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

Area of Science:

• Geosciences
• Artificial Intelligence
• Environmental Science

Background:

• Data-driven landslide susceptibility prediction models heavily depend on data quality, model structure, and parameter tuning.
• Few existing methods incorporate landslide priors (knowledge or statistics of landslide occurrence) to improve understanding of landslide mechanisms.
• Enhancing model transferability and stability in landslide prediction remains a challenge.

Purpose of the Study:

• To couple landslide priors with a deep learning model to improve prediction transferability and stability.
• To develop a knowledge-aware methodology for landslide susceptibility mapping.
• To identify key landslide causal factors through model interpretation.

Main Methods:

• Selected non-landslide samples based on landslide statistics.
• Disentangled landslide features using a variational autoencoder.
• Crafted a loss function with physical constraints.
• Utilized the SHAP method for deep learning model interpretation.
• Integrated MT-InSAR data for landslide susceptibility map augmentation and cross-validation.

Main Results:

• The integrated model demonstrated superior performance over other data-driven methods in accuracy, precision, recall, F1-score, AuROC, and Cohen Kappa.
• Slope was identified as the most influential factor in landslide occurrence.
• Incorporating extreme rainfall priors improved the feature ranking for rainfall, especially in regions like Hong Kong.
• MT-InSAR data augmentation enhanced cross-validation efficiency.

Conclusions:

• The knowledge-aware deep learning methodology significantly enhances landslide susceptibility prediction models.
• The approach improves model generalization and mitigates training bias from unbalanced datasets.
• Integrating landslide priors offers a promising direction for more robust and reliable landslide prediction systems.