Landslide susceptibility analysis in the Bhilangana Basin (India) using GIS-based machine learning methods

Landslides are frequent natural hazards in mountainous regions, and harshly upset people's lives and livelihoods. In the present study, we have carried out an analysis of seven GIS-based machine-learning techniques; and asses their performance for landslide susceptibility mapping (LSM) in the Bhilangana Basin, Garhwal Himalaya. A landslide inventory consisting of 423 polygons was prepared using repeated field investigations, and multi-dated satellite images for the periods between 2000 and 2022. The landslide dataset was classified into two groups: training (70%) and test dataset (30%), and 12 predictive variables were used for the LSM. The methods used to produce LSM are boosted regression tree (BRT), Fisher discriminant analysis (FDA), generalized linear model (GLM), multivariate adaptive regression splines (MARS), model-architect analysis (MDA), random forest (RF) and support vector machine (SVM). The sensitivity and performance of these models to predict landslide susceptible areas were carried out using the area under the curve (AUC) method. The RF model (AUC = 0.988) has given the highest precision indicating the best performance. Though MARS (0.974), SVM (0.965) and MDA (0.952) models have also performed adequately for the LSM (all have AUC values above 0.95), however, it is recommended that the RF model is highly suitable for LSM in the mountainous region.