SFEL a machine learning framework for forecasting radar backscatter based ground deformation

Accurate prediction of soil deformation is crucial for mitigating landslide hazards in mountainous regions, such as the Himalayas, where terrain instability continually threatens infrastructure and livelihood. While previous models have often relied on either geotechnical data or satellite-based observations in isolation, this study presents an integrated and interpretable framework for forecasting radar-backscatter-based deformation. The proposed Stacked Forecasting Ensemble Learner (SFEL) merges geotechnical parameters, traditional soil indicators, and Synthetic Aperture Radar (SAR)-derived ΔVV deformation proxies processed using the Google Earth Engine into a unified ensemble learning pipeline. Using dual feature selection Pearson correlation and mutual information, the model achieved strong predictive performance for both the 1-month and 6-month horizons, with low root mean squared errors of 0.01 and 0.03 ΔVV/year, respectively. Shapley additive explanations confirmed the complementary roles of traditional and scientific features, revealing distinct shifts in feature importance across various temporal scales. While the results indicate the robustness of the SFEL model, they also highlight certain limitations, including regional data constraints and the proxy nature of radar backscatter as a deformation indicator. Nonetheless, this study demonstrates that integrating experiential soil knowledge with remote sensing and machine learning can advance both predictive reliability and interpretability, offering a scalable pathway for informed slope management and regional-hazard monitoring.