Urban expansion cellular automata model based on multi-structures convolutional neural networks

Based on multi-structures convolutional neural networks, this paper proposes an urban expansion cellular automata model (MSCNN-CA) considering the multi-scale neighborhood information to explore the problem of traditional cellular automata (CA) models merely accounting for a single pixel's factors while mining the urban development suitability. This paper took the main urban zone of Wuhan and Pudong New District of Shanghai as examples to simulate the urban expansion process of the two study areas from 2005 to 2015. The experimental results show that, compared with two traditional CA models (LR, ANN), the three single-structure CNN-CA models constructed in this paper have different degrees of improvement in Kappa coefficient, FoM coefficient, hit rate (h) and miss rate (m). In particular, the FoM coefficient is increased by 23.3%~29.4% in the main urban zone of Wuhan and 20.3%~28.5% in Pudong New District of Shanghai. In addition, compared with the three single-structure CNN-CA models, the MSCNN-CA model is also improved in various indicators. Such as, the FoM coefficient is increased by 0.8%~4.8% in the main urban zone of Wuhan and 2.8%~7.8% in Pudong New District of Shanghai. The two study areas' simulation results show that, compared with the traditional CA models, the urban expansion cellular automata model based on multi-structures convolutional neural network (MSCNN-CA) can effectively improve the accuracy of urban expansion simulation and more realistically reflect the evolution process of urban expansion. Besides, both the stability and the accuracy of the MSCNN-CA model are improved comparing with the single-structure convolutional neural network CA model.