A novel deep learning based spatial ensemble approach and segment anything model for landslide risk assessment in...

This study aims at introducing novel "Spatial Ensemble" approach for landslide risk assessment (LRA). For this purpose, the present work focuses on three different algorithms that includes a state-of-the-art deep learning DenseNet Neural Network, and 2 popular benchmark algorithms of MLPNN and XGBoost. The spatial ensemble approach in LRA focuses on generating two distinct types of "Landslide Impedance Composite Maps", where the first one predicts the broadest possible danger zones covering the maximum extent with potential landslide occurrences across the study area. The second map targets the places in all probability to encounter landslides, by demarcating specific locations with the greatest risk of landslide occurrences to prioritize accuracy and precision, making sure that areas with highest levels of risk receive detailed attention. The outcome of the second map has helped during field verification, leading to identification of a new landslide in close proximity of human settlements and buildings. It is a crucial finding as this study applies state-of-the-art auto-detection algorithm of Segment Anything Model (SAM) to detect buildings automatically in close proximity which is accurate and time-saving alternative to landslides, for risk assessment. And finally, a landslide risk zone is developed to delineate the building vulnerability around the landslide, which is validated against cracked buildings and presence of fault lines in the area. Thus, this research combines cutting-edge technologies with robust field observations, to introduce a novel and stepped approach of landslide risk assessment.