Assessing landslide susceptibility mapping in the Sikkim Himalayas using an ensemble machine learning approach

Abstract

Prediction of landslides is a burning issue on the global map especially in areas of the mountainous region where the occurrence of landslides may be a major setback in the development of sustainability in such areas. This paper presents the mapping of landslides susceptibility in Namchi-Sikkim region of India, which is implemented via the latest machine learning techniques. A variety of 6 models of machine learning were used, Support Vector Machine (SVM), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), Random Forest (RF), Extra Trees (ET) and Logistic Regression (LR) to determine the regions prone to landslides in this seismically affected high-risk area. The database containing 1020 points, including the presence of landslides as well as their non-occurrence and other 23 environmental factors like terrain, lithology, and rainfall pattern were used to train the models. The Random Forest model and the Meta Classifier model demonstrated best accuracy (0.941) and F1 score (0.941) indicating that they predict very well. Gradient Boosting model showed the accuracy of 0.935 and Area Under Curve (AUC) 0.982, which points to its effectiveness in labeling the risky areas and the areas that remain stable. Other models such as SVM and XGBoost also gave useful outputs whereby SVM model achieved accuracy of 0.886 and XGBoost recorded a precision of 0.94. The results emphasize the importance of environmental and anthropogenic factors in landslide prone conditions. The study is a sophisticated workable model of predicting hazard being scalable, which has a major implication to disaster management, land-use planning and environmental protection especially in areas with high landslide susceptibility.