Causal Machine Learning for predicting Flood induced residential property damage with limited data

Abstract Submission: Society is firmly in the age of Machine Learning (ML), a new tool for addressing complex problems across natural sciences, and flood risk is not exception. Various approaches exist to determine flood damages, including regressing historic flood damage data, engineering-based models derived from structural behavior principles, and heuristic models that solely rely on expert opinion, or some combination of these three. In this study, we propose a novel data driven approach, using causal machine learning, algorithms that combine ML with causal inference to understand the cause-and-effect relationship and predict outcomes. The advantage of causal ML is that it helps identify the true drivers of an event and can be robust in providing predictions under changing conditions, which is essential for modelling different and unseen flood events. We use 4,163 flood damage claims data from Hurricanes Ivan (2004) and Katrina (2005), comparing causal relationships derived from algorithms, domain expertise, and traditional ML approaches. Causal discovery methods are used to uncover the relationships between flood hazard, vulnerability, and exposure factors, which are then utilized by causal ML models for more accurate predictions. The results demonstrate the advantages of incorporating causal reasoning into flood damage prediction, improving model robustness and transferability across various contexts.

Learning Objectives/Expected Outcome (Optional) : Flood risk prediction and assessment