Landslide susceptibility mapping for western coastal districts of India using geospatial techniques and eXplainable artificial intelligence

Abstract

The west coast of India is more vulnerable to landslides due to high rainfall and hilly topography. To identify the landslide susceptible areas and the most important landslide triggering factor in the western coastal districts of India a landslide susceptibility mapping (LSM) was carried out using fourteen landslide triggering factors. LSM assists in identifying probable zones for future landslide occurrences within a given location by considering various landslide-triggering factors. For locating landslide-susceptible areas and to identify the best preforming model, a comparison between frequency ratio (FR), logistic regression (LR), machine learning (ML) models was performed. ML models used in this study were random forest (RF), support vector machine (SVM), extreme gradient boosting (XGB) and deep neural network (DNN). Most of the area was covered by very low class, i.e., 60.12% followed by low (13.50%), moderate (10.54%), high (8.04%) and very high (7.79%) classes, respectively. From the variable importance plots, it was found that factors such as slope, TRI, LS-factor, distance to road and rainfall were the most significant landslide-triggering factors. The results of the area under the ROC curve (AUC) revealed that the RF model achieved an excellent accuracy rate of 0.993 surpassing the other models. The ranking based on multiple model evaluation parameters using validation dataset revealed DNN as the best-performing model. The partial dependence plots (PDP) of the DNN model revealed that factors such as TRI, rainfall, slope, elevation and TWI were positively related to the landslide occurrences. It was concluded that the performance of ML models was excellent compared to the statistical model. The results of this study could help to identify landslide-vulnerable areas and adopt suitable preventive measures for mitigating the likely occurrence of future landslide events.