FDRL: a data-driven algorithm for forecasting subsidence velocities in Himalayas using conventional and traditional soil features.

Landslides are a frequent geohazard within the Himalayas, threatening human lives, infrastructure, and indigenous economies. Traditional subsidence velocity forecasting models, however, typically rely on either satellite remote sensing data or geotechnical parameters in isolation, which limits their predictive power and applicability. This work bridges this gap by suggesting an interpretable data-driven model that systematically integrates traditional soil information with geotechnical features for improved prediction. A stacking ensemble regression model called Forecasting Data-Driven Regression Learning (FDRL) was developed on the basis of the last machine learning breakthroughs, including feature selection techniques such as Pearson correlation and mutual information scores. The model combined both quantitative variables (e.g., specific gravity and plasticity index) and qualitative indicators based on conventional soil evaluation procedures (e.g., water retention, odor, and soil color). The FDRL model outperformed baseline regression models with a training Root Mean Squared Error (RMSE) of 1.11 mm/year and a test RMSE of 1.32 mm/year. Explainability analysis with SHAP showed that geotechnical as well as traditional soil characteristics significantly contributed to model predictions, confirming the utility of this hybrid combination. By demonstrating the explanatory potential of traditional soil indicators, typically excluded from scientific models, this study bridges local knowledge systems with modern data science. The method provides a scalable, interpretable, and locally implementable approach to early warning of slope creep and long-term deformation trends, facilitating proactive landslide risk management.