HGAT-VCA: Integrating high-order graph attention network with vector cellular automata for urban growth simulation.

Since urban growth results from frequent spatial interaction between urban units, adequate representation of spatial interaction is important for urban growth modeling. Among urban growth models, vector-based cellular automata (VCA) excels at expressing spatial interaction with realistic entities, and has accordingly been used extensively in recent studies. However, two issues with VCA modeling still remain: 1) inefficient manual selection of interaction targets with various neighborhood configurations; 2) inaccurate quantification of interaction intensity due to ignorance of spatial heterogeneity in entity interaction. To address these two limitations, this study proposed a novel VCA model with high-order graph attention network (HGAT-VCA). In this model, a graph structure is first built from the topology adjacency relationship between cadastral parcels. In terms of the HGAT components, the original 1 st -order parcel neighborhood is extended to high-order to capture the distant dependency, while graph attention is applied to quantify the heterogeneous interaction intensity between parcels. Finally, the conversion probability obtained by HGAT is integrated with VCA to simulate urban land use change. Land use data from the Moreton Bay Region in Queensland, Australia from 2005 to 2009 are selected to verify the proposed HGAT-VCA model. Experimental results illustrate that HGAT-VCA outperforms four classical CA models and achieves the highest simulation accuracy (e.g., the increase of FoM is about 40.7%). In addition, extensive neighborhood configuration experiments show that with HGAT only tuning discrete topological order can generate similar accuracy results compared with the repetitive buffer-based neighborhood configuration, and this can significantly improve the calibration efficiency of VCA models. • VCA is integrated with graph neural network to accurately model spatial interaction in urban growth. • High-order neighborhood extension is conducted to extract distant dependency between parcels. • Graph attention is applied for adaptive target selection and intensity quantification of spatial interaction. • HGAT with discrete topological order tuning can highly improve the efficiency of model calibration. [ABSTRACT FROM AUTHOR]