Hybrid Machine Learning and Optimum Feature Selection Based Landslide Susceptibility Analysis

Abstract

Landslide susceptibility analysis of Kullu district, India, is carried out using optimum feature selection and Frequency Ratio-Logistic Regression (FR-LR) and Frequency Ratio –Random Forest (FR-RF) hybrid models. 981 historical landslide events of the Kullu district were inventoried using information from various documented sources, high-resolution satellite imagery, and data collected by handheld GPS during field visits were analyzed and used for susceptibility analysis. Based on expert opinions, topography, hydrological setting, and data availability a set of 14 landslide causative factors (LCF) were identified for the study area. A hybrid approach for the selection of optimum LCF has been developed using feature ranking and statistical significance methods. 70% landslide events were used for training and selection of optimum LCFs, and the remaining 30% events were used for the validation of landslide susceptibility mapping. The landslide susceptibility maps (LSM) generated using two machine learning-based hybrid models i.e. FR-LR and FR-RF were compared using ROC curves and confusion matrix. Both FR-LR and FR-RF hybrid models have excellent prediction performance for landslide susceptibility analysis of the study area. Among the models, FR-RF performed better than FR-LR model.