Enhancing landslide disaster prediction by evaluating non landslide area sampling in machine learning models for Spiti Valley India

This study focuses on assessing the effectiveness of two novel sampling methods: Buffer Zone Safe Points (BZSP) and Slope Buffer Safe Points (SBSP). Detailed susceptibility zonation maps were created employing advanced statistical techniques, specifically Extreme Gradient Boosting (XGBoost), Random Forest (RF), and K-Nearest Neighbors (KNN), allowing for an in-depth comparison of their predictive performance.

The findings clearly indicate the advantages of the SBSP technique, showcasing notable improvements in performance across all metrics. In the analysis of Category-II, XGBoost showed a significant rise in the Area Under Curve (AUC) from 0.91 to 0.97, RF increased from 0.89 to 0.97, and KNN improved from 0.87 to 0.94, with corresponding enhancements in accuracy, sensitivity, kappa values, and F1-scores. These advancements highlight the capability of the SBSP method to enhance susceptibility predictions and reduce overestimation in areas of high vulnerability. The Landslide Density Index (LDI) supports these findings, as Category-II sampling provides more dependable estimates across all susceptibility classes, minimizing variability and improving interpretive confidence. This study emphasizes the essential importance of sophisticated sampling techniques in enhancing the dependability of LSM and establishes a fundamental framework for upcoming research focused on reducing landslide hazards in intricate landscapes. The results highlight the importance of integrating various conditioning factors and flexible approaches to enhance regional hazard evaluations and strengthen disaster readiness.