GateSTNet: a novel model for mapping spatiotemporal forest cover

Forest resources have important environmental and economic values. Although many models have been developed to analyze the dynamic changes of forest cover, significant challenges remain in explaining the relationships between geographically non-contiguous regions with similar forest cover changes. To address this issue, we propose Gate Spatiotemporal Network (GateSTNet), a new model for predicting multi-region, multi-feature, and multi-network datasets. The GateSTNet comprises the gate GCAN (Graph Convolutional Attention Network) information focus module and the ConvOut spatiotemporal predict module. The network effectively combines the capabilities of graph convolutional networks (GCNs) for processing non-Euclidean spatial data and convolutional LSTM (ConvLSTM) networks for spatiotemporal data prediction. Experimental results show that GateSTNet performs well in both prediction accuracy and efficiency. Compared with existing models, including LSTM, CNN-LSTM, and ConvLSTM, GateSTNet exhibits excellent prediction accuracy and stability. This enhanced performance is attributed to the network's ability to effectively integrate spatial and temporal information, making it a powerful tool for predicting forest cover dynamics and informing environmental management and policy decisions.