Integrating Deep Learning and Superpixel-based Graph Convolution for Ecological Constraint Calibration in Urban Growth Modeling: The ANN-SGCCA Approach

Spatial zoning is an effective way to protect ecological resources and guide sustainable urban growth. However, existing urban growth simulation studies have the following limitations: 1) the heterogeneous neighborhood spatial interactions of ecological factors at multiple scales are often overlooked; 2) deep learning algorithms are less used to learn the characteristics of these complex interactions and capture ecological constraints. Therefore, we develop an novel approach to explore the calibration process of ecological constraints in urban growth modeling. The model extracts development suitability through artificial neural networks (ANNs), and further extracts pixel-level and super-pixel-level ecological features through convolutional neural networks (CNNs) and graph convolutional networks, respectively. The proposed model was used to simulate the urban dynamics in Quanzhou, a city in southeastern coastal China. The results show that compared with the urban cellular automaton model based on ANNs, the proposed model has achieved 5.60% and 7.0% improvements in Figure of Merit and Recall-based Efficacy respectively. Meanwhile, compared with the urban cellular automaton model that only uses CNNs to obtain ecological constraints, The above two metrics obtained by the proposed model increased by 2.37% and 2.46%. The developed model can potentially be applied to simulate urban dynamics under complex ecological resource constraints.