Resumen
Exploring spatial anisotropy features and capturing spatial interactions during urban change simulation is of great significance to enhance the effectiveness of dynamic urban modeling and improve simulation accuracy. Addressing the inadequacies of current cellular automaton-based urban expansion models in exploring spatial anisotropy features, overlooking spatial interaction forces, and the ineffective expansion of cells due to traditional neighborhood computation methods, this study builds upon the machine learning-based urban expansion model. It introduces a spatial anisotropy index into the comprehensive probability module and incorporates a gravity-guided expansion neighborhood operator into the iterative module. Consequently, the RF-CNN-SAI-CA model is developed. Focusing on the 21 districts of the main urban area in Chongqing, the study conducts comparative analysis and ablation experiments using different models to simulate the land use changes between 2010 and 2020. Different model comparison results show that the recommended model in this study has a Kappa value of 0.8561 and an FOM value of 0.4596. Compared with the RF-CA model and the FA-MLP-CA model, the Kappa values are higher by 0.0407 and 0.1577, respectively, while the FOM values are improved by 0.0529 and 0.0654, respectively. Ablation experiment results indicate that removing gravity, SAI, and expansion neighborhood operators leads to a decrease in both Kappa and FOM values. These findings demonstrate that the RF-CNN-SAI-CA model, based on the expanded neighborhood iteration algorithm, effectively integrates spatial anisotropy features, captures spatial interaction forces, and resolves neighborhood cell failure issues, thereby significantly improving simulation effectiveness.