Principled XAI analysis of the deep learning-based landslide susceptibility prediction model

Abstract Research on applying machine learning (ML) and deep learning (DL) techniques to landslide susceptibility analysis is widespread, with increasingly accurate analyses through novel models. Predicting landslide susceptibility using ML models involves analyzing relationships between conditioning factors and landslide occurrences. Unlike traditional methods, ML models do not explicitly incorporate geotechnical or hydrological theories, raising concerns about result reliability despite high accuracy. This “black-box” limitation has prompted research applying eXplainable Artificial Intelligence (XAI) algorithms to interpret relationships between conditioning factors (digital elevation models (DEM), forest characteristics, soil properties, and geological features) and landslide susceptibility, thereby validating proposed ML models. In this paper, landslide susceptibility prediction models were developed using 20 conditioning factors and multiple architectures, including Support Vector Machine (SVM), Random Forest (RF), Multilayer Perceptron (MLP), and Convolutional Neural Networks (CNNs). Quantitative performances and XAI outcomes were compared. Specifically, the quantitative evaluation showed that the traditional point-based models (RF, SVM, and MLP) achieved Accuracies of 0.6931, 0.6621, and 0.7034, respectively, while the image-based CNNs achieved a higher Accuracy of 0.7586. Furthermore, regarding Recall—a critical metric for disaster management to minimize false negatives—the CNNs (0.8138) significantly outperformed the RF (0.6345), SVM (0.6276), and MLP (0.6828). These results underscore that capturing spatial context through image-wise inputs is far more effective for landslide susceptibility mapping than conventional pixel-level analysis. Because CNNs process input data differently, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied alongside SHapley Additive exPlanations (SHAP) for CNNs, whereas only SHAP was applied to the other models. Results indicated specific patterns associated with certain conditioning factors in landslide susceptibility prediction. CNNs’ Grad-CAM heatmap effectively illustrated these patterns by treating data as images, improving interpretability and reliability of ML outputs.