Integrating landslide magnitude in the susceptibility assessment of the City of Doboj, using machine learning and heuristic approach

Abstract

ABSTRACT In this work, landslide assessment of the Doboj City area was modeled by combining machine learning and heuristic tools. The machine learning part was used to map the Morphometric factor. i.e. probability of landslides based on relation between the magnitude of events and morphometric parameters: elevation, distance to streams, slope, profile curvature, and aspect. The Random Forest and Support Vector Machines algorithms were implemented in the learning protocol, which included several strategies: balancing of the training/testing set size, algorithm optimization via cross-validation, and cross-scaling. The best performing Morphometric factor ap was created by learning on 50 m and testing on 25 m dataset. The heuristic part was used for modeling of Lithological factor and Land Cover factor maps, by expert-driven scoring of their units, within 0-1 range of values. The final Susceptibility map was obtained by multiplying all three factor maps resulting in a high-performing model with AUC=0.97 and acc=92%.