Title:XAI Model for Accurate and Interpretable Landslide Susceptibility

Abstract:Landslides are notoriously difficult to predict. Deep neural networks (DNNs) models are more accurate than statistical models. However, they are uninterpretable, making it difficult to extract mechanistic information about landslide controls in the modeled region. We developed an explainable AI (XAI) model to assess landslide susceptibility that is computationally simple and features high accuracy. We validated it on three different regions of eastern Himalaya that are highly susceptible to landslides. SNNs are computationally much simpler than DNNs, yet achieve similar performance while offering insights regarding the relative importance of landslide control factors in each region. Our analysis highlighted the importance of: 1) the product of slope and precipitation rate and 2) topographic aspects that contribute to high susceptibility in landslide areas. These identified controls suggest that strong slope-climate couplings, along with microclimates, play more dominant roles in eastern Himalayan landslides. The model outperforms physically-based stability and statistical models.