Explainable AI (xAI) for Landslide Susceptibility Modeling: A Comparative Analysis of Machine Learning and Deep Learning Approaches

A landslide is a geo-hazard that occurs due to the dislocation of soil or rock mass from the parent mass in the direction of gravitational force. More than 6. 7 million is spent annually in the Midwest region on infrastructural repairs due to slope failures and landslides. Identifying landslide-susceptible zones can improve urban management and help prioritize areas requiring immediate remediation. However, no full regional-scale landslide susceptibility (LS) map is available for Minnesota (MN), which is a formerly glaciated region. Hence, this study addresses this gap by leveraging machine learning (ML) and deep learning (DL), combined with explainable artificial intelligence (xAI), to develop a high-resolution LS map. Additionally, the study demonstrates how counterfactuals can be used as a preliminary tool for identifying mitigation measures. Five quantitative methods, namely, Logistic regression, Random Forest (RF), Multi-layer perceptron, TabNet, and TabKANet, were trained on a balanced dataset of landslide and non-landslide points. The total dataset was divided into training (70%) and validation (30%) subsets. The validation accuracy and field verification of the LS map demonstrate the superiority of RF and TabKANet. SHapley Additive exPlanations (SHAP) was employed to address the limitations of past studies in practically interpreting the model’s decision. The SHAP model indicates that slope, elevation, land use and land cover, and rainfall have the greatest impact on model predictability globally in the MN map. However, the field validation study also emphasized that the dominant factors at the global scale may not necessarily govern local slope failures, where localized conditions such as drainage concentration or anthropogenic factors can play a more critical role. Overall, this study demonstrates how integrating xAI with ML and DL enhances both accuracy and interpretability, offering a transparent, data-driven framework for making landslide risk-informed decisions and for its global application in geotechnical hazard management. The graphical abstract illustrates a workflow in which explainable artificial intelligence (xAI) serves as the central component of a clear, trustworthy landslide susceptibility (LS) map for Minnesota, a state that faces recurring slope failures. However, there is no full-scale LS available for the state. To address this gap, data on landslide causative factors were collected, along with a comprehensive landslide inventory compiled from various sources, and these were used to build a high-resolution LS map. Five machine learning (ML) and deep learning (DL) methods, namely, logistic regression (LR), random forest (RF), multi-layer perceptron (MLP), TabNet, and TabKANet, were trained for LS modeling. TabKANet, a modified version of TabNet, was used for the first time in the literature to develop an LS map. The obtained results highlight the excellent performance of RF and TabKANet during validation and field verification. To enhance the interpretability of the LS map produced using black-box ML and DL methods, xAI, specifically SHapley Additive exPlanations (SHAP), was used. The SHAP analysis shows that slope angle and elevation of the area were the top influencing factors. Furthermore, for the first time, the present analysis uses counterfactuals in landslide analysis to explore mitigation measures that can be adopted in the field to stabilize a landslide-prone area. Overall, the study not only delivers Minnesota’s first detailed LS map but also introduces an xAI-centered framework that can help state and federal agencies make informed decisions when dealing with potentially catastrophic landslide hazards. Developed the first statewide landslide susceptibility (LS) map for Minnesota. Utilized Explainable Artificial Intelligence (xAI) to understand the LS results. SHAP analysis shows slope angle and elevation as key LS contributors in Minnesota. Validated LS maps with field data and generated local SHAP-based site insights. Counterfactuals guided informed decisions for stabilizing unstable slopes.