Revealing Hidden Deformation Patterns in Shallow Creeping Landslides: A Data-Driven InSAR Phase Filtering Method Addressing Geometric Distortions

Landslides, prevalent in mountainous regions, pose significant risks to both human life and infrastructure, making accurate monitoring crucial for risk assessment and early warnings. interferometric synthetic aperture radar (InSAR) has emerged as a key technology for dynamic, all-weather monitoring of landslide deformation, with phase filtering being a critical step for restoring true surface phases. Traditional phase filtering methods, such as homogeneous pixel filtering, often face challenges in complex terrains like mountainous regions, where environmental noise, thermal interference, and SAR-induced geometric distortions make it difficult to identify homogeneous pixels, frequently obscuring critical information. This study introduces a data-driven phase filtering method for mountainous regions. It considers the geometric distortions of distributed scatterers. We explore the relationship between terrain characteristics, radar imaging geometry, and InSAR deformation phases, utilizing the maximum detectable deformation gradient (MDDG) to distinguish deformation units in complex terrains. The method applies deep learning techniques to effectively denoise interferometric phases, enhancing the accuracy of deformation monitoring in areas with low coherence and high geometric distortion. We demonstrate the applicability of this approach through the Hongyanzi landslide in the Hanyuan Reservoir area, Sichuan Province, China, where traditional filtering methods fail due to geometric distortions and vegetation noise. Experimental results show that our method significantly improves phase filtering accuracy compared to existing model-driven and data-driven methods, with a root mean square error (RMSE) smaller than conventional techniques when compared to global navigation satellite system (GNSS) observations. Furthermore, the time-series deformation monitoring accuracy improved by approximately 53%, providing a more efficient and reliable solution for InSAR-based landslide monitoring in mountainous regions.