Spatiotemporal Dynamic Graph Isomorphism Network for Satellite Image Time Series Classification

The evolution patterns of land cover types exhibit similarities, which increases the difficulty of high-precision temporal land cover classification tasks. Utilizing satellite image time series (SITS) data can effectively capture the spatiotemporal features of land surface changes. This paper proposes a novel spatiotemporal dynamic graph isomorphism network (STDGIN) for the classification of SITS data. The STDGIN consists of the temporal graph network module and the temporal continuity module. First, each band of SITS is combined into nodes using aggregation convolution, and the dynamic adj acency matrix is designed to enable information transmission between graphs. Then, the hierarchical information of the graph and the local spatiotemporal features of SITS are captured using the node aggregation mechanism. Concurrently, the Bi-LSTM-based module is constructed to capture the temporal continuity features of SITS. Finally, the two features are fused into spatiotemporal representations, and land categories are predicted using a linear classifier. Simulation results indicate that the STDGIN can achieve superior performance on a public SITS dataset (TiSeLaC).