This item has the following downloads:
1 QUANTITATIVE ANALYSIS OF DEFORESTATION PATTERN DYNAMICS: DEVELOPING FOREST RESILIENCE METRICS IN AN AMAZON FRONTIER By J ING SUN A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFI LLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2013
2 2013 Jing Sun
3 To my parents
4 ACKNOWLEDGMENTS Many people helped sustain, enrich and strengthen this dissertatio n. Without them this document would not exist. Jane Southworth provided clear, helpful critiques on many manuscripts and greatly assisted my development as a scientist by continually confronting me with new ideas and introducing me to exciting scientists. Beyond her practical wisdom and incredible academic achievements, I came away enriched and refreshed from her humility, kin dness and boundless curiosity. My dissertation committee along with a number of other professors provided me with a wide variety of assistance. Stephen Perz introduced me to the NSF deforestation project in the Amazon, which greatly enriched my thinking on human dimensions of resilience and disturbance. Stephen Perz is an excellent sociologist, and I am honored to have had the opportun ity to work so closely with him. Timothy Fik gave me a different perspective of statistics and encouraged me to think more critically. Peter Waylen always gave insightful comments and suggestions. I enjoyed working with them all. The collegial atmosphere of the Department of Geography encouraged discussions that developed this research as well as providing the friendship necessary to support my work, and for this I would like to thank my fellow graduate students: Zhuojie Huang, Yang Yang, Forrest Stevens, Erin Bunting, Jessica Steel, Tim Fullman Huiping Tsai, Likai Zhu, Ying Wang, Qiuyin Qi, Yibin Xia, Mario Mighty, and Caroline Staub. Also, I would like to thank my friends in Gainesville, whose friendship and support helped me research and then write this dissertation. Finally, I would like to acknowledge that this work was supported by an NSF Infrastructure Change, Human Agency, and Resilience in Social
5 #05 sensitivity and uncertainty analysis in the evaluation of social ecological resilience:
6 TABLE OF CONT ENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 8 LIST OF FIGURES ................................ ................................ ................................ .......... 9 LIST OF ABBREVIATIONS ................................ ................................ ........................... 10 ABSTRACT ................................ ................................ ................................ ................... 11 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .... 13 2 RESILIENCE AND DEFORESTATION IN THE AMAZON ................................ ...... 16 What Is Resilience? ................................ ................................ ................................ 16 The Promise of Ecological Resilience as a Theoretical Framework for Land Use and Land Cover Change ................................ ................................ ...................... 16 Understanding Deforestation: the Importance of Modeling ................................ ..... 19 ................................ .................. 21 3 DATA SOURCE: FOREST/NON FOREST COVER MAPS ................................ .... 24 Landsat Satellite D ata ................................ ................................ ............................. 24 Image Pre proce ssing ................................ ................................ ............................. 25 Image Classification: Object oriented Approach ................................ ..................... 26 Forest Cover Change in the MAP Region ................................ ............................... 27 4 DEFORESTATION ................................ ................................ ................................ 31 Background ................................ ................................ ................................ ............. 31 Moving Win dow Analysis across Multiple Spatial Scales ................................ ........ 35 Results ................................ ................................ ................................ .................... 36 ................................ ................................ ................................ .. 36 Clustering across scales ................................ ................................ ................... 38 Focus area study: Rio Branco ................................ ................................ .......... 39 Summary ................................ ................................ ................................ ................ 40 5 INDICATING STRUCTURAL CONNECTIVITY USING MORPHOLOGICAL IMAGE PROCESSING ANALYSIS ................................ ................................ ......... 53 Background ................................ ................................ ................................ ............. 53
7 Morpholo gical Spatial Pattern Analysis ................................ ................................ ... 55 Existing MSPA ................................ ................................ ................................ .. 55 Proposed additional MSPA classes: hole and outer background ..................... 55 MSPA Results ................................ ................................ ................................ ......... 57 Core and Boundary: Edge and Perforation ................................ ....................... 57 Islet ................................ ................................ ................................ ................... 58 Connector: Bridge, Loop, and Branch ................................ .............................. 58 Background: Hole and Outer background ................................ ........................ 59 Discussion ................................ ................................ ................................ .............. 60 Scale problems in MSPA applications ................................ .............................. 60 Inferences: A conceptual connectivity model ................................ .................... 61 Application: A deforested landscape in the Amazon ................................ ........ 62 6 MAPPING FRACTALITY DURING THE PROCESS OF DEFORESTATION IN AN AMAZON TRI NATIONAL FRONTIER ................................ ............................. 70 Background ................................ ................................ ................................ ............. 70 Fractal Geometry: A Promised Approach to Study Deforested Landscapes ........... 74 Fractal dimension ................................ ................................ ............................. 74 Box counting method ................................ ................................ ....................... 75 Cartographic Representation of Fractal Structure across the MAP Region ............ 76 Summary ................................ ................................ ................................ ................ 78 7 CONCLUSIONS ................................ ................................ ................................ ..... 84 What Can We Tell the Peopl e from This Research? ................................ .............. 84 Resilience Prompts the Understanding of Unexpected Shifts ................................ 86 Embracing Resilience Thinking ................................ ................................ ............... 88 LIST OF REFERENCES ................................ ................................ ............................... 90 BIOGRAPHICAL SKETCH ................................ ................................ .......................... 102
8 LIST OF TABLES Table page 3 1 Summary of forest cover changes from 1986 to 2010 in the MAP region* ......... 29 4 1 P region* ................................ ................................ ................................ ................ 44 4 2 Summary of forest/non forest areas from 1986 to 2010 around Rio Branco, Acre. ................................ ................................ ................................ ................. 45 4 3 as in percentage from 1986 to 2010 around Rio Branco, Acre. ................................ ................................ ................................ ... 46 5 1 S Fragstats indices compared to the MSPA class metrics ................................ 65 5 2 Statistical summary of MSPA classes of the three sides of the MAP region from 1986 to 2010. ................................ ................................ ............................ 6 5 5 3 over thre e states in the MAP region from 1986 to 2010. ................................ ... 66
9 LIST OF FIGURES Figure page 2 1 Map of the study area with major roads a nd state capitals superi mposed ........ 23 3 1 Forest cover maps showing fragmentation dynamics in the MAP region fo r ..... 30 4 1 A simulated landscape (spatial res olution: 30m) was con voluted by 9 different filters and these such 9 filtered maps were then stacked. .................... 47 4 2 The results of the clustering analysis in the MAP region from 1986 to 2010, using t he migrating means clustering algorithm. ................................ ................. 48 4 3 Cluster means at all scales in Acre, Brazil from 1986 to 2010 ............................ 49 4 4 Cluster mea ns at all scales in Pando, Boliva from 1986 to 2010. ....................... 50 4 5 Cluster means at all scales in Madre de Dios, Peru from 1986 to 2010. ............ 51 4 6 The results of the clustering analysis around Rio Branco, Capital of Acre, Brazil, from 1986 to 2010, using the migrating means clustering algorithm ........ 52 5 1 Example of morphological se gmentation of binary patterns. .............................. 67 5 2 Example of MSPA patterns. Left: MSPA segmentation result. Right: Proposed hole and outer background additions. Adapted with permission. ....... 68 5 3 Proposed conceptual model of forest structural connectivity and composition/amount versus mat hematical morphological indices ...................... 69 6 1 Spatio temporal fractal analysis of the cleared areas in the MAP region.. .......... 82 6 2 The evolving pattern of development as indicated by density sliced fractal dimension in the MAP region ................................ ................................ ............. 83
10 LIST OF ABBREVIATIONS G IS Geographic Information System I SO Iterative self organizing
11 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy QUANTITATIVE ANALYSIS OF DEFORESTATION PATTERN DYNAMICS: DEVELOPING FOREST RESILIENCE METRICS IN AN AMAZON FRONTIER By Jing Sun August 2013 Chair: Jane Southworth Major: Geography Deforestation, the process of clearing the Earth's forests, has occurred globally on a massive scale and negatively impacts the resilience of forest landscapes. Aimed to support the understanding of deforestation dynamics, the dissertation first reviews resilience theory, and then demonstrates possible solutions to infer the resilience of deforested landscapes with particular attention to the forest spatial pattern changes via an example in the Amazonian rainforest. The solutions were based on the application of three different methods, and integrating these with remote sensing (RS) and geographic information system (GIS) techniques. Data inputs were fine-grained forest/non-forest classifications of the years 1986, 1991, 1996, 2000, 2005 and 2010 interpreted from Landsat satellite images. The three different methods presented in this dissertation are: first, the combination of a moving window approach and an iterative self-organizing deforestation. These patterns are revealed by forest cover changes, which are initiated from developed areas and then extend into forest areas across multiple spatial scales. Second, the local changes of forest spatial patterns were illustrated, where the dynamics of forest cover were obtained by applying the morphological spatial pattern
12 analysis (MSPA). T created to enhance the appreciation of the spatial patterns of developed areas. Third, fractal geomet ry was introduced to further prompt the understanding of the spatial patterns of developed areas. A fixed grid scan strategy was adopted to pixelize the entire landscape and to cartographically represent the spatial patterns of developed areas by fractal d imension. Moreover, a configuration framework was proposed to partition the level of deforestation characterized by fractal dimensions and as such can potentially be used to better understand and manage the evolution of forest clearings. These three metho ds each have their own set of advantages and limitations, and yet work in a complementarily manner and cast particular insight on spatial pattern change during the process of deforestation. Their useful value added contributions towards resilience theory i s discussed at the end, as the ontogenetic stage of resilience theory is
13 CHAPTER 1 INTRODUCTION Geographers study the spatial and temporal distribution of phenomena, processes, and features as well as the interaction of humans an d their environment. Because space and place affect a variety of topics, such as economics, health, climate, plants and animals, it is necessary to introduce a framework to integrate multiple topics in an interdisciplinary system. The recently discussed resilience theory might be able to tackle and build critical thinking towards this question, studies on social ecological resilience in geography (Agder, 2000, 2003) as well as on more or less successful re sponses to disasters, such as Hurricane Katrina in 2005, have greatly raised interest in the use of resilience theory. Operating within the paradigm of resilience theory, this dissertation investigates the resilience of deforested landscapes from a combin ation of theoretical, technique and empirical perspectives. The approach followed is to build a set of models to understand the resilience of forest landscapes to deforestation activities, which has been completed in this dissertation, in order to then pro vide the information to conservation programs and then guide real world forest management. The dissertation is structured into seven chapters. This brief introduction to the organization of the d issertation is C hapter 1 and the remaining c hapters are outli ned below: In C hapter 2 resilience theory and its development are briefly introduced. The necessity of applying resilience theory in deforestation studies is discussed and then proposed as the research paradigm of this dissertation. C hapter 3 introduces the reader to the data used throughout the upcoming analyses and describes the procedure followed for the satellite images classification
14 and processing. An object oriented classification method, a decision tree approach, was applied, which largely improv es the classification accuracy compared with more traditional techniques. The products are multiple forest/non forest maps of the years 1986, 1991, 1996, 2000, 2005 and 2010, which are utilized as the only data source for C hapters 4, 5 and 6 which are them selves outlined and discussed below. The pattern and process of deforestation are not restricted to cleared areas but rather they exhibit strong influences beyond the clearings physical boundaries. Such that, C hapter 4 presents a moving window technique w ith an ISO supervised clustering algorithm, this combination is designed to analyze landscape information across multiple spatial scales by varying window sizes, which help think through the structure, function and dynamics of deforested landscapes from lo cal scales to regional scales. Deforestation affects the spatial layout of all components of the forest system, and spatial patterns have been shown to influence many processes that are ecologically important. Therefore, quantitative methods are required to detect spatial patterns of deforestation, identify significant changes through time, and relate these patterns to important ecological functions. So that, C hapter 5 presents an image processing tool, MSPA, as used here to describe the process of landsca pe fragmentation as deforestation proceeds. It is described in detail and compared with the frequently used are created and proposed to further classify the spatial patterns of developed areas. The incorporation of these two new tools improves the MSPA algorithms, expands their potential applications, and provides a new perspective to understand the spatial patterns of developed areas as well as the entire landscape.
15 Chapter 6 introduces the concept of fractal geometry as an application for studying the spatial patterns of developed areas. Specifically, the heterogeneous landscape is pixelized by a fixed grid scans approach and then the value of each pixel is characterized by the fractal dimension calculated by the box counting method. The results indicate that deforestation has expanded across the landscape but with different patterns and rates. Such differences are summarized within a proposed framework which indicates the l evel of deforestation or development, and as such, can potentially be used to apprehend and regulate deforestation over time. Chapter 7 summarizes and concludes the dissertation. Critically, a new development in the philosophy of science inspired by resil ience theory is discussed, utilizing the research presented.
16 CHAPTER 2 RESILIENCE AND DEFORESTATION IN THE AMAZON What Is Resilience ? Development of ecological resilience theory began in the 1960s with attempts to mathematically model ecosystem dynamics (Gunderson et al., 2010). Early models focused on the stability of systems, such as Lewontin (1969) and May (1977), where they assumed ecological systems were stable from both theoretical and practical perspectives. Usually, their research emphasized a gl obal or stable equilibrium, such as carrying capacity in ecosystem, and usually led to policies of maximum. Holling (1973) introduced the word resilience to describe three aspects of changes that occur in an ecosystem over time. The first was to describe the second concept recognized the occurrence of alternative and multiple states as op posed to the assumption of a single equilibrium and global stability; hence, resilience was the amount of disturbance a system can absorb before it shifted into alternative c ecological systems and how practitioners have attempted to manage these systems. The Promise of Ecological Resilience as a Theoretical F ramework for L an d Use and Land Cover C hange Resilience has two faces, engineering and ecological (Holling, 1996). Engineering resilience is defined as the rate or speed of recovery of a system following a shock (Pimm
17 the magnitude of a disturbance that triggers a shift between alternative states (Holling, 1996). In this sense, a regime shift occurs when the controlling variables in a system (including feedbacks) result in a qualitatively different set of structures and dynamics of these systems (Walker et al., 2004). It has been four decades since the term resilienc e was introduced, and various documentations still maintain momentum. Scientists, working in disturbance driven ecosystems, found that resilience theory was the only theory that helped explain the complex changes that they were studying (Walker et al., 200 4). Scheffer et al. (2001) described two alternative states, clear water with rooted aquatic vegetation and turbid water with phytoplankton, in shallow lake systems. Coral reef systems may shift from a coral dominated state to a macroalgae dominated state (Hughes, 1994). Terrestrial systems have traditionally been studied in much more detail than most freshwater or marine systems; and as such, there is a large collection of compelling examples of critical transitions between alternative states. Walker (1981 ) studied semiarid rangelands and found dramatic shifts between grass dominated and shrub dominated systems. Those shifts were mediated by interactions among herbivores, fires, and drought cycles. Gragson (1998) argued that disturbances derived from land u se and land cover changes are related to the extent of the transformation, and he recognized two types: modification, a change within a given cover type (e.g., from closed to open forest), and conversion, that is, a change from one cover type to another (e .g., forest to crop), which involves a more profound transformation and, therefore, an associated change in resilience. Foster et al. (1998) suggested that forest cover loss has a
1 8 substantial effect not only on the structure and functioning of ecosystems, but also on resilience and management. Shifts in forest cover, associated with management of fire regimes, in the well drained soils of the southeastern United S tates reflect alternate states. Among these examples, not all threshold crossings are sudden. Evidently, most This does not indicate these gradual changes are not important. Conversely, in many cases, regime shifts may be largely irreversible (Carpenter et al., 2 009; Scheffer, 2009). For example, Wilson and Agnew (1992) demonstrated that the forests were established under a wetter rainfall regime thousands of years previously. Necessary moisture is supplied through condensation of water from clouds intercepted by the canopy. If the trees are cut, this water input stops and the resulting conditions can be too dry for recovery of the forest. Similarly, extinction of mega fauna at the end of the Pleistocene in Siberia, possibly through improvement in hunting technolog y, may have triggered an irreversible shift from steppe grassland to tundra. The resulting increase in mosses led to cooler soils, less decomposition, and greater carbon sequestration in peat (Zimov et al., 1995). Deforestation in large forest areas, like the Amazonian rainforest, coupled with savanna or desert), with catastrophic al., 2001, 2007). In order to prevent the (undetectable, unforeseeable) irreversible shift
19 in this forest system, it is necessary to build models to appreciate deforestation dynamics and most importantly, system resilience. Understanding D eforestation: the I mportance of M odeling Deforestation is the process of clearing the Earth's forests and has occurred globally on a massive scale which can negatively impact the resilience of forest systems, resulting in numerous repercussions. Understanding deforestation is important to regional biodiversity and global climate. The large scale estimate of deforestation has been feasible since the analysis of satellite imagery which began in 1972, and there is a large col lection of compelling examples of deforestation studies (Southworth and Tucker, 2001; Brook et al., 2003; Nagendra, 2004). Deforestation, most often, goes with economic development and is continuing today into the remaining areas of undisturbed forest glob ally. Time series satellite data confirms that due to human land use, deforestation is already underway in the Amazon and will probably continue well into the future. Road building is a key determinant of deforestation in the Amazon (Perz et al., 2008). state as part of the development policy to meet geopolitical and economic objectives. Paving of official roads raises land values, which provides the incentive to exploit natural resources farther out from official road corridors and consequently motivates unofficial road building. Unofficial roads were built by local interest groups speci fically to gain access to land, timber and other natural resources. Unlike official roads which were published on maps or regulation layouts, the building of unofficial roads does not have official documentation or planning. Unofficial roads keep growing r apidly, and are not
20 sustainable (Perz et al., 2005, 2007b). Unofficial roads form dense networks and these fragment forests into small, irregular shapes, and importantly diminish the resilience of forest systems (Peterson, 2002). Deforestation processes, especially when these processes occur inside the forested areas, can disconnect conterminous forest areas which largely change the spatial properties of forest systems, such as forest connectivity. The consequences affect the forest health and forest succe ssion. Forests are habitats to animals and wildlife and therefore the reduced forest connectivity would limit the food sources and the behaviors of these species, and probably lead to extinctions in a few decades. The highly isolated forests were not able to maintain material and energy flow, which would inevitably threaten the function of ecosystems and human society. However, deforestation keeps expanding with population growth and economic development in Amazonian nations, hence, we need to analyze this increasingly fragmented forest landscape, otherwise the whole forest landscape with its ecosystem may be going to a catastrophic shift which will inevitably cause economic and ecological losses. When viewing the spatial pattern of a certain landscape, we look at its composition and spatial configuration: the elements present and how these elements are arranged (Li and Reynolds, 1994; Turner et al., 2001). In deforestation landscapes, we may observe complex patterns of forest and non forest surfaces (like f ish bone or chess board), we may observe gaps in an otherwise continuous forest, and we may Deforestation patterns affect the spatial patterning of all components of the sys tem, and have altered the resilience of historical landscapes ( Gunderson and Pritchard, 2002).
21 Understanding forest resilience is critical for managing deforestation, and for restoration and preservation of historical landscapes. But, it is difficult to d irectly measure resilience (Carpenter et al., 2005), so that a set of models is proposed in this dissertation to study deforested landscapes and demonstrate it as an indirect measure to indicate forest resilience (to deforested disturbances). In order to a ccomplish this, the dissertation provides three measurements to evaluate the spatial patterns of forest landscapes from different perspectives. The modeling would help to gain an understanding of how deforestation affects the composition and configuration of forest landscapes, as well the function and services of the forest, and as an indirect measure objective threading through the whole research, as the results are expected to provide useful knowledge in forest management and conservation programs. Amazonian R ainforest Over a 25 year period, as a function of distance to major roads, forest conversion occurred in the Amazon region of South America the largest remaining tract of contiguous tropical forest in the world. The study area of this dissertation is restricted to the heart of southwestern Amazon, located near the shared borders of the Peruvian state Madre de Dios, Acre in Brazil, and the Bo livian department of Pando (termed the 2 see F igure 2 1). This tri national frontier was historically covered by humid tropical forest and remains roughly 90% forested. It is estrial biodiversity and is considered a global hand, MAP is a relatively isolated area in three countries and is featured by a composite of rich natural resources and fa irly poor economic development. Such differences in the
22 MAP frontier and elsewhere in South America prompted presidents of several countries to develop the Initiative for the Integration of Regional Infrastructure in South America (IIRSA). IIRSA prioritize s cross border infrastructure projects such as road paving as a means of economic integration and economic development. The MAP frontier is among the first wave of IIRSA projects, and is currently being integrated via paving of the Inter Oceanic highway. The paving in Acre was completed in 2002 and therefore forest clearings are more extensive and numerous in this region, while the paving in Madre de Dios began around 2005 and is only now being completed, so that the deforestation in Madre de Dios is relat ively patchier than that of Acre. On the other hand, the road in Pando has not been paved yet, and so here most forest clearings occur around major towns and along the Brazilian border (Southworth et al., 2011). Thus, there is considerable spatial variabil ity in land cover changes and the three regions exhibit a pathway of development from most developed (Acre), to developing (Madre de Dios), and currently much less developed (Pando). This is a critical historical moment at which to obtain estimates of past and present land cover change in order to see how the MAP social ecological system responds to external disturbances brought by migration, investment and new technologies in the wake of infrastructu re upgrades (Brown et al., 2002 ; Perz et al., 2011).
23 Figure 2 1 Map of the study area with major roads and state capitals superimposed. The region encompasses the tri national frontier of the Peruvian state of M adre de Dios, the Brazilian state of A cre, and the Department of P ando,
24 CHAPTER 3 DATA SOURCE: FOREST/NON FOREST COVER MAPS A land cover database from digital maps of the land surface, documents precisely where land cover change has occurred, which is used for thousands of applications in diverse investigations s uch as ecosystem status and health, spatial patterns of biodiversity, indications of climate change, and best practices in land management (Turner et al., 2001). In the MAP region, deforestation is currently expanding and therefore, it is important to make a set of land cover databases to record the detailed information of deforestation processes and then to analyze their spatial patterns. Since deforestation is a dynamic process, the cost of gathering timely and accurate information becomes prohibitive if done by field work alone. In addition, deforestation is a spatial temporal process with categorical change (forest and non forest) and requires studying through temporal ordering, which made satellite data become a major data source in the large scale fore st research. Landsat Satellite D ata Single layered binary deforestation index maps are derived from Landsat TM 4, 5 (TM is the abbreviation for thematic mapper) and ETM + 7 (ETM + is the abbreviation for enhanced thematic mapper plus) satellite images acr oss six near anniversary months with less than 10% cloud for detection (July or August, 1986 1991 1996 2000 2005 2010, corresponding to the dry season that permits minimal cloud cover and lower aerosol concentrations). Landsat 4 is the fourth satellite of the Landsat program and was launched in 1982. Landsat 4 science operations ended in 1993 when it lost the ability to transmit science data, far beyond its designed life expectancy of five years. Landsat TM 5 is the
25 fifth satellite of the Landsat program and was launched in 1984. Landsat TM 5 is still sending back data now, and in the year 2012, it is 25 years over its 3 year mission. The satellite was said by the USGS (USGS is the abbreviation for United State Geological Survey) to be nearing the end of i ts life, after more than 28 years in space. Landsat 7, launched in 1999, is the latest satellite of the Landsat program. Landsat 7 is the most accurately calibrated Earth observing satellite, i.e., its measurements are extremely accurate when compared to t he same measurements made on the ground. The Landsat 7 mission went flawlessly until May 2003 when a hardware component failure left wedge remainder of the ETM+ sensor, including the primary mirror, continues to operate, radiometrically and geometrically, at the same high level of accuracy and precision as it did before the anomaly; therefore, image pixels are still accurately geolocated and calibrated. Two data providers were used, t Science Center (EROS, http://glovis.usgs.gov ) and Divis o de Gerac o de Imagens ( http://www.dgi.inpe.br ), all scenes were freely downloaded from two websites to comprise the MAP region (path/row of the scenes: 1/67, 1/68, 2/67, 2/68, 2/69, 3/67, 3/68, 3/69) for the target years. Image P re processing For each image, we performed the atmospheric and radiometric calibration, standardization of map projection, coordinate system, and datum; and geometric correction to the Global Land Cover Facility Geocover product for 1999/2000. The image preprocessing follows the CIPEC protocol (Moran and Ostrom, 2005) and was standardized for each image, providing a robust preprocessing methodology
26 comparable across dates and country boundaries and cut to exclude clouds and rivers and undertook an object oriented classification to produce forest/non forest maps. Due to the similarity of spectral signatures, some transitional woodlands and agro forestry areas are difficult to separate and have been classified as forest. Agriculture cleared areas, major road, urban fabrics, and shrub were aggregated to create a non forest class. Image C lassificat ion: O bject oriented A pproach Image mosaics were created for 1986, 1991, 1996, 2000, 2005 and 2010 with eight images for each year. Explic atory variables: tasseled cap brightness, greenness, and wetness indices were calculated along with a mid infrared vegetation index and a three by three moving window variance for each mosaic date as input to a decision tree classification. Due to striping in the visible and thermal bands in many of the available Landsat images, only the near and mid infrared bands were used along with the derived products to create a forest and non forest classification for each mosaic date (see Marsik et al., 2011 for mo re detail on processing of imagery). The image striping problem limited the spectral data available, and due to this problem, we created derived products, rather than use much of the raw image bands, as these derived products did not have the striping prob lem (e.g., TCA analysis which is a form of Principal Components Analysis (PCA) itself a technique commonly used to destripe images, and so the first three components of the TCA were striping free, and used in concert with the unaffected image bands). The f orest/non forest classification was used in this analysis due to the focus on deforestation processes as a function of road building. Differences in timing and type of agriculture made a set agricultural class inaccurate given the available imagery. As suc h pasture, bare fields and built cover types were grouped into
27 the non forest class. The forest class included all dense vegetated covers, which would by default include secondary succession as a cover type once a dense canopy was achieved (considered to b e 3 5 years within this region). A decision tree classifier was used for the analysis to allow the incorporation of a series of rules and derived image products within the classification. Compumine Rule Discovery System (RDS) software was used, which is a data mining software which predicts the specified land cover classes, and was established here using a split sample validation, where we used 85% of our training sample points to create the decision rules and tree classifier and the remaining 15% were used to test the tree output. The rules, once developed and tested for accuracy (each year was analyzed separately and percentage accuracy of the rules were 98 99.8% accurate, see Marsik et al. 2011 and Southworth et al. 2011 for more detailed information) w ere then incorporated into the ERDAS Knowledge Engineer rule was assessed using over 350 training samples collected during fieldwork from 2005 to 2006 and Kappa coef ficient and overall percent accuracy for each class and for the overall classification, with resulting accuracies for 2005 of greater than 90%. In addition the 2000 image data were independently checked using ASTER images for the year 2000, and an overall accuracy of 96% was achieved across these products. The 2010 image data were also independently examined using Google Earth, and an overall accuracy of 95% was reached across the products. So that, these six forest/non forest binary maps are the inputs for the sections outlined and discussed below. Forest Cover Change in the MAP R egion With relatively uniform temporal interval and fine spatial resolution, the created time series forest/non forest cover maps (F igure 3 1) will be able to reflect the change
28 t rajectories of forest covers, forest areas, fragmentation patterns (Tole, 2006) and other forest landscape properties. Specifically, from 1986 to 2010, we observed an increase in non forest area for all three states (Acre, Madre de Dios, and Pando), at the expense of forest areas. Across all three regions, the forest is still really dominant and the landscape is quite static, such that, in 2010 the forest cover percentage is 80.0% in Acre, 97.4% in Madre de Dios, and 98.0% in Pando (Table 3 1). Correspondin gly, the area percentage of the non forest area is relatively small, with Acre at 20.0%, Madre de Dios at 2.6%, and Pando at 2.0%.
29 Table 3 1 Summary of forest cover changes from 1986 to 2010 in the MAP region* Region Cover 1986(%) 1991 (%) 1996 (%) 2000(%) 2005(%) 2010(%) Acre Non forest 2.5 6.6 7.5 9.6 14.4 20.0 Forest 97.5 93.4 92.5 90.4 85.6 80.0 Madre de Dios Non forest 2.2 1.7 0.9 1.8 2.0 2.6 Forest 97.8 98.3 99.1 98.2 98.0 97.4 Pando Non forest 0.4 0.7 0.8 1.3 1.9 2.0 Forest 99.6 99.3 99.2 98.7 98.1 98.0 *where 1% in Acre represents 622.79 km2, 1% in Madre de Dios represents 493.48km2, and 1% in Pando represents 542.87 km2.
30 Figure 3 1. Forest cover maps showing fragmentation dynamics in the MAP region for A ) 1986 B ) 1991 .C ) 1 996 D ) 2000 E ) 2005 F ) 2010
31 CHAPTER 4 IMPACTS OF DEFORESTATION Background Land use and land cover changes play a major role in the study of global change, which have largely resulted in deforestation, biodiversit y loss, global warming and increase of natural hazards (Mas et al., 2004; Dwivedi et al., 2005; Reis 2008). Significant attention has been drawn to areas of tropical forest cover, and specifically the Amazon region across the past few decades due to the al arming rates of clearing which have been occurring here (Nepstad et al., 1999, 2001; Anderson et al., 2002; Giles and Burgoyne 2008; Marsik et al., 2011; Southworth et al., 2011). Laurance et al. (2004) pointed out that infrastructure development projects are significant contributors to deforestation in the Brazilian Amazon, which have caused numerous complex problems to social ecological systems (Cumming et al., 2012). In forest landscapes, like the Amazonian rainforest, a new and pervasive spatial pattern has been created by road paving (Riitters and Wickham, 2003) and then been enhanced by the expansion of human settlements. The r oad paving encourages economic developments via regional integrations, such as facilitating ties of remote areas to local and regional markets (van de Walle, 2002) and opens up previously inaccessible and intact rainforest to these more developed regions. To access the natural resources, the paving of major roads stimulates the construction of secondary and tertiary level roads, often temporary logging roads which continue to be used well after their initial creation and often become permanent landscape features (Perz et al., 2007a, b, 2008), and these roads form dense and hierarchical networks where the l and cover changes along them exhibit some recursive structures.
32 For years, economic, social, and ecological research has drawn contrasting conclusions about the consequences of such new patterns created by road paving (Perz et al., 2012). This infrastruct ure is notorious for intensifying social conflicts over newly accessible natural resources (Schmink and Wood, 1992). For example, lack of participation raises issues of exclusion of indigenous and other minority groups, often whereby newcomers make legalis tic claims to natural resources in order to discredit traditional claims (Perz et al., 2012). However, both macroeconomic and microeconomic literatures demonstrate the positive impacts of road paving on the economy, specifically empirical studies show sign sub national administrative units (Demurger, 2001; Fan et al., 2004), local communities (Gunasekera et al., 2008), or individual households (Gibson and Rozelle, 2003). On the other hand, there ar e significant ecological changes which can occur as, both road networks and developed areas could fragment a continuous forest into discrete habitats and thereby reduce landscape connectivity (Estrada and Bodin, 2008). Connectivity is a vital element of la ndscape structure (Taylor et al., 1993), and one to which the persistence of spatially structured species populations, and meta populations, is strongly related (Hanski and Ovaskainen, 2003). Lack of connectivity reduces the capability of species movement (i.e., local extinctions of certain species could not be compensated by the immigrant population from the neighbor patches) and interferes with pollination, seed dispersal, wildlife migration and breeding (Estreguil and Mouton, 2009). As such, via road net works, the impacts of social conflicts, economic benefits, and ecological damages extend from a small scale region (cleared area) to an entire social
33 pattern and process of def orestation are not restricted to those cleared areas but rather they exhibit strong influences beyond the clearings physical boundaries. As such, a multiple scale analysis that includes larger surrounding areas beyond just the clearings themselves becomes a key factor to better understand the social ecological systems in deforested landscapes. corridor matrix model (Forman, 1995) that simplifies the landscape structure as a network wit h individual patches connected by corridors (Hanski, 1999; Opdam, 2002). Thus, a landscape can be described as an area of land containing a mosaic of habitat patches, often within which a particular focus habitat patch is embedded (Dunning et al., 1992). Z s and corridors within the boundaries of the conservation area where each area is assumed to be spatially homogeneous according to the corresponding descriptions. In lan dscapes that may exhibit significant impacts on interiors within the conservation boundaries (Wiens, 2002). At a landscape level, the context can be taken as the buffer zone around a target area, which exhibits either negative or positive impacts on the co ntent, as different land uses and land covers may promote or block ecological function and services, such as population migration (Hansen and Rotella, 2002). To further illustrate the interactions between the content and context, Riitters et al. (2009) pro
34 defined the forest security as the likelihood that a pixel of forest remains as forest over time, which depends on the surrounding landscape mosaic. The security level of c ontext that surrounds the content is not uniformly distributed, in fact, the security level declines when the distance to the content increases. During the expansion process of developed areas, the interfaces between forest areas and non forest areas, e.g. forest edges, become sensitive areas that usually have a high likelihood to be converted into non forest. Such interface was not emphasized in the classic landscape analysis before (Cumming et al., 2012), but it has been estimated that the amount of Amazo nian rainforest modified by the forest edge exceeded the area that had been cleared ( Skole and Tucker, 1993 ). As such, content, context, interface (i.e., forest edge), and forest security level should become important considerations in any multi scale anal ysis of deforested landscapes. Landscapes do not exist in isolation. Landscapes nest in a panarchic structure (Gunderson and Holling, 2002; Zurlini et al., 2006), that is, focus landscapes embed within larger landscapes that embed within even larger lands capes, and so on. This panarchic representation makes the multi scale analysis imperative for thinking through structure, function and dynamics of deforested landscapes (Wu et al., 2000). In answering such a requirement, moving window operations provide a useful technique to derive multi scale information in landscape research (e.g. Riitters et al., 2000, 2002; Southworth et al., 2006; Zurlini et al., 2006; Zaccarelli et al., 2008; Wade et al., 2009 ), which would be able to appreciate and then link the loca l pattern and process of cleared areas. In this section, we will apply a moving window operation combined with
35 the MAP region. Moving Window Analysis across Multiple Spatial Scales Advancements in satellite technology combined with mathematical and/or spatial statistical methods have been widely used in landscape ecology. To measur impacts on forest/non forest binary maps from 1986 to 2010, we used a moving window algorithm to locate a set of 9 fixed area windows around each pixel. The window sizes are 0.15, 0.39, 0.81, 1.65, 3.21, 6.45, 9.75, 12.87, and 16.11 kms on a s ide, where the window side lengths span two orders of magnitude (from 0.15 km to 16.11 km) to ensure, as best as possible, accurate estimates of the scale dependent behaviour of percentage fragmentation (Wade et al. 2009). We measured the amount of fragme ntation within each window by the proportion of non forest pixels (P n ) on each binary map, and then produced a stack of these raster layers, where each raster was the specific output for a certain window size. Next, we performed the ISO clustering algorith m (also known as the migrating means technique) to run a clustering analysis according to the similarity of P n values over all 9 window sizes, where ISO is an abbreviation for the iterative self organizing way of performing clustering (Zurlini et al., 2006 ; Zaccarelli et al., 2008). The ISO Cluster algorithm is an iterative process for computing the minimum Euclidean distance when assigning each candidate cell to a cluster. The process starts with arbitrary means being assigned by the software (ArcGIS 10.0) one for each cluster (the cluster number is defined by the user). Every cell is assigned to the closest of these means (all in the multidimensional attribute space). New means are recalculated for each cluster based on the attribute distances of the cell s that belong to the cluster after the first iteration. The process is repeated: each cell
36 is assigned to the closest mean in multidimensional attribute space, and new means are calculated for each cluster based on the membership of cells from the iteratio n. The number of iterations of the process is specified by the user, and this number should be large enough to ensure that, after running the specified number of iterations, the migration of cells from one cluster to another is minimal, and thus all the cl usters become stable. Corresponding to the previous discussion of content, context, forest edge, and forest security level, we defined five clusters in this research: core forest area, sub forest area, transition area, sub developed area, and core develop ed area. For the forest areas (content), the security level of the core forest area is the highest and then the security level decreases from the sub forest area to the most insecure level, which is the transition area. Meanwhile, we can interpret cleared areas (context) using a similar gradient level: the core developed area where forest clearing is permanent, and then the sub developed area where the forest clearing is new, and again the transition area where the forest clearing is most likely underway. deforestation start from the core developed area, spread over the sub developed area, via the transition area, and finally reach the sub forest area (see F igure 4 1 for a detailed illustration of the method described). To val idate this classification framework for the real world landscape we will implement this moving window operation to study deforestation dynamics in the MAP region from 1986 to 2010. Results al scales, the results of the cluster ing analysis are summarized in F igure 4 2 and Table 4
37 impacts of the cleared areas in the MAP region (core developed area, sub developed area, transition area, and sub forest area) expand rapidly from 1986 to 2010, and the developed areas increase substantially among them. In Acre, the core developed area increases from 1.7% in 1986 to 13.3% in 2010 (where 1% in Acre represents 622.79 km2), which may signal the beginning or initialization of the opening up (urbanization) of this predominantly forested landscape. Subsequently, its impacts spread over the surrounding classes, where the sub developed area increases from 3.0% in 1986 to 6.7% in 2010. In addition, the transition area has a slight incre ase from 4.4% in 1986 to 6.7% in 2010, since this interface area usually has elevated temperature, reduced humidity, and increased wind speed relative to forest interiors (Cumming et al., 2012) the expansion of this area would bring many uncertain factors to the forest system. The sub forest area decreased from 15.9% in 1986 to 14. 7% in 2010 and the core forest area changes from 75.0% in 1986 to 58. 6% in 2010, and such significant reduction of these two stable forest classes indicates the entire forest l andscape in Acre may be undergoing a critical transition. In Pando, the core developed area increases from 0.2% in 1986 to 2.0% in 2010 and the sub developed area changes from 0.5% in 1986 to 2.5% in 2010 (where 1% in Pando represents 542.87 km2), and the small percentage of these two non forest clusters (4.5%) until 2010 characterizes the less developed situation in Pando. However, there is a re allocation of the forest area among the three clusters, where both the transition area (0.6% in 1986 to 3.3% in 2010) and the sub forest area (5.5% in 1986 to 18.5% in 2010) increase significantly, while the core forest area decreases from 93.3% in 1986 to 73.7% in 2010. Though the total forest area (transition, sub forest, core forest) is still high in 2010 (95.5 %), there is a large part of
38 the forest area transformed from stable core forest to the unstable clusters: transition and sub forest. On the other hand, Madre de Dios experienced a reforestation deforestation process during the past 25 years and therefore the percentage area of the core forest starts to increase from 75.1% in 1986, reaches the peak at 87.0% in 1996, and finally decreases to 75.7% in 2010. Consequently, the other four clusters: core developed, sub developed, transition and sub forest have a U shape change from 1986 changes in forest cover across these regions (as illustrated in Table 4 1) that if studied in a more traditional or static analysis is completel y obscured. Omission of material that unites components to form a complete manuscript. Clustering across scales For the three states in the MAP region, the profiles of the five cluster means along multip le window sizes were depicted (F igures 4 3, 4 4, 4 5 ). For a given location, the profile of Pn with different window sizes fluctuates with respect to the deforestation experienced by that location at different spatial lags. For example, a small window with high Pn embedded within a large window with low Pn implies a locally intensive deforestation (content) nested in a larger region with less forest loss (context). Therefore, the resultant profiles describe deforestation dynamics from local to regional scale (here, from 0.0225 km 2 to 259.5321 km 2 ), and could therefore be incorporated into conservation programs to monitor system dynamics across multiple spatial scales. In F igures 4 3, 4 4, 4 5, the asymptotic behaviours of the five clusters of all three states suggest the scale dependent pattern of percentage developed areas was adequately described with the multiple window sizes used, ranging two magnitudes from 0.15 km to 16.11 km. Though the exact cluster means are different, the relative
39 trend of these profiles behaved similarly. The core developed areas t ypically have higher densities at local scales (small windows) and then decrease as the window size increases, so do the sub developed areas. Both the core forest areas and the sub forest areas have the converse pattern, that is, their densities increase w ith window size increases, where the sub forest areas increase more rapidly than the core forest areas. The transition areas have an interesting profile, the densities increase quickly at relatively small window sizes and then maintain at a relatively stab le state. In sum, the profiles described characterize the fact that intensive fragmentation occurs locally in the MAP region, while a large part of this region is still covered by forest. Focus area study: Rio Branco s of deforestation, a region around Rio Branco, the capital city of Acre, was subset as an illustration (Figure 4 6 ). Rio Branco was one of the first settlements to appear on the banks of the Acre River, and has become the administrative centre of the econ omic and cultural region. According to Table 4 2 the non forest area increases from 4.71% in 1986 to 34.80% in 2010 (where 1% in Acre represents 622.79 km 2 ) and consequently, the forest area shrinks from 95.29% in 1986 to 65.20% in 2010, where 65.20% is close to a critical percolation threshold of 59.28% as proposed by Gardner (1999) with any value below this threshold representing an evident loss of landscape connectivity. The most recent demographic census reported that the population in Rio Branco was 305,954 in 2009 (Brazilian Institute of Geography and Statistics). According to the definition of the US census, urbanized areas are defined as densely settled territories that contain 50,000 or more people. In Table 4 3 the increase of the core develope d area around Rio Branco, including major towns and small settlements, can be re defined as urban expansion
40 which accounts for 3.3% of the subset area in 1986 and increases to 24.6% in 2010. The impacts of urban expansion, including suburbanization area (i .e. sub developed area), transition area, and sub forest area, grows from 40.1% in 1986 (5.6% + 8.8% + impacts stretch over large areas that are almost twice that of the actual u rban fabric and this confirms that urbanization exhibits strong influences beyond the extent of its boundaries, and thus this improved across scale analysis works better than the traditional methods that present forest/non forest only. Summary Window anal ysis is a widely used technique in forest fragmentation analyses, as users can define various convolution kernels to capture composition and configuration information as needed (e.g. Riitters et al., 2000, 2002). Convolution kernels have regular geometric shapes, such as rectangle, circle, hexagon, while fragmented areas or patches in the landscape are not regular and such mismatch between the boundary of window and target object may limit the applications. To address this boundary problem, morphological sp atial pattern analysis was proposed to map forest morphology, such as corridors, scale pattern, and connectivity (MSPA, Soille, 2003; Vogt et al., 2007a, 2007b). Specifically, MSPA is used to solve the problem of ambiguity between forest exterior boundary (edge) and interior fragmentation boundary (perforation) and therefore, as a pixel level analysis MSPA can be used as a complementary reference of the window analysis that prompts the understanding of deforestation and its related problems. There is no si ngle correct scale or level at which to describe a system, nor does that mean all scales serve equally well (Levin, 1992). In the spatial domain, grain and
41 extent are inversely correlated, where the detailed information of fine grain would be inevitably sa crificed at large scales. A heterogeneous landscape at a relatively small scale may become homogenous at a large scale, and vice versa. The moving window n, however various spatial patterns analysed from multiple window sizes may lead to different statistical relationships that are not able to provide consistent references. To overcome such problems, the unsupervised ISO algorithm was used here to integrate various spatial patterns analysed from different window sizes and then expresses such information by profiles ranging from local to regional scales, as such, under study o r the process behind it over an appropriate range of scales. In this paper, the smallest scale of the profile is 0.15 km and such fine resolution would be able to provide a practical guide for conservation programs, and the largest scale of the profile is 16.11 km, where the spatial patterns analysed by such large windows provides an understanding of deforestation at a regional level. Additional window sizes can be added to calculate the information at higher levels, and the largest size should be less than the entire landscape. In recent decades, initialized by the road paving and then enhanced by the urbanization, forest landscapes in the MAP region has been largely affected by rapid land cover change, such that with respect to many social and ecological concerns, the mapping and interpretation of forest cover changes are needed. However, due to spatial autocorrelation (or the first law of Geography), the forest security of the deforested landscape is not uniformly distributed, but rather, exhibits a gradi ent that decreases
42 from the interior forest to the interface between forest areas and cleared areas. Such forest within the moving window and has actually been used as a pattern (configuration) indicator in the paper. Based on such indicators, the five calculated clusters can be used to characterize the complex interaction between forest and non forest. The in crease of the core developed area represents the need to satisfy the growth of urban populations, and the increase of the sub developed area is used to accommodate more urban populations or is being modified to create more agricultural area to feed the lar ger numbers of people. And both the core developed and the sub developed areas are closely related to the urban growth, social and economic developments. The transition area is the most unstable cluster identified, as for the most part either deforestation or reforestation would be occurring here, so the transition area becomes a hotspot that should be considered as an ideal area to measure biodiversity loss and/or urban sprawl. The sub developed area, and its change could indicate a very early signal of either deforestation or reforestation. Finally, the core forest represents the most stable forest distribution and therefore any substantial reduction of this cluster would signal a funda mental transition of the forest area as well as the entire forest system. The accumulated time series forest cover maps from 1986 to 2010 (Cumming et al., 2005, Cumming et al., 2008, Southworth et al., 2011, Marsik et al., 2011, Cumming et al., 2012) reve al that the MAP region is still largely forested, but a fundamental shift in
43 analysis. The fluctuation of these five clusters along both spatial and temporal scales, and part icularly the different dynamics among the three regions, requires policy regulation to be adaptive to different clusters and be flexible for their evolving properties, as well as acting in a timely manner. Hence, such mapping may have an influence in eithe r ecological conservation or economic planning, depending on government policies, either addressing sustainable environme nt or economic growth. the combination of GIS/remote sensing techniques and spatial statistics, and such integration would be useful to provide a framework to study the influences of land use and land cover problems that are not restricted by visible borders. The method used here is applicable and has been verified by a deforested example in the Amazon and hence, such a simple framework should be widely adopted to implement the automatic classification of deforested landscapes, as well be used to monitor and address forest/urban regulation and sustainability
44 Table 4 1 Region Cluster 1986 (%) 1991(%) 1996(%) 2000(%) 2005(%) 2010(%) Acre Core Developed 1.7 3.4 4.0 5.6 9.0 13.3 Sub Developed 3.0 4.2 4.4 5.8 6.7 6.7 Transitio n 4.3 3.9 5.6 5.7 6.6 6.7 Sub Forest 16.0 16.5 16.9 16.1 15.2 14.7 Core Forest 75.0 72.0 69.3 66.8 62.5 58.6 Madre de Dios Core Developed 1.5 1.1 0.5 0.9 0.9 1.7 Sub Developed 2.5 2.0 1.1 2.1 1.9 2.5 Transition 3.5 2.6 1.2 2.1 2.6 2. 5 Sub Forest 17.4 16.8 10.2 20.1 19.4 17.3 Core Forest 75.1 77.5 87.0 74.2 75.2 75.7 Pando Core Developed 0.2 0.3 0.4 0.8 1.0 2.0 Sub Developed 0.5 0.5 0.7 1.4 1.6 2.5 Transition 0.6 3.1 0.8 1.0 1.3 3.3 Sub Forest 5.5 3.7 8.5 10.4 11.2 18.5 Core Forest 93.2 92.4 89.6 86.5 84.9 73.7 where 1% in Acre represents 622.79 km 2 1% in Madre de Dios represents 493.48km 2 and 1% in Pando represents 542.87 km 2
45 Table 4 2 Summary of forest/non forest areas from 1986 to 2010 a round Rio Branco, Acre. Region Cluster 1986 (%) 1991(%) 1996(%) 2000(%) 2005(%) 2010(%) Rio Branco Forest 95.3 88.5 85.0 82.9 73.2 65.2 Non forest 4.7 11.5 15.0 17.1 26.8 34.8 where 1% in Acre represents 622.79 km 2
46 Table 4 3 Region Cluster 1986 (%) 1991(%) 1996(%) 2000(%) 2005(%) 2010(%) Around Rio Branco Core Developed 3.3 8.3 9.9 11.7 19.0 24.6 Sub Developed 5.6 5.3 7.2 8.6 10.4 12.2 Transition 8.8 8.5 11.4 10.8 11.9 11.1 Sub Forest 25.7 26.1 28.8 28.1 26.9 24.5 Core Forest 56.6 51.8 42.7 40.8 31.8 27.6 where 1% in Acre represents 622.79 km 2
47 Figure 4 1. A simulated landscape (spatial resolution: 30m) was convolut ed by 9 different filters, ranging from 55 to 537 537, and these such 9 filtered maps were then stacked to produce a multi band layer stack image and to then run an ISO classification with user defined cluster numbers (5 clusters here from c1 to c5).
48 Figure 4 2. The results of the clustering analysis in the MAP region from 1986 to 2010, using the migrating means clustering algorithm (also known as the ISO cluster) which is based on the similarity of P n (fragmentation proportion) within 9 window sizes. A ) 1986 B ) 1991 .C ) 1996 D ) 2000 E ) 2005 F ) 2010
49 Figure 4 3. Cluster means at all scales in Acre, Brazil from 1986 to 2010. X axis is the side length of different windows (unit: km) and Y axis is P n (fragmentation proportion within each window) A ) 1986 B ) 1991 .C ) 1996 D ) 2000 E ) 2005 F ) 2010
50 Figure 4 4. Cluster means at all scales in Pando, Boliva from 1986 to 2010. X axis is the side length of different windows (unit: km) and Y axis is P n (fragm entation proportion within each window) A ) 1986 B ) 1991 .C ) 1996 D ) 2000 E ) 2005 F ) 2010
51 Figure 4 5. Cluster means at all scales in Madre de Dios, Peru from 1986 to 2010. X axis is the side length of different windows (unit: km) and Y axi s is P n (fragmentation proportion within each window) A ) 1986 B ) 1991 .C ) 1996 D ) 2000 E ) 2005 F ) 2010
52 Figure 4 6. The results of the clustering analysis around Rio Branco, Capital of Acre, Brazil, from 1986 to 2010, using the migrating me ans clustering algorithm (also known as the ISO cluster) which is based on the similarity of P n (fragmentation proportion) within 9 window sizes. A ) 1986 B ) 1991 .C ) 1996 D ) 2000 E ) 2005 F ) 2010
53 CHAPTER 5 INDICATING STRUCTURAL CONNE CTIVITY USING MORPHOLOGICAL IMAGE PROCESSING ANALYSIS Background Large scale deforestation studies became feasible since 1972 with the analysis of satellite imagery (Giles and Burgoyne, 2008), and there is a large collection of compelling examples of de forestation studies (Nepstad et al., 1999, 2001; Southworth and Tucker, 2001; Laurance et al., 2001, 2002; Brook et al., 2003; Nagendra, 2004). When viewing a landscape, we look at its composition and spatial configuration (Li and Reynolds, 1994), that is, both the elements present and how these elements are arranged (Turner et al., 2001; Wu and Hobbs, 2002). Correspondingly, forest fragmentation refers to two processes the first is the area of forest cleared and the second is the pattern of this clearing Both of these factors are important for a variety of ecological reasons species survivability, species dispersal, species migration patterns etc. Deforestation affects the spatial layout of all components of the forest system. Many satellite based trop ical surveys have treated deforestation as a temporal change in forest extent, and not as a spatial property of forests (Foody and Curran, 1994). A given amount of forest can be arranged in many patterns (Riitters et al., 2002), and spatial patterns have b een shown to influence many processes that are ecologically important (Turner, 1989). Substantially different spatial arrangements with the same abundance of forest cover (Vogt et al. 2007a, 2007b) could have totally different ecological functions. Therefo re, quantitative methods are required to detect spatial patterns of deforestation, Reprinted with permission from: Sun, J., & Southworth, J. Indicating structural connectivity in Amazonian Rainforests from 1986 to 2010 using morphological image processing analysis Interna tional Journal of Remote Sensing 34 (14): 5187 5200
54 identify significant changes through time, and relate these patterns to important ecological functions. Based on mathematical morphology theory (Soille, 2003), morphologica l spatial pattern analysis (MSPA) identifies spatial patterns from a single layer binary map, and describes the size, shape and other geometric characteristics of spatial entities. For a forest/non forest map, the existing MSPA algorithm focuses on the spa tial information of forest areas without giving enough consideration to non forest areas. The spatial distributions of non forest areas should be emphasized, because the perforated fragmentation inside the continuous forest and the exterior fragmentation o n forest boundaries would exhibit different influences on landscape connectivity. During the deforestation process, with the same amount of forest loss, the increased perforated fragmentation may split a continuous forest into discrete patches, while exter ior fragmentation may only remove forest areas from the frontier but still maintain a contiguous forest area. It follows logically that widely connected perforated fragmentation has a stronger early warning signal on connectivity loss than that of exterior fragmentation. Therefore, to better understand the impacts on connectivity from both forest and non forest areas, it is necessary to further identify non forest areas and then add the new patterns to the existing MSPA classes. C hapter 5 aims to assess fo rest/non forest changes and examine how these spatial patterns change and how these changes themselves influence landscape connectivity. Such quantitative mapping of forest/non forest patterns at a pixel level would be able to support forest condition eval uation and long term conservation programs in the Amazon.
55 Morphological Spatial Pattern Analysis Existing MSPA MSPA uses a series of image processing routines to identify different spatial patterns at a pixel level. The value of each forest pixel in the output image is based on the connectivity with its neighbors. Connectivity can be set to either four (pixels are connected if their edges touch) or eight neighbors (pixels are connected if their edges or corners touch). Dilation and erosion are two fundame ntal morphological operations in MSPA that traverse the whole image. Based on connectivity, dilation adds pixels to the boundaries of objects in an image, while erosion removes pixels on object boundaries (Soille 2003). The pattern classes identified by MS PA include core, islet, bridge, loop, branch, edge, and pe rforation, which are listed in T able 5 1 (MSPA colu mn) and illustrated in Figure 5 1 (Soille and Vogt, 2009). To stress the advantages that can only be achieved by MSPA, we compare the MSPA classes with the commonly used landscape metrics provided by Fragstats (McGarigal et al., 2002) in T able 5 1. The MSPA operations described are implemented by the software package GUIDOS (version 1.3) Proposed additional MSPA classes: hole and outer background Th e existing seven MSPA classes described previously focus on the configuration of foreground pixels, that is, the spatial arrangement of forest areas, where the background pixels, e.g. non forest areas were simply treated as a single class. MSPA was first p roposed specifically to solve the problem of ambiguity between edge and perforation (Soille, 2003; Vogt et al., 2007a, 2007b), which corresponds to the outer boundaries and inner boundaries. Perforation is the result of the process of making holes in an ar ea, or rather, the interruption of land cover continuity by the
56 formation of openings or gaps in canopy cover (Bogaert et al., 2004). A hole is defined as a connected component of the background that does not contain any pixel of the border of the image, w here perforation pixels of a given connected component (patch) are defined as its boundaries that contact with the hole (Soille and Vogt, 2009). Thus, perforation pixels of a given connected component (patch) are defined as its boundary pixels that are wit hin certain distance to a hole of this connected component where a hole is defined as a connected component of the background that does not contain any pixel of the border of the image (Soille and Vogt, 2009). Perforations introduce fragmentation impacts d eeper into the forest, in comparison to removing the same amount of forest on the exterior boundary of a large forest tract (Riitters et al., 2002). The hole is delineated by perforation pixels, but perforation pixels (object perimeter) and hole pixels (ob ject area) are not a 1:1 relationship and thus using only the perforation feature we cannot provide the information regarding hole extent. Given such perforated holes exhibit rapid impacts to fragmentation of the connected forest by breaking it into discre te patches this is of paramount importance to both measure and understand. In this respect, we propose that background pixels (non forest class) could be further F igure 5 2, where the hole is d elineated by perforation pixels, and outer background is delineated by edge pixels (outer boundaries). In this way we expand upon the use of MSPA in fragmentation and conservation studies by focusing not only on the dominant landscape features (here set a s forest) but also on the background matrix (here non forest). Here we have divided forest cover found within n
57 forest cover found on perimeter of existing forest regions). As such we now locate our forest) class on the landscape, and specifically locate it within the forest matrix thus allowing us to better understand our landscape. MSPA Results We applied MSPA to the three states of the MAP region and set MSPA connectivity rules to the eight neighbor with one pixel as edge width, where we hypothesized that the eight neighbor rule is closer to the real world representation of neighbor influences in this landscape than a four neighbor rule. The MSPA results (without the two newly proposed backgro und classes) are displayed in T able 5 2. Over the past 25 years, we observed an increase in non forest cover for all three states (Acre, Madre de Dios, an d Pando), at the expense of forest areas. Overall, our results across all three regions reveal a landscape still dominated by forest cover up through 2010 (Acre 80.0%, Madre de Dios 97.40%, and Pando 98.02%). There are different trends in the various morphological statistics across the three states, and the different morphological groups will be discussed below, in terms of percentage of state area in each class. Core and Boundary: Edge and Perforation Core represents the interior area of a forest pat ch. For Acre (95.69 % in 1986 to 74.86% in 2010) and Pando (99.21% in 1986 to 95.21% in 2010), the core area has declined across the study period. For Madre de Dios a peak in forest core area was observed in 1996 (97.49%), but overall core area has increa sed from 92.56% in 1986 to 95.01% in 2010. For both Acre and Pando the increase in edge area is simultaneous with the decline in core area. Interestingly, Madre de Dios follows this same pattern up to 2010
58 when, despite an increase in forest core area, th ere is an increase in edge area. In Acre, the area representing the perforated pixels was at a maximum in 1991 (3.89%) and has declined from 1991 through 2005 (1.19%) until 2010 where this value increased slightly to 1.97%. In Pando, with much lower rates of perforation (and forest clearing) overall, 1986 has only 0.24% perforation and this stays below 1% until 2010, where an increase to 1.56% perforation and an increase in edge (from 0.03% in 1986 to 0.52% in 2010) together illustrate a beginning or initia lization of the opening up of this predominantly forested landscape. Madre de Dios has a perforation rate of 2.26% in 1986, this drop to 1% by 1996 and then increases to around 2.53% in 2000 and 2.45% in 2005. This declines once more to around 1.07% by 201 0, although this same period sees a slight decrease in edge. Islet Islets are areas of forest cover which are not connected to any core areas. Islet pixels are extremely low for all three states, Pando (0.01% in 1986 to 0.05% in 2010), Madre de Dios (arou nd 0.04%), and Acre (0.05% in 1986 to 0.22% in 2010). Connector: Bridge, Loop, and Branch Connector pixels within our landscape are here designated as bridge, loop, and branch. For Pando the presence of loop (0.07 in 1986 to 0.32% in 2010), bridge (0.02 in 1986 to 0.17% in 2010) and branch (0.02 in 1986 to 0.19% in 2010) remains very low although these values do increase from the minimum in 1986 to their maximums in 2010. For Madre de Dios the lowest values of loop, bridge and branch features occur in 1996 (0.25%, 0.09%, and 0.06%, respectively), higher values occur in 1986, 2000 and 2005. By 2010 the values are still low (0.36%, 0.20%, and0.15%, respectively) again illustrating these really are quite minor landscape features at this time. Finally Acre,
59 whi ch is much further along in the forest fragmentation process than Pando or Madre de Dios, has higher values in general when compared to the other states but these are still quite low overall (loop 0.22% in 1986 to 0. 45% in 2010, bridge 0.12% in 1986 t o 0.63% in 2010, and branch 0.18% in 1986 to 0.47% in 2010). Background: Hole and Outer background In a perforated landscape, a large (relative to landscape size) and non compact fore st land cover statistic results of hole areas as well as ou ter background areas in Table 5 3. The hole pixels experienced an increase from 2.23 % in 1986 till 2.96% in 20 00, and then decreased to 1.85% in 2010. For both Madre de Dios and Pando, their hole profiles behave similarly: increasing from 1986 (Madre de Dios 0.81%, Pando 0.35%) to 2010 (Madre de Dios 0.94%, Pando 1.37%), with minimal values occurring in 1996 (Madr e de Dios 0.47%, Pando 0.52%). The outer background in Acre has been increasing from 1986 (0.31%) to 2010 (18.12%), while in both Madre de Dios (1.34% in 1986 and 1.66% in 2010 with lowest value 0.44% occurred in 1996) and Pando (0.05% in 1986 and 0.61% in 2010), the profiles of outer background are relatively stable. A typical inverse U shape profile of hole class was observed in Acre. When the total forest cover decreased monotonically over the 25 years in Acre, this U shape profile could indicate that di screte fragmentations first occurred inside the continuous forest at a small scale, and as the deforesting areas continued increasing, some holes started to connect with neighbor holes. During this process, the hole areas had grown large enough to break th rough forest outer boundaries (edge class here), and thus these holes transformed into outer background. Thus, it is naturally to conclude that the
60 spatial information of background pixels needs to be emphasized and classified, as hole and outer background exhibit different impacts during the deforestation process. Specifically, the hole class represents the fragmentation occurs inside forest and its typical inverse U shape profile described reveals a deforestation process: from a dispersed fragmented area to a highly discrete area. Discussion Scale problems in MSPA applications It is noteworthy that F igures 5 1 and 5 2 are illustrations where we have focused on a very small sub region, at a pixel level as MSPA is constrained by image resolution. On the co ntrary, the study here is a real landscape and thus may cause misinterpretations of MSPA classes. Among the defined MSPA classes, islet pixels and connector pixels, including bridge, loop, and branch may not be able to serve any ecological function. Islet class could be understood as small patch; however, they are of little value here when looking at a landscape level. With a spatial resolution of 30 meter, an islet would simply be a very few number of trees remaining isolated in the landscape. The almost n egligible proportion of loop, bridge and branch pixels may indicate that deforestation is not selective but happening in large or more extensive areas. In any case, the very low number of these pixels shows that connectivity is essentially provided through the core areas. Connector pixels alone are not representative for connectivity. For example, a connector may widen over time and consequently turn into a core area; here the connector pixels decreased or were even lost completely while connectivity has in creased. Forest regrowth may result in loop and branch features disappearing but connectivity still increasing.
61 Inferences: A conceptual connectivity model In contrast to traditional fragmentation analyses (Southworth et al., 2001, 2006; Riitters et al., 2 000, 2002; Nagendra et al., 2003, Marsik et al., 2011), the research approach taken here is based on mathematical morphological methods (Soille, 2003, Soille and Vogt, 2009) and provides simultaneous input on the spatially explicit information regarding th e landscapes configuration (i.e. pattern) and composition (i.e., area and number of fragments). The use of mathematical morphologies provides an excellent suite of meaningful and interpretable indices as illustrated in this research. MSPA results provide concrete data to infer connectivity dynamics in the MAP region, and we summed up this inference graphically with a conceptual model ( F igure 5 connectivity completely due to decreased forest percentage. At the beginning of this deforestation process, the landscape is dominated by the core class, where the forest is closely connected. When the total forest percentage continues decreasing, the hole, perforation, and edge classes are dominant, as fragmentation causes more forest to be exposed to non forest surfaces. Such forest landscape is still connected though forest function and service ar e now being lost. From the transition stage to the final stage, the landscape has changed qualitatively from predominantly forest area to non forest area due to extremely low forest cover, where loop, bridge, branch, and islet are the dominant classes left and most non forest areas are outer non forest. At this final stage a landscape is dominated by non forest cover and so as such the forested landscape would be considered lacking in connectivity. Our results reveal that all three regions are still relati vely early in their landscape transitions, and based on the statistical summary
62 (Figure 5 1, T able 5 2), the conceptual model indicates that both Madre de Dios and Pando are in the onset stage, while Acre is transforming into the transition stage ( F igure 5 3). Application: A deforested landscape in the Amazon MSPA has been quite widely used for such uses as mapping spatial patterns, landscape corridors, scale pattern, landscape functional connectivity, and to assess green infrastructure (Vogt et al., 2007a, 2007b, 2009; Riitters et al., 2007, Ostapowicz et al., 2008, Wickham et al., 2010). In both the Europe and the US, forest maps have been morphologically classified to indicate forest ecosystem quality and to assess the amount of fragmentation present in a landscape as well its potential impact. deforestation is growing rapidly and will likely continue to do so in the future, and this in turn, results in the development of a su bstantial body of publications. However, fine or medium resolution land cover products are not available for South America, especially at the scale of the whole Amazon, and consequently, there is no spatial pattern analysis that extends to such a large sco pe. Due to continuous work in the MAP region (Cumming et al., 2005, Cumming et al., 2008, Southworth et al., 2011, Marsik et al., 2011, Cumming et al., 2012), we have accumulated a time series of forest/non forest maps from 1986 through 2010. Thanks to thi s set of data, we can interpret the indications of different MSPA classes from their profiles, and based on the results, build a conceptual model to represent landscape connectivity. Moreover, this paper enriches the paucity of MSPA applications in the Ama emphasizes that the spatial properties of background are as important as those of
63 foreground (forest pixels), such that should be included into conservation programs to promote the understanding of forest dynami c complementarily.
64 Table 5 1. Fragstats indices compared to the MSPA class metrics. MSPA Class MSPA Fragstats Edge External perimeter of core area Perimeter of core area Islet Small patch that is too small to contain a core area. No dif ference between patch size, all patches are treated the same Branch A state between connected and disconnected. A degraded corridor in deforestation, or a construction corridor in reforestation No similar index Loop Fragmentation that happened next to the forest edge No similar index Bridge Connection to different core areas. Patch cohesion (COHESION) is computed from the information contained in patch area and perimeter, which is the proportional to the area weighted mean perimeter area ratio divid ed by the area weighted mean patch shape index. Connectance (CONNECT) is defined on the number of functional joinings, where each pair of patches is either connected or not based on some criterions, and FRAGSTATS computes connectance using a threshold dist ance specified by the user and reports it as a percentage of the maximum possible connectance given the number of patches. Also, traversability (TRAVERSE), based on the idea of ecological resistance, can measure the connectivity. Besides the Fragstats, the re are some papers discuss landscape connectivity and build indices to measure it. Perforation Internal perimeter of core area. There is no index to represent perforation, which occurs inside the forest. Some indices combination can indicate fragmentat ion, such as, total edge (TE), edge density (ED), and landscape similarity index (LSI) can measure the fragmentation. Core Interior area determined by edge width Interior area determined by the edge width
65 Table 5 2. Statistical summary of MSPA classes of the three sides of the MAP region from 1986 to 2010. Region Class 1986(%) 1991 (%) 1996 (%) 2000(%) 2005(%) 2010(%) Acre Developed 2.5 6.6 7.5 9.6 14.4 20.0 Core 95.7 86.8 88.0 84.7 80.9 74.9 Edge 0.2 0.8 1.0 1.4 1.7 1.8 Islet 0.1 0.1 0.2 0.3 0.3 0.2 Branch 0.2 0.4 0.5 0.6 0.6 0.5 Loop 0.2 1.1 0.5 0.6 0.4 0.5 Bridge 0.1 0.3 0.3 0.6 0.4 0.3 Perforation 1.0 3.9 2.1 2.1 1.2 2.0 Madre de Dios Developed 2.2 1.7 1.0 1.8 2.0 2.6 Core 92.6 95.1 97.5 93.6 93.6 95.0 Edge 0.9 0.6 0.2 0.5 0.5 0.6 Islet <0.1 <0.1 <0.1 <0.1 <0.1 <0.1 Branch 0.3 0.2 0.1 0.2 0.2 0.2 Loop 0.8 0.5 0.3 0.8 0.8 0.4 Bridge 1.1 0.6 0.1 0.6 0.5 0.2 Perforation 2.3 1.3 1.0 2.5 2.5 1.1 Pando Developed 0.4 0.7 0.8 1.3 1.9 1.9 Core 99.2 98.6 98.0 97.5 96 .5 95.2 Edge <0.1 0.1 0.1 0.1 0.2 0.5 Islet <0.1 <0.1 <0.1 <0.1 <0.1 0.1 Branch <0.1 <0.1 0.1 0.1 0.1 0.2 Loop 0.1 0.1 0.2 0.2 0.3 0.3 Bridge <0.1 <0.1 0.1 0.1 0.1 0.2 Perforation 0.2 0.4 0.8 0.6 0.9 1.6 where the total area for Acre is about 62279km 2 the total area for Madre de Dios is about 49348 km 2 the total area for Pando is about 54287 km 2
66 Table 5 region from 1986 to 2010. Region Class 1986(%) 1991 (%) 1996 (%) 2000(%) 2005(%) 2010(%) Acre Hole 2.2 2.9 2.8 3.0 2.1 1. 9 Outer background 0.3 3.7 4.7 6. 7 12. 3 18.1 Madre de Dios Hole 0.8 0. 8 0. 5 1.2 1. 2 1.0 Outer background 1.3 0.9 0.4 0.6 0. 9 1. 7 Pando Hole 0. 4 0 5 0.5 1.0 1. 4 1. 4 Outer background 0. 1 0.2 0. 3 0.3 0.5 0.6 where the total area for Acre is about 62279km 2 the total area for Madre de Dios is about 49348 km 2 the total area for Pando is about 54287 km 2
67 Figure 5 1. Example of morphological segm entation of binary patterns. Left: input binary pattern (foreground forest, background developed area). Right: Resulting segmentation. Based on the spatial arrangement, forest pixels are reclassified into seven classes: core, islet, loop, bridge, perf oration, edge, and branch, while developed areas pixels were simply treated as a single class. Note: and Vogt, P., 2009, Pattern Recogn. Lett. 30 (4), 456 459. Adapted with permiss ion.
68 Figure 5 2. Example of MSPA patterns. Left: MSPA segmentation result. Right: Proposed hole and outer background additions. Note: Adapted from 2009, Pattern Reco gn. Lett. 30 (4), 456 459. Adapted with permission.
69 Figure 5 3. Proposed conceptual model of forest structural connectivity and composition/amount versus mathematical morphological indices, as indicators of forest landscape change and traje ctory. Our three study regions are located within this model for Stages I (Pando and Madre de Dios) and Stage II for Acre.
70 CHAPTER 6 MAPPING FRACTALITY DURING THE PROCESS OF DEFORESTATION IN AN AMAZON TRI NATIONAL FRONTIER Background Human society is undergoing a rapid technique revolution that will eventually enter a post industrial era and this drastic transition has already produced threatening changes in every segment of social ecological systems (Gunderson and Holling, 2002) where abrupt global en vironmental changes can no longer be excluded (Scheffer et al., 2001, 2009). In coping with such multi face problems, Rockstrm et al. (2009a, b) numerically estimate a safe opera ting space for humanity, in which they argued that people must stay within the boundaries where transgressing the boundaries could be in land use across the globe may soon be approaching the upper limit of safe operating boundaries, so that much attention should be given in this section to avoid potential catastrophic shift (Carpenter et al., 2009, Scheffer, 2009). While land use and land cover changes are occurring globall y at an increasingly unprecedented rate, impacting almost all major biomes worldwide (Lambin and Geist 2006; Gutman et al., 2004), and among these terrestrial dynamics, deforestation has received considerable attention (Nepstad et al. 1999, 2001; Laurance et al., 2001, 2002; Andersen et al., 2002; Perz et al., 2007a, b, 2008, 2012; Pfaff et al., 2007; Giles et al., 2008; Marsik et al., 2011; Southworth et al., 2011, Sun et al., 2013). Sun, J., Huang, Z., Southworth, J., & Qiu, Y. (2013). Mapping fractality during the process of deforestation in an Amazon tri national frontier. Remote Sensing Letter s 4:6, 589 598.
71 Deforestation has a substantially negative influence on regional and glo bal climates (Malhi et al., 2008). For example, if the trees are removed, evapotranspiration of moisture from the canopy/vegetation back to the atmosphere will be reduced, as such more solar energy will go into the heating of the land, less into evapotrans piration, resulting in a warmer and drier environment (Wilson and Agnew, 1992). The consumption of the cleared forest produces greenhouse gases, such as carbon dioxide, methane and nitrous oxide, which play an important role in exacerbating global warming (Fearnside, 2000; Fearnside et al., 2004). Deforestation may also signal an impending biodiversity loss or even a biotic collapse (Wilcox and Murphy, 1985; Steininger et al., 2001; Nobre et al., 2009) caused by the loss of landscape connectivity, both stru ctural and functional. This is because persistence of spatially structured species populations, meta populations, is strongly related to landscape connectivity (Hanski and Ovaskainen, 2003), and the fragmentation breaks continuous forests into discrete pat ches and then increases the level of patch isolation. Lack of connectivity would reduce the capability of species movement, for example, local extinctions of certain species could not be compensated by the immigrant population from the neighbor patches, an d would interfere with pollination, seed dispersal, wildlife migration and breeding (Estreguil and Mouton, 2009). Deforestation changes forest spatial patterns which are readily apparent to the human, and such spatial patterns can be analyzed by their com position and configuration (Li and Reynolds, 1994), that is, both the elements present and how these elements are arranged (Turner et al., 2001; Wu and Hobbs, 2002). Correspondingly, deforestation refers to two processes the first is the area of forest c leared and the
72 second is the patterns of this clearing. Both of these factors are important for a variety of ecological reasons species survivability, species dispersal, species migration patterns and therefore quantitative methods are required to measur e these two factors, identify their change, and incorporate them during the process of deforestation. Since 1972, large scale deforestation analyses become possible with the support of satellite imagery (Giles and Burgoyne, 2008). The rapid developments o f geo techniques, such as remote sensing and geographic information system (GIS), have advanced the collection and analysis of forest data. With the assistance from remote sensing and other tools, forest studies have progressed rapidly, and in turn, forest ed landscapes have provided many important testing grounds for the development and application of landscape ecological principles, tools and methods (Perera et al., 2007), like Island Biogeography, Meta population. Recently, Kupfer (2012) pointed out that the widely used package FRAGSTATS released two decades ago have revolutionized the landscape analysis and entrenched landscape pattern indices in minds and statistical tool boxes of many landscape ecologists and biogeographers. On the other hand, FRAGSTATS as well as FRAGSTATS like metrics suffer limitations, like the mismatch between patterns and processes (Cardille et al., 2005) where the true relationship between them is important for uncovering the controlling mechanisms of the systems. Hence, in order to better understand the spatial pattern during the process of deforestation, a large collection of techniques and methodologies have been developed to facilitate measurements. Spatial enhancement performed by spatial filters provides a new way to measure the spatial patterns across multiple scales (e.g., Riitters et al., 2000, 2002; Zurlini et al.,
73 2006, 2007). Graph theory, as a well suited method, can provide spatially explicit knowledge, and a growing genre of graph theory applications (Urban and Keitt 2001; Bunn, et al., 2000; Pascual Hortal and Saura, 2006 Minor and Urban, 2008, Minor and Gardner, 2011, Baggio et al., 2011) and software (Pajek, UCINET, LQGraph) have been published or released during the past decades. Moreover, a promising image pro cessing technology, morphological spatial pattern analysis (MSPA) was proposed to understand the spatial patterns at a pixel level, which uses a series of image processing routines to identify different spatial patterns (Vogt et al., 2007a). MSPA has been widely employed in landscape analysis, such as corridor mapping, scale pattern measurement, landscape connectivity mapping, and green infrastructure detection (Ostapowicz et al., 2008; Riitters et al., 2007; Vogt et al., 2007b; Wickham et al., 2010). An en tropy related n ) was introduced to describe the degree of regularity (orderliness) of an ecological time series based on its power spectrum (Zaccarelli et al., 2012, Zurlini et al., 2012). Particularly, by layer stack ing time series NDVI (normalized difference vegetation index) images, Hs n can be used to indicate vegetation dynamics and most importantly, display this information in maps. Sun (2013) introduced an image processing technique, the Canny Deriche filter, to evaluate edge complexity of satellite images, which greatly promote the understainding of landscape heterogeneity across mu ltiple spatio temporal scales. Based on the short review described, it is logical to conclude that there is a clear trend to move fr om simple (FRAGSTATS and FRAGSTATS like) landscape metrics applied to landscapes to the use of graphs (Kupfer, 2012), which provide a more functional approach to landscape quantification and to understand landscape dynamics.
74 Fractal G eometry: A Promised A pproach to Study Deforested L andscapes In deforested landscapes, fragmentation can be envisag ed as a surface growth process and such complex, irregular patterns cannot be described by traditional Euclidean geometry but can be measured by fractal geometry. In this research, the fixed grid scans a pproach is adopted to cartographically represent the fractal structures of developed areas in the MAP region from 1986 to 2010. We will (i) produce time series raster maps of the study area along a 25 year span which display the spatial distribution of fragmentation characterized by fractal dimension at a pixel level; and (ii) based on the created raster maps, a classification scheme will be summarized to interpret fragmentation progress from the less developed areas to the most developed ones. The products of this research will be able to support coarse scale deforestation evaluations and can be incorporated into conservation planning to serve as important references in forest management. Fractal dimension A fractal s tructure is self similar if it looks like itself at different scales (Mandelbrot, 1977, 1983) and can be described by a scaling law. Consider an object in Euclidean dimension D, if we reduce its linear size by 1/r in each spatial direction, its measure wou ld increase to N = r D times the original. Next, taking the log of both sides, log (N) = D log (r), and solving for D, we have equation (6 1) D = log (N) / log (r) (6 1 ) where D is the dimension of the scaling law and it needs not be an integer (The iler, 1990) especially for the real world objects. D can be a fraction, as it is in fractal geometry, known as Hausdorff dimension that is a commonly used fractal dimension.
75 Fractal dimension is used to characterize the fractal objects, where fractal dimen sions are between 1 and 2 for curves and between 2 and 3 for surfaces. In a 2 dimensional plane, the fractal dimension has been characterized as a measure of the space filling capacity of certain patterns. For instance, for an object residing in a 2 dimen sional plane, the more it fills a plane, the closer it approaches the dimension 2. In addition, values of less than 1 are also possible, indicating very small and isolated objects occurring across the 2 dimensional plane. As the value increases the size of the objects increases and also their consolidation with each other such that as we approach the maximum values (2 would be an absolute maximum, but generally the range 1.8 2 approaches this maximum) we are often looking at very consolidated, compact reg ions (Theiler, 1990). Box counting method Box counting is a widely used method to calculate fractal dimension. As a sampling process, the basic procedure is to recursively superimpose a series of regular grids of decreasing box sizes over a target object ( developed areas here), and then record the counting for each successive box, where the counting equals to how many of the boxes are occupied by the target object. Theoretically, the logarithm of N, the number of occupied cells, versus the logarithm of 1/r, where r is the size of one cell (the length of one side), gives a line whose gradient corresponds to the box dimension D. For developed areas in the interpreted images, the fractal dimensions can be used to indicate their spatial filling capacity, that is to what extent developed areas can occupy the entire box area. The more clearings and conversion into developed areas, the closer the fractal dimension approaches the value 2 within each grid. In practice, a series of grids is recursively generated withi n an AOI box which has been laid over developed
76 areas. The size of the grid is taken as r = 1/(2 n ), where n is an integer ranging from 1 to m, and we set the iteration number as m = 7 based on the finest grid, r = 1/(2 7 ) or 1/128 of the original box size, which is fine enough to approximate developed areas in the MAP region and thus able to yield stable results. In this research, we superimposed a matrix of 9 km 9 km non overlapping fixed grids on the developed areas in the MAP region for the year 1986, 1 991, 1996, 2000, 2005, and 2010 and then calculated the fractal dimension by the box counting method within each grid. 9 km is an arbitrary choice, but is based on careful considerations: the forest fragmentation maps are interpreted from Landsat images wi th the smallest pixel being 30 m and thus after recursively dividing the maps (2 7 = 128) to calculate the fractal dimension, the finest grid should be larger than the map resolution 30 m. Moreover, for calculation convenience, 9 km could be divided exactly by 30 m without remainders. Finally, the estimation of the linear regression, lnN = D ln r + c, is implemented in R (the R foundation for Statistical Computing), where c is a constant, both the coefficient D and the coefficient of determination R 2 are recorded to measure the fractal dimension and to describe how well the regression fits. Carto graphic Representation of Fractal Structure across the MAP R egion The maps of fractal dimension in F igure 6 1 support that the landscape is heterogeneous, as the r ange of fractal dimensions are not evenly distributed, and such a property thus indicates that a sing le fractal dimension calculated by a single large AOI (box) would blur much of this information. The resultant figure reveals that the cleared areas in the MAP region have become increasingly compact from 1986 to 2010, where the low fractal dimensions typically represent little to no forest clearing, and higher fractal dimensions are associated with more highly developed areas.
77 The fractal dimension 1.7 is a worldwide average of urban built up areas (Batty and Lon gley 1994, Alfasi and Portugali, 2004), and based on the findings of Chen (2000) who employed a wave spectrum analysis to estimate the fractal dimension and concluded that the closer the fractal dime nsion is to 1.7, the better estimated the results will be, so that the boxes with 1.7 or higher fractal dimensions are highlighted to indicate these regions of highly developed areas. The total highlighted areas increases from 8 grids (each grid is 81 km 2 ) in 1986 to 217 grids in 2010, and this process in terms of timeline was seen first in Rio Branco (capital of Acre) in 1986, and then Cobija (capital of Pando) in 1991, and then finally in Puerto Maldonado (capital of Madre de Dios) by 1996. By 2010, there are 193 highlighted grids in Acre, 15 in Madre de Dios, and 9 in Pando, which corresponds to the different levels and rates of development and infrastructure across the three sides of the border. In addition, to estimate the quality of the fractal analysi s, goodness of fit (R 2 ) is computed and summarized in F igure 6 1, where the large R 2 distributions can be seen to support the robust regression results and fractal structures. Previous research across this landscape (Southworth e t al. 2011; Perz et al. 20 12) has highlighted the increasing patterns of deforestation, which do vary across the national frontiers, but all of which are dominated by forest clearing leading to forest fragmentation, and only extremely limited amounts of forest regrowth. While these levels of deforestation are still quite low (compared to global rates, over 80% forest cover is considered a high amount of forest cover) it is still of great importance to better understand the spatial patterns of these changes, especially across such dy namic landscapes. Intending to understand forest fragmentation as a development process,
78 we propose a framework to classify the continuous fractal dimensions into four categories and this is su mmarized in F igure 6 2. The classification scheme interprets fr actal dimension into four levels based on the amounts, patterns and arrangements of the forest clearings or developed areas. dispersed, not revealing any substantial patterns. O nce we move beyond the level of D we found that developed areas change from qu ite disperse to compact as these regions observe clump consolidation as a common feature, and globally, this level is considered to equate to urbanized regions i.e. hi ghly developed. For this research we also chose to add a Level IV, in part as the data highlighted differences here which we felt were mostly consolidated regions of develop ed areas. Using this classification scheme, developed from global studies, facilitates the comparison of this region both internally and also with other global models of deforestation development. Summary Amazon deforestation has often been seen as the mo st prominent galvanizing image of environmental change during the last several decades, and thus it is not easy to position this research within the literature. Focusing on the MAP region, however, Peralta and Mather (2000) interpreted multi temporal fores t/non forest maps in Acre and then calculated landscape metrics to characterize the impacts of social and economic processes on the development of the forest landscape. Millington et al. (2003)
79 measured different spatial patterns of forest fragmentation in Bolivia, and explained these patterns with reference to specific government policies. Moreover, their results also indicated that most landscape metrics were highly dependent on pixel size. Similarly, Bradley and Millington (2006) discussed the mismatch b etween the spatial scale of biomass burning and the satellite products in Bolivia and Peru and found that an entire component of the fire region in the study region is omitted, despite its importance in the farming systems. The MAP region provides an idea l case for study because it contains high ecological and cultural diversity, while entering a period of rapid change triggered by the paving of the Inter Oceanic highway. The road paving in Brazil has been completed to the Peruvian and Bolivian borders, wh ich has greatly stimulated local economic development. New legislation for logging concessions in Peru and Bolivia has fostered increased logging, and consequently, land values have risen (Wadt et al., 2005). However, this tri national frontier is still la rgely forested and has been designated a global biodiversity hotspot (Myers et al., 2000). In the face of increasing environmental changes, we emphasize that landscape management should focus on the local solutions instead of a global optimization and the refore foster local conservation management as a way to withstand disturbances and to maintain system stability. Currently, the expansion of clearings in the MAP region is accompanied by some early urbanized areas, however, the binary forest/non forest map s produced cannot reflect this change. The theoretically supported fractal dimension of 1.7 (Chen, 2010) serves as a useful value to slice the fragmentation stages, but it just describes the spatial patterns of developed areas and cannot indicate
80 urbanizat ion here, as the input data is not city boundaries but rather is cleared/developed areas which include human settlements, roads, agriculture, bare soil, etc. To better understand the interaction between deforestation a nd urbanization (DeFries et al., 2010 ) in this landscape, independent statistical data, such as economic, social, or demographic data, should be incorporated into any future analysis. The cartographic representations of fractal dimension and classification schemes presented are simple enough to be applied in any fragmented area, which provides an alternative perspective to the spatial methods devised to date. The proposed fixed grid scans methodology pixelizes the entire landscape into a matrix of grids and calculates the fractal dimension w ithin each grid, which expands the application of this approach in coarse scale studies and prompts the understanding of fragmentation by fractals across this heterogeneous landscape. Since forest clearings are not homogeneous across the MAP region, the cl assification scheme proposed further partitions the continuous fractal dimensions into four categories, that is, from no or very small developed areas occurring in a highly forested landscape (level I), to where developed areas have initiated but are still quite patchy (level II), then to a process of increasing forest clearings, where fragmentation converts over to clumping of cleared patches (level III) and then finally into clearing consolidation (level IV). As such, we define the fragmentation/developme nt level according to fractal dimension and this could be utilized to provide useful context to conservation and management programs. Development of these ideas and models could be further developed and tested, as just one potential indicator of significa nt landscape changes. This could then provide important information to conservation or management groups of the types of changes
81 already in progress, their rates of change and ideally an idea of future trajectories (assuming no change in current patterns/t rends). In addition, different levels of deforestation identified by the model require conservation management to be flexible and adaptive. Such planning tools seem imperative in these fragile landscapes and biodiversity hotspots.
82 Figur e 6 1. Spatio temporal fractal analysis of the cleared areas in the MAP region for the years ( A ) 1986 B ) 1991 .C ) 1996 D ) 2000 E ) 2005 F ) 2010 The boxes with fractal dimensions equal or above 1.7 are highlighted in light blue to indicate intensively de veloped areas and their evolving processes. To maintain the data quality and result accuracy, some grids were removed due to cloud influence. In addition, the distributions of R 2 of the linear regression in the box counting calculation within each grid are displayed, where a large R 2 suggests a robust fractal property.
83 Figure 6 2. The evolving pattern of development as indicated by density sliced fractal dimension in the MAP region for the years A ) 1986 B ) 1991 .C ) 1996 D ) 2000 E ) 2005 F ) 2010 where the level I decreases from 63.4% in 1986 to 37.6% in 2010, and level II, III, and IV increase from 12.7%, 0.4%, 0.0% in 1986 to 28.5%, 4. 6%, and 5.8% in 2010 and 1% represents 1655.9 km2 in the entire MAP region.
84 CHAPTER 7 CONCLUSIONS Operating w ithin the resilience paradigm, this study assesses and reports spatial pattern changes during the process of deforestation in the Amazonian rainforest, aimed to guide forest evaluation and conservation programs. Since forest patterns are readily detectable from remotely sensed images, and are thus amenable to study at broad scales, they have provided many opportunities for studying the complex relationship between patterns and processes in many areas. After a brief review of knowledge on resilience theory t o address spatial pattern process likely to have ecological effects during the process of deforestation, three measures were illustrated including their roles as proxies to indirectly assess forest landscape resilience. What Can We Tell the People from Th is R esearch? Changes in forest spatial patterns are readily apparent to the human eye, which alters the composition and configuration of forest landscapes and this, in turn, influences many species and ecosystem processes. Therefore, the importance of land scape level considerations in the management and conservation of forested landscapes has become increasingly important, and a variety of stakeholders are involved. The discipline of landscape ecology has developed concepts and methodologies that can be dir ectly applied in forested landscapes, but to be most useful, these need to be more widely available. Included are principles and theory that relate spatial patterns at multiple spatial and temporal scales to ecological and anthropogenic drivers; methods fo r quantifying and evaluating spatial patterns in both discrete and continuous variables; and models for projecting the consequences of alternative scenarios.
85 on the spatial pa ttern changes and spatial distribution of the forest landscape, including forest areas and developed areas, which is not only useful for the MAP region and the Amazon, but also applicable for deforested landscapes around the world. However, it can be obse rved that the difficulties faced by forestry professionals in their attempts to apply these landscape ecological concepts, models real world problems. It is important to develop and synthesize landscape ecological knowledge and transfer it to users to ensu re appropriate application: The product of C hapter 4, measured by a combination of moving window analysis and ISO unsupervised classification, can be used to evaluate the influences of forest clearings expressed by land cover changes across multiple spati al scales. It notifies people that forest clearings are not limited by the physical boundaries but decreasingly spread over a larger scope, and the mappings of these impacted areas and their associated impact levels help make forest conservation programs b etter able to apply these data, both locally and effectively. In C hapter 5, measures of forest spatial patterns at a pixel level were based on the application of MSPA methods, and the existing MSPA method was improved by adding two classes which focus on the spatial properties of developed areas. This product strengthens the understanding of forest configuration by emphasizing spatial properties. In addition, the proposed conceptual model can be used to indicate the dynamics of landscape structural connec tivity based on the MSPA results. Besides the two additi onal tools of MSPA proposed in C hapter 5, the spatial patterns of developed areas were further characterized by fractal geometry in C hapter 6. The fractal dimensions were calculated across the entir e landscape by adopting a fixed grid scans approach and summarized by a classification cartographic representation of fractal structures provides flexible and adaptive informa tion about forest clearings to forest managers and to more regional land use planning. By providing three measures and illustrating them with the real world data in the MAP region, this research makes a significant contribution to the applications of
86 lan dscape ecology in deforested landscapes as well as promoting the understanding of forest resilience with particular attention being paid to the spatial patterns. These results should encourage the application of the methods and measures for large areas ass essment and reporting. For biodiversity purposes, it will be essential to further complement local change in forest pattern with field based assessments of forest quality. This was not feasib le to undertake in this study. Knowledge transfer is necessary b ut not sufficient for applications of landscape ecology to be successful. The chapters within demonstrate the potential possibility to translate the science of landscape ecology into practice. This research should be of broad interest to all those interest ed in understanding and managing forested landscapes, and in particular to current and future researchers committed to making the knowledge they develop available to a wider audience of users. Resilience Prompts the Understanding of Unexpected S hifts Huma n society is undergoing a rapid technical revolution that will eventually enter a post industrial global information age and this drastic transition has already produced threatening changes in the every segment of social ecological systems where abrupt glo bal environmental changes can no longer be excluded. In coping with such multi faceted prob lems, Rockstrm et al. (2009 ) proposed an earth system framework of a holisti c perspective (Naveh, 2000, 2001), and they argued that people must stay within the boundaries, as transgressing the boundaries could be dangerous and catastrophic. Therefore, plans to understand complex problems of social ecological systems should always consider system boundaries (Carpenter et al., 2009) and regime shift prevention.
87 Regime shifts are large, sudden changes in ecosystems that last for substantial periods of time (Scheffer et al., 2001, Carpenter, 2003, Biggs et al., 2009), caused by the com bination of the internal resilience of the system and the magnitudes of external forces (Folke et al., 2004). Regime shifts in ecosystems can cause large losses of requi re drastic and expensive intervention, which is known as system hysteresis. From a practical point of view, hysteresis is important, as it implies this kind of catastrophic transition is not so easy to reverse (Scheffer et al., 2001, 2009; Scheffer, 2009) and therefore, managers should be especially vigilant about trying to prevent or avoid regime shifts. Based on experiments, historical data, and theoretical models, in an attempt to signal the surprising in the systems, some early warning indicators have b een obtained to announce the coming regime shifts in advance (Carpenter et al., 2011), such as rising variance (Carpenter and Brock, 2006), rising autocorrelation (Dakos et al. 2008), and extreme changes in skewness and variance (Guttal and Jayaprakash, 20 08, 2009). But, when these indicators become statistically significant and are able to be detected and verified, systems have already lost a certain amount of resilience. For those systems with severely degraded resilience, they may not be able to withstan d external disturbances (could be a slight disturbance) even after human interventions (such as disturbance prevention managements), and could therefore be tipped into an alternative regime, often less desirable. Several studies have shown that conventiona l approaches to resource and ecosystem management are not always working, like disturbance control strategy (Scheffer et al., 2001), and may indeed make
88 (Walker and Salt, 2006) in our creation of tools and techniques developed to monitor and warn of just such potentially catastrophic and irreversible regime shifts or tipping points. Embracing Resilience T hinking The concept of resilience has evolved considerably since Holl seminal paper, and while Walker et al. (2004) pointed out that different interpretations of what is meant by resilience can cause confusion, they then attempted to clarify this cs. However, a lot of research uses resilience now, and it seems to mean different things, depending on the science (e.g., Vienna Circle, Karl Popper, Frankfurt school) that declares conceptual clarity as essential for scientific research. After comparing with the traditional school of ommunication and pragmatic problem solving. Strunz (2012) proposed that resilience thinking can be framed to take advantage of its vagueness without forgoing precision where it matters, which greatly enriches and solidifies the paradigm of resilience. To a ppreciate the complex ecosystem problems like the potential regime shift described, the vagueness of resilience thinking may provide a road map that is able to promote interdisciplinary integration which addresses the dynamics and development of complex so cial ecological systems and solves the problems from a holistic perspective (Naveh, 1998, 2000; Folke et al., 2010; Walker and Salt, 2006). In the face of increasing environmental changes, it is important to emphasize that ecosystem management should conc entrate on not only the disturbance preventions
89 but also the gradual changes that affect system resilience and thus building resilience to withstand disturbances and to maintain a desired system state is likely to be an alternative pragmatic and effective system dynamics and the way to enhance resilience, where the system is a ball in the basin, and the resilience is the maximum perturbation that can be taken without tipping the ball to an alternative basin (Walker et al., 2004; Scheffer et al., 2001). To increase system resilience equals making the desirable basin wider and deeper, consequently, reducing the probability of the ba ll b eing tipped outside the basin. As has become clear, the dissertation is critical towards using the resilience framework in explaining deforestation, and it has answered key questions of our subject. Therefore, it is useful to take up resilience theory in geography study, which has added numerous values to geography, both physical geography and human geography.
90 LIST OF REFERENCES Achard, F., Eva, H.D., Stibig, H. J., Mayaux, P., Gallego, J., Richards, T., and Malingreau, J. f deforestation rates of the world's Science 297 (5583), 999 1002. in time versus planning just in Cities 21, 29 39. Andersen, L.E., Granger, C.W.J., Reis, E.J., Weinhold D., and Wunder, S. (2002), The dynamics of deforestation and economic development in the Brazilian Amazon Cambridge: Cambridge University Press. connectivity and predatorpr Landscape Ecology 26, 33 45. Batty, M., and Longley, P. (1994), Fractal Cities Waltham: Academic Press. detecting an impending regime shift in tim Proceedings of the National Academy of Sciences 106, 826 831. Environmental management 33( 1) 62 73. International Journal of Remote Sensing 27 2221 2253. atastrophic extinctions follow Nature 424 (6947), 420 426. change in corals across shallow reef flats in Ko Phuket, Thailand, due to the effe Marine Biology 141, 24 29. Journal of Environmental Management 59 ( 4), 265 278. Bioscience 55, 983 988.
91 Carpenter, S. Models in Ecosystem Science eds. C. Canham, J. Cole and W. Lauenroth, Princeton: Princeton University Press. Ecology Letters 9, 311 318. http://dx.doi.org/10.1111/j.1461 0248.2005.00877.x Carpenter, S.R., Cole, J.J., Pace, M.L., Batt, R.D., Brock, W.A., Cline, T., Coloso, J., Hodgson, J.R., Kitchell, J.F., Seeke ll, D.A., Smith, L., and Weidel, B. (2011), Science 332, 1079 1082. Carpenter, S., Mooney, H., Agard, J., Capistrano, D., DeFries, R., Diaz, S., Dietz, T., Duriappah, A., Oteng Yeboah, A., Pereira, H. M., Perrings, C., Reid, W. V., Proceedings of the National Academy of Sciences 106 1305 1312. Carpente social Ecosystems 8, 941 944. Chen, Y. (2010 ), Exploring the fractal parameters of urban growth and form with wave spectrum analysis Discrete Dynamics in Nature Society Article ID 974917, 20 page. doi: 10.1155/2010/974917. Cumming, G.S., Barnes, G., Perz, S., Schmink, M., Sieving, K., Southworth, J., Binford, the empirical m Ecosystems 8, 975 987. in Complexity Theory for a Sustainable Future (Complexity in Ecological Systems) eds. J. Norberg, and G.S.Cumming, New Y ork: Columbia University Press. complexity in fragmenting Amazonian rainforests: Do feedbacks from edge Ecological Complexit y 11 67 74. Dakos, V., Scheffer, M., van Nes, E. H., Brovkin, V., Petoukhov, V., and Held, H. Proceedings of the National Academy of Sciences 105 (38), 14308 14312. DeFries, R. S., Houghton, R. A., Hansen, M. C., Field, C. B., Skole, D., and Townshend, Proceedings of the National Academy of Sciences 99 ( 22), 14256 14261.
92 urban population growth and agricultural trade in the twenty Nature Geoscience 3 178 181. conomic growth: An explanation for regional Journal of Comparative Economics 29 95 117. Oikos 65 169 175. use/land cover International Journal of Remote Sensing 26, 1285 1287. Ecolog ical Applications 18 (7), 1810 1825. Measuring and reporting on forest landscape pattern, fragmentation and connectivity in Europe: methods and indicators Research Centre, European Communities Scientific and Technical Research series EUR 23841EN; ISSN 1018 5593, pp. 69. Economic Development and Cultural Change 52 395 421. and global warming: carbon stocks in Forest Ecology and Management 80 (1 3), 21 34. use change: Greenhouse gas emissions from biomass burning, decompos ition and soils in forest Climate Change 46 115 145. Climatic Change 66 (2 1), 1 8. http://dx.doi.org/10.1023/B:CLIM.0000043174.02841.23 nd Conservation Biology 19 (3), 680 688. Folke, C., Carpenter, S.R., Walker, B., Scheffer, M., Chapin, T., and Rockstrm, J. Ecology and Society 15(4 ): 20. [online] URL: http://www.ecologyandsociety.org/vol15/iss4/art20/
93 Folke, C., Carpenter, S.R., Walker, B., Scheffer, M., Elmqvist, T., Gunderson, L., and shifts, resilience, and biodiversity in ecosystem Annual Review of Ecology, Evolution, and Systematics 35 557 581 regenerative stage using remotely sensed data, Journal of Biogeography 21, (3), 223 244. Forman, R.T.T. Landscape Ecology 10 133 142. use history as long term broad scale di Ecosystems 1 96 119. Landscape ecological analysis: issues and applications ed s. J.M. Klopatek, R.H. Gardner, New York: Springer Verlag, pp 280 303. Economic Development and Cultural Change 52 159 185. ole, D.L. and Tucker, C.J. 1993: Tropical deforestation and habitat fragmentation in the Amazon: satellite data from 1978 to 1988. Science 260, 1905 Progress in Physical Geography 32 (5), 575 580. ion: Human behavior in a Advances in Historical Ecology eds. W. Ballee, New York: Columbia University Press, pp. 213 231 induced development: Evidence from Sri L World Development 36 2371 2389. Ecology and Society 15 (2): 18. [online] URL: ht tp://www.ecologyandsociety.org/vol15/iss2/art18/ Gunderson, L, and Holling. C.S. (2002), Panarchy: Understanding Transformations in Human and Natural Systems 1st Edition. Washington D.C.: Island Press. Gunderson, L., and Pritchard L. (2002), Resilience and the behavior of large scale systems Washington D.C.: Island Press.
94 Gutman, G., Janetos, A., Justice, C., Moran, E., Mustard, J., Rindfuss, R., Skole, D., and Turner, B. L. (2004), Land change science: Observing, monitoring, and understanding trajec the Netherlands: Kluwer Academic. Theoretical Ecology 2 (1), 3 12. Ecology letters 11 (5), 450 460. viability in and aroun Conservation Biology 16 1 12. Hanski, I. (1999), Metapopulation ecology Oxford, Oxford University Press. Theoretical Population Biology 64 1 19 127. Annual Review of Ecology and Systematics 4, 1 23. Holling, C. Engineering within ecological constraints, ed s. P. Schulze, Washington, D.C.: National Academy Press, pp. 31 44. scale degradation of a Science 265, 1547 1551. strategies to mitigate impacts Philosophical transactions of the Royal Society of London. Series B, Biological sciences 363 (1498), 1881 1888. Kupfer, J. The SAGE Handbook of Biogeography eds. A.Millington, M. Blumler, and U. Schickhoff, Newbury Park: Sage Press, pp. 57 74. Lambin, E. F., and Geist, H.J. (2006), Land use and land cover change: Local processes and global impacts Berlin: Springer Environmental Conservation 28 (04), 305 311. Laurance, W.F., Albernaz, A.K., Fearnside, P. M., Vasconcelos H. L., and Ferreira l. V. (2004), Science 304 21.
95 Laurance, W.F., Albernaz, A.K., Schroth, G., Fearnside, P. M., Bergen, S., Venticinque, Journal of Biogeography 29( 5 6), 737 748. Levin, S.A. Ecological Applications, 2 103 106. Lewontin, R.C. (1969), The Meaning of Stability, Diversity and Stability in Ecological Systems New York: Brookhave n National Laboratory. Ecology 75, 2446 2455. Malhi, Y., Roberts, J.T., Betts, R.A., Killeen, T. J., Li, W., and Nobre, C.A. (2008), Science 319, 169 172. Mandelbrot, B.B. (1977), Fractals, Form, Chance and Dimension San Francisco: W.H. Freeman and Company. Mandelbrot, B.B. (1983), The Fractal Geometry of Nature New York: Freema n and Company. Marsik, M., Stevens, F. and patterns of land cover change and fragmentation in Pando, northern Bolivia, Progress in Physical Geography 35(3), 353 374. Mas, J. F., Ve lzquez, A., Daz Gallegos, J.R., Mayorga Saucedo, R., Alcntara, C., Bocco, G., Castro, R., Fernndez, T., and Prez International Journal of Applied Earth Observation and Geoinformation 5, 249 261. Nature 267 (5628): 471 477. http://dx.doi.org/10.1038/269471a0 Computer software program produced by the authors at the University of Massachusetts Amherst. Available at the following w eb site: http://www.umass.edu/landeco/research/fragstats/fragstats.html. Theory framework for evaluating Conservation Biology 22( 2), 297 307.
96 Minor, E.S., and Ecological Applications 21 (3), 739 749. Moran, E.F., and Ostrom, E. (2005), Seeing the Forest and the Trees: Human envi ronment Interactions in Forest Ecosystems Cambridge: MIT Pres s, pp: 456. Myers, N., Mittermeier, R. A., Mittermeier, C.G., da Fonseca, G.A.B., and Kent, J. Nature 403( 6772), 853 858. Nagendra, Agriculture, Ecosystems and Environment 101 (2 3), 111 115. of Landscape Ecology 18 (2), 141 158. evolution towards a Post Industrial symbiosis between human society and Restoration Ecology 6 (2), 135 143. Landscape and Urban Planning 50 (1 3), 7 26. ltifunctional Landscape and Urban Planning 57, 26 284. Nepstad, D., Carvalho, G., Cristina Barros, A., Alencar, A., Paulo Capobianco, J., Bishop, J., Moutinho, P., Lefebvre, P., Lopes Silva, U. and Prins, E. (2001), Forest Ecology and Management 154, 395 407. Nepstad, D ., Verssimo, A., Alencar, A., Nobre, C., Lima, E., Lefebvre, P., Schlesinger, P., Potter, C., Moutinho, P., Mendoza, E., Cochrane, M., and Brooks V. (199 9), Nature 398 505 508. Nobre, P., Malagutti, M., Urbano, D.F., de Almeida, R.A.F., and Giarolla, E. (2009), Journal of Climate 22 (21), 5686 5697. L., Waide, J.B., and Allen, T.F. H. (1986), A hierarchical concept of ecosystems Princeton: Princeton University Press.
97 in Applying landscape ecology in biological conservation eds. K. Gutzwiller, and T.T. Forman, New York: Springer, pp. 381 404. scale on morphological spatia Landscape Ecology 23( 9), 1107 1117. Pascual based landscape connectivity indices: towards the priorization of habitat patches Landscape Ecology 21 (7), 959 967. extractive reserves of Acre, Amazonia from satellite imagery: A landscape International Journal of Remote Sensing 21, 2555 2570. Perera, H.A., Buse, J.L. and Crow, R.T. (2007), Forest Landscape Ecology: Transferring Knowledge to Practice Berlin: Springer Verlag in the pan Amazon B asin: A comparison of Brazil, Bolivia, Colombia, Ecuador, Environment, Development and Sustainability 7 (1), 23 49. Perz, S.G., Brilhante, S., Brown, F., Caldas, M., Ikeda, S., Mendoza, E., Overdevest, C., Reis, V., Reyes, J. F., Rojas D., Schmink, M., Souza, C., Walker, R., May 2008. Philosophical Transactions of the Royal Society B: Biological Sciences 363 (1498), 1889 1895. Perz, S.G spatial process of unofficial road building in the Amazon: Socioeconomic and biophysical Development and Change 38( 3), 529 551. Perz, S.G., Overdevest, C., Arima, E.Y., Caldas, M. M., and Walker, R.J. (2007b), Environmental Conservation 34, 112 121. Perz, S.G., Shenkin, A., Barnes, G., Cabrera, L., Carvalho, L. A., and Castillo, J. (2012), Social Indicators Research 106, 259 285. Pfaff, A., Kerr, S., Lipper, L., Cavatassi, R., Davis, B., Hendy, J., and Sanchez Azofeifa, G.A. (200 Land Use Policy 24, 600 610. Pimm, S.L. (1984), The Balance of Nature Chicago: University of Chicago Press, Chicago.
98 ges using remote sensing and GIS in Rize, North Sensors 8, 6188 6202. Environmental management 35( 4), 483 492. Riitters, K.H., Vogt, P., S Landscape Ecology 22 1033 1043. Frontiers in Ecology and the Environment 1, 125 129. Conservation Ecology 4 (2): 3. Riitters, K.H., Wickham, J.D., O'Neill, R.V., Jones, B.K., Smith, E. R., Coulston, J.W., Ecosystems 5, 815 822. E cological Indicators 9, 107 117. Rockstrom, J., Steffen, W., Noo ne, K., Persson, A., Chapin, F.S., Lambin, E.F., Lenton, T. M., Scheffer, M., Folke, C., Schellnhuber, H.J., Nykvist, B., de Wit, C. A., Hughes, T., van der Leeuw, S., Rodhe, H., Sorlin, S., Sn yder, P. K., Costanza, R., Svedin, U., Falkenmark, M., Karlberg, L., Corell, R. W., Fabry, V. J., Hansen, J., Walker, B., Liverman, D., Richardson, K., Crutzen, P., and Foley, J. A. (2009), A safe operating space for humanity Nature 461(7263), 472 475. S aura, S., Estreguil, C., Mouton, C. and Rodrguez to assess landscape connectivity trends: application to European forests (1990 Ecological Indicators 11, 407 416. Scheffer, M. (2009), Critical transitions in n ature and society Princeton; Princeton University Press. Scheffe r, M., Bascompte, J., Brock, W. A., Brovkin, V., Carpenter, S.R., Dakos, V., Held, H., van Nes, E. H., Rietkerk, M., and Sugihara, G. (2009), Early warning signals for critical transitions Nature 461(7260), 53 59. Scheffer, M., Carpenter, C.R., Foley, J., Folke, C., and Walker, B. Nature 413, 6856. Scheffer, M., and Jeppesen, Regi me shifts in shallow lakes Ecosystems 10( 1): 1 3.
99 Schmink, M., and Wood, C.H. (1992), Contested frontiers in Amazonia New York: Columbia University Press the Ama Science 260 (5116), 1905 1910. Soille, P. (2003), Morphological image analysis: principles and applications 2 nd edition. Berlin: Springer Verlag. Pattern Recognition Letters 30( 4), 456 459. The Professional Geographer 58 (4), 406 420. Southworth, J., Marsik, M., Qiu, Y., Perz, S., Cumming, G.S., Stevens, F., Rocha, K., across the Tri Remote Sensing 3, 1047 1066. and socioeconomic factors on forest cover change in the mountains of western Mountain Research and Development 21 (3), 276 283. Steininger M.K., Tucker, C.J., Townshend, J.R.G., Killeen, T.J., Desch, A., Bell, V., and Environmental Conservation 28( 02), 127 134. from philosophy of Ecological Economics 76 112 118. sensing vegetation index http://dx.doi:10.1080/1747423X.2012. 756071 process of deforestation in an Amazon tri Remote Sensing Letters http://dx.doi:10 .1080/2150704X.2013.771827 Oikos 68, 571 572. Journal of the Optical society of Ame rica A 7, 1055 1073. Nature 367, 363 365.
100 induced patterns of forest Environmental Ma nagement 37, 788 801. Annual Review of Ecology and Systematics 20 (1), 171 197. Landscape Ecology in the theory and practice Berlin: Springer Verlag. Ecology 82( 5), 1205 1218. World Development 30 575 58 9. Vogt, P., Ferrari, J.R., Lookingbill, T.R., Gardner, R.H., Riitters, K.H., and Ostapowicz, Ecological Indicators 9 (1), 64 71. Vogt, P., Riitters, K., Estreguil, C., Kozak, J., Wade, T., and Wickham, J. (200 7a), Landscape Ecology 22( 2), 171 177. Vogt, P., Riitters, K., Iwanowski, M., Estreguil, C., Kozak, J., and Soille, P. (2007b), Ecological Indicators 7( 2), 481 488. scale Environmental Modelling and Software 24 1252 1256. Wadt, L.H.O., Kainer, K.A., and Gomes cture and Forest Ecology and Management 211, 371 384. Forest succession: Concepts and application e ds. D.C. West, H. Shugart and D.B. Botkin New York: Springer adaptability and transformability in social Ecology and Society 9 (2), 5. [online] URL: http://www.ecologyandsociety.org/vol9/iss2/art5/ Walker, B., and Salt, D. (2006), Resilience Thinking: Sustaining Ecosystems and People in a Changing World Washington D.C.: Island Press. Wear, Ecology and Society 9( 5): 4.
101 change in forest fragmentation at multi Landscape Ecology 22: 481 489. http://dx.doi: 10.1007/s10980 006 9054 6 of green infrastructure and change for the conterminous United States using morph Landscape and Urban Planning 4 (3 4), 186 195. Wiens, J. Applying Landscape Ecology in Biological Conservation eds. K.J. Gutzwiller, New York: Springer, pp. 3 21. feedback switches in plant Advances in Ecological Research 33 263 336. The Americ an Naturalis t 125 879 887. Landscape Ecology 17 (4), 355 365. erview Information Sciences 6, 1 5. Zaccarelli, N., Li, B ecological time Ecological Indicators http://dx.doi:10.1016/j.ecolind.2011.07.008 Ecological Indicators 8, 841 853. Zurlin disorder in social ecological landscapes to foster adaptive capacity and Landscape Ecology http://dx.doi: 10.1007/s10980 012 9763 y. Zurlini, G., Riit Landscape Ecology 22 (5), 705 721. Zurlini, G., Riitters, K.H., Zaccarelli, N., Petrosillo, I., Jones, B.K., and Ros si, L. (2006), Ecological Complexity 3, 119 128. Zimov, S.A., Chuprynin, V.I., Oreshko, A.P., Chapin III, F.S., Reynolds, J.F., and tundra transition: An herbivore driven biome shift The American Naturalist 146, 765 794.
102 BIOGRAPHICAL SKETCH study Geographic Information System at Northwest University, China and received B.A.Sc. in GIS in 2005. Jing Sun entered the Master program of Human Geography at Northwest University, China in 2005. Jing worked with Xin jun Yang and Jun Wang, two scientist s who combine a strong quantitative approach with an equal devotion to focus on applied questions. Jing studied resilience theory which introduced him to a new arena and he conducted fieldwork many times in Loess Plateau region and published one of the fir st resilience papers in China (been cited 27 times as of April 2012). In 2008, Jing Sun ecological systems to drought in semi ng this time convinced Jing that he need attend a Ph.D. program to continue resilience study. Jing Sun entered the Ph.D. program in the Department of Geography, University of Florida to continue and expand upon his work with Jane Southworth. He received h is Ph.D. from the University of Florida in the summer of 2013.