Enhancing probabilistic machine-learned guidance and verification of excessive rainfall and flash floods project goal: this project will enhance a machine learning (ML)-based excessive rainfall…

The National Oceanic and Atmospheric Administration (NOAA) awarded the University of North Carolina at Chapel Hill a $243,178 Cooperative Agreement under the Weather and Air Quality Research program (CFDA 11.459) effective August 1, 2025, through July 31, 2028. This project enhances a machine learning (ML)-based excessive rainfall forecast system by developing improved historical event datasets leveraging insurance claim data and integrating new geospatial predictors to better characterize atmospheric conditions and antecedent factors contributing to flood risk. The initiative addresses operational deficiencies in how excessive rainfall events are defined in model training and improves predictions for extreme rainfall events through the incorporation of impact-based datasets.

Primary deliverables include enhanced historical event datasets for ML model training, improved ML models with corresponding forecasts and diagnostic tools, and explainable artificial intelligence (AI) analyses to increase forecaster understanding and trust in model predictions. These products are designed for operational integration at NOAA's Weather Prediction Center and evaluation through NOAA's testbed infrastructure. By combining rainfall and flood damage information, the project advances understanding of vulnerability conditions and flood impacts, ultimately providing more skillful and valuable ML-based guidance to operational forecasters for excessive rainfall and flood risk prediction.

Enhancing probabilistic machine-learned guidance and verification of excessive rainfall and flash floods project goal: this project will enhance a machine learning (ML)-based excessive rainfall forecast system by improving and extending the historical datasets used to train and verify the model forecasts, as well as the predictors that characterize the larger-scale atmospheric environment and antecedent conditions that contribute to enhanced flood risk. Problem/opportunity statement: operational ML-based guidance of excessive rainfall suffers from regional frequency biases, in part due to deficiencies in how excessive rainfall events are defined in model training and insufficient antecedent conditions contained in the model predictors. The existing forecast system also has limitations in its predictions of the mostextreme rainfall events. Methodology/activities to be performed: we will devise new event datasets by leveraging observations of historical flood damage derived from insurance claims and incorporate new geospatial predictors into ML model training to produce more skillful forecasts that will be evaluated in a NOAA testbed, while also using explainable AI tools to better understand the ML predictions. Primary project products/outputs: improved and enhanced historical events datasets will be generated for ML model training. In turn, ML models and corresponding forecasts and diagnostics will be output to support forecast operations at the weather prediction center. Expected results, outcomes, and benefits: the integration of impact-based datasets into an ML-based excessive rainfall and flooding impact prediction system will yield more valuable ML guidance for forecasters and enhance trust among forecasters. Combining rainfall and flood damage information will better characterize vulnerability and enhance our understanding of conditions that have led to past flood impacts. Intended beneficiaries or recipients: the anticipated products and forecasts generated in this project will benefit the weather prediction center and other NOAA entities (e.g., national water center) tasked with predicting floods and flood impacts.