Adaptive prototype-guided multi-temporal landslide extraction from long-term landsat imagery

Reliable multi-temporal landslide extraction is essential for understanding long-term landslide dynamics, identifying seismic or climatic triggers, and supporting large-scale hazard monitoring. However, accurate extraction remains challenging due to cross-temporal feature inconsistency and high intra-class variability caused by changes in surface conditions, illumination, and imaging sensors. To address these issues, we propose an adaptive prototype-guided framework that combines a PCA-based candidate extraction module with a DINOv2-enhanced Adaptive Prototype-Guided Model (APGM). The APGM introduces two synergistic modules, including an Adaptive Prototype Learning Module for dynamic and diverse sub-class prototype generation to capture complex landslide patterns while maintaining temporal stability, and a Progressive Prototype-Guided Decoder for progressive refinement of spatial boundaries across hierarchical feature scales refinement. Additionally, a Prototype Consistency Loss is designed to constrain feature variations across time. Applied to annual Landsat composites from 2001 to 2020 over the Hengduan Mountains region, our method achieves consistent and accurate long-term extraction results, improving IoU and F1-score by 2.21% and 2.86%, respectively, compared with seven state-of-the-art methods. This approach enhances the robustness of large-scale multi-temporal landslide mapping and provides a foundation for dynamic inventory construction and regional disaster risk analysis.