A Cellular Automata Approach to Land Use Change Modeling in Northeastern Thailand

Transcription

1 A Cellular Automata Approach to Land Use Change Modeling in Northeastern Thailand Scenarios of Policy Intervention in the Cultivation of Upland Field Crops and Patterns of Land Use/Land Cover Change using Spatial Simulation Models Stephen J. WALSH, Barbara ENTWISLE, Ronald R. RINDFUSS and Philip H. PAGE KEYWORDS The cultivation of upland field crops, primarily cassava and sugar cane, in Nang Rong district, northeast Thailand, beginning in the mid- to late-1960s, has helped transform a once forest-dominated landscape to one dominated by agriculture. Today, paddy rice is cultivated throughout the lowlands, field crops and a fragmented forest matrix comprise the uplands, and fruit trees, rubber plantations, and vegetable gardens are among the crops dispersed around nuclear villages. Distributed along a topographic terrace system, upland and lowland crops are cultivated relative to environmental and economic opportunities and geographic access, although crops grown in marginal settings may not be sustainable. Relying upon a remote sensing image time-series, a longitudinal social survey, and GIS coverages, a cellular automata (CA) model is described that is used to characterize land use/cover (LULC) change patterns through specified initial conditions, neighborhood associations, and transition or growth rules. Results of four scenarios or experiments are described that perturb the base LULC change model of cassava, forest, and rice by imposing production quotas in the cultivation of cassava. Derived for the period , CA model results for the scenarios are compared to a time-series of Landsat satellite classifications of LULC using images of simulation runs and plots that describe trends in the composition and spatial organization of cassava, forest, and rice for each scenario. Results are interpreted within a population-environment context in which people, place, and environment are integrated in complex ways. The demise of forest at the expense of expanded lowland paddy rice and upland field crops is a central story of the region. Quotas on the production of cassava alter the trajectories of forest change and result in more consolidated stands over the period of the simulations. Cassava consolidation and in-filling in the extensive uplands of the southwest portion of Nang Rong district is a persistent outcome of the simulations for the various scenarios tested. Complexity theory, cellular automata, scenarios of land use/land cover change, model validation and uncertainty, Northeast, Thailand. INTRODUCTION Nang Rong district, northeast Thailand has experienced rapid social and environmental change over the past five decades. While the region has been cultivated for over 2,000 years, deforestation and agricultural extensification in upland settings primarily began in the mid to late 1960s, driven by a European demand for calorie-rich animal feed. Extensive tracts of land were converted from forest to cassava. Large fields, bordering forests, and small to medium forest remnants now characterize the uplands, while small rice paddies in the lowlands have coalesced into an extensive rice producing region that is interspersed by isolated trees, small patches of forest, and riparian forest corridors, as well as by population settlements, ponds, orchards, and other mixed agricultural uses. The topography of the region features a series of shallow hills interspersed by lowlands. Upland agriculture predominates in hillier, drier areas, while rain-fed paddy rice dominates in moister, lowland areas. The topographic terrace structure of the region influences the ecological gradients and the resulting environmental suitability of the landscape for lowland versus upland agriculture. Functional territories of nuclear villages are defined by the location of land parcels cultivated by villagers, the spatial distribution and establishment dates of other villages, the nature of administrative village splits, and by other social and geographical factors that affect a village s site and situation. While the majority of villagers in the region are lowland, paddy rice cultivators, a variety of alternative upland crops (e.g., cassava and sugar cane) provide forms of cash income, although they are subject

2 2 to dramatic price fluctuations, particularly cassava, as a consequence of market globalization, as well as the setting of production quotas on the cultivation of cassava that are periodically implemented by the Thai government. Climate variation is a another important factor that affects land use patterns and human behavior, as well as the nature of their feedbacks. Monsoonal rains, normally extending from June through November, are subject to substantial variations in their timing and precipitation levels that have implications for population migration patterns, land use practices, and the redefinition of resource marginality and land suitability for agricultural practices. In short, LULC patterns in Nang Rong district are characterized by complex interactions between population and the environment. The basic intent of this research is to: (a) model LULC change for northeast Thailand within a CA modeling environment that examines deforestation and agricultural extensification in Nang Rong district by focusing on the cultivation of upland cassava and plausible scenarios of agricultural production in which government policies restrict the areal extent and timing of cassava production through imposed quotas, (b) considers hypothesized drivers of LULC change by perturbing a base model of LULC change derived through CA approaches using a longitudinal social survey, a rich remote sensing image time series, and terrain descriptors to set conditions and run the models; and (c) describes the nature of uncertainty in characterizing the composition and spatial organization of modeled (expected) versus mapped (observed) LULC patterns across space and through time as a consequence of the selected scenarios that vary production quotas in the cultivation of cassava. Using a classified land use/land cover (LULC) image time-series of Landsat TM (Thematic Mapper) and MSS (Multi-Spectral Scanner) images, we have created growth or transition rules for upland field crops, and lowland paddy rice. We have relied upon GIS spatial coverages to categorize geographic accessibility, terrain conditions, hydrography, resource endowments, and village location and characteristics at the 30-m cell that corresponds to the pixel dimension of the remote sensing images. Finally, we have used social and demographic survey data from 1984, 1994, and at the household and village levels to represent the human dimension of landscape dynamics for model building and validation. We examine the effects of agricultural production quotas on the cultivation of cassava within Nang Rong district. Initial conditions are set to image dates contained within the Landsat TM and MSS (resampled to 30-meter cells to correspond to the Landsat TM images and the cell size of the models) time-series. Neighborhood conditions are described using a 3 x 3 focal filter, and the model iterates on an annual time step. A stochastic element is used to represent LULC changes occurring through chance on the landscape, for instance, by new household formation in locations close and far from the nuclear village centers. GIS is used to score each LULC type relative to its suitability or propensity for change through a dilation process. This model constitutes our base LULC change model. Using this base model, we perturb it through a set of scenarios or experiments. We generate model runs for these scenarios and assess their corresponding LULC composition and spatial structure for each time-step using ecological pattern metrics computed at the landscape (i.e., Nang Rong district) and class (i.e., cassava) levels to compare the observed (i.e., satellite LULC classifications) versus expected (i.e., CA model runs) patterns of LULC change. In scenario 1, the base model is run from to assess LULC change of forest, rice, and cassava with no imposed production quota of cassava. Scenario 2 imposes a cassava production quota using classified Landsat imagery and production limits as of 1985 (291,832 cells were classified as cassava) as the production cap of cassava for subsequent years in the simulation. Scenario 3 imposes a production quota on the cultivation of cassava using 1975 imagery and the corresponding amount of cassava being cultivated as of that date in Nang Rong (i.e., 73,420 cells were classified as cassava). Scenario 4 imposes a cassava production quota using 1985 cultivation levels (i.e., 291,832 cells were classified as cassava) and a quota period after which cassava production is allowed to increase as the quota is relaxed. In each of the scenarios, LULC change is simulated through a cellular automata approach using the same initial condition (i.e., 1972), LULC types (i.e., cassava, forest, lowland paddy rice), transitions or growth rules (except for those having to do with the imposition of production quotas on the cultivation of cassava), and the same neighborhood conditions being applied using a 3 x 3 focal filter. The number of cells that are permitted to change based on stochastic processes is held constant, although the areal distribution varies. STUDY AREA Nang Rong district is in a rural portion of Buriram province, northeast Thailand that has experienced rapid social and environmental change over the past few decades (see Figure 1). Settlement and

3 3 subsequent agricultural extensification of lowland paddy rice and upland cassava and sugar cane has transformed the region from one dominated by forest to a landscape dominated by agriculture. The district is approximately 1,300 sq. km in size and is positioned within the Korat Plateau. The region is characterized by relatively infertile soils, poor drainage, and inconsistent precipitation levels caused by a highly variable monsoonal rainfall pattern, where over 80 percent of the average annual precipitation occurs between April and November [16, 32]. Figure 1. Study area: Nang Rong district, northeast Thailand. The terrace structure of the region influences the environmental suitability of the landscape for lowland versus upland agriculture. Rice production is generally associated with the alluvial plains and low terraces, while upland crops are associated with the upper terraces. The middle terrace is a zone of transition where rice production and upland crops are influenced by site accessibility and land suitability, moisture availability, subsistence demands, alternate land management strategies, and crop prices. The region is one of the poorest in the country. The most fertile land in the region has been under cultivation for some time, in some cases millennia, with major expansion of agriculture following WW II. At the beginning of the study period, 1972, there was little remaining land for agricultural extensification. What land there was for extensification is marginal in nature and is composed of primarily upland sites. The potential of the land to support agricultural extensification or increase production through intensification was subject to site and situation constraints. The biophysical, social, and geographic domains combine to influence the relative attractiveness or suitability for agricultural expansion or intensification. Road building and access to water, the topologic relationships of villages to each other and to the hydrographic and transportation networks, availability of land for conversion to agriculture and its location relative to the nuclear village centers where people live, as well as the competition of land from people living in other villages are some of the factors shaping LULC practices within the region and the trajectories of LULC change. DATA A broad array of data has been assembled to support the study of Nang Rong district. The data come from many sources demographic surveys, administrative records, maps, satellite images, aerial photographs, and field observations. The data are multilevel. On the social side, they cover individuals (including migrants), households, and villages for the period from 1984 to /01. On the spatial side, they cover pixels, plots, village territories, and the district for the period from 1954

4 4 to the present. The data are linked, over time, between scales, and between social, biophysical, and spatial domains. The following describes the highlights of the data of key interest in this research (see Figure 2). Survey Data The Nang Rong CEP-CPC surveys consist of three waves of data collection 1984, 1994, and. We begin with two surveys conducted in 1984: a community survey in 51 study villages; and a complete household census conducted in the study villages, with the census obtaining information on all members of all households. The 1994/95 data collection included: a community survey, done in all villages in Nang Rong, including but not limited to the original 51; a household survey, again a complete census of all households in each of the 51 villages; and a migrant follow-up that collected data from out-migrants from 22 of the original 51 villages who had gone to one of four urban destinations (Bangkok and surrounds; the Eastern Seaboard, a focus of rapid growth and development; Korat, a regional city; or Buriram, the provincial city). The /01 data collection included: a community survey, again done in all villages in Nang Rong; a household survey, again a complete census in the 51 study villages; the collection of locational data for dwelling units and agricultural plots, linked to the household survey; a migrant follow-up that surveyed migrants from 22 villages to the four urban destinations and to rural villages within Nang Rong district. The surveys are linked, such that migrants can be attached to home households, which can be placed in their village contexts. Data for individuals, households, and villages are also linked over time. Because of administrative splits of villages (villages are generally administratively split when the total number of households are greater than 100), the original 51 villages expanded to 76 in 1994 and 97 by year. Also, because of village splits as well as new village formation, the total number of district villages expanded from 310 in 1994 to 346 in. Data Source 51 Villages Migrants Remotely Sensed Imagery Household Surveys Village Surveys Maps Ikon/ETM+ Radar AVHRR SPOT TM MSS Aerial photo Remote Sensing Data X X X XX X X X X XXXX (+Aerial photos 1954, 67, 68, 69) Residents 1984 Individuals and Households Residents 1994 X(N=51) X(N=51) X(N=310) X(N=346) X X X XXXXXXX X X XX XXXXXXXXX Figure 2. Nang Rong project data sets. An aircraft and satellite image time-series has been assembled that extends from 1954 to the present (see Figure 2). Panchromatic aerial photography at scales ranging from 1:6,000 to 1:50,000 have been acquired for 1954, 1968, 1969, 1974, 1976, 1982, 1983/84, 1985, and Landsat TM and MSS data have been obtained for the period of to represent intra-annual, interannual, and decadal periods. The satellite data were classified using a hybrid approach and a hierarchical mapping scheme. The hybrid classification approach was designed to be repeatable across all images within the time-series, rely upon image characteristics through derived statistical measures, limit the reliance on in-situ data in the classification process, and exclude the use of nonspectral data in the process. The approach involved the use of the ISODATA decision-rule to define 100 naturally occurring spectral classes that were reduced to about 30 classes through the interpretation of the transformed divergence and divergence statistics generated as output from the classification process [24]. Then, a supervised classification was applied using the maximum X Individuals and Households Return Migration Residents X X

5 5 likelihood classifier to relate unclassed pixels to the 30 spectral classes (i.e., the training data) defined through the unsupervised classification. The approach allows for the generalization of classes to a few key LULC types as well as the expansion of cover types to additional classes as details warrant. The images were recoded to represent only cassava (an upland field crop), lowland paddy rice, and forest. GIS Data GIS was used to represent base coverages and to derive value-added products for assessing resource endowments and geographic accessibility of field sites as well as communities and households in our studies of human-environment interactions. Most fundamental was the generation of a digital elevation model (DEM). Using a 1:50,000 scale 1984 base map from the Thai Ministry of Defense, contour lines and spot elevations were digitized. A 10-m contour interval was used on the 1984 map and spot elevations were maintained to a 1-m vertical resolution. Using the contour lines, spot elevations, topographic sinks, and linear surface drainage patterns (i.e., perennial and intermittent rivers and streams and ponds/reservoirs), a DEM was developed along with a number of value-added terrain products (e.g., topographic curvature, topographic convergence or wetness index, and solar radiation potential). The DEM was also used to map important landforms within the study area particularly uplands, lowlands, and low-medium-high terraces that serve as areas of topographic and LULC transitions. Also fundamental to our studies was the development of a road network generated by digitizing road types from the 1984 base maps. Roads were described on the Thai base map (and subsequently digitized) as paved-surface all weather roads, loose-surface all weather roads, fair-dry weather roads, and trails and footpaths. The district outline was also captured from the 1984 base map, as well as district villages and regional market towns. In the 1994 and surveys, villages (and households) were also geographically referenced by using differentially corrected GPS coordinates. COMPLEXITY THEORY: BACKGROUND & CONTEXT Complexity theory holds that systems cannot be suitably understood without a focus on feedbacks and consequent nonlinearity that leads to emergent characteristics and multi-scale phenomena [21, 25]. A complexity theory analysis of LULC change aims at understanding these feedbacks and changes in state space through nonlinearities and thresholds, and in relation to a dynamic environment and a changing population-environment system [14, 18]. Emergent behavior is seen at a regional scale as an outcome of actions and patterns at local settings [19]. Spatially extended systems can self-organize to generate order as seen in frontier environments [20]. In the LULC context, change occurs around development fronts shaped by geographic accessibility into a region and constrained by resource endowments [20, 27]. Complexity emerges as a result of the patterns of interactions between elements [7]. For example, the complex structure of social inequalities emerges from unstructured interactions of households, resource endowments of the land, and land uses at local scales. Research in complex systems attempts to identify corresponding morphology of patterns and processes in social and natural systems [4]. Complex systems have the following characteristics: they consist of a large number of interactive and dynamic elements, interactions between elements are nonlinear, and positive and negative feedback mechanisms affect future trajectories, implying that small changes in a system can result in large impacts (and vice versa) in system behaviors [7]. Complex systems are also far from equilibrium, because there is constant interactions through feedback mechanisms that maintain the organization of the system (negative feedbacks) or alter subsequent alternatives in state space (positive feedbacks). A complex system does not only evolve through time, its past is co-responsible for its present behavior. Interactions are restricted to limited knowledge and responds to the information locally available. Spatially-explicit modeling methods such as cellular automata (CA) [27, 28] are highly suited to the exploration of spatial factors in land-use systems within the concepts of complexity [29, 38]. For a CA model, the state of the cell and rules of the cellular automaton define the transition functions, emergence occurs in generated systems, and patterns may be persistent with changing components [7]. Complexity theory concepts and hierarchical relationships are infused into the CA models for generating simulations to match observed states or for future periods by allowing the model to iterate within the expected bounds of the defined rules. The models allow us to spatially simulate LULC patterns, assess likely future LULC states, and examine how social, demographic, and biophysical factors have altered trajectories of LULC change resulting in possible shifts in the

6 6 composition and spatial structure of the landscape, and with it, pattern-process relations among people, place, and environment. CA belongs to a family of discrete, connectionist techniques being used to investigate fundamental principles of dynamics, evolution, and self organization [40]. CA models are examples of mathematical systems constructed from many simple identical components that together are capable of complex behavior. CA approaches can be used to develop specific models for particular systems, and to abstract general principles applicable to a wide variety of complex systems [42]. CA is not used to describe a complex system with complex equations (e.g., differential equations, multilevel statistical modeling), but allows the complexity to emerge from interactions of basic building blocks of systems (e.g., individuals and households) that follow simple rules. The essential properties of a CA are: a regular n-dimensional lattice is where each cell of the lattice has a discrete state, and a dynamical behavior described through growth or transition rules. These rules describe the state of a cell for the next time step, depending upon the states of the cells in the defined spatial neighborhood. The essential components of a CA are: (1) the cell -- the basic element of a CA that is capable of storing defined states, (2) the lattice, or cells arranged in a spatial matrix form, commonly represented in one- or two-dimensions, and (3) neighborhoods defined by growth or transition rules that perform changes to the state of the cells depending upon neighboring cells and their conditions [1]. Four classes of behavior are recognized in CA models: fixed, periodic, chaotic, and complex [42]. Lambda (λ) is used to relate the nature of the rules to the overall behavior of CA models [17]. CA approaches have been used as tools for natural resources management, including fisheries [1], urban systems [8, 9], forest fires [3], nutrient cycle dynamics [15], climate change [36], vegetation modeling [31], biological modeling [12], LULC change [27], urban land use simulation [2, 39, 40, 41], and tropical deforestation [24, 35]. The spatially explicit nature of the CA and its compatibility with geographic information systems and remote sensing has supported the development of various LULC CA models that follow a diversity of techniques. The structure of urban expansion producing fractals or bi-fractals land structures has been modeled, finding similar patterns to those found in real urban environments [40]. Other studies show the plausibility of using Bayesian statistics within the CA context to simulate LULC change [10], and the use of statistical data to parameterize a CA model to predict traffic flows for various times of the day [37]. Development alternatives and urban planning objectives can be simulated using a combination of CA and neural networks [43], and the use of transition probabilities in simulating deforestation patterns can be examined as part of land change science [28]. CA MODEL DESIGN: SIMULATING LULC CHANGE IN NORTHEAST THAILAND The intent of this research is to examine the integrative effects of people, place, and environment on LULC change patterns in Nang Rong district, northeastern Thailand by specifying the initial conditions, neighborhood associations, and transition rules associated with the development of CA models that are used to simulate LULC change according to a set of defined land management scenarios. Here we use a Landsat TM and MSS image time-series beginning in 1972 to set the starting date for the simulations of LULC change across the 1,300 sq. km district. A LULC classification of the 1972 image is used to set the initial conditions of the three primary LULC types being modeled forest, rice, and cassava. Neighborhood associations within a 3 x 3 cell kernel are used to assess the propensity of focal cells to transition to alternate LULC types depending upon LULC frequencies and change scores occurring within the kernel window. Transition rules are used to describe relationships among LULC types and terrain settings, geomorphic conditions, distance to water, topographic relative moisture patterns, and soil suitability scores. The model iterates on annual time-steps. At the beginning of each iteration, individual LULC types are extracted from the model output from the previous year. For each class, a score is created, based on one or more data inputs that indicate the suitability or likelihood of a LULC assignment to one of the modeled classes. For each class, a stochastic process is simulated by defining a small number of cells that are randomly seeded as new cells of that class. The random seeding can be set by the analyst to represent hypothesized processes of LULC change for the entire district or for defined landscape strata. Dilation is also represented by passing a 3 x 3 kernel over the LULC input data and using a threshold to determine whether the focal cell changes to an alternate LULC class based upon the frequency of their cells occurring within the kernel [28, 34]. Dilation is used to fill holes or add pixels along a region s boundary. Dilation is also used to change an OFF cell to an ON cell if the number of neighborhood ON cells exceeds a pre-defined threshold. Following each model iteration, LULC classes are recombined based upon their likelihood of change to an alternate class using a scaled model input of scores derived for each class. If their differences exceed a defined decision-threshold, output class

7 7 possibilities are randomly selected to represent an uncertainty interval in household decisionmaking and their LULC change patterns. Figure 3 is a generalized schematic of how our CA model operates. START Time = 0 Initial conditions from satellite land use classification Class suitability scores derived from GIS inputs, scaled 0-1 Extract RICE, CASSAVA, & FOREST classes RICE, UFC, FOREST class growth sub-models Add stochastic growth (uniform random seeding) - Growth may be stratified within study area Dilate/grow existing class areas - Where neighborhood class frequency exceeds kernel threshold Reconcile/combine class sub-model output For cells in competition between multiple classes - Inter-class score differences greater than inter-class uncertainty interval High score wins - Inter-class score differences within uncertainty interval Random selection Time + 1 Figure 3. CA Model schematic. END; Go back to start for Time + 1 CA MODEL RUNS: SCENARIOS OF CASSAVA CULTIVATION Recall that four scenarios were developed to examine the effects of imposing a production quota on the cultivation of cassava in Nang Rong district during the period A CA model was used to represent initial conditions, neighborhood associations, and growth or transition rules for representing changes in the composition and spatial pattern of cassava, forest, and rice. A 1972 Landsat MSS LULC classification was used to set initial conditions. Scenario 1 is the base model in which no quota or policy intervention is imposed on the cultivation of cassava during the simulation of LULC change during Scenario 2 imposes a production quota on cassava by setting the limits of the area in production equal to that which was represented by the Landsat TM classification for 1985, i.e., 291,832 cells. Note that each cell measures 30 x 30 meters, which is the spatial resolution of the model and that of the Landsat TM system used to categorize LULC throughout the district. Scenario 3 imposes a production quota on the cultivation of cassava equal to that which was mapped by the Landsat MSS classification in 1975, i.e., 73,420 cells. Finally, Scenario 4 imposes a cassava production quota equal to the amount categorized by the 1985 Landsat TM classification, i.e., 291,832 cells, as in Scenario 2. The quota in the amount of cassava cultivation is retained for the period of , after which cassava production is allowed to increase unconstrained by the quota amount or period. Note that our simulations for the four scenarios describe the area of cassava in cultivation and not necessarily the associated yields. Figures 4-6 show the percent of land in Nang Rong district that was modeled for cassava, rice, and forest respectively for the period of for Scenarios 1 4, as well as the percent of each LULC type characterized by Landsat MSS and TM LULC classifications for periods represented in our remote sensing image time-series. The mapped amounts represent the observed, whereas the modeled amounts represent the expected area of land in cassava for the respective scenarios and years. The satellite image time-series was assembled by selecting images from a Landsat MSS and TM archive. As an optical system, Landsat views of the landscape are best achieved during cloud-free conditions. In the Thailand setting, long periods of time are characterized by partially or completely clouded skies as a consequence of monsoonal rains. The monsoons bring much needed moisture that nourishes the rain-fed paddy rice, refills reservoirs following a pronounced period of dryness, and alters the calculus of population and environment interactions by signally the return of out-migrants to engage in activities surrounding the cultivation of rice within the district. Routinely, young adults temporarily out-migrate to Bangkok, the Eastern Seaboard, and other places to engage in off-farm employment following the harvest of rice in November or December. Migrants may return to their villages in May or June to help with planting rice, and also towards the end of the water-year to assist with the November or December harvest. Finally, the assembled remote sensing time-series

8 8 reflects various phenological periods in the annual landscape cycle from green-up to browndown. While an attempt was made to select only a single image to characterize the Nang Rong landscape each year for model validation purposes, it is important to note that the assembled images do not represent exact anniversary dates, and contain inherent variability in environmental conditions across years. Images used to generate an observed composition and spatial structure of the landscape are subject to variation seen in the annual measures and the overall trends Area (hectares) Observed S1 S2 S3 S Figure 4. Trends in the percent of land in upland field crops, primarily cassava cultivation, for Nang Rong district from for the satellite-based classifications (observed) and Scenario 1 (S1), Scenario 2 (S2), Scenario 3 (S3), and Scenario 4 (S4) Year Area (hectares) Observed S1 S2 S3 S4 Year Figure 5. Trends in the percent of land in rice for Nang Rong district from for the satellite-based classifications (observed) and Scenario 1 (S1), Scenario 2 (S2), Scenario 3 (S3), and Scenario 4 (S4).

9 Area (hectares) Observed S1 S2 S3 S Year Figure 6. Trends in the percent of land in forest for Nang Rong district from for the satellite-based classifications (observed) and Scenario 1 (S1), Scenario 2 (S2), Scenario 3 (S3), and Scenario 4 (S4). Figures 7 9 show the number of patches of cassava, forest, and rice respectively that were modeled in Nang Rong district for for Scenarios 1 4, as well as the number of patches characterized by Landsat MSS and TM LULC classifications for images within our assembled timeseries. The number of patches for each LULC type, for each of the scenarios, as well as for the satellite classifications were generated using ecological pattern metrics algorithms that compute the spatial structure of LULC classes at the landscape, class, and patch levels. Here, the spatial organization of cassava, forest, and rice were computed at the class level for the district. While a host of pattern metrics can be computed, such as contagion, edge density, and mean patch size, we only report the number of patches for simplicity # Patches Observed S1 S2 S3 S4 Year Figure 7. Trends in the number of upland field crop patches, primarily cassava, for Nang Rong district from for the satellite-based classifications (observed) and Scenario 1 (S1), Scenario 2 (S2), Scenario 3 (S3), and Scenario 4 (S4).

10 # Patches 1500 Observed S1 S2 S3 S Year Figure 8. Trends in the number of rice patches for Nang Rong district from for the satellite-based classifications (observed) and Scenario 1 (S1), Scenario 2 (S2), Scenario 3 (S3), and Scenario 4 (S4) # Patches Year Observed Figure 9. Trends in the number of forest patches for Nang Rong district from for the satellite-based classifications (observed) and Scenario 1 (S1), Scenario 2 (S2), Scenario 3 (S3), and Scenario 4 (S4). A closer examination of the trend lines in Figures 4-9 provide more detail about the operation of the model. Figures 4-6 compare the relative landscape prevalence of the three modeled classes in each of the scenarios with the observed prevalence derived from the satellite imagery. In Figure 4, the trend line for upland field crops in the unconstrained base model (Scenario 1) shows them consistently and rapidly increasing to fill the land areas suitable for their growth. The trend continues upward at the end of the model period, indicating there is still room for upland field crop expansion at the expense of rice and forest. In Figures 5 and 6, the Scenario 1 trend lines for rice and forest show a corresponding steady decline. The Scenario 4 trend line for upland field crops in Figure 4 shows a steady increase similar to Scenario 1, but delayed by several years due to the temporary 5-year cap on cassava production simulated in this scenario. Figures 5 and 6 show the same delayed effect: the same overall trend, but delayed in time. The Scenario 2 trend line for upland field crops under a simulated 1985 quota level generally tracks the observed percentage for that class in Figure 4, indicating that this simulation is getting the percentage of cassava correct at the landscape level. The upland field crop trend line bounces up and down as the model iterates around the quota level. Figure 4 shows the low, steady percentage of upland field crops to be expected for the 1975-level cassava quota simulated in Scenario 3, and Figures 5 and 6 show a compensatory expansion of rice and forest. The forest S1 S2 S3 S4

11 11 increase is greater than rice in this cassava-restricted scenario due to the larger overlap between forest and cassava in terms of areas of land suitability. Figures 7 through 9 compare the number of patches modeled for the three classes with the observed number of patches derived from the satellite imagery. For the 1985-level quota simulated in Scenario 2, Figure 7 shows the pattern of alternating increases and decreases in the number of upland field crop patches expected as the model iterates around the cassava cap. The same is true for the trend lines for the limited 5-year quota period in Scenario 3. Figures 8 and 9 show the corresponding opposite alternating patterns for rice and forest (increases where cassava decreases, and vice versa). This exchange is especially noticeable between rice and upland field crops classes and takes place spatially in the transitional terrain areas suitable for both crop regimes. Figures 7, 8, and 9 indicate that the model is generally underestimating the number of patches for all three modeled classes relative to the observed data. This is due in large part to the neighborhood frequency criterion imposed within the 3x3 kernel, which makes it harder for more isolated patches of a class to survive and grow. This is especially noticeable with the forest class, and is a result of this simplified model not being designed to capture some areas of forest re-growth, areas of forest retention within agricultural fields, as well as the forest areas surrounding village settlements. In short, we do a better job of modeling LULC composition than spatial pattern, a consistently difficult task for CA models, and hence not always addressed in the literature. DISCUSSION & INTERPRETATION The topographic-settings throughout the district and the spatial organization of the region s landforms -- floodplains, low-middle-high terraces, and dissected hills and erosional surfaces -- strongly affect the environmental and land use gradients that influence LULC patterns throughout the district. Figure 10 is a DEM of the study area, with primary roads and rivers superimposed, that shows the pronounced upland areas in the southern portion of Nang Rong district, principally the broad upland settings of the southwest and the more areally restricted remnant volcanoes located in the south-central and southeastern portion of the district. The elevation in the district decreases in a south-to-north direction, but local variations, as indicated in the enlargement box shows obvious and subtle elevational gradients that characterize the terrace system and serves as areas of important topographic, environmental, and LULC transitions. For instance, the high hills and uplands are infrequently or never flooded, the upper and middle terraces are variably flooded, and the low paddies and alluvial plains are nearly always flooded [11]. In addition, soil fertility is highest at the lower paddy and lowest at the hills and dry uplands. Drought-prone areas are organized in the opposite direction a greater propensity in the uplands and less so at the low terrace and flood plains. Finally, land clearing is more difficult in the uplands where a relatively dense forest resides (although less so and more fragmented as a consequence of the cultivation of the upland field crops of cassava and sugar cane beginning in the mid to late 1960s) on unevenly sloping land and far less difficult on the relatively flat, rain-fed rice paddies. In the uplands the primary ecological difficulties for cultivation of crops are low organic matter, poor fertility, and relatively high soil erosion. In the upper terraces, insufficient water is the central problem; in the middle terraces, it is poor soil fertility; and in the low terraces and alluvial plain, it is persistent flooding, sometimes excessive even for water-demanding paddy rice. The terrain and ecological settings in Nang Rong district combine to influence LULC dynamics, particularly, on the terraces where crop cultivation responds to a set of complex and interacting socio-economic and environmental factors that affect household decision-making and subsequent uses of the land. Areas that are converted to rice-paddies generally stay in that land use for extended periods of time, because of the significant effort associated with the construction of bunds that outline the paddies and serve as earthen-dikes to retain water during periods of cultivation. The medium and high terraces lack the general soil, terrain, and hydrologic characteristics for the continued cultivation of rice, and poor drainage further exacerbates long-term cultivation of rice in such settings. These areas may be more suitable for upland field crops, but only when balanced against competing economic and environmental opportunities. We hypothesize that during periods of above-average rainfall, particularly for extended water years, farmers may choose to cultivate rice on middle and high terraces, that is, in normally marginalized settings where water is generally the limiting resources. High market prices for rice over multiple years, even without above average rainfall, may encourage farmers to plant rice in medium and high terraces because of a favorable reward to risk ratio.

12 12 Figure 10. DEM of Nang Rong district, northeastern Thailand. In general, our spatial simulations show the expansion of cassava in the high and middle terraces throughout the district, and most centrally in the core upland areas. Riparian forest corridors are retained, but only in minimal extents and then only for the most dominant river courses that flow through the region. Our base model, scenario 1, begins with a substantial amount of forest in 1972, but increasingly over time is affected by the expansion of paddy rice in lowland settings and cassava and sugar cane in upland settings. The representation of cassava, forest, and rice in 2001 shows a landscape essentially dominated by two cover types cassava and rice -- with only some riparian forest and forest that rings the ancient volcanoes and fringes the upland-high terraces in the southwest. The extensive, rice paddy plains of the east-central portion of the district is complemented with cassava dominating in the uplands and higher terrace landforms of the west, southwest, and south-central portions of the district. Scenario 2 uses 1985 cassava production levels (i.e., area in cultivation, not crop yield) to constrain further cultivation of cassava beyond those levels for the remainder of the simulation period. As a consequence, far more forest is represented in the simulation for the later years, culminating with a 2001 image that shows a substantial amount of forest that is consolidated around the ancient volcanoes in the south and the zone that marks the contact between the uplands and the highmiddle terraces. Cassava in the western portion of the district and on the middle and high terraces is areally restricted as a consequence of the quota used to constrain production linked to 1985 levels. In Scenario 3, 1975 production limits (i.e., area in cultivation, not crop yield) are used to constrain the expansion of cassava cultivation throughout the district. As a consequence, forest assumes a more extensive and dominant signature throughout the district, and does so far earlier in the simulation period than in previous scenarios. Forest consolidation occurs in the higher areas, those most removed from lowland paddy fields, that have also become more dominant with only smaller patches of forest being represented at the expense of cassava occurring in more marginal, wetter sites. In short, cassava is areally constrained, forest is expansive, and rice dominates the landscape in even more obvious ways. Finally, in Scenario 4 the cultivation of cassava is limited to 1985 production levels (i.e., area in cultivation, not crop yield) and the quota is restricted to , at which time the quota is relaxed and cassava is allowed to increase as the rules and patterns dictate. The quota and the period of implementation constrains the growth of cassava and allows the retention of forest in

13 13 upland sites in greater amounts and in more extensive patches than in other scenarios. But once the quota is relaxed, forest is again removed and cassava cultivation is expanded in upland settings, affecting high and medium terrace areas towards the end of the simulation period, as upland forest has mostly been subsumed by the cassava in earlier years. Small patches of forest are also transformed to rice in the lowland areas, and cassava is represented at somewhat lower sites that might otherwise have supported forest or even transitional rice at these marginal locations. Regardless of the scenario, the consolidation of cassava production occurs in the southwest portion of the district over time as the model runs for each scenario iterated to the year In-filling in the spatial pattern of cassava occurs as a fragmented pattern of cassava throughout the district gives way to a more homogenous pattern characterized by large areal extents of cultivation in the southwest portion of the study area. Primarily forested during the late 1960s and early 1970s, the uplands were somewhat external to the rain-fed, paddy rice economy that dominated the lives of the people and which occurred in lowland settings throughout the district. The demand for high calorie animal feed in Europe was an opportunity for Thai farmers, and the national government more generally, to become further integrated into the global economy. Farmers sought lands to be transformed into cassava that met the resource needs of the crop, were relatively nearby to their villages, or closely situated to other lands that they used or owned. Deforestation and agricultural extensification was the process in which relatively unused land was converted to extensive agriculture in support of upland field crops. Site conditions for the cultivation of rice and cassava are quite different. So different, in fact, that competition for the same parcel of land for their cultivation generally is not an issue of debate nor concern. In some instances, however, cassava was planted in areas too low and wet resulting in the crop rotting in the field, and so too, paddy rice was on occasion planted in too high and dry a site that prevented the accumulation of sufficient moisture to satisfy its continued cultivation. In such situations, farmers learned the functional limits of the extensification of rice and cassava in marginalized settings. While land parcels accessible to nuclear villages or spatially associated with other parcels that farmers used or owned could encourage planting of rice or cassava in less than optimal conditions, sustainability of their cultivation was limited to favorable periods of environmental circumstances and/or periods of high crop prices to justify the risk of crop failure and the costs associated with their cultivation. Another set of factors favoring the cultivation of upland field crops is the fact that cassava (as well as sugar cane) can remain in the field for periods of time well beyond the optimum harvest period without serious consequences. That flexibility affords farmers the opportunity to consider fluctuations in market prices, as well as to gauge labor availability. Rice is normally harvested in December, the date set by the variety used and the planting mode and date, whereas upland field crops are harvested in subsequent periods throughout the year as time and opportunities dictate. Finally, upland field crops are normally planted in large, extensive fields in which large tractors are used during planting and harvesting. Sharing of tractors among villages is a common practice in large upland sites dominated by field crops [13]. In the southwest, a number of nuclear villages occur at the fall-line of the uplands were the outlets of springs and some surface water sustain settlement patterns and other forms of agriculture. A road built to facilitate access to the southwest uplands, the position of nearby villages, and the existence of broad tracts of upland forest combined to help consolidate cassava in large fields at the expanse of forest. Forest has also been the counter-weight for the expansion of lowland paddy rice as only isolated trees, small patches, and some riparian forest still exist in lowland paddy lands.

14 Figure 11a. Model output for Scenarios 1, 2, 3, and 4 (S1, S2, S3, and S4) for 1973 and Green (or darkest greyshade) represents forest, pink (or medium greyshade) represents upland field crops, and yellow (or lightest greyshade) represents rice. 14

15 Figure 11b. Model output for Scenarios 1, 2, 3, and 4 (S1, S2, S3, and S4) for 1993 and Green (or darkest greyshade) represents forest, pink (or medium greyshade) represents upland field crops, and yellow (or lightest greyshade) represents rice. 15

16 16 Figures 11a and b show model outputs from four time points in the model runs: 1973 (first iteration), 1983, 1993, and 2001 (last iteration). For the 1973 time step, the only differences between the scenarios is the random seeding applied in the growth cycle. At this presentation scale, outputs for all the scenarios (Figure 11a) look the same for that year. However, differences are apparent in the 1983 output, most notably in Scenario 3 (1975-level cassava quota): the lack of upland field crops in the eastern one-half of the district, an area much more suitable for rice than upland field crops because of terrain morphology. This scenario also shows restricted upland field crops in the southwest portion of the district. Scenario 1 (no quota) shows an overabundance of upland field crops by 1993 (Figure 11b) relative to our satellitebased observed data. The broad scale spatial patterning of upland field crops (prominent in the southwest highlands, and scattered elsewhere) and rice (predominant elsewhere) are correct in the Scenario 2 outputs for 1993 and At a finer scale of observation, in Figure 11b, we see the lack of upland field crop patchiness in 1993 and 2003 for all scenarios relative to the observed data that corresponds to the results presented in Figure 7 and described earlier. All the scenarios outputs for 1993 and 2001 correctly maintain some of the riparian forest that runs from the southwest-central to the northeast edge of the district. These riverine corridors score very high on forest suitability and low for both rice and upland field crops because of the near-annual inundation during the monsoon season, and thus the forest is never overcome by rice or field crops in these very moisture areas, regardless of the number of iterations the model is allowed to run. CONCLUSIONS We demonstrate the use of cellular automata models to examine LULC change scenarios by emphasizing the impacts of production quotas on the cultivation of cassava in Nang Rong district, northeast Thailand. While initial conditions were consistently represented in the scenarios by using a 1972 classified Landsat MSS image to set the initial LULC conditions, the spatial and compositional patterns of cassava, forest, and rice varied in interesting ways for the scenarios examined for the nearly 30-year simulation period. The ability to perturb the base model relative to plausible scenarios of LULC change is a useful way to consider landscape alternatives relative to policy and environmental conditions. Data visualization approaches and what if scenarios can engage analysts, as well as policy-makers and local stakeholders in assessing the implications of cultivation alternatives on LULC change patterns and trajectories. Once the base model is developed, addressing an array of possible factors hypothesized to affect LULC path dependencies is possible. Stability of rules through time can be a concern in CA modeling particularly when simulations are run for long periods of time for coupled human-natural systems in which population-environment interactions can be robust with important feedback mechanisms influencing pattern-process relations. Judging how reasonable and plausible the simulation outcomes are can be of concern [30]. The true measure of spatial complexity as applied in a complex model context is one not yet fully realized in the literature [27]. One of the challenges is to use spatial simulations in general and complexity-based methods such as cellular automata and agent-based models in particular, in answering the question of what is a good fit? when spatial simulations are developed for antecedent and future time periods. Unanswered questions about the effects of the ecological fallacy and the modifiable areal unit problem can influence model outcomes by affecting the apparent strength and magnitude of relationships between variables [5, 33]. Beyond composition and pattern of model outcomes is the need to understand complex processes and their characteristics. The complexity as property defines a system as complex if it exhibits certain characteristics of complexity, such as fractal dimension or scale invariance [22]. Here we assess the outputs of our scenarios and model simulations by comparing them to satellite observations. In essence, we are concerned about our ability to replicate observed spatial and compositional patterns, and hence trend lines and pattern metrics are used to assess the certainty or plausibility of our model outcomes. However, other approaches to assess model performance are under development [23] including the study of pattern invariant areas in simulations [6], and emergent patterns and the creation of development fronts through the actions of individuals or some set of base actors on the landscape [20]. In this research, we have thus far only infused our CA models with dilation characteristics and not erosion processes. Therefore, the antagonism between classes is focused upon their expansion across the landscape through stochastic processes, neighborhood effects, and derived site suitability scores. The contraction of a LULC type as a response to some social and/or biophysical factor or event is not yet explicitly accounted for in the operation of our CA model. Future development plans will include erosion processes, but will also include more social effects into our model runs. Using our rich longitudinal social survey, we anticipate creating a number of scenarios that more explicitly consider