Optimized machine learning based model for the large-scale spatial prediction of landslides at Western Ghats in the State of Karnataka, India

The Western Ghats (WG) of India, as per National Disaster Management Authority (NDMA) statistics, is one of the most landslide-prone regions, with frequent monsoon-triggered landslips. Landslide susceptibility Mapping (LSM) is crucial tool for supporting risk mitigation strategies. This study develops large-scale LSM using an optimized Machine Learning (ML) model for the six districts of Karnataka state including Chikmagalur, Uttara Kannada, Dakshina Kannada, Udupi, Shimoga and Kodagu within WG, all with a history of recurrent landslides. A dataset of 1267 landslides (2014-2020) from the Geological Survey of India (GSI) and twenty landslide conditioning factors (LCF) from diverse sources were used. Feature selection using Pearson correlation, Chi-square identified thirteen key LCFs, with geomorphology, slope, lineament density and rainfall as the most significant. Fourteen ML classifiers including single and ensemble models such as Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT), K-Nearest Neighbours (KNN), Gaussian Process (GP), Linear Discriminant Analysis (LDA), Light Gradient Boosting (LGBM), Random Forest (RF), Artificial Neural Networks (ANN), Extreme XGBoost (XGB), ADAboost, Categorical Boosting (CatB), stacking and bagging were trained using LCFs and LIM. XGB, LGBM and CatB classifiers performed well. Further, hyper-parameter tuning techniques including grid search (GS), random search (RS), Bayesian optimization (BO) and nature inspired Particle Swarm Optimization (PSO) were applied. PSO tuned CatB demonstrated superior performance with the values of 0.9854, 0.952, 0.958, 0.937 and 0.948 for AUC-ROC, accuracy, precision, recall and F1-score respectively, making it as the most effective approach. Final LSM maps were classified into five susceptibility levels: very high, high, moderate, low and very low. Results show that Chikmaglur, Kodagu and Uttara Kannada districts have 5.43% , 4.10% and 1.21% of the highly vulnerable areas. Dakshina Kannada, Shimoga and Udupi districts have less portion of vulnerable areas. These maps provide crucial input for the policymakers to differentiate between vulnerable and safe regions, enabling informed development planning and minimizing landslide risk. • The research mainly focuses on creating the large-scale landslide susceptibility map for the six landslide prone districts of Karnataka state that fall within the boundaries of Western Ghats covering the total geographic area of 38513 km 2 . • Out of twenty initially selected landslide triggering factors, thirteen key factors were identified by applying different feature selection methods and used in creating the dataset. • Fourteen ML models were applied including both the simple and the advanced techniques like CatBoost, XGBoost, LightGBM to predict the landslide susceptibility. • CatBoost integrated with PSO achieved the highest performance with an AUC-ROC of 0.9854 and high accuracy, recall, precision and F1-score. • Separate LSMs of 30 m resolution have been prepared to for each six districts considered in the study area. The LSMs have shown that Chickmaglur, Uttara Kannada and Kodagu districts have the highest landslide susceptibility.