Spatial constraints in cellular automata-based urban growth models: A systematic comparison of classifiers and input urban maps

Spatial constraints are fundamental to integrating the spatial suitability to urbanization into Cellular Automatabased (CA) urban growth models, but there is a lack of consensus on the optimal methods for this purpose. This study compared the performance of three probabilistic classifiers to generate suitability surfaces for CA-based urban growth models: Logistic Regression using Generalized Linear Model (LR-GLM), Logistic Regression using Generalized Additive Model (LR-GAM), and Random Forest (RF). The study also evaluated the sensitivity of these classifiers to the input urban map adopted as a dependent variable. For this analysis, seven maps were tested: the historical urban map containing the entire extent of the urban footprint, and six additional maps containing only the recently urbanized areas over timeframes ranging from one year up to two decades. The comparison evaluated the goodness of fit of the suitability surfaces and the spatial accuracy of the urban growth simulations, using five large Brazilian cities as case study areas. The results revealed that the RF classifier significantly outperformed the LR-based classifiers. However, this overperformance was more prominent when incorporating the new urban cells over the last one to two decades of growth as input urban maps. In addition, the sensitivity analysis of the input urban maps emphasized the benefits of calibrating the classifier using the recently urbanized cells rather than the historical urban extent. We consistently observed these results concerning classifiers and input urban maps across all five case study areas. Thus, the RF classifier combined with a training dataset containing the newly urbanized areas over at least the last 10 years systematically resulted in the suitability surfaces with the highest predictability among all tested scenarios.