ABSTRACT Precise landslide susceptibility mapping (LSM) is essential for landslide hazard assessment, especially in geologically fragile areas. However, many LSM studies mainly use random splitting, which exploits spatial autocorrelation, leading to overly optimistic models. To overcome this issue, a spatially explicit landslide susceptibility model was developed for the Chandra‐Bhaga Basin in Himachal Pradesh, India, employing six machine learning algorithms (KNN, RF, SVM, GaussianNB, LR, and LDA). Factors redundancy and distributional skewness were mitigated through feature engineering approach, resulting in the selection of 14 conditioning factors. Global Moran's I, computed using a row‐standardized k ‐nearest neighbor ( k = 8) spatial weights matrix, revealed significant spatial clustering in both the landslide inventory ( I = 0.71, p < 0.01) and conditioning factors. To minimize spatial overfitting, models were trained and evaluated using K‐fold spatial block cross‐validation. Analysis of ML‐based models, pertaining to AUC‐ROC scores, delineated that RF (99.5%) outperformed other models and showed the highest landslide predictive capability, followed by SVM (93.7%), and KNN (93.1%). SHAP‐based interpretation revealed lithology, aspect, slope, and distance from lineament as dominant drivers, revealing both magnitude and direction of feature influence. The manuscript identifies clear spatial variation of landslide vulnerability in the Chandra‐Bhaga basin, with strong instability located in the Southern and Eastern sectors owing to fragile rock structure. Overall, this framework integrates spatial autocorrelation and spatial block cross‐validation to get a reliable ML‐based landslide susceptibility assessment, which provides a solid scientific foundation for conducting land‐use planning and infrastructure development in this geodynamically sensitive Himalayan region.
Dhiman et al. (Wed,) studied this question.