Deep learning and benchmark machine learning based landslide susceptibility investigation, Garhwal Himalaya (India)

Abstract

Garhwal Himalaya is the worst affected landslide prone region in Indian subcontinent mainly due to its complex geological settings and active tectonic activities. The data showed that every year, around 400 fatalities occur in Himalayan terrain due to landslide. In the current study, we have mapped the landslide susceptibility zones in the segment of Garhwal Himalaya using robust machine and deep learning algorithms. Individual machine and deep learning models have its own limitations like low generation capacity with nonlinear functions to describe the intricate relationship among predictors. In this study total five models i.e., SVM (Support Vector Machine), RF (Random Forest), bagging, ANN (Artificial Neural Network), DLNN (Deep Learning Neural Network) have been used along with twenty landslide controlling factors. Here, the principal objective of the study is to precisely delineate landslide susceptibility zones of the Garhwal Himalaya. The selecting factors have been considered through multi-collinearity test and information gain ratio statistics and the previous landslide points have been taken as training (70%) and testing (30%) dataset. According to area under curve value (AUC), the DLNN technique has high capability (AUC = 0.925) and accuracy for landslide area demarcation. The approach of integrated physical and social factors creates more precise prediction aptitude that can support large scale landslide management. These high precision models identified most of the parts of Rudraprayag and Tehri Garhwal as a very high landslide susceptibility zone. The generated maps can assist to policy makers for micro scale landslide management and sustainable land use planning particularly in Himalayan terrain.