Abstract: This paper proposes a cellular automata model for land use evolution that integrates graph convolutional neural network and geographical condition-driven effects, in view of the fact that existing land use evolution modeling methods are limited by Euclidean spatial constraints and thus can not effectively model the emerging phenomenon of land with no historical data in the spatial neighborhood. Firstly, the spatial neighborhood multiscale effect of the cell is modeled by introducing a dilated convolution layer. Then the geographic condition similarity network is constructed based on the geographic condition vector of the cell, applying the graph convolutional neural network to extract the regional potential features in it. Finally, land use evolution is simulated by fusing the artificial neural network and the cellular automata. This paper improves the ability of the model to the emerging phenomenon of land with no historical data in the spatial neighborhood by modeling the geographical condition-driven effects, and successfully achieves effective simulation of urban land use evolution. The experiments were conducted in three research regions including Xuanwu district of Nanjing, Furong district of Changsha, and Nanchang. The results show that our method improves in overall accuracy (OA), Kappa coefficient, and figure of merit(FoM), allowing for more accurate modeling of land use evolution than the classical land use evolution modeling method considering only Euclidean spatial neighborhood features.

Key words: land use evolution, cellular automata, graph convolutional networks, geographic simulation, geographical condition-driven effects