Landslide Susceptibility Assessment Using a Physics-informed Deep Learning Model

Abstract

While deep learning techniques have demonstrated considerable potential in modeling complex, nonlinear relationships among environmental and geological variables, their application often neglects the incorporation of physical principles, thereby limiting interpretability and compromising the reliability of negative sample selection. To address this limitation, the present study introduces a physics-informed deep learning framework that integrates a physically based slope stability model with a one-dimensional convolutional neural network (1D-CNN). In contrast to conventional approaches that rely on randomly selected negative samples, the proposed method utilizes a physically justified criterion extracting non-landslide samples from areas exhibiting low probability. The model is applied to the 2018 landslide event in Hiroshima, Japan, to evaluate its effectiveness. The results reveal that the physics-informed framework yields a substantial improvement in predictive performance, as evidenced by an increase in the Area Under the Receiver Operating Characteristic Curve (AUC) from 91.0% to 93.3%, relative to standard data-driven approaches. These findings highlight the advantages of integrating domain knowledge into deep learning models to enhance the accuracy of landslide susceptibility assessments.