Human Guided Causal Learning For Predicting Residential Damage Due to Flooding: A Case Study For Coastal Alabama

Science and society are firmly in the age of Machine Learning (ML). ML has become a pivotal tool in addressing complex problems across various natural sciences, and flood risk is not an exception to this trend. Traditional methods, such as depth-damage functions, have been commonly used to predict flood damage. However, recent studies have exposed their limitations, showing that multivariable supervised learning techniques can provide more accurate flood damage estimates. Despite these advancements, data-driven models often rely on correlations within the data, which may not fully capture the underlying causal relationships essential for accurate predictions. Consequently, these models are generally as reliable as the data they are trained on and can be highly vulnerable to distributional changes. To overcome these limitations, this paper explores the integration of causal discovery and causal machine learning techniques for enhanced flood damage prediction. Causal discovery methods are employed to identify the causal relationships among various flood hazard, vulnerability, and exposure parameters influencing flood damage. Subsequently, causal machine learning models leverage these identified relationships for predictive purposes. Our analysis utilizes a dataset comprising 4,163 cases of reported flood damage to residential properties from Hurricanes Ivan (2004) and Katrina (2005). Finally, this study concludes by comparing the causal relationships identified through algorithmic causal discovery with those derived from domain knowledge and traditional machine learning approaches. Our findings underscore the potential benefits of incorporating causal reasoning into flood damage prediction processes. This approach not only addresses issues related to limited data but also enhances the robustness and transferability of flood damage models across different spatial and temporal contexts.