Nationwide Prediction of Flood Damage Costs in the Contiguous United States Using ML-Based Models: A Data-Driven Approach

Flooding remains one of the most disruptive and costly natural hazards worldwide. Conventional approaches for estimating flood damage cost rely on empirical loss curves or historical insurance data, which often lack spatial resolution and predictive robustness. This study develops a data-driven framework for estimating flood damage costs across the contiguous United States, where comprehensive hydrologic, climatic, and socioeconomic data are available. A database of 17,407 flood events was compiled, incorporating approximately 38 parameters obtained from the National Oceanic and Atmospheric Administration (NOAA), the National Water Model (NWM), the United States Geological Survey (USGS NED), and the U.S. Census Bureau. Data preprocessing addressed missing values and outliers using the interquartile range and Walsh tests, followed by partitioning into training (70%), testing (15%), and validation (15%) subsets. Four modeling configurations were examined to improve predictive accuracy. The optimal hybrid regression–classification framework achieved correlation coefficients of 0.97 (training), 0.77 (testing), and 0.81 (validation) with minimal bias (−5.85, −107.8, and −274.5 USD, respectively). The findings demonstrate the potential of nationwide, event-based predictive approaches to enhance flood-damage cost assessment, providing a practical tool for risk evaluation and resource planning.