A Spatiotemporal Fusion Network Based on InSAR Time-Series Evolution for Landslide Detection in Noise-Dominated Mountainous Areas

Abstract

Rapid and accurate identification of potential landslides is essential for disaster prevention and mitigation. Interferometric Synthetic Aperture Radar (InSAR) provides an effective tool for detecting slow-moving landslide deformation. However, in mountainous areas with dense vegetation and strong atmospheric delays, InSAR noise and deformation can be highly intermingled in both amplitude and spatial morphology. Traditional image-segmentation models tend to rely primarily on visual similarity rather than physically plausible temporal evolution, which limits their ability to achieve accurate automated recognition when transferred to complex, noise-dominated environments. To address these challenges, this article proposes a Fusion Spatial–Temporal Network (FuSTNet). FuSTNet captures evolution patterns from landslide deformation time series, while combining deformation-rate maps and digital elevation model data to encode slope geometry constraints. In addition, a geospatial continuity constraint is introduced to incorporate the prior that landslide deformation generally exhibits local spatial continuity, thereby reducing false detections caused by isolated noise. Experimental results show that FuSTNet achieves 93.70% mIoU, 96.69% recall, and 96.75% F1-score on the test set. Its detection results are highly consistent with expert interpretation and further identify potential landslides with low-amplitude deformation that may be overlooked during manual mapping. The proposed framework shifts the emphasis from image segmentation to dynamics-aware discrimination, improving the mechanism-level transparency and geological consistency of the predictions and providing a reliable approach for InSAR-based landslide-deformation recognition in mountainous areas affected by strong noise.