Abstract: Landslide susceptibility assessment is an important tool for landslide risk early warning. To address the issue of insufficient capability in extracting and utilizing the physical effects and intrinsic correlations of environmental factors during the assessment process, which makes it difficult to effectively integrate multiple factors for accurate prediction, this study proposes a landslide susceptibility assessment method that combines physics-informed information with deep learning. First, environmental factors are divided into natural conditions and driving factors based on their physical roles in landslides. Then, based on deep learning methods, a natural condition assessment module and a driving factor assessment module are designed to evaluate potential landslide risks and external driving intensity, respectively. Finally, the assessment results are integrated by considering the sequence and causal relationships of the factors to accurately predict landslide probability. The results show that the new method significantly outperforms seven commonly used methods in terms of AUC (0.898 4), overall accuracy (0.829 7), precision (0.795 8), and recall(0.863 6). The self-attention mechanism within the method can accurately identify the key driving factors of landslides, providing precise guidance for disaster prevention and mitigation efforts.

Key words: landslide susceptibility assessment, deep learning, transformer, self-attention mechanism, physical information