LogISEL: an interpretable ensemble framework for real-time landslide prediction in the Indian Himalayas

Abstract Landslides remain a significant hazard across the Indian Himalayan regions, where steep slopes, episodic rainfall, and man-made constructions contribute to slope instability. Accurate prediction remains a challenge due to data imbalance, environmental variation, and the limited interpretability of predictive models. This study presents LogISEL, an interpretable ensemble framework designed for real-time landslide prediction at a ten-minute horizon using multi-sensor field data. The proposed framework is tested at two geologically different sites—Griffon Peak and Ghora Farm in Himachal Pradesh. LogISEL integrates six machine learning classifiers through a logistic regression meta-learner, allowing transparent and temporally coherent predictions. Compared with state-of-the-art baselines, LogISEL achieved up to 15% improvement at Griffon Peak and over 20% at Ghora Farm, without requiring artificial balancing of the inherently imbalanced dataset. SHAP-based interpretability analysis revealed that location-specific precursors, such as temperature lags in alpine terrain and light–pressure variations in cultivated slopes, dominate the prediction of slope movements. These features control predictions with respect to movement. These findings underscore the model’s adaptability and interpretability, addressing critical shortcomings of the existing ML and physics-based approaches. Overall, LogISEL represents a significant step toward designing efficient early-warning systems within landslide-prone regions, offering a transferable framework for integrating data-driven prediction with decision support in complex geophysical environments.