Integrating cellular automata with long short-term memory neural network to simulate urban expansion using time-series data

Abstract

Cellular automata (CA) have been widely used to simulate urban expansion. However, previous studies rarely explored how to extract transition rules from time-series urban extent maps, hence modeling the temporal dependency of urban expansion process. To solve this issue, this study proposes a novel CA model based on the long-short term memory method (LSTM-CA). The LSTM-CA model combines artificial neural network (ANN) and LSTM to calculate transition potential of cells using information derived from time-series urban extent maps and other spatial variables. A case study was conducted in Foshan, China, to evaluate the performance of the proposed model. Experiments were carried out to compare the performance between LSTM-CA and the other two urban expansion simulation models of FLUS and GeoSOS. The results indicated that LSTM-CA outperforms FLUS and GeoSOS in terms of figure of merit (FoM) and landscape metrics. Sensitivity analysis was also performed, revealing that the size of neighborhood variables derived from time-series urban extent maps has a moderate impact on LSTM-CA model's performance, and the output size of LSTM model affects the simulations when the neighborhood size is 3 × 3. Moreover, the Shapley algorithm was employed to explain how different variables influence the transition potential. The results of Shapley analysis suggest that the neighborhood variables derived from the latest input year is the most important factor. Overall, the results of the case study confirmed the effectiveness of the proposed LSTM-CA model to simulate urban expansion.

Keywords

• Urban expansion simulation
• Time-series urban extent maps
• Cellular automata
• LSTM
• Shapley algorithm