An interpretable deep learning model to map land subsidence hazard.

Area of Science:

• Geosciences
• Artificial Intelligence
• Environmental Science

Background:

• Land subsidence (LS) poses significant geohazards, necessitating accurate predictive models.
• Deep learning (DL) models offer powerful tools for mapping LS susceptibility but often lack interpretability.
• Understanding the factors driving LS is crucial for effective risk management and mitigation strategies.

Purpose of the Study:

• To interpret the outputs of deep learning models (CNN and LSTM) for land susceptibility to subsidence hazard mapping.
• To identify and rank the most influential features controlling land subsidence using advanced interpretation techniques.
• To enhance the transparency and reliability of DL-based geohazard assessments.

Main Methods:

• An inventory map of land subsidence (LS) was created using fieldwork and presence points.
• Particle Swarm Optimization (PSO) identified 11 key features for DL models (CNN and LSTM).
• Six interpretation methods, including SHAP and PFIM, were employed to analyze model outputs.

Main Results:

• Both CNN and LSTM models achieved excellent accuracy (AUC > 0.90) in mapping LS hazard.
• Distance from the well, GDR, and DEM were identified as the top three most impactful features influencing DL model predictions.
• Waterfall plots revealed that distance from wells and coarse fragments negatively impacted LS, while land use and DEM positively contributed.

Conclusions:

• The applied interpretation techniques effectively address the 'black box' nature of DL models in geohazard assessment.
• Feature importance analysis provides valuable insights into the drivers of land subsidence.
• This research demonstrates the utility of interpretable DL for robust land subsidence hazard mapping.