ABSTRACT
A landslide is a geohazard that occurs when rock or soil mass moves downslope under gravity, often triggered by a combination of natural and human factors. Landslides cause fatalities, property damage, traffic disruptions, and annual infrastructure losses of around $6.7 million in Minnesota, Wisconsin, and Michigan. Identifying vulnerable areas is critical for prioritizing restoration and improving urban management. Landslide depends on multiple causative factors with nonlinear relationship with its occurrence, which can be modeled using artificial neural networks (ANNs). TabNet, a modification of ANNs with features of random forests, effectively captures these nonlinearities. This study develops a landslide susceptibility (LS) map for Minnesota using TabNet, considering ten influencing factors and a dataset of 7,710 landslide and 7,710 non-landslide events, split 70% for training and 30% for validation. The model showed 96% accuracy. Additionally, a post-hoc explainable artificial intelligence (XAI) approach was employed to interpret the results, enhancing state agencies’ understanding of slope stability and guiding mitigation strategies.
