Fuzzy Cellular Automata Approach for Urban Growth Modeling

This work focuses on adapting artificial intelligence techniques for urban growth modeling using multitemporal imagery. Fuzzy set theory and cellular automata are used for this purpose. Fuzzy set theory preserves the spatial continuity of the growth process through allowing a test pixel to be partially developed unlike the binary crisp system (developed/undeveloped). The development level identified by the fuzzy system is incorporated in cellular automata transition rules definition to identify the urban neighborhood threshold, beyond which a test pixel is allowed to develop. The fuzzy system parameters are defined based on the effect each input variable has on the urban growth potential. Synthetic data are used to present and test the principles of using fuzzy cellular automata for artificial city growth modeling. Using the developed fuzzy cellular automata algorithm, the growth of Indianapolis over three decades is studied. In addition to satellite images, digital elevation model, road networks and population data are included in the modeling. The transition rules of the cellular automata are calibrated spatially and over time using multitemporal images. Modeling results are evaluated on a township basis for accuracy assessment. Results indicate the role of fuzzy system in preserving the spatial continuity of the growth process. Partial development concept introduced by the fuzzy system takes into account the contribution of the test pixel in identifying the required urban neighborhood for its development. The results indicate also the importance of spatial calibration based on township specific features in improving the prediction results. Temporal calibration of the model adapts the dynamic growth pattern over time to improve the modelling accuracy. Careful definition of fuzzy parameters is crucial for accurate modeling. Stable accuracy results for prediction are achieved over time. Average accuracies of 95.81% for short term prediction (5 years) at 1992 and 95.49% for long term prediction (11 years) at 2003 are obtained.