Integrated Ministack-InSAR Monitoring and Multi-Source-Factor-Informed CNN-LSTM Prediction of Reservoir-Bank Landslide Deformation: A Case Study of the Xiaolangdi Reservoir, China

Reservoir-bank landslides in canyon-type reservoirs commonly show slow creep and episodic acceleration driven by rainfall infiltration, groundwater fluctuation, reservoir regulation, and engineering disturbance. Sparse coherent targets and the high computational burden of long synthetic aperture radar image stacks limit operational time-series interferometric synthetic aperture radar monitoring (TS-InSAR). Moreover, effectively linking long-term deformation monitoring with mechanism interpretation and short-term prediction remains challenging. This study develops an integrated framework for the Xiaolangdi Reservoir, China, combining Ministack-InSAR, interpretable machine learning, and multi-source deep learning. Sentinel-1A images acquired from 2018 to 2024 were processed using Ministack-InSAR, while random forest (RF) and extreme gradient boosting (XGBoost) combined with Shapley additive explanations (SHAP) were employed to identify the dominant conditioning factors controlling deformation. Based on the identified factors and historical deformation information, multi-source deep learning models were further developed for short-term deformation prediction. Ministack-InSAR improved the spatial continuity of monitoring points (MPs) and preserved phase quality in vegetated reservoir-bank slopes. The RF/XGBoost–SHAP results identified groundwater storage, rainfall, distance to rivers, overburden thickness, and road density as the dominant controls on the spatial variability of deformation. Among the tested prediction models, the multi-source-factor convolutional neural network–long short-term memory (MSF-CNN-LSTM) model achieved the best overall performance, with a mean absolute error (MAE) of 3.0 mm, a root mean square error (RMSE) of 4.5 mm, and a coefficient of determination (R2) of 0.885. These results demonstrate that the proposed framework can effectively integrate deformation monitoring, mechanism interpretation, and short-term prediction, providing practical support for active-zone identification and early warning of reservoir-bank landslides.