Publication

Abstract : Globally, landslides impact in a site-specific and regional scale, and have affected 4.8 million human beings during1998–2017. Landslides, being a highly complex phenomenon involving real-time and near real-time interactions between hydrological, geomorphological, climatological as well as anthropological factors impacting large spatial areas, demand the development of regional-scale warning systems. Even though an extensive body of research already exists in the field of landslide early warning, the prediction of the actual location and the time of landslide initiation is still a major challenge. In the current study, we compare the performance of ten machine learning algorithms useful for landslide early warning. The current study is performed in the Idukki district of Kerala state in India. A database with landslide incidents is created using research literature, reports from the Geological Survey of India as well as news articles. Landslide causative factors, as indicated by previous literature, have been mapped using Geographic Information System . The different ML algorithms considered for this study are Decision Tree, Logistic Regression, K Nearest Neighbor, Gaussian Naive Bayes, Support Vector Machine, and its different kernel functions such as linear, Polynomial, Gaussian and ensemble algorithms namely Random Forest and AdaBoost. The performance of the different algorithms is quantified and compared utilizing established statistical metrics such as G-Mean, F1 score, and ROC-AUC score. Our results follow the similar ones in literature where the machine learning techniques provide an efficient tool for landslide susceptibility mapping. All the algorithms considered produce reasonable results.