Flooding has increased by 50%, and for the past 25 years, the United States has experienced a flooding event every 2 to 3 days [1]. The Federal Emergency Management Agency (FEMA) is the only resource people utilize to determine whether their homes are at risk for flooding. FEMA maps are costly, outdated, and lack predictive accuracy, leaving homeowners unaware and unprepared for potential dangers [2]. Additional tools are needed to ascertain accurate estimates of flood damages, but computational methods require rich datasets, which are difficult to obtain for flooding events [3]. We create a model that predicts monetary damages using various factors related to the flood event, such as duration, location, and cause. We utilize a flood dataset from the National Oceanic and Atmospheric Administration (NOAA) and data pre-processing techniques to handle missing values and remove irrelevant features [4]. We generate aggregate predictions by leveraging linear regression, random forest regression, XGBoost regression, and neural networks. Our findings show we can effectively use upsampling techniques to combat limited flood data. We show that floods with higher monetary damages are easier to predict, which is important because these floods inflict greater hardship on communities. The results improve preparedness for flood-related risks, property value assessments, and the accuracy of insurance policy underwriting [2]. Ultimately, the model provides a preliminary study on how individuals can make better-informed decisions and prepare for the impact of flooding in their communities. We hope this work encourages further machine learning applications to help prepare citizens for natural disasters.
Read full abstract