Improving disaster resilience with causal machine learning for flood damage estimation

Area of Science:

• Environmental Science
• Computer Science
• Data Science

Background:

• Machine Learning (ML) models are crucial for flood risk management but struggle with scarce data and correlational reliance.
• Traditional models are vulnerable to data distribution shifts and lack causal understanding.

Purpose of the Study:

• To introduce a Causally Informed Neural Network (CINN) framework for enhanced flood damage prediction.
• To improve model adaptability to unseen data distributions in flood damage modeling.

Main Methods:

• Utilized Deep End-to-End Causal Inference (DECI) to identify causal relationships and estimate average treatment effects.
• Integrated causal insights into neural networks via causal weight initialization and regularization.
• Validated the CINN framework using Hurricane Katrina NFIP claims data and benchmarked against six ML models.

Main Results:

• Discovered causal relationships aligned with domain knowledge, confirming model credibility.
• Achieved an average 22% error reduction compared to traditional ML models.
• Feature attribution confirmed model decisions align with identified causal relationships, enhancing interpretability.

Conclusions:

• The CINN framework demonstrates superior robustness and predictive accuracy in flood damage modeling.
• Integrating causality into ML models enhances adaptability and trustworthiness for disaster risk assessment.
• The findings support the development of more resilient and generalizable disaster risk frameworks.