White-Boxed Landslide Susceptibility Mapping Using Unsupervised Causal Model

Currently, in landslide susceptibility mapping (LSM), data-driven models mostly utilize association to present the connection between landslides and conditioning factor (CF). However, an association cannot reflect the structural interaction or the underlying causal mechanism. This leads to black-box results and reduces interpretability. To address this problem, this study introduces an unsupervised causal model to LSM. The methodology consists of two parts: causal discovery with noncombinatorical optimization via trace exponential augmented Lagrangian structure learning (NOTEARS) algorithm, and causal inference with Bayesian networks (BNs). By applying structural causality to LSM, we “white-boxed” the process of data-driven models, thereby ensuring the interpretability and data utility. To comprehensively evaluate the results, comparison experiments with other machine learning (ML) models were performed. The results showed that BN achieved the highest accuracy (approximately 0.877) and the second highest [area under curve (AUC)] (0.907). Moreover, BN outperformed other models in ground-truth validation. The high and extremely high susceptibility zones in the BN-LSM results covered the most actual landslides (more than 71.07% of the total, and more than 80% after excluding areas not used in model training). This showed the generalizability and stability of structural causality-guided BN. The proposed methodology provides a link between data-driven causal science and natural hazard analysis, which can help intelligent hazard management. Data and codes that support the findings of this study are openly available at doi.org/10.6084/m9.figshare.26395420