An Attribution Deep Learning Interpretation Model for Landslide Susceptibility Mapping in the Three Gorges Reservoir Area

Deep learning (DL) models are increasingly used for landslide susceptibility mapping (LSM) due to their higher accuracy. However, due to the lack of explanations of the influence of input contributing factors by current DL models, accurately identifying the cause of each landslide remains challenging. This study proposes a novel interpretable DL model named Deep-Attention-LSF, which assigns significance scores to contributing factors at local levels for attributing landslide susceptibility. This model considers the significance scores of input factors to more accurately predict landslide occurrence. DeepLIFT is used as an attribution branching network for interpreting the relationship between input factors and each landslide event. Subsequently, a landslide classification network formed by combining convolutional neural network and long short-term memory is used to predict the landslide occurrence in the entire study area. The performance of Deep-Attention-LSF is tested using the landslide inventory map of Three Gorges reservoir area and the associated maps of 18 landslide-related factors. The accuracy, recall, precision, and F1-score of our model were 0.9645, 0.9583, 0.9676, and 0.9522, respectively. These suggest that our model outperformed the compared models, including self-attention LSM, frequency-ratio-attention LSM, bagging and random subspace naive bayes tree, gradient boosting decision tree, random forest, information value model and enhanced C5.0 decision tree model. Deep-Attention-LSF provided reasonable explanations for landslide attributions by comparison with field investigation reports for four specific landslide cases. Combining the interpretation of Deep-Attention-LSF with field investigations can provide more comprehensive information for evaluating specific landslides, providing a useful tool for landslide prevention and management.