A Comparative Study of Landslide Susceptibility Mapping Using Bivariate Statistical and Artificial Intelligence Models in the Upper Beas Valley, Himachal Pradesh, India

Landslides rank among the most devastating geological hazards in mountainous areas. One such region in Himalaya is the Upper Beas Valley in Himachal Pradesh, India having history of number of destructive landslides. Therefore, landslide susceptibility maps (LSM) become important tool for future land use planning and infrastructure development of the area. In the present study, two quantitative methods namely weight of evidence (WoE) which is a bivariate statistical method and state-of-the art artificial intelligence method namely random forest (RF) were compared for the preparation of LSM of the area. Landslide inventory and various conditioning factor including elevation, slope aspect, slope angle, lithology, plan curvature, profile curvature, distance to road, distance to thrust, distance to drainage, topographic wetness index and rainfall are prepared using remote sensing satellite images and extensive field survey. The prepared dataset was randomly divided into 70:30 as a training dataset and validation dataset, respectively. The RF model was trained repeatedly using training data till the optimized results are obtained. For WoE method, relative weights of each class of conditioning factors were calculated using weight of evidence statistics. Subsequently, cumulative weight of each factor was used for the preparation of susceptibility map. The susceptibility maps created indicate that approximately 40–43% of the entire area is classified as being in high and very high landslide-susceptible zones using the Weight of Evidence (WoE) and Random Forest (RF) methods, respectively. In the study area llithology, slope, drainage, and rainfall are major contributors of landslide. The maps have been validated using ROC (Receiver Operating Characteristic) curve showing the accuracy around 84% and 89% for WoE and RF respectively. Hence, it can be concluded that artificial intelligence methods have better predictive capability than the primitive methods.