A knowledge-aware deep learning model for landslide susceptibility assessment in Hong Kong

Despite the success of the growing data-driven landslide susceptibility prediction, the model training heavily relies on the quality of the data (involving topography, geology, hydrology, land cover, climate, and human activity), the structure of the model, and the fine-tuning of the model parameters. Few data-driven methods have considered incorporating 'landslide priors', as in this article the prior knowledge or statistics related to landslide occurrence, to enhance the model's perception in landslide mechanism. The main objective and contribution of this study is the coupling of landslide priors and a deep learning model to improve the model's transferability and stability. This is accomplished by selecting non-landslide sample grounded on landslide statistics, disentangling input landslide features using a variational autoencoder, and crafting a loss function with physical constraints. This study utilizes the SHAP method to interpret the deep learning model, aiding in the acquisition of feature permutation results to identify underlying landslide causes. The interpretation result indicates that 'slope' is the most influential factor. Considering the extreme rainfall impact on landslide occurrences in Hong Kong, we combine this prior into the deep learning model and find feature ranking for 'rainfall' improved, in comparison to the ranking result interpreted from a pure MLP. Further, the potency of MT-InSAR is utilized to augment the landslide susceptibility map and promote efficient cross-validation. A comparison of InSAR results with historical images reveals that detectable movement before their occurrence is evident in only a minority of landslides. Most landslides occur spontaneously, exhibiting no precursor motion. Comparing with other data-driven methods, the proposed methods outperform in accuracy (by 2 %-5 %), precision (by 2 %-7 %), recall (by 1 %-3 %), F1-score (by 8 %-10 %), and AuROC (by 2 %-4 %). Especially, the Cohen Kappa performance surpasses nearly 20 %, indicating that the knowledge-aware methodology enhances model generalization and mitigates training bias induced by unbalanced positive and negative samples.