Active landslides pose a persistent threat to life and infrastructure worldwide, underscoring the need for reliable displacement monitoring and forecasting. Interferometric synthetic aperture radar (InSAR) offers a cost-effective means of measuring surface deformation, yet its application to displacement prediction is hindered by the line-of-sight (LOS) viewing geometry and irregular acquisition intervals. To overcome these constraints, we develop a data-driven framework that fuses multi-temporal (MT) InSAR, global navigation satellite system (GNSS) observations, and machine learning models to predict three-dimensional landslide displacement time series. Using dual-track Sentinel-1 data from 2018, we first reconstruct full three-dimensional displacement fields through weighted inversion constrained by GNSS measurements. The resulting time series are then decomposed using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to separate monotonic gravitational trends from oscillatory environmental responses. A third-order polynomial model captures the long-term trend, whereas the periodic component is predicted with a long short-term memory (LSTM) network whose hyperparameters are automatically tuned for complex periodic patterns. Applied to coastal landslides in California, the framework achieves mean prediction errors of 1.603 mm, 3.266 mm, and 6.278 mm for east-west, north-south, and vertical components respectively, with correlation coefficients exceeding 0.83 for horizontal displacements. The mean absolute percentage errors below 0.4% for horizontal components demonstrate the framework’s capability to capture essential deformation patterns while maintaining operational accuracy. By utilizing the complementary strengths of geodetic observations and machine learning, this approach advances landslide forecasting from traditional numerical or analytical analysis to comprehensive three-dimensional prediction, providing a practical tool for early warning systems in landslide-prone regions.
Bi et al. (Sun,) studied this question.