Landslide susceptibility mapping using tree-based machine learning classifiers and remote sensing derived conditioning factors: A case study of Chikmagalur District, Western Ghats, India

Chikmagalur district of Karnataka state, situated within the boundaries of the Western Ghats is highly susceptible to landslides, especially during the monsoon season. Despite the recurring nature of these slope failures, limited research has been conducted to assess and mitigate the risk of landslides in the region. Existing studies often lack a comprehensive analysis of the triggering elements and rely on basic machine learning (ML) techniques, even though there are several advanced techniques that are being adopted across the world. A comprehensive dataset of the study area was prepared by integrating twenty different Landslide Conditioning Factors (LCFs) sourced from different remote sensing techniques and the information of 197 historical landslide events acquired from the Geological Survey of India (GSI). The 5-fold stratified cross validation method was applied to generate training and testing dataset in different iterations. Four different tree-based ML classifiers including Decision Trees (DT), Random Forest (RF), Extreme Gradient Boosting (XGBoost) and Categorical Boosting (CatBoost) were employed to prepare the models to predict the landslide prone areas of the district. These classifiers were specifically chosen because they have the capability to handle feature importance and do not require separate feature selection methods, which are often subjective and difficult to standardize. These ensemble models were then evaluated using different performance metrics that are generally used to evaluate classification models. CatBoost classifier exhibited superior performance, achieving an accuracy of 87.93%, with a precision of 0.85, recall of 0.913, F1-score of 0.88, and an AUC-ROC value of 0.95. Although the RF model also demonstrated strong and competitive performance across all the evaluation metrics, CatBoost was selected for the final preparation of landslide susceptibility map (LSM) due to its comparatively higher recall and AUC-ROC values, which are critical for reliably identifying landslide-prone areas. Consequently, the final LSM was generated using the CatBoost model. According to the LSM, approximately 20.53% of the total district falls within the range of high susceptibility with prediction probability values ranging from 0.6 to 1.0. A major portion, around 47.04% lies within moderate susceptibility zones, and the remaining percentage corresponds to places that are relatively safe from slope failures. Furthermore, the feature importance scores extracted from CatBoost model revealed that slope, rainfall, soil type and distance to road are the main factors that contribute to triggering slope failures in the study area. The application of reverse geocoding techniques on the final LSM indicated that, southwestern and southern taluks including Mudigere, Sringeri, border regions of Koppa and the southern parts of Chikmagalur exhibit a high concentration of landslide prone areas compared to other places. This map serves as a critical tool for early warning systems and informed decision-making to reduce landslide risks in the district.