Application of ML- based approach for co-seismic landslides susceptibility mapping and identification of important controlling factors in eastern Himalayan region

Abstract

Co-seismic landslides pose a significant concern for the Himalayas and its nearby area due to high seismic activity in the region, coupled with steep slopes and heavy rainfall, responsible for substantial socioeconomic losses. Accurate and reliable Co-seismic landslide susceptibility maps are vital in highlighting high-risk zones where proactive measures can be taken to minimise the risk. Despite numerous machine learning (ML) models and landslide controlling factors being explored for susceptibility mapping, uncertainty remains about removing irrelevant factors and identifying optimal controlling factors for an ML based model. Further earlier research highlights that the performance of ML based models improves when optimal controlling factors are utilized for training the model. This study aims to evaluate the efficiency of Random Forest (RF), Logistic Regression (LR), and Naive Bayes (NB) for co-seismic landslide susceptibility mapping and identifying most important controlling factors in the eastern Himalayan region. The landslide inventory of the 2011 Mw 6.9 Sikkim earthquake and a spatial database comprising 16 landslide-controlling factors have been utilised. A novel approach is proposed for selecting the optimal controlling factors for an ML model. Susceptibility maps for the Indian state of Sikkim are prepared by each model using optimal controlling factors. Peak Ground Acceleration (PGA), river distance, slope, fault distance, and elevation are identified as the most important factors, with the RF model showing superior performance. The outcomes of this study provide valuable insights for policymakers and engineers for land use planning and proactive measures to minimize losses.