Landslide Hazard Identification and Prediction in Complex Mountainous Areas Using Ascending and Descending Orbits...

Time-series InSAR is an important means for early identification and monitoring of landslides. However, in complex mountainous areas, it still faces challenges such as significant geometric distortions and complicated deformation mechanisms. To address these issues, this paper proposes a landslide identification and prediction framework that integrates ascending and descending orbits InSAR observations with physics-guided deep learning. Taking Yangbi County, Yunnan Province, as a case study, we combined ascending and descending Sentinel-1A data and employed the SBAS-InSAR method to identify potential landslides, detecting a total of 41 hazardous sites. The cumulative displacement time series of typical landslides were further extracted along the slope aspect to more realistically reflect landslide movement characteristics. On this basis, wavelet decomposition was introduced to separate the displacement series into trend and periodic components. Gray relational analysis was then used to select influencing factors such as precipitation and temperature, and a stepwise prediction model based on LSTM (WT-LSTM) was constructed. The results indicate that the model achieves significantly higher prediction accuracy at characteristic points of the representative landslide (RMSE = 1.16-2.19 mm) compared to standalone LSTM and SVR models. These findings demonstrate its effectiveness and potential applicability in landslide deformation monitoring and prediction in complex mountainous areas, while also providing a useful reference for landslide risk early warning.