temporal da ta (McDonnell et al. 2007). Challenges arise from the lack of effective empirical monitoring of key hydrologic variables linked to dominant watershed processes at multiple scales. Much of the existing hydrologic knowledge stems from intensive field work in sma ll, experimental watersheds, the increased use of remote sensing, simulation modeling, and sp atial analyses can advance hydrology (Grayson and Bloschl 2001) by linking hydrologi c pattern and process, partic ularly at regional scales, which would compliment current land cover change research agendas. Recommendations Conceptually speaking, landscape ecology has do ne well with analyzing spatial patterns with numerous methods, linking pattern and proce ss, treatment of scale, and providing guiding

PAGE 135

135 concepts and theories. Remote sensing offers 30 plus years of multi-scalar and multiple platform data for natural, seamless integration into many disciplines. Individual di sciplinary studies that incorporate remotely sensed data use hypothe sis testing based on comparisons between the imagery and the empirical data. Hydrology offers a history of process identification and the modeling of key watershed processes in a compartmentalized fashion by grouping core watershed processes that generate observable patte rns. The goal here is to combine the strengths of each discipline to form intedisciplinary tena nts for spatial-temporal analyses for land cover change. The primary question is this: What can each discipline learn from each other? The following are recommendations for the incorpora tion of current research directions for the development of robust methods to tackle the complexities of land c over change research. Landscape ecology should identify and comp artmentalize key ecological process to continue pattern-process linkages across multiple scales. Remote sensing should continue to develop data products and refine analysis methods directed towards hydrology and ot her subsurface disciplines. Spatial-temporal analyses of land cover and habitat change in bot h landscape ecology and remote sensing should use trajectory, or pa nel analysis methods to account for changes over time. Land cover change research us ing hydrologic models should in corporate spatial land cover data, either from remotely sensed imagery or existing spatial data. Hillslope hydrology needs to incorporate spatial structure, scale, dominant processes, connectivity and critical thresh olds to help identify key watershed processes that are impacted by land cover change. Develop a classification system for catalo ging of watershed patterns and hydrologic parameters controlled or produced by land cover change and develop hypotheses incorporating these field data testable by simulation models run at multiple scales. A way forward for land cover change research is to inventory current research agendas and draw methodological strengths from complimentar y disciplines. For example, the incorporation of remote sensing and field data, spatial anal yses and simulation modeling for hydrologic studies

PAGE 136

136 investigating land cover change at multiple scales to link pattern and process would yield an understanding of how changes in land cover within the watershed impact runoff and the resulting hydrologic responses. Interdisciplin ary research, like land cover cha nge, would be best served by complimentary methods of analysis and research agendas, and the incorp oration and combination of theoretical concepts to be tter identify and understand the co mplex processes and outcomes of global change in human-environment systems. Th is research makes a good start along this approach and can be built upon in my future rese arch and by other researchers who choose to use similar, interdisciplinary methods.

PAGE 137

137 APPENDIX A LANDSAT IMAGES USED IN LAND COVE R CHANGE ANALYSIS FOR PANDO, BOLIVIA Table A-1 Landsat image platform, path, row, and acquisition date information. Platform Path Row Year Month Day TM 1 67 1986 7 30 TM 1 68 1986 7 30 TM 2 67 1986 8 6 TM 2 68 1986 8 6 TM 2 69 1986 9 7 TM 3 67 1986 7 28 TM 3 68 1986 7 12 TM 3 69 1986 7 12 TM 1 67 1991 7 28 TM 1 68 1991 7 28 TM 2 67 1991 7 27 TM 2 68 1991 7 27 TM 2 69 1991 7 27 TM 3 68 1991 10 14 TM 3 69 1991 10 14 TM 3 67 1992 6 18 TM 1 67 1996 7 25 TM 2 67 1996 8 1 TM 2 68 1996 7 16 TM 2 69 1996 8 17 TM 3 67 1996 7 23 TM 3 68 1996 7 23 TM 3 69 1996 7 23 TM 1 68 1997 9 14 ETM+ 2 67 1999 8 2 ETM+ 1 67 2000 7 28 ETM+ 1 68 2000 8 13 TM 2 67 2000 7 27 ETM+ 2 68 2000 11 24 TM 2 69 2000 7 27 ETM+ 3 67 2000 5 23 ETM+ 3 68 2000 7 26 ETM+ 3 69 2000 5 23 TM 1 67 2005 6 16 TM 1 68 2005 8 3 TM 2 67 2005 8 10 TM 2 68 2005 6 7 TM 2 69 2005 8 10 TM 3 67 2005 6 30 TM 3 68 2005 9 18 TM 3 69 2005 9 18

PAGE 138

138 APPENDIX B ASTER IMAGES USED IN LAND COVER CHANGE ANALYSIS FOR PANDO, BOLIVIA

PAGE 139

139Table B-1. ASTER images used for classifi cation and trajectory accuracy assessment. GLOVIS L1A ID Acquistion Date Sun Azimuth Sun Elevation Scene Center Latitude Scene Center Longitude VNIR Look angle TIR Look angle SWIR Look angle Cloud Condition SC:AST_L1A.003:2003396555 6/29/2001 33.368 48.730 -11.209794 -67.211437 0.000 0.000 0.000 None SC:AST_L1A.003:2003396560 6/29/2001 33.097 48.421 -11.744631 -67.328358 0.000 0.000 0.000 None SC:AST_L1A.003:2003642025 7/29/2001 38.642 52.641 -10.738689 -69.746039 0.000 0.000 0.000 None SC:AST_L1A.003:2003642028 7/29/2001 38.286 52.109 -11.273812 -69.861367 0.000 0.000 0.000 None SC:AST_L1A.003:2003768861 8/7/2001 41.235 55.309 -9.710410 -67.680141 0.000 0.000 0.000 None SC:AST_L1A.003:2003768862 8/7/2001 40.859 55.032 -10.245577 -67.794400 0.000 0.000 0.000 None SC:AST_L1A.003:2003768864 8/7/2001 40.601 54.570 -10.780825 -67.908885 0.000 0.000 0.000 None SC:AST_L1A.003:2003768865 8/7/2001 40.349 54.105 -11.315893 -68.023553 0.000 0.000 0.000 None SC:AST_L1A.003:2004493410 9/17/2001 59.158 64.988 -11.242403 -66.987314 0.000 0.000 0.000 None SC:AST_L1A.003:2004493415 9/17/2001 58.356 64.728 -11.777346 -67.103513 0.000 0.000 0.000 None SC:AST_L1A.003:2009936839 8/20/2000 42.400 58.602 -11.272874 -68.322753 0.000 0.000 0.000 Some SC:AST_L1A.003:2010879893 9/21/2000 60.067 68.409 -10.686346 -68.562147 0.000 0.000 0.000 Some SC:AST_L1A.003:2010879953 9/21/2000 59.131 68.147 -11.221333 -68.678395 0.000 0.000 0.000 Few SC:AST_L1A.003:2010879955 9/21/2000 58.222 67.712 -11.756269 -68.794933 0.000 0.000 0.000 Few SC:AST_L1A.003:2010879956 9/21/2000 57.338 67.440 -12.291173 -68.911780 0.000 0.000 0.000 Few SC:AST_L1A.003:2010879958 9/21/2000 56.480 67.164 -12.825864 -69.028914 0.000 0.000 0.000 Few SC:AST_L1A.003:2018642412 9/5/2000 50.055 64.982 -9.750218 -67.401382 8.588 8.567 8.000 Few SC:AST_L1A.003:2018642447 9/5/2000 48.023 63.859 -11.355908 -67.743708 8.588 8.567 8.000 None

PAGE 140

140Table B-1. Continued GLOVIS L1A ID Acquistion Date Sun Azimuth Sun Elevation Scene Center Latitude Scene Center Longitude VNIR Look angle TIR Look angle SWIR Look angle Cloud Condition SC:AST_L1A.003:2018642453 9/5/2000 48.671 64.368 -10.820633 -67.629394 8.588 8.567 8.000 None SC:AST_L1A.003:2018642458 9/5/2000 49.352 64.676 -10.285447 -67.515305 8.588 8.567 8.000 None SC:AST_L1A.003:2018683275 9/12/2000 52.365 66.298 -10.821039 -69.172012 8.578 8.567 8.031 Some SC:AST_L1A.003:2018683283 9/12/2000 50.041 65.165 -12.426434 -69.515532 8.578 8.567 8.031 Some SC:AST_L1A.003:2018683286 9/12/2000 50.758 65.468 -11.891353 -69.400821 8.578 8.567 8.031 Some SC:AST_L1A.003:2018683287 9/12/2000 51.496 65.768 -11.356222 -69.286320 8.578 8.567 8.031 Some SC:AST_L1A.003:2018683295 9/12/2000 49.358 64.666 -12.961587 -69.630494 8.578 8.567 8.031 Some SC:AST_L1A.003:2029535916 6/8/2005 34.646 48.435 -11.208756 -67.219635 -2.829 -2.853 -2.824 None SC:AST_L1A.003:2029535919 6/8/2005 34.366 47.834 -11.743515 -67.337063 -2.829 -2.853 -2.824 None SC:AST_L1A.003:2030305185 8/2/2005 40.864 50.068 -12.846886 -68.883442 -0.025 0.004 -0.006 None SC:AST_L1A.003:2030653656 8/27/2005 52.149 57.363 -10.569448 -67.813325 -8.583 -8.558 -8.575 Some SC:AST_L1A.003:2030653658 8/27/2005 51.595 56.850 -11.103915 -67.931617 -8.583 -8.558 -8.575 Few SC:AST_L1A.003:2030653669 8/27/2005 51.063 56.573 -11.638459 -68.050323 -8.583 -8.558 -8.575 Few SC:AST_L1A.003:2030807072 9/3/2005 55.504 60.762 -9.643626 -68.142435 0.016 0.005 -0.082 Haze SC:AST_L1A.003:2030807073 9/3/2005 54.493 60.128 -10.178508 -68.258275 0.016 0.005 -0.082 Haze SC:AST_L1A.003:2030807075 9/3/2005 53.861 59.856 -10.713612 -68.374434 0.016 0.005 -0.082 None SC:AST_L1A.003:2030807077 9/3/2005 53.246 59.356 -11.248532 -68.490856 0.016 0.005 -0.082 None SC:AST_L1A.003:2030807078 9/3/2005 52.643 59.080 -11.783402 -68.607585 0.016 0.005 -0.082 None

PAGE 141

141Table B-1. Continued GLOVIS L1A ID Acquistion Date Sun Azimuth Sun Elevation Scene Center Latitude Scene Center Longitude VNIR Look angle TIR Look angle SWIR Look angle Cloud Condition SC:AST_L1A.003:2030807088 9/3/2005 52.054 58.802 -12.318226 -68.724621 0.016 0.005 -0.082 None SC:AST_L1A.003:2034741183 6/25/2006 35.470 47.977 -10.708547 -69.953730 -0.017 0.004 -0.033 Few SC:AST_L1A.003:2034741185 6/25/2006 35.181 47.679 -11.243494 -70.070299 -0.017 0.004 -0.033 Few SC:AST_L1A.003:2034876294 7/4/2006 37.495 47.961 -10.565638 -69.383491 -8.578 -8.558 -8.509 None SC:AST_L1A.003:2034876305 7/4/2006 37.189 47.370 -11.100249 -69.501921 -8.578 -8.558 -8.509 None SC:AST_L1A.003:2034876310 7/4/2006 36.889 46.777 -11.634780 -69.620733 -8.578 -8.558 -8.509 None SC:AST_L1A.003:2035845602 8/5/2006 43.593 52.296 -10.613081 -69.062602 -5.724 -5.700 -5.674 None SC:AST_L1A.003:2035845603 8/5/2006 43.189 51.731 -11.147795 -69.180395 -5.724 -5.700 -5.674 Few SC:AST_L1A.003:2035845656 8/5/2006 42.794 51.162 -11.682440 -69.298541 -5.724 -5.700 -5.674 None SC:AST_L1A.003:2035845657 8/5/2006 42.417 50.871 -12.217016 -69.417054 -5.724 -5.700 -5.674 None SC:AST_L1A.003:2035845659 8/5/2006 42.041 50.298 -12.751515 -69.535951 -5.724 -5.700 -5.674 None SC:AST_L1A.003:2036830079 9/13/2006 57.381 63.627 -10.852487 -68.949709 8.580 8.567 8.492 Yes SC:AST_L1A.003:2036830096 9/13/2006 55.903 62.886 -11.922671 -69.179032 8.580 8.567 8.492 Few SC:AST_L1A.003:2036830098 9/13/2006 56.626 63.362 -11.387604 -69.064272 8.580 8.567 8.492 Yes SC:AST_L1A.003:2036830102 9/13/2006 55.184 62.614 -12.457686 -69.294003 8.580 8.567 8.492 Few SC:AST_L1A.003:2036896039 9/15/2006 59.572 63.142 -11.248567 -66.945133 -0.022 0.004 -0.011 Few SC:AST_L1A.003:2036896072 9/15/2006 58.817 62.888 -11.783425 -67.061923 -0.022 0.004 -0.011 Few

PAGE 142

142 APPENDIX C TEST STATISTICS FOR FRAGMENTATION METRICS AT THE PANDO EXTENT Table C-1. Kruskall-Wallis statis tics for Mean Patch Size and Pe rimeter-Area Corrected metrics. Under the Bonferroni Test, to adjust fo r multiple tests, medians are significantly different if the z-value > 2.8070 as indicated in bold. Under the regular Z test medians are significantly different if the z-va lue > 1.9600. Metric types are defined where MPS is the mean patch size and PAC is the perimeter-area ratiocorrected. Cover and metric Year 1986 1991 1996 2000 2005 Forest MPS 1986 0.000 1991 1.270 0.000 1996 1.292 0.023 0.000 2000 0.721 0.549 0.572 0.000 2005 0.049 1.319 1.342 0.770 0.000 Non-forest MPS 1986 0.000 1991 0.767 0.000 1996 3.647 4.414 0.000 2000 2.637 1.870 6.284 0.000 2005 1.887 2.654 1.760 4.524 0.000 Forest PAC 1986 0.000 1991 1.447 0.000 1996 1.631 0.184 0.000 2000 5.521 4.074 3.890 0.000 2005 0.422 1.025 1.209 5.099 0.000 Non-forest PAC 1986 0.000 1991 0.860 0.000 1996 3.012 3.872 0.000 2000 6.621 5.761 9.633 0.000 2005 1.114 1.975 1.898 7.735 0.000

PAGE 143

143 LIST OF REFERENCES Alencar, A.A.C., L.A. Solorzano, and D.C. Neps tad. 2004. Modeling forest understory fires in an eastern Amazonian landscape. Ecolog ical Applications 14(4):S139-S149. Alvarado, A. and W. Forsythe. 2005. Densidad aparente en suelos de Costa Rica. Agronoma Costarricense 29(1):85-94. Anderson, E.P., C.M. Pringle, and M. Ro jas. 2006. Transforming tropical rivers: an environmental perspective on hydropow er development in Costa Rica. Aquatic Conservation-Marine and Freshwater Ecosystems 16(7):679-693. Aplin, P. 2005. Remote sensing: ecology. Progress in Physical Geography 29(1):104-113. Arima, E.Y., R.T. Walker, S.G. Perz, and M. Caldas. 2005. Loggers and forest fragmentation: Behavioral models of road building in the Amazon basin. Annals of the Association of American Geographers 95(3):525-541. Armenteras, D., G. Rudas, N. Rodriguez, S. Sua, and M. Romero. 2006. Patterns and causes of deforestation in the Colombian Amazon. Ecological Indicators 6(2):353-368. Arnold, J.G. and N. Fohrer. 2005. SWAT2000: current capabilities and resear ch opportunities in applied watershed modelling. Hydrological Processes 19(3):563-572. Arnold, J.G., R. Srinivasan, R.S. Muttiah, and J.R. Williams. 1998. Large area hydrologic modeling and assessment Part 1: Model development. Journal of the American Water Resources Association 34(1):73-89. Asner, G.P., D.E. Knapp, E.N. Broadbent, P.J. C. Oliveira, M. Keller, and J.N. Silva. 2005. Selective logging in the Brazilian Amazon. Science 310(5747):480-482. Ault, T.R. and C.R. Johnson. 1998. Spatially and temporally predictabl e fish communities on coral reefs. Ecological Monographs 68(1):25-50. Bedient, P.B. and W.C. Huber. 1988. Hydrology and floodplain analysis. Addison-Wesley: Reading, Mass. Bellehumeur, C. and P. Legendre. 1998. Multiscale sources of variation in ecological variables: modeling spatial dispersion, el aborating sampling designs. Landscape Ecology 13(1):15-25. Beven, K., A. Calver, and E.M. Morris. 1987. Th e Institute of Hydrol ogy Distributed Model. Report UK Institute of Hydrology 98 (1987) p33p:33p. Beven, K., R. Warren, J. Zaoui, ) Danish Hydr aulic Institute, and G. SOGREAH. 1980. SHE: towards a methodology for physically-based distributed forecasting in hydrology. In Hydrological forecasting, 133-7.

PAGE 144

144 Beven, K.J., M.J. Kirkby, N. Schofield, and A.F. Tagg. 1984. Testing a physically-based flood forecasting model ( TOPMODEL) for three UK catchments ( Crimple Beck, Hodge Beck, Wye). Journal of Hydrology 69(1-4):119-143. Blomquist, W.A., 1957-, A. William, and B. Wo rld. 2005. Institutional and policy analysis of river basin management the tarc oles river basin, Costa Rica /. Bonell, M. 1998. Possible impacts of climate vari ability and change on tropical forest hydrology. Climatic Change 39(2-3):215-272. Bonell, M. 1999. Tropical forest hydrology a nd the role of the UNESCO International Hydrological Programme. Hydrology and Earth System Sciences 3(4):451-461. Bonell, M. and J. Balek. 1993. Recent scientif ic developments and research needs in hydrological processes of the humid tropics. In Hydrology and water management in the humid tropics, eds. M.Bonell, M.M.Hufschmidt, and J.S.Gladwell. 167-260. Cambridge University Press. Borcard, D. and P. Legendre. 2002. All-scale spa tial analysis of ecological data by means of principal coordinates of neighbour matrices. Ecological Modelling 153(1-2):51-68. Bosch, J.M. and J.D. Hewlett. 1982. A Review of Catchment Experiments to Determine the Effect of Vegetation Changes on Wa ter Yield and Evapo-Transpiration. Journal of Hydrology 55(1-4):3-23. Boyd, D.S. and F.M. Danson. 2005. Satellite remote se nsing of forest resour ces: three decades of research development. Progress in Physical Geography 29(1):1-26. Boyd, D.S. and F. Petitcolin. 2004. Remote sensing of the terrestrial environment using middle infrared radiation (3.0-5.0 mu m). International Journal of Remote Sensing 25(17):33433368. Breiman, L. 1984. Classification and regressi on trees /. Belmont, Calif. : Wadsworth International Group, Wadswo rth International Group. Brown, D.G., B.C. Pijanowski, and J.D. Duh. 20 00. Modeling the relationships between land use and land cover on private lands in the Upper Midwest, USA. Journal of Environmental Management 59(4):247-263. Bruijnzeel, L.A. 1990. Hydrology of moist tropical forests a nd effects of conve rsion: a state of knowledge review. Netherlands IHP Committee: Netherlands IHP Committee. Bruijnzeel, L.A. 2004. Hydrological functions of tropical forests: not seeing the soil for the trees? Agriculture Ecosystems & Environment 104(1):185-228. Butler, D.R., G.P. Malanson, and L.M. Resler. 2 004. Turf-banked terrace treads and risers, turf exfoliation and possible relationshi ps with advancing treeline. Catena 58(3):259-274.

PAGE 145

145 Centro Cientfico Tropical (CCT). Tipo de Su elos en Costa Rica. Scale 1:200,000. 1989. Costa Rica. Chan, J.C.W., N. Laporte, and R.S. Defries. 2003. Texture classification of logged forests in tropical Africa using mach ine-learning algorithms. International Journal of Remote Sensing 24(6):1401-1407. Chen, K. 2002. An approach to linking remote ly sensed data and areal census data. International Journal of Remote Sensing 23(1):37-48. Chomitz, K.M. and D.A. Gray. 1996. Roads, land us e, and deforestation: A spatial model applied to belize. World Bank Economic Review 10(3):487-512. Cihlar, J. 2000. Land cover mapping of large areas fr om satellites: status and research priorities. International Journal of Remote Sensing 21(6-7):1093-1114. Cohen, W.B. and C.O. Justice. 1999. Validati ng MODIS terrestrial eco logy products: Linking in situ and satellite measurements. Remote Sensing of Environment 70(1):1-3. Colby, J.D. 2001. Simulation of a Costa Rican wa tershed: Resolution e ffects and fractals. Journal of Water Resources Planning and Management-Asce 127(4):261-270. Coops, N. and D. Culvenor. 2000. Utilizing local variance of simulated high spatial resolution imagery to predict spatial pattern of forest stands. Remote Sensing of Environment 71(3):248260. Coops, N.C. and P.C. Catling. 1997. Utilising air borne multispectral videography to predict habitat complexity in eucalypt forests for wildlife management. Wildlife Research 24(6):691702. Costa, M.H., A. Botta, and J.A. Cardille. 2003. E ffects of large-scale changes in land cover on the discharge of the Tocantins River, Southeastern Amazonia. Journal of Hydrology 283(14):206-217. Coxhead, I., A. Rola, and K. Kim. 2001. How do national markets and pri ce policies affect land use at the forest margin? Evidence from the Philippines. Land Economics 77:250-267. Crews-Meyer, K. 2006. Temporal Extensions of Landscape Ecology Theory and Practice: Examples from the Peruvian Amazon. Professional Geographer 58(4):421-435. CUMAT. Desbosque de la Amazonia Bolivia na. 1992. La Paz, Bolivia, Centro de Investigaciones de la Capacidad de Uso Mayor de la Tierra. Cumming, G.S. 2003. Measures of "eroading" ecology. Frontiers in Ecology and the Environment 1(5):233. Currie, J. 1981. The Economic Theory of Agricultural Land Tenure. Cambridge University Press: Cambridge.

PAGE 146

146 Dale, M.R.T., P. Dixon, M.J. Fortin, P. Lege ndre, D.E. Myers, and M.S. Rosenberg. 2002. Conceptual and mathematical relationshi ps among methods for spatial analysis. Ecography 25(5):558-577. Dingman, S.L. 2002. Physical hydrology. Prentice Hall, Upper Saddle River, NJ. Droogers, P. and G. Kite. 2002. Remotely sensed data used for modelling at different hydrological scales. Hydrological Processes 16(8):1543-1556. Duncan, B.W., S. Bolye, D.R. Breininger, and P.A. Schmalzer. 1999. Coupling past management practice and historic landscape change on John F. Kennedy Space Center, Florida. Landscape Ecology 14:291-309. Dunne, T. 1978. Field studies of hillslope processes. In Hillslope Hydrology, ed. M.J.Kirkby. 227-94. John Wiley and Sons: Chichester. Dykes, A.P. and J.B. Thornes. 2000. Hillslope hydrology in tropical rainforest steeplands in Brunei. Hydrological Processes 14(2):215-235. Elsenbeer, H. 2001. Hydrologic fl owpaths in tropical rainfore st soilscapes a review. Hydrological Processes 15(10):1751-1759. Elsenbeer, H. and D.K. Cassel. 1991. The mechan isms of overland flow generation in a small catchment in wetsern Amazonia. In Water Management of the Amazon Basin Symposium: Proceedings of Manaus Symposium August 1990, eds. B.P.F.Braga and C.A.FernandezJauregui. 275-88. UNESCO: Montevideo, Uruguay. Elsenbeer, H., K. Cassel, and J. Castro. 1992. Spa tial-Analysis of Soil H ydraulic Conductivity in A Tropical Rain-Forest Catchment. Water Resources Research 28(12):3201-3214. Elsenbeer, H., B.E. Newton, T. Dunne, and J.M. de Moraes 1999. Soil hydraulic conductivities of latosols under pasture, forest and teak in Rondonia, Brazil. Hydrological Processes 13(9):1417-1422. Etter, A., C. McAlpine, K. Wilson, S. Phi nn, and H. Possingham. 2006. Regional patterns of agricultural land use and deforestation in Colombia. Agriculture Ecosystems & Environment 114(2-4):369-386. Evans, T.P., G.M. Green, and L.A. Carlson. 2001. Multi-scale analysis of landcover composition and landscape management of public and private lands in Indiana. In GIS and Remote Sensing Applications in Biogeography and Ecology, eds. A.C.Millington, S.J.Walsh, and P.E.Osborne. 271-87. Kluwer Academic Publications: Boston, MA. Fearnside, P.M. 2002. Avanca Brasil: Environmen tal and social consequences of Brazil's planned infrastructure in Amazonia. Environmental Management 30(6):735-747.

PAGE 147

147 Foley, J.A., G.P. Asner, M.H. Costa, M.T. Coe, R. Defries, H.K. Gibbs, E.A. Howard, S. Olson, J. Patz, N. Ramankutty, and P. Snyder. 2007. Amazonia revealed: forest degradation and loss of ecosystem goods and services in the Amazon Basin. Frontiers in Ecology and the Environment 5(1):25-32. Freeze, R.A. 1972. Role of Subsurface Flow in Generating Surface Runoff .2. Upstream Source Areas. Water Resources Research 8(5):1272-&. Freeze, R.A. and R.L. Harlan. 1969. Blueprint for a physically-based, digitally-simulated hydrologic response model. Journal of Hydrology 9(3):237-258. Gehlke, C.E. and K. Biel. 1934. Certain eff ects of grouping upon the size of the correlation coefficient in census tract material. Journal of the American Statistical Association 29(185A):169-170. Giambelluca, T.W. 2002. Hydrology of altered tropical forest. Hydrological Processes 16(8):1665-1669. Goodchild, M.E. 2004. GIScience, geography, form, and process. Annals of the Association of American Geographers 94(4):709-714. Grayson, R.B., G. Bloschl, A.W. Western, a nd T.A. McMahon. 2002. Advances in the use of observed spatial patterns of cat chment hydrological response. Advances in Water Resources 25(8-12):1313-1334. Green, C.H. and A. van Griensven. 2008. Au tocalibration in hydrologic modeling: Using SWAT2005 in small-scale watersheds. Environmental Modelling & Software 23:422-434. Green, G., C. Schweik, M. Hanson, L. Weber, C. Agarwal, G. Bullman, L. Carlson, M. Reisinger, S. Shields, J. Hanson, A.C. Barro s, and B. Davis. Radiometric Calibration for Landsat 5 TM scenes (Equation 6) Version 1.0. 1999. Bloomi ngton, IN, CIPEC: Center for the Study of Institutions, Population, and E nvronmental Change, Indiana University. Ref Type: Data File Green, G., C. Schweik, M. Hanson, L. Weber, C. Agarwal, G. Bullman, L. Carlson, M. Reisinger, S. Shields, J. Hanson, A.C. Barro s, and B. Davis. Radiometric Calibration for Landsat 7 ETM+ scene using CPF (Equati on 8) V. 1. 2001. Bloomington, IN, CIPEC: Center for the Study of Institutions, Popul ation, and Envronmental Change, Indiana University. Gustafson, E.J. 1998. Quantifying landscape spatial pattern: What is the state of the art? Ecosystems 1(2):143-156. Gutman, G. 2004. Land change science : observing, monitoring and understanding trajectories of change on the Earth's surface /. Dordrecht: Kluwer Academic Publishers. Habeeb, R.L., J. Trebilco, S. Wotherspoon, a nd C.R. Johnson. 2005. Determining natural scales of ecological systems. Ecological Monographs 75(4):467-487.

PAGE 148

148 Hamilton, D. 2006. Interoceanic highway: The road to riches, the road to ruin. 26-32. www.geographical.co.uk Date last accessed: January 3, 2008. Hastenrath, S. 2002. The Intertropical Convergen ce Zone of the eastern Pacific revisited. International Jour nal of Climatology 22(3):347-356. He, H.S., S.J. Ventura, and D.J. Mladenoff. 2002 Effects of spatial aggregation approaches on classified satellite imagery. International Journal of Geogr aphical Information Science 16(1):93-109. Herbert, M.E. and F.P. Gelwick. 2003. Spatial vari ation of headwater fish assemblages explained by hydrologic variability and ups tream effects of impoundment. Copeia(2):273-284. Hewlett, J.D. and A.R. Hibbert. 1967. Factors affe cting the the response of smalll watersheds to precipitation in humid areas. In Forest Hydrology, eds. W.E.Sopper and H.W.Lull. 275-90. Pergamon Press: Oxford, New York. Hill, M.F., J.D. Witman, and H. Caswell. 2002. Spatio-temporal variation in Markov chain models of subtidal community succession. Ecology Letters 5(5):665-675. Horton, R.E. 1933. The role of inf iltration in the hydrologic cycle. Transactions of the American Geophysical Union 14:446-460. Houghton, R.A. 1994. The Worldwide Extent of Land-Use Change. Bioscience 44(5):305-313. Imbernon, J. and A. Branthomme. 2001. Characteriz ation of landscape patter ns of deforestation in tropical rain forests. International Journal of Remote Sensing 22(9):1753-1765. INPE. Monitoramento da Floresta Amaznica Brasil eira por Satlite. 20 08. Instituto Nacional de Pesquisas Espaciais. Projeto PRODES. Instituto de Pesquisa Espaciais. Isobe, T., E.D. Feigelson, M.G. Akritas, and G.J. Babu. 1990. Linear-Regression in Astronomy .1. Astrophysical Journal 364(1):104-113. Jensen, J.R. 2005. Introductory digital image proce ssing : a remote sensing perspective. Prentice Hall, Upper Saddle River, N.J. Kaimowitz, D., G. Thiele, and P. Pacheco. 1999. The effects of structural adjustment on deforestation and forest degr adation in lowland Bolivia. World Development 27(3):505-520. Kaimowitz, D., A. Angelsen, and Center for International Forestry Research. 1998. Economic models of tropical de forestation : a review. CIFOR, Bogor, Indonesia. Kalnay, E., M. Kanamitsu, R. Kistler, W. Collins, D. Deaven, L. Gandin, M. Iredell, S. Saha, G. White, J. Woollen, Y. Zhu, M. Chelliah, W. Ebisuzaki, W. Higgins, J. Janowiak, K.C. Mo, C. Ropelewski, J. Wang, A. Leetmaa, R. Reynolds, R. Jenne, and D. Joseph. 1996. The NCEP/NCAR 40-year reanalysis project. Bulletin of the Americ an Meteorological Society 77(3):437-471.

PAGE 149

149 Kauth, R.J., B.J.S.C. Lyndon, Eart h Observation Division., and Envi ronmental Research Institute of Michigan. 1976. System for analysis of Lands at agricultural data : automatic computerassisted proportion estimati on of local areas : final re port /. Ann Arbor, Mich. : Environmental Research Institute of Michigan, Environmental Research Institute of Michigan. Killeen, T.J., V. Calderon, L. Soria, B. Quezada, M.K. Steininger, G. Harper, L.A. Solorzano, and C.J. Tucker. 2007. Thirty years of land-cover change in Bolivia. Ambio 36(7):600-606. Krishnaswamy, J., P.N. Halpin, and D.D. Richte r. 2001. Dynamics of sediment discharge in relation to land-use and hydroclimatology in a humid tropical watershed in Costa Rica. Journal of Hydrology 253(1-4):91-109. Lambin, E.F. 1999. Monitoring forest degradation in tropical regions by re mote sensing: some methodological issues. Global Ecology and Biogeography 8(3-4):191-198. Lambin, E.F. 1997. Modelling and monitoring landcover change processes in tropical regions. Progress in Physical Geography 21(3):375-393. Lambin, E.F., H.J. Geist, and E. Lepers. 2003. D ynamics of land-use and land-cover change in tropical regions. Annual Review of Environment and Resources 28:205-241. Lambin, E.F. and A.H. Strahler. 1994. Indicato rs of Land-Cover Change for Change-Vector Analysis in Multitemporal Space at Coarse Spatial Scales. International Journal of Remote Sensing 15(10):2099-2119. Lambin, E.F. and H.J. Geist. 2006. Land-use and land-cover change : local processes and global impacts. Berlin: Springer. Lambin, E.F., X.Baulies, N.Bockstael, G.Fische r, T.Krug, R.Leemans, E.F.Moran, R.R.Rindfuss, Y.Sato, D.Skole, B.L.Turner II, and C. Vogel. 1999. Land-Use and Land-Cover Change (LUCC) : implementation strategy /. Stoc kholm : International Geosphere-Biosphere Programme, International Geos phere-Biosphere Programme. Laurance, W.F., A.K.M. Albernaz P.M. Fearnside, H.L. Vasconcelos, and L.V. Ferreira. 2004. Deforestation in Amazonia. Science 304(5674):1109. Legendre, P. and E.D. Gallagher. 2001. Ecologica lly meaningful transformations for ordination of species data. Oecologia 129(2):271-280. Legendre, P. and J. Legendre. 1998. Numerical ecology. Elsevier, Amsterdam ; New York. Lehner, B., K. Verdin, and A. Jarvis. Hydr oSHEDS Technical Documentation Version 1.0. 2006. Levin, S.A. 1992. The Problem of Pattern and Scale in Ecology. Ecology 73(6):1943-1967.

PAGE 150

150 Loehle, C. and B.L. Li. 1996. Statistical prope rties of ecological and geologic fractals. Ecological Modelling 85(2-3):271-284. Loh, W.Y. and Y.S. Shih. 1997. Split sele ction methods for classification trees. Statistica Sinica 7(4):815-840. Lorup, J.K., J.C. Refsgaard, and D. Mazvimavi. 1998. Assessing the effect of land use change on catchment runoff by combined use of statistic al tests and hydrologi cal modelling: Case studies from Zimbabwe. Journal of Hydrology 205(3-4):147-163. Lu, D., A. Batistella, P. Mausel, and E. Mo ran. 2007. Mapping and monitoring land degradation risks in the Western Brazilian Amazon using multitemporal landsat TM/ETM plus images. Land Degradation & Development 18(1):41-54. Lu, D.S., P. Mausel, M. Batistella, and E. Mo ran. 2004. Comparison of la nd-cover classification methods in the Brazilian Amazon Basin. Photogrammetric Engineering and Remote Sensing 70(6):723-731. Magana, V., J.A. Amador, and S. Medina 1999. The midsummer drought over Mexico and Central America. Journal of Climate 12(6):1577-1588. Malanson, G.P., D.R. Butler, D.M. Cairns, T.E. Welsh, and L.M. Resler. 2002. Variability in an edaphic indicator in alpine tundra. Catena 49(3):203-215. Marceau, D.J. 1999. The scale issue in social and natural sciences. Canadian Journal of Remote Sensing 25(4):347. Marshall, E. and T. Randhir. 2008. Spatial mode ling of land cover change and watershed response using Markovian cellular automata and simulation. Water Resources Research 44(4):W04423. Marsik, M. and P. Waylen. 2006. An application of the distributed hyd rologic model CASC2D to a tropical montane watershed. Journal of Hydrology 330(3-4):481-495. McDonnell, J.J., M. Sivapalan, K. Vache, S. D unn, G. Grant, R. Haggerty, C. Hinz, R. Hooper, J. Kirchner, M.L. Roderick, J. Selker, a nd M. Weiler. 2007. Moving beyond heterogeneity and process complexity: A new vision for watershed hydrology. Water Resources Research 43(7). Mertens, B. and E.F. Lambin. 2000. Land-cover-c hange trajectories in southern Cameroon. Annals of the Associatio n of American Geographers 90(3):467-494. Ministerio de Agricultura y Ganadera (MAG). Us o de la Tierra en Costa Rica en 1992. Landsat image (30m x 30m). 1992. Ministerio de Recursos Naturales, E.y.M. Institu to Meterolgico Nacional Ao del Centenario 1888-1988: Catastrode las Series de Precipitaciones Medida s en Costa Rica. 1-361. 1988. San Jos, Costa Rica, MINAE.

PAGE 151

151 Mladenoff, D.J. and B. DeZonia. APACK 2.14 Us ers Guide. 2000. Madison, WI, Department of Forest Ecology and Management University of Wisconsin. Moody, A. and C.E. Woodcock. 1995. The influence of scale and the spatia l characteristics of landscapes on land-cover ma pping using remote sensing. Landscape Ecology 10(6):363-379. Moran, E., E. Ostrom, and J.C. Randolph. A Multilevel Approach to Studying Global Environmental Change in Forest Ecosyste ms. 1-26. 1998. Bloomington, IN, Center for the Study of Institutions, Population, and Environmental Change. Moran, E.F. 1993. Deforestation and Land-Use in the Brazilian Amazon. Human Ecology 21(1):1-21. Mostacedo, B., J. Balcazar, and J.C. Montero. 2006. Tipos de bos que, diversidad y composicin florstica en la Amazonia sudoeste de Bolivia. Ecologa en Bolivia 41(2):99-116. Munroe, D.K., J. Southworth, and C.M. Tu cker. 2004. Modeling spatia lly and temporally complex land-cover change: The case of western Honduras. Professional Geographer 56(4):544-559. Myers, R.L. 1990. Ecosystems of Florida. University of Central Florida Press, Orlando. Nagendra, H., J. Southworth, and C. Tucker. 200 3. Accessibility as a determinant of landscape transformation in western Honduras : linking pattern and process. Landscape Ecology 18(2):141-158. Nash, J.E. and J.V. Sutcliffe. 1970. River flow fo recasting through conceptual models part I -A discussion of principles. Journal of Hydrology 10(3):282-290. Naughton-Treves, L. 2004. Deforesta tion and carbon emissions at tr opical frontiers: A case study from the Peruvian Amazon. World Development 32(1):173-190. Neitsch, S.L., J.G. Arnold, J.R. Kinry, and J.R. Williams. Soil and Water Assessment Tool. Version 2000. Theoretical Documentation. TW RI Report TR-192, 1-412. 2002. Texas Water Resources Institute, College Station, Texas. Nelson, G.C. and D. Hellerstein. 1997. Do roads cau se deforestation? Using satellite images in econometric analysis of land use. American Journal of Agricultural Economics 79(1):80-88. Nepstad, D., G. Carvalho, A.C. Barros, A. Alencar, J.P. Capobianco, J. Bishop, P. Moutinho, P. Lefebvre, U.L. Silva, and E. Prins. 2001. Road paving, fire regime feedbacks, and the future of Amazon forests. Forest Ecology and Management 154(3):395-407. Nortcliff, S. and J.B. Thornes. 1981. Seasonal Vari ations in the Hydrology of a Small Forested Catchment near Manaus, Amazonas, and the Implications for its Management. In Tropical Agricultural Hydrology, eds. R.Lal and E.W.Russell. 37-57. John Wiley & Sons Ltd.: Chichester, U.K.

PAGE 152

152 O'Neill, R.V., C.T. Hunsaker, S.P. Timmins, B.L. Jackson, K.B. Jones, K.H. Riitters, and J.D. Wickham. 1996. Scale problems in reporting la ndscape pattern at the regional scale. Landscape Ecology 11(3):169-180. O'Neill, R.V. and A.W. King. 1998. Homage to St Michael: Or why are there so many books on scale? In Ecological Scale, Theory and Applications., eds. D.L.Peterson and V.T.Parker. 315. Columbia University Press: New York, NY. Ojima, D.S., K.A. Galvin, and B.L. Turner. 1994. The Global Impact of Land-Use Change. Bioscience 44(5):300-304. Oliveira, P.J.C., G.P. Asner, D.E. Knapp, A. Almeyda, R. Galvan-Gildemeister, S. Keene, R.F. Raybin, and R.C. Smith. 2007. Land-use allo cation protects th e Peruvian Amazon. Science 317(5842):1233-1236. Oliveira Filho, F. and J. Metzger. 2006. Thres holds in landscape structure for three common deforestation patterns in the Brazilian Amazon. Landscape Ecology 21(7):1061-1073. Openshaw, S. The Modifiable Areal Unit Pr oblem. Concepts and Techniques in Modern Geography. 38, 1. 1984. Norwich, UK., Geobooks. Openshaw, S. 1977. Geographical Solution to Sc ale and Aggregation Problems in RegionBuilding, Partitioning and Spatial Modeling. Transactions of the Institute of British Geographers 2(4):459-472. Openshaw, S. 1978. Empirical-Study of Some Zone-Design Criteria. Environment and Planning A 10(7):781-794. Panayotou, P. and S. Sungsuwan. 1994. An econometr ic analysis of the causes of tropical deforestation: the case of Northeast Thailand. In The Causes of Tropi cal Deforestation, eds. K.Brown and D.W.Pearce. 192-207. University of British Columbia Press: Vancouver, British Columbia. Pearson, D.M. 2002. The application of local meas ures of spatial autocorrelation for describing pattern in north Australian landscapes. Journal of Environmental Management 64(1):85-95. Perz, S., S. Brilhante, F. Brown, M. Caldas, S. Ik eda, E. Mendoza, C. Overdevest, V. Reis, J.F. Reyes, D. Rojas, M. Schmink, C. Souza, a nd R. Walker. 2008. Road building, land use and climate change: prospects for envi ronmental governance in the Amazon. Philosophical Transactions of the Royal Society B-Biological Sciences 363(1498):1889-1895. Petit, C.C., T. Scudder, and E.F. Lambin. 2001. In tegration of multi-source remote sensing data for land cover change detection. International Journal of Geographical Information Science 15(8):785-803. Poveda, G., P. Waylen, and R. Pulwarty. 2006. A nnual and Interannual Va riability of Present Climate in Northern South America and Southern Mesoamerica. Paleogeography, Paleoclimatology, Paleoecology 234:3-27.

PAGE 153

153 Qi, Y. and J.G. Wu. 1996. Effects of changing spatial resolution on the results of landscape pattern analysis using spa tial autocorrelation indices. Landscape Ecology 11(1):39-49. Quattrochi, D.A. and M.F. Goodchild. 1997. Scale in remote sensing and GIS. Lewis Publishers, Boca Raton, Fla. Refsgaard, J.C., International Hydrological Program, and Unesco. 1989. Methodology for distinguishing between man's influence and clim atic effects on the hydrological cycle. Paris : Unesco, Unesco. Repblica de Bolivia. Ley del Servicio Naci onal de Reforma Agraria. Ley no. 1715. 10-18-1996. La Paz, Bolivia. Riitters, K.H., R.V. Oneill, C.T. Hunsaker, J.D. Wickham, D.H. Yankee, S.P. Timmins, K.B. Jones, and B.L. Jackson. 1995. A Factor-Ana lysis of Landscape Pattern and Structure Metrics. Landscape Ecology 10(1):23-39. Rindfuss, R.R., S.J. Walsh, B.L. Turner, J. Fox, and V. Mishra. 2004. De veloping a science of land change: Challenges a nd methodological issues. Proceedings of the National Academy of Sciences of the United States of America 101(39):13976-13981. Roberts, D.A., M. Keller, and J.V. Soares. 2003. Studies of land-cover, land-use, and biophysical properties of vegetation in the Large Scale Bi osphere Atmosphere experiment in Amazonia. Remote Sensing of Environment 87(4):377-388. Rogers, P. 1994. Hydrology and water quality. In Changes in land use and land cover : a global perspective, ed. William B.Meyer and B.L.Turner II. 231-57. Cambridge University Press: New York, NY. Rojas, D.M., I. Martnez, W. Cordero, and F. C ontreras. Tasa de Deforestacin de Bolivia 19932000. 2003. Santa Cruz, Bolivia, BOLFOR ; Superintendencia Forestal. Rundle, H.D. and D.A. Jackson. 1996. Spatial an d temporal variation in littoral-zone fish communities: A new statistical approach. Canadian Journal of Fisheries and Aquatic Sciences 53(10):2167-2176. Schaap, M.G., F.J. Leij, and M.T. van Ge nuchten. 2001. ROSETTA: a computer program for estimating soil hydraulic parameters with hierarchical pedotransfer functions. Journal of Hydrology 251(3-4):163-176. Schroder, B. and R. Seppelt. 2006. Analysis of pattern-process in teractions based on landscape models-Overview, general concep ts, and methodological issues. Ecological Modelling 199:505-516. Schulz, K., R. Seppelt, E. Zehe, H.J. Vogel, and S. Attinger. 2006. Importance of spatial structures in advanci ng hydrological sciences. Water Resources Research 42(3).

PAGE 154

154 Siriwardena, L., B.L. Finlayson, and T.A. McMahon. 2006. The impact of land use change on catchment hydrology in large catchments: The Co met River, Central Queensland, Australia. Journal of Hydrology 326(1-4):199-214. Skole, D., W.H. Chomentowski, W.A. Salas, and A.D. Nobre. 1994. Physical and human dimensions of deforestation in Amazonia. Bioscience 44(5):314-322. Soares, B.S., D.C. Nepstad, L.M. Curran, G.C. Cer queira, R.A. Garcia, C.A. Ramos, E. Voll, A. McDonald, P. Lefebvre, and P. Schlesinger. 2006. Modelling conser vation in the Amazon basin. Nature 440(7083):520-523. Southworth, J., D. Munroe, and H. Nagendr a. 2004a. Land cover change and landscape fragmentation comparing the utility of continuous and disc rete analyses for a western Honduras region. Agriculture Ecosystems & Environment 101(2-3):185-205. Southworth, J., H. Nagendra, L.A. Carlson, and C. Tucker. 2004b. Assessing the impact of Celaque National Park on forest fragmentation in western Honduras. Applied Geography 24(4):303-322. Southworth, J., H. Nagendra, and C. Tucker. 200 2. Fragmentation of a landscape: incorporating landscape metrics into satellite analyses of land cover change. Landscape Research 27(3):253-269. Southworth, J., G.S. Cumming, M. Marsik, and M.W. Binford. 2006. Linking Spatial and Temporal Variation at Multiple Scal es in a Heterogeneous Landscape. The Professional Geographer 58(4):406-420. Steininger, M.K., C.J. Tucker, J.R.G. Townshend, T.J. Killeen, A. Desch, V. Bell, and P. Ersts. 2001. Tropical deforestation in the Bolivian Amazon. Environmental Conservation 28(2):127-134. Turner, B.L., D. Skole, G. Sanderson, G. Fischer, L.O. Fresco, and R. Leemans. Land-Use and Land-Cover Change Science/Research Pl an. 1-132. 1995. Stockholm and Geneva, The International Geosphere-Biosphere Programme : A Study of Global Change (IGBP) of the International Council of Scientific Unions (ICSU) and the Human Dimensions of Global Environmental Change Programme (HDP) of the International So cial Science Council (ISSC). Turner, M.G. 2005. Landscape ecology: What is the state of the science? Annual Review of Ecology Evolution and Systematics 36:319-344. Turner, M.G. 1989. Landscape Ecology th e Effect of Pattern on Process. Annual Review of Ecology and Systematics 20:171-197. Turner, M.G. 1990. Spatial and temporal analysis of la ndscape patterns. Landscape Ecology 4(1):21-30.

PAGE 155

155 Turner, M.G., V.H. Dale, and R.H. Gardner. 1989 Predicting across scales : theory development and testing. Landscape Ecology 3(3-4):245-252. van Griensven, A. and T. Meixner. Dealing with unidentifiable sources of uncertainty within environmental models. Pahl-Wostl, Claudia, Schm idt, Sonja, Rizzoli, A. E., and Jakeman, A. J. 2[1], 1045-1052. 2004. University of Osnabrc k, Germany. Complexity and Integrated Resources Management Transactions of the 2nd Biennial Meeting of the International Environmental Modelling and Software Society. 6-14-2004. van Loon, E.E. and P.A. Troch. 2002. Tikhonov regular ization as a tool for assimilating soil moisture data in distributed hydrological models. Hydrological Processes 16(2):531-556. Varlyguin, D.L., R.K. Wright, S.J. Goetz, and S.D. Prince. Advan ces in land cove r classification for applications research: a case study from the Mid-Atlantic RESAC. 2001. St. Louis, MO. Proceedings, ASPRS 2001: Gateway to the New Millennium. 2001. Vasquez, A. Metodologia para la determinacion de la capacidad de uso de la tierra. 1-60. 1980. San Jose, Costa Rica., Direccion de Investigaciones Agricolas, Ministerio de Agricultura y Ganaderia. Vitousek, P.M. 1994. Beyond global warm ing: Ecology and global change. Ecology 75:18611876. Wachholtz, R., J.L. Artola, R. Carmargo, and D. Yurca. Avance de la deforestacin mecanizada en Bolivia: Tasa anual de deforestacin mecanizada en los aos 2004 y 2005. 2006. Santa Cruz de la Sierra, Bolivia, Superinte ndencia Forestal, Repblica de Bolivia. Walker, R. 2004. Theorizing land-cover and land-us e change: The case of tropical deforestation. International Regional Science Review 27(3):247-270. Walsh, S.J. and K.A. Crews-Meyer. 2002. Linking people, place, and policy : a GIScience approach. Kluwer Academic Publishers ; Norwell, MA : Distributors for North, Central and South America, Kluwer Academic Publishers, Boston. Ware, S., C. Frost, and P. Doerr. 1989. Southe rn mixed hardwood forest: the former longleaf pine forest. In Biodiversity of the southeastern United States: Lowland terrestrial communities., eds. W.H.Martin, S.G.Boyce, and A.C.Echternacht. 447-93. John Wiley & Sons, Inc: New York, NY. Waylen, P. and M.E. Quesada. 2002. The E ffect of Atlantic and Pacific Sea Surface Temperatures on the Mid-Su mmer Drought of Costa Rica. In Environmental Change and Water Sustainability, eds. J.M.Garcia-Ruiz, A.A.Jones, and J.Arnez. 197-209. Instituto Pirenaico de Ecologa, Consejo Superior de Investigaciones Cientficas.: Zaragoza, Spain. Waylen, P., M. Quesada, C. Caviedes, G. P oveda, and O. Mesa. 1998. Rainfall distribution and regime in Costa Rica and its response to the El Nino-Southern Oscillation. Conference of Latin Americanist Geographers Yearbook 24 (1998) p75-84:75-84.

PAGE 156

156 Western, A.W., G. Bloschl, and R.B. Gr ayson. 2001. Toward capturing hydrologically significant connectivity in spatial patterns. Water Resources Research 37(1):83-97. Western, A.W., T.A. McMahon, R.B. Grayson, G. Bloschl, and G.R. Willgoose. 1999. Observed spatial organization of soil moisture and its relation to terrain indices. Water Resources Research 35(3):797-810. Western, A.W., S.L. Zhou, R.B. Grayson, T.A. McMahon, G. Bloschl, and D.J. Wilson. 2004. Spatial correlation of soil mo isture in small catchments and its relationship to dominant spatial hydrological processes. Journal of Hydrology 286(1-4):113-134. Wiens, J.A. 1989. Spatial scaling in ecology. Functional Ecology 3:385-397. Wilk, J., L. Andersson, and V. Plermkamon. 2001. Hydrological impacts of forest conversion to agriculture in a large river basin in northeast Thailand. Hydrological Processes 15(14):27292748. Williams, J.B. 1982. Temporal and Spatial Patterns of Abundance of the Chydoridae (Cladocera) in Lake Itasca, Minnesota. Ecology 63(2):345-353. Winchell, M., R. Srinivasan, M. Di Luzio, and J.G. Arnold. ArcSWAT interface for SWAT2005 User's Guide. 1-431. 2007. Temple, Texa s., Blackland Research Center, Texas Agricultural Experiment Sta tion and Grassland, Soil and Wa ter Research Laboratory, USDA Agricultural Research Service. Woodcock, C.E. and S. Gopal. 2000. Fuzzy set theo ry and thematic maps : Accuracy assessment and area estimation. International Journal of Geog raphical Information Systems 14(2):153172. Woodcock, C.E. and M. Ozdogan. 2004. Trends in Land Cover Mapping and Monitoring. In Land Change Science: Observing, Monitoring and Unde rstanding Trajectories of Change on the Earth's Surface., ed. G.Gutman. 367-77. Woodcock, C.E. and A.H. Strahler. 1987. Th e Factor of Scale in Remote-Sensing. Remote Sensing of Environment 21(3):311-332. Woods, C.H. and D. Skole. 1998. Linking satellite, census, and survey data to study deforestation in the Brazilian Amazon. In People and Pixels: Linking remote sensing and social science, eds. D.M.Liverman, E.Moran, E.F.Rindfuss, and P.C.Stern. 70-93. National Academy Press: Washington, D.C. Wu, J., D.E.Jelinski, M.Luck, and P.T.Tuelle r. 2000. Multiscale analysis of landscape heterogeneity: Scalevariance and pattern metrics. Geographic Information Sciences 6(1):619. Wu, J.G. and J.L. David. 2002. A spatially explic it hierarchical approach to modeling complex ecological systems: theory and applications. Ecological Modelling 153(1-2):7-26.

PAGE 157

157 Wu, W., C.A.S. Hall, and F.N. Scatena. 2007. Modelling the impact of recent land-cover changes on the 25 stream flows in northeastern Puerto Rico. Hydrological Processes 21(21):2944-2956. Ziegler, A.D., T.W. Giambelluca, L.T. Tran, T.T. Vana, M.A. Nullet, J. Fox, T.D. Vien, J. Pinthong, J.F. Maxwell, and S. Evett. 2004. Hydrological consequences of landscape fragmentation in mountainous northern Vietna m: evidence of accelerated overland flow generation. Journal of Hydrology 287(1-4):124-146. ZONISIG. Zonifcacin Agroecol gica y Socioeconmica y Perfl Ambiental del Departamento de Pando. 1997. La Paz, Bolivia, Proy ecto de Zonifficacin Agroecolgica y Establecimiento de Una Base de Datos y Re d de Sistema de Informacin Geogrfica en Bolivia.

PAGE 158

158 BIOGRAPHICAL SKETCH Matt Marsik earned an Associate in Arts degree in December 1997 from Valencia Community College, Orlando, Florida, and receiv ed a Bachelor of Science degree with high honors in December 2000 from the Department of Ge ological Sciences, University of Florida. His masters work, from January 2001 to May 20 03, focused on the consequences of 30 years of land cover change on stream flow in a sma ll watershed in San Ramn, Costa Rica. While maintaining interests in waters hed hydrology and land cover cha nge, his PhD work expanded to incorporate remote sensing a nd landscape ecology to develop and apply methods toward improved spatial and temporal analyses of land cover change. Focused on intedisciplinary applications of land cover cha nge, Matt has free range to pursue a career either in agency, non-governmental, consulting firms, or academics, or any combination of the previous.