Spatiotemporal prediction of reservoir landslide deformation by coupling multi-orbit InSAR with a hydrometeorology-modulated deep learning model

Reservoir impoundment and water-level fluctuations often reactivate slow-moving landslides, but reliable three-dimensional deformation prediction remains difficult at the hillslope scale. This is mainly because single-orbit InSAR line-of-sight measurements cannot reveal true downslope kinematics or hydrologically driven processes. We investigated two landslides, L1 (Xiapingzi) and L2 (Maocaofang), in the Baihetan Reservoir area, China. Multi-temporal Sentinel-1 InSAR from ascending and descending orbits was combined with a Surface Parallel Flow (SPF) model to reconstruct slope-aligned displacement. Using these deformation fields, we developed a Feature-wise Linear Modulation-based Convolutional Gated Recurrent Unit (FiLM-ConvGRU), coupling ConvGRU-based spatiotemporal learning with precipitation and reservoir water-level conditioning. Lagged dependence analysis revealed site-dependent associations: L1 showed stronger and more consistent relationships with reservoir water-level fluctuations, while L2 showed weaker and more heterogeneous responses, with characteristic lags of 0–24 days. The FiLM-ConvGRU achieved PCC values of 0.871 and 0.735 and R2 values of 0.732 and 0.490 for L1 and L2, respectively. Compared with ConvGRU, PCC increased from 0.858 to 0.871 for L1 and from 0.691 to 0.735 for L2, indicating site-dependent predictive information from environmental conditioning. These results show that integrating multi-orbit InSAR, terrain-constrained displacement retrieval, and hydrometeorology-conditioned deep learning provides a route for short-term full-field deformation forecasting of reservoir landslides.