Advanced Machine Learning Models for Landslide Susceptibility Mapping in South Sikkim Himalayas, India

ABSTRACT Landslides pose a serious threat, especially in mountainous regions, where they can severely impact human life, property and economic activities. This hazard has been particularly challenging in the Himalayan region, particularly in the Sikkim Himalayas. In this study, hybrid models are used to generate landslide susceptibility maps (LSMs) in the South Sikkim Himalaya region of India that could significantly help in reducing the risk. To construct an accurate susceptibility map, a detailed 163 landslide inventories were created using historical data, Google Earth and field investigations. A Random Forest (RF) model was employed to identify relevant factors that revealed 15 significant variables, out of which, distance from road and rainfall emerged as the primary drivers of landslide risk. The Stacking Ensemble model combines three classic machine learning models, such as Artificial Neural Networks (ANNs), Support Vector Machines (SVMs) and Gradient Boosting Decision Trees (GBDTs) to avoid overfitting and improve generalisation. For modelling, landslide data were randomly divided into training (70%) and validation (30%) subsets. The models were evaluated using the receiver operating characteristic (ROC) curve and statistical metrics. The results showed that the ANN had the highest individual prediction performance with an area under curve (AUC) value of 0.92, followed by GBDT (0.89), SVM (0.83) and the Stacking method (0.89). However, the Stacking model performed best in generalisation and class‐specific accuracy, particularly in very high (0.54%) compared to other models. The results of this study highlight the Stacking model's promise as a reliable tool for landslide susceptibility mapping by showing that it performs better than alternative methods in precisely identifying landslide‐prone locations. The combination of many machine learning methods inside a hybrid framework markedly improves the dependability and accuracy of these maps. The utilisation of hybrid models enhances understanding of landslide dynamics and establishes a scientific foundation for formulating efficient risk reduction methods. This improvement offers critical insights for policy‐making focused on reducing landslide risks and supports long‐term regional growth and sustainability, especially in geologically vulnerable and mountainous areas.