Predicting Flood Damages using Machine Learning and National Flood Insurance Program Data

Precise predictions of flood damage location, timing, and extent can enhance mitigation and governmental response efforts. While the Federal Emergency Management Agency’s flood insurance rate maps can estimate where damage is probable to occur overall, they are outdated and potentially underestimate flood damages. This project presents a framework for improving the ability of random forest models to predict National Flood Insurance Program insured losses as a function of hydrometeorological and economic variables across California at a 1-degree resolution. Two models, a random forest regression and binary classification, were developed and verified against a test dataset using error statistics and a confusion matrix. While both models overpredicted the occurrence of flood damage, the classification model showed a higher accuracy rate of 87%. Random forest variable importance showed that runoff, number of insurance policies, soil moisture, and precipitation were the main determinants of model accuracy.