Improving disaster resilience with causal machine learning for flood damage estimation

Abstract

Machine Learning (ML) models have become a pivotal tool in the scientific community, successfully addressing complex problems across various domains, including flood risk management. Despite these advancements, traditional data-driven models often struggle when training data is scarce and primarily rely on correlation rather than causal relationships, making them vulnerable to shifts in data distribution. This paper introduces a Causally Informed Neural Network (CINN) framework that integrates causal prior knowledge as an inductive bias to improve flood damage predictions for residential properties and address these limitations. The proposed approach enhances model adaptability to unseen data distributions—an essential requirement for flood damage modeling. First, Deep End-to-End Causal Inference (DECI) is used to discover causal relationships and estimate their average treatment effects. These causal insights are then embedded into the neural network through causal weight initialization and causal regularization. The framework is validated using an enhanced National Flood Insurance Program (NFIP) claims dataset from Hurricane Katrina, and its performance is benchmarked against six widely used ML models from previous studies. Results show that the discovered causal relationships align with domain knowledge, reinforcing the approach's credibility. The proposed CINN model achieves an average 22 % error reduction compared to traditional ML models, demonstrating its superior robustness and predictive accuracy. Additionally, a feature attribution experiment confirms that the model's decision-making process is consistent with the identified causal relationships, increasing interpretability and trust in its predictions. These findings highlight the potential of integrating causality into ML-based flood damage modeling, paving the way for more resilient and generalizable disaster risk assessment frameworks.