Coupling cellular automata with area partitioning and spatiotemporal convolution for dynamic land use change simulation

Urbanization processes have accelerated over recent decades, prompting efforts to model land use change (LUC) patterns for decision support and urban planning. Cellular automata (CA) are extensively employed given their simplicity, flexibility, and intuitiveness when simulating dynamic LUC. Previous research, however, has ignored the spatial heterogeneity among sub-regions, instead applying the same transition rules across entire regions; moreover, most existing methods extract neighborhood effects with only one data time slice, which is inconsistent with the nature of neighborhood interactions as a long-term process exhibiting obvious spatiotemporal dependency. Accordingly, we propose a hybrid cellular automata model coupling area partitioning and spatiotemporal neighborhood features learning, named PST-CA. We use a machine-learning-based partitioning strategy, self-organizing map (SOM), to divide entire regions into several homogeneous sub-regions, and further apply a spatiotemporal three-dimensional convolutional neural network (3D CNN) to extract the spatiotemporal neighborhood features. An artificial neural network (ANN) is then built to create a conversion probability map for each sub-region using both spatiotemporal neighborhood features and factors that drive the LUC. Finally, the dynamic simulation results of entire study area are generated by fusing these probability maps, constraints and stochastic factors. Land use data collected from 2000 to 2015 in Shanghai were selected to verify our proposed method. Four traditional models were implemented for comparison, including logistic regression (LR)-CA, support vector machine (SVM)-CA, random forest (RF)-CA and conventional ANN-CA. Results illustrate that the proposed PST-CA outperformed four traditional models, with overall accuracy increased by 4.66%~6.41%. Moreover, three distinctly different "coverage rate-growth rate" composite patterns of built-up areas are shown in the SOM partitioning results, which verifies SOM's ability to address spatial heterogeneity; while the optimal time steps in 3D CNN generally maintained a positive correlation with the growth rate of built-up areas, which implies longer temporal dependency should be captured for rapidly developing areas.