Explainable Machine Learning for Multi-Temporal Landslide Detection along the Mechi– Tamor Corridor, Eastern Nepal"

This study explores a multi-temporal, multi-sensor framework for landslide detection along the Mechi Highway–Tamor River corridor in Ilam, Panchthar, Taplejung and Dhankuta districts of eastern Nepal. The framework is evaluated on two contrasting rainfall-triggered events, a late-monsoon withdrawal event in early October 2025 and a peak-monsoon event on 20 July 2026, which differ in rainfall intensity, duration, structure and antecedent wetness. An event-based inventory of 116 landslide scars was compiled through visual interpretation of Sentinel-2 imagery and cross-checked against a regional baseline of 255 landslide records. This inventory was combined with Sentinel-1 SAR backscatter change, optical spectral indices, terrain derivatives, antecedent rainfall and lithology to create a feature table of 410 units with 820 observations (unit × event). Random Forest and XGBoost classifiers were trained and evaluated under a naive random split and two spatially independent validation schemes (leave-one-municipality-out and spatial block cross-validation); the spatial schemes yielded F1 scores of 0.47–0.56 and ROC-AUC of 0.80–0.84, substantially lower than the optimistic random-split estimate (F1 = 0.77–0.80, ROC-AUC = 0.95), demonstrating the necessity of spatial validation for this problem. SHAP analysis showed that SAR backscatter change was the most important predictor overall, but that optical spectral indices, especially the Normalized Burn Ratio (NBR) difference, carried more signal wherever cloud-free images were available. SAR's value in this setting lies in cloud penetration rather than in superior predictive power. Combining predicted susceptibility with proximity to schools, health facilities and road bridges identified 35 critical corridor segments.