A unified deep learning framework coupling InSAR-derived spatiotemporal kinematic features for robust landslide displacement prediction

Accurate displacement prediction is a prerequisite for assessing landslide stability. However, slow-moving landslides in remote regions often lack ground-based monitoring infrastructure, leading to significant gaps in in situ displacement measurements and triggering factor data. To address this data scarcity, this study presents a unified spatiotemporal prediction framework by coupling time-series interferometric synthetic aperture radar (TS-InSAR) analysis with kinematic feature learning. Variational mode decomposition was applied to the displacement time series derived from 190 Sentinel-1 scenes, resolving the total displacement into monotonic gravitational trends and seasonal fluctuations. Spatial grey correlation analysis suggested that the TS-InSAR displacement sequences of individual monitoring points exhibited strong associations, which were further amplified following the signal decomposition process. To compensate for missing environmental triggers, spatial feature vectors were constructed based on Euclidean distance as a robust physical proxy for the kinematic consistency of the landslide mass. By integrating these spatial kinematic features with a long short-term memory network, a spatiotemporal prediction framework termed SFV-LSTM was established for nonlinear displacement forecasting. In comparative experiments, the SFV-LSTM framework consistently outperformed ten benchmark models encompassing both traditional machine learning and modern deep learning architectures, achieving root mean square error (RMSE) reductions ranging from 22.74% to 58.44%. To evaluate the robustness and cross-regional applicability of the framework, more than 1200 Sentinel-1 scenes were analyzed across six additional landslides characterized by distinct driving mechanisms, including seismic-triggered instability, rainfall-induced flows, and coastal water-level-driven creep, among others. The framework demonstrated consistent performance with RMSE reductions of 13.28% to 30.28%. Furthermore, the associations between triggering factors and nonlinear displacement were evaluated using partial correlation and wavelet analysis. Additionally, phase synchronization analysis was conducted to investigate the asynchrony among multiple drivers. The results indicate that, in landslides influenced by multiple triggers, temporal misalignment between drivers may act as feature noise, thereby degrading predictive performance. In contrast, the proposed framework bypasses this interference by leveraging stable spatial kinematic consistency, offering a physics-informed and robust solution for landslide motion modeling. • Unified framework integrating InSAR spatiotemporal features for landslide prediction. • Coupling spatiotemporal kinematic features effectively improves prediction accuracy. • VMD decomposes landslide displacement into a creep trend and seasonal fluctuations. • Asynchronous triggers potentially act as noise, affecting prediction accuracy. • Generalization capability validated across six landslides with diverse mechanisms.