Novel Spatiotemporal ConvLSTM-Based Cellular Automata Model for Simulating Urban Expansion

Accurately capturing and simulating the spatiotemporal dynamics of urban expansion remains a persistent challenge in urban planning and environmental management. Traditional machine learning methods often struggle to capture the complexity of spatial features and the multi-scale spatiotemporal dynamics involved in deriving the transition rules of cellular automata (CA), making it difficult to balance detailed local expansion with broader overall trends during the urban expansion simulation process. To address these challenges, we propose a novel spatiotemporal CA model based on the convolutional long short-term memory (ConvLSTM) neural network. This model, termed ST-ConvLSTM-CA, simultaneously enhances critical spatial features and incorporates multi-scale spatiotemporal features by integrating the shuffle attention mechanism and a spatiotemporal expression (SE) block into the ConvLSTM architecture. First, the channel dimensions of spatial features are grouped and channel spatial attention is performed in parallel, allowing for efficient fusion and weighting of key spatial features with minimal computational cost. Second, the SE module is used to integrate the multi-scale information within ConvLSTM cells, enhancing the representation of neighborhood details across different spatial granularities and capturing dynamic temporal features. Finally, the improved model generates conversion probability maps, which are used by the CA model to simulate time series urban expansion maps. Using the urban agglomerations of the Guangdong-Hong Kong-Macao Greater Bay Area, the Yangtze River Delta, and Beijing-Tianjing-Tangshan as case studies, results show that ST-ConvLSTM-CA achieves the highest accuracy across all three regions. This demonstrates its robust adaptability to complex spatiotemporal data and superior predictive performance.