3D deformation retrieval and LSTM prediction for landslides in the upper Yellow River based on aspect-constrained SBAS-InSAR

The Lijiaxia–Jishixia reach of the upper Yellow River, characterized by deeply incised gorges and a cascade of major hydropower stations, is highly prone to landslides that pose severe threats to critical infrastructure and public safety. However, systematic three-dimensional (3D) deformation monitoring, kinematic mechanism decoupling, and reliable short-term forecasting remain limited in this region. We propose an integrated framework that combines SBAS-InSAR-based 3D deformation inversion, spatial pattern recognition using Independent Component Analysis (ICA), and Long Short-Term Memory (LSTM) predictive modeling. Using 113 Sentinel-1A ascending and descending image pairs acquired between 2021 and 2023, we reconstructed high-precision surface deformation velocity fields and time series. At the regional scale, 123 active landslides were identified, predominantly clustered along both riverbanks. Their spatial distribution is closely associated with long-term river incision and lateral scouring by the Yellow River, which actively steepens valley slopes. In a detailed single-slope analysis, incorporating a slope-aspect constraint into a dual-orbit fusion model enabled high-precision resolution of the 3D displacement field of the Suozi landslide, revealing dominant east–west horizontal sliding accompanied by minor, spatially heterogeneous vertical subsidence. To further elucidate the underlying kinematic mechanisms, Principal Component Analysis (PCA) and spatial-domain ICA were applied to the east–west displacement series, decomposing the signal into three key deformation modes: deep-seated rock creep, broad regional background deformation, and rainfall-sensitive shallow sliding. This decomposition provides quantitative evidence for kinematic analysis and hazard zonation. Finally, to capture short-term landslide evolution, we developed an LSTM network based on cumulative deformation features, which outperformed Support Vector Machine (SVM) and Back Propagation (BP) neural networks, achieving superior predictive accuracy (R² > 0.70). Collectively, these findings enhance the accuracy and timeliness of 3D landslide monitoring and offer a reproducible technical framework for multi-scale early warning, risk assessment, and mitigation planning at the basin level.