4th International Conference on Integrating GIS and Environmental Modeling (GIS/EM4):
Problems, Prospects and Research Needs. Banff, Alberta, Canada, September 2 - 8, 2000.

A GIS/Spatial Statistical Predictive Model of Land Use Change in a Developing Country Context Using a Tree Classification Approach

GIS/EM4 No. 52

John S. Felkner

Extended Abstract

Nowhere has the global conflict between economic and population growth on the one hand and protection of vital ecological resources (including forest and biodiversity resources) on the other been as acute as in the tropical developing world (Turner, Skole et al. 1995). Often, conditions of rapid, uncontrolled economic and population growth have been coupled with a regulatory inability to enforce conservation policies and with a paucity of reliable land analysis data (Gallup and Sachs 1998). Thus, there exists a great need for rigorous developing country land planning models that can utilize available data inputs from both economic and environmental factors (Deichmann 1993). The use of Geographic Information Systems and remote sensing data in particular offer tremendous potential for rapid, comprehensive and highly accurate land use change and land planning models (Turner, Kasperson et al. 1990). Integration of remote sensing data into automated analytic GIS models which also incorporate socioeconomic data regularly collected by developing countries could provide illuminative models for the accurate description and prediction of land use change (Skole 1994).

Figure 1. Overall Structure of the Model.

For this research, a computer predictive model was created in a GIS framework with the goal of explaining and predicting land use change in a developing country at a spatial level as a function of a set of economic and environmental input factors (Figure 1). The model considered land use change in two Provinces of Thailand from 1979 to 1989 (Figure 2), and then predicted future land use for 1999. The two Thai Provinces - Chachoengsao and Sisaket - were chosen because they are representative of the range of both economic and environmental variability among all Thai Provinces (Binford, Biesboer et al. 1998). Eight input factors - four economic and four environmental - were modeled within a GIS for each Province and then inputted into a spatial statistical predictive model that used a tree classification approach (Breiman, Friedman et al. 1984; Venables and Ripley 1999). The dependent variable for the tree classifications (performed in Splus software) were maps of land use change derived from processed Landsat satellite imagery from 1979 and 1989 for both Provinces. Predicted land use maps were generated using GIS modeling based on the predicted probabilities of change from the tree classification analyses. Finally, these predicted land use maps were systematically compared using GIS techniques with actual maps of land use derived from 1999 Landsat 7 satellite imagery for accuracy evaluation (Congalton and Mead 1983; Rosenfield and Fitzpatrick-Lins 1986).

The basic research question was: which economic and environmental factors were most influential in driving new landcover change, particularly new urban or "built" change? The fundamental hypothesis was that economic and environmental factors together better explain land use change than either set independently.

Thailand is an ideal country for such an analysis, as it has experienced both enormous economic and population growth as well as extensive deforestation and degradation of natural resources (including very high plant biodiversity) in recent decades (Hirsch 1993). For this study, a cooperative arrangement was established with a research group at the University of Chicago, headed by economist Robert Townsend of the Department of Economics. Professor Townsend is currently heading a five-year research study of Thai economic and environmental factors, funded by the NSF and the NIH (Townsend and Lim 1998).

Figure 2. Study area - Sisaket and Chachoengsao Provinces, Thailand.

For the predictive model, eight continuous spatially variable representations of selected economic and environmental factors for each of the Provinces modeled in a GIS served as input factors. The economic factors were: spatially variable human development index (HDI) (McKinley 1997) (Figure 3); proximity to infrastructure networks (Chomitz and Gray 1995); proximity to local and regional markets; and a population density model. The environmental factors were: topography; soil moisture spatial variability (Barling, Moore et al. 1994) (Figure 4); proximity to water resources; and proximity to forest edges (Deininger and Minten 1997). The economic factors utilized an extensive database of socioeconomic variables - including per capita income, population, household agricultural holdings, and proximity to local facilities - collected at the village-level by the Thai government and the Townsend research group (http://www.src.uchicago.edu/users/robt/). This socio-economic data was combined with detailed digital geographic data on village locations, infrastructure networks and topography to create the four economic spatial inputs. For the environmental factors, historical Thai rainfall and climatic data was combined with the digital topographic models.

Figure 3. Spatially varying Human Development Index (HDI) Model for Sisaket Province.

Figure 4. Soil moisture index model for Sisaket Province.

A time-series of Landsat TM and MSS satellite imagery from the late 1970s to the early 1990s was used to derive basic landcover maps (Figure 5) and to create maps of land use change (Figure 6) using several methods including an unsupervised classification of a multi-temporal dataset, image differencing and supervised classification (Liu and Ridd 1998).

Figure 5. Sisaket Province 1979 landcover.

Figure 6. 1979-1989 land use change for Sisaket Province.

The GIS HDI, population density and proximity to markets models were created in a series of steps. First, road networks obtained from Thai government military maps were vectorized and merged into a continuous digital GIS road network for each Province, under the assumption that population distribution, development distribution and accessibility to markets occurs primarily along road networks. The roads were categorized into several levels depending on their quality - highways, paved roads, unpaved, etc. - and assigned travel cost values which approximated their average travel speed (e.g. 65 miles per hour along highways, 40 miles per hour along small paved roads, etc.). Income and population values for several hundred villages in each Province (obtained from Thai government national surveys) were then mapped to the nearest node in the road network. The values at these nodes were then spatially interpolated to all nodes in the road network - taking into consideration different road travel costs - using a negative exponential gravity model (Figure 7) (Dijkstra 1959; Fotheringham and O'Kelly 1989). Finally, a continuous raster map was interpolated from the point values at these nodes using an Inverse Distance Weighted (IDW) spatial interpolation. The same method was used in the proximity to markets model by assigning values to villages, towns and cities reflecting their importance as markets.

Figure 7. Population values interpolated along road network by road speed using negative exponential gravity model.

All eight environmental and economic input factors, as well as maps of land use change from 1979-1989, were created as raster GIS grids with cell spatial resolutions of 120 meters. The values of these cells were then fed into Splus spatial statistical software and a tree classification algorithm was run to create explanatory tree models. The predicted probability values of conversion for each cell (observation) were then used to create predicted maps of land use change (Figure 8). The predicted maps were compared with actual maps of 1999 landcover derived from processed 1999 Landsat 7 imagery.

A tree approach is particularly appropriate because the assignment of relative influence to the predictor variables and the partitioning of the observations into homogenous groupings is done through an automatic binary recursive partitioning procedure, rather than by requiring the designer to hypothesize beforehand what the possible inter-relationships of the variables will be. The tree algorithm recursively splits the dataset into binary partitions by calculating the reductions in deviance that would occur with every possible split of the data along all variables. Each subsequent partition is then examined and another split is made. The result is that the data is progressively partitioned into areas within which exists a high degree of homogeneity.

Tree models specifically allow more general interactions between predictor variables, in contrast to linear combinations of variables typical of more traditional classical models, such as logistic and linear regression analyses (Breiman, Friedman et al. 1984; Clark and Pregibon 1992). Thus, the trees are more adept at capturing nonadditive behavior, in contrast with the standard linear model which does not allow interactions between variables unless they are prespecified and of a particular multiplicative form (Venables and Ripley 1999). In addition, because most of the reduction in deviance is often captured in the first series of binary splits, consideration of a certain number of the initial splits only allows for the effective screening of variables, since the tree will list which variables are most important in allowing for large reductions in overall deviance (Venables and Ripley 1994).

Figure 8. Sisaket Province predicted landcover based on tree classification model for all eight inputs (economic and environmental).

Initial results show that, for both Provinces, the overall model (using all eight inputs, economic and environmental) does a superior job of prediction for future deforestation and future agriculture than does either the economic model or the environmental model. The economic model, however, does a superior job of prediciton of future "built" areas than do the overall model or the environmental model. Numerical results of the comparisons between the predicted landcovers and the actual landcover is summarized in Figure 9.

Figure 9. Numerical results of several systematic comparison measures between predicted maps from tree classification model probability surfaces with actual 1999 landcover.

The research is valuable in at least three separate ways. First, it comprises a detailed record of land use change in these two Provinces from the late 1970s to the late 1990s. Second, it utilizes spatial statistical analysis techniques to specifically evaluate, explain and predict land use change in a GIS from a primarily spatial perspective. Finally, it improves understanding of the drivers of new urbanization and land use change in rapidly developing tropical countries, specifically focusing on the relative influences of environmental versus economic factors.

This research constituted my Doctoral Dissertation work at the Harvard University Graduate School of Design. Fieldwork in Thailand in 1998 and 1999 was supported by a Harvard Kennedy Fellowship.

References used

Barling, R. D., I. D. Moore, et al. (1994). "A Quasi-Dynamic Wetness Index for Characterizing the Spatial Distribution of Zones of Surface Saturation and Soil Water Content." Water Resources Research 30(4): 1029-1044.

Binford, M. W., D. D. Biesboer, et al. (1998). Collaborative Research: Agroecological and Socioeconomic Systems Interactions in Thailand.

Breiman, L., J. H. Friedman, et al. (1984). Classification and Regression Trees. Monterey, Wadsworth and Brooks/Cole.

Chomitz, K. and D. Gray (1995). Roads, Land, Markets and Deforestation: A Spatial Model of Land Use in Belize, The World Bank: Policy and Research Department, Environment, Infrastructure and Agriculture Division (PRDEI).

Clark, L. A. and D. Pregibon (1992). Tree-Based Models. Statistical Models in S. J. M. Chambers and T. J. Hastie. London, Chapman & Hall: 608.

Congalton, R. G. and R. A. Mead (1983). "A Quantitative Method to Test for Consistency and Correctness in Photointerpretation." Photogrammetric Engineering & Remote Sensing 49(1): 69-74.

Deichmann, U. (1993). Issues in the Integrated Spatial Analysis of Socio-Economic and Environmental Data. Developing Large Environmental Databases for Sustainable Development, Nairobi, Kenya.

Deininger, K. and B. Minten (1997). Determinants of Forest Cover and the Economics of Protection: An Application to Mexico. Washington, D.C., World Bank.

Dijkstra, E. W. (1959). "A Note on Two Problems in Connexion with Graphs." Numerische Mathematik 1: 269-271.

Fotheringham, A. S. and M. E. O'Kelly (1989). Spatial Interaction Models: Formulations and Applications. Dordrecht, Kluwer Academic Publishers.

Gallup, J. and J. Sachs (1998). Geography and Economic Development. Cambridge, Harvard Institute for International Development.

Hirsch, P. (1993). Political Economy of Environment in Thailand. Manila, Journal of Contemporary Asia Publishers.

Liu, J. and M. K. Ridd (1998). "A Comparison of Four Algorithms for Change Detection." Remote Sensing of Environment 63: 95-100.

McKinley, T. (1997). Beyond the Line: Implementing Complementary Methods of Poverty Measurement. Technical Support Document, Poverty Reduction, Module 3 - Poverty Measurement: Behind and Beyond the Poverty Line. R. L. Dessallien, United Nations Development Program (UNDP).

Rosenfield, G. H. and K. Fitzpatrick-Lins (1986). "A Coefficient of Agreement as a Measure of Thematic Classification Accuracy." Photogrammetric Engineering & Remote Sensing 52(2): 223-227.

Skole, D. L. (1994). Data on Global Land-Cover Change: Acquisition, Assessment, and Analysis. Changes in Land Use and Land Cover: A Global Perspective. W. B. Meyer and B. L. Turner. Cambridge, Cambridge University Press: 437-472.

Townsend, R. and Y. Lim (1998). "General Equilibrium Models of Financial Systems: Theory and Measurement in village Economies." Review of Economic Dynamics 1(1).

Turner, B. L., II, R. E. Kasperson, et al. (1990). "Two types of global environmental change: Definitional and spatial-scale issues in their human dimensions." Global Environmental Change 1: 14-22.

Turner, B. L., D. Skole, et al. (1995). Land-Use and Land-Cover Change (LUCC): Science/Research Plan. Stockholm, IGBP: 132.

Venables, W. N. and B. D. Ripley (1994). Modern Applied Statistics with S-PLUS. New York, Springer-Verlag.

Venables, W. N. and B. D. Ripley (1999). Modern Applied Statistics with S-PLUS. New York, Springer-Verlag.

Authors

John S. Felkner, Doctoral Candidate
Harvard University
Graduate School of Design
48 Quincy Street
Cambridge, MA 02138
Email: gsd97jjf@gsd.harvard.edu, Tel: 617-495-8807, Fax: 617-495-9146.