Spatial–Temporal Collaborative Network for Satellite Image Time Series Semantic Change Detection

Semantic change detection (SCD) aims to generate both semantic segmentation maps and binary change (BC) maps from binary temporal remote sensing images and plays a crucial role in remote sensing image interpretation. While learning-based methods have shown strong performance on standard benchmarks, the more challenging satellite image time series (SITS)-based SCD task remains underexplored. Existing SITS-SCD approaches often overemphasize temporal cues while neglecting effective spatial–temporal interaction and are stuck in low-efficiency feature fusion, limiting their representation capacity. To this end, we propose ST3DUNet, a spatial–temporal collaborative network for SITS-SCD. ST3DUNet adopts a U-shaped architecture that extracts hierarchical spatial–temporal representations through a bidirectional spatial–temporal module (BSTM) and recursively fuses them via a collaborative fusion block (CFB) based on dynamically updated prototype matrices. This collaborative fusion enhances decoding by leveraging informative spatial–temporal embeddings. Moreover, we evaluate the robustness of our model under spatial domain shifts, which adhere to real-world remote sensing image decryption conditions. Extensive experiments on three SITS datasets demonstrate that ST3DUNet consistently outperforms independent temporal processing, pairwise temporal modeling, and existing SITS-SCD methods, as well as three Earth observation foundation models (FMs), achieving consistent improvements of approximately +5%, +3%, and +5% in mean intersection-over-union (mIoU) across DynamicEarthNet-RGB, DynamicEarthNet-Sentinel2, and Muds, respectively.