A hybrid cellular automaton model integrated with 3DCNN and LSTM for simulating land use/cover change

Abstract

Accurate simulation of land use/cover change (LUCC) is crucial for societal development. LUCC is a nonlinear spatiotemporal process with complicated relationships and latent dependencies on spatial and temporal neighborhoods. It is a challenge for conventional statistical or machine learning methods to efficiently obtain high-level representations of spatiotemporal information and time series features at the same time. To address this issue, we introduced a hybrid model integrating deep spatiotemporal networks and cellular automata, named DST-CA. This model uses a 3D Convolutional Neural Network (3DCNN) to capture local short-term spatiotemporal features and Long Short-Term Memory (LSTM) to extract long-term chronological featurereferences, thereby more comprehensively capturing the nonlinear spatiotemporal characteristics of LUCC. We employed DST-CA to simulate LUCC in Guangdong Province from 2015 to 2020. The results indicate that DST-CA outperforms traditional models and the four temporal-feature models across four evaluation metrics, including Overall Accuracy (OA), F1-score, Figure of Merit (FoM), and Kappa coefficient. Compared to the temporal-feature models 3DCNN-CA and LSTM-CA, these metrics improved by 1.5%, 1.61%, 14.36%, and 2.75%, respectively. This implies that the DST-CA possesses outstanding global simulation capabilities and a superior ability to capture the spatiotemporal characteristics of LUCC. Finally, we forecasted LUCC for Guangdong Province in 2025.