Interpretable susceptibility assessment of coseismic landslides using a Superposable Neural Network

Coseismic landslides pose significant threats to seismically active mountainous regions, where the interplay of topographic, geological, and seismic factors jointly controls slope failures. Conventional susceptibility models often fail to capture nonlinear feature interactions while maintaining physical interpretability. To address this issue, we introduced an interpretable deep learning framework—Superposable Neural Network (SNN)— to seismic landslide susceptibility modeling for the first time. This innovative modeling framework employs additive, independent sub-networks to represent both individual predictor influences (Level-1 features) and their pairwise interactions (Level-2 features) with transparent interpretability. Such a configuration was used to quantify the contribution of each factor and their combinations, and to evaluate the improvements in predictive performance when accounting for feature interactions. The results indicate that peak ground acceleration (PGA), slope angle, and distance to the seismogenic fault (dis2fault) are the most influential Level-1 predictors. Furthermore, composite interactions such as “slope × PGA” and “rainfall × lithology” have substantially improved the model’s interpretation capability. The SNN effectively captures the spatial heterogeneity of landslide distribution and delineates high-susceptibility zones where ground motion amplification exacerbates slope instability. Compared with conventional models, the proposed approach delivers superior predictive performance and enhanced interpretation capability. This study validates the SNN as a robust and effectively explainable tool for seismic landslide hazard assessment in complex tectonic settings. • An interpretable SNN model is proposed to assess coseismic landslide susceptibility. • The model quantifies the effects of different input features on landslide susceptibility. • Composite features (e.g., slope × PGA, rainfall × lithology) improve model interpretability in landslide prediction.