Landslide Susceptibility Zonation (LSZ) Using Machine Learning Approach for DEM Derived Continuous Dataset

Abstract

Landslide-prone areas can be shown by depicting occurrence of landslides; landslide susceptibility zonation map (LSZ), landslide hazard zonation map (LHZ) and landslide risk zonation map (LRZ). However, for the preparation of LRZ map, we need LHZ map and for LHZ map, we need LSZ map. In this work Logistic Regression (LR), Fisher Discriminant Analysis (FDA) associated with the weighted linear combination (WLC) and ANN are used for the preparation of LSZ maps. Seven causative factors that give continuous dataset are taken into consideration i.e. aspect, slope, digital elevation model (DEM), topographic wetness index (TWI), tangential curvature, profile curvature and plan curvature for the part of Mandakini river basin in Garhwal Himalayas. Geology, geomorphology, soil type, thrust/fault buffer, road buffer, drainage buffer etc., which are also important landslide governing factors, were not used as they give a discrete/classified dataset. The study area is spread in 275.60 km$^2$ area where total 122 landslides occurred between 2004 and 2017. The landslides occurred from 2004 to 2012 (46 landslides with 1203 pixels) have been used for training of the models and from 2013 to 2017 (76 landslides) have been used for testing of the models. The susceptibility maps were classified/categorized into five different zones (very low, low, moderate, high and very high) based on the natural break in data. The landslide locations with the index value greater than 0.55 have been considered for the validation of the maps. The assessment of accuracy is done based on the Heidke Skill Score (HSS). The HSS score for FDA, LR and ANN is obtained as 0.89, 0.98 and 0.96. Based on the HSS score, the LR method can be selected as the best method amongst the three.