Understanding Training Data Components for Excessive Rainfall Machine-Learning Models: A look inside the Unified Flooding Verification System

The challenge of forecasting excessive rainfall stems from not only needing to resolve the magnitude and location of the event but also taking into consideration antecedent rainfall and current hydrologic conditions. The Colorado State University – Machine Learning Probabilities System (CSU-MLP) creates random forest-based probabilistic forecasts for excessive rainfall. A method for verifying these forecasts and to train particular algorithms is to use the Unified Flooding Verification System (UFVS) dataset which was created to serve as a more comprehensive collection of flooding proxies and observations (Erickson et al. 2019).

In this study, the internal components of the UFVS consisting of 5-year average recurrence interval (ARI), flash flood guidance exceedances (FFG), local NWS storm reports, and USGS stream gauge reports were analyzed. This spatial, statistical, and temporal analysis looked to identify biases in differently trained versions of the CSU-MLP. It was observed that distinct geographic regions of the CONUS were partial to higher exceedances in one of the flooding proxies, being either ARI or FFG. In conjunction with this finding, these geographic regions often experience a higher probabilistic forecast for excessive rainfall in models that were trained with the proxy that was more often exceeded. This study also identified case studies in California and the southeastern United States that highlight these biases. This work leads to many questions as to why these proxies are exceeded at different rates. It also questions the magnitude of the individual internal UFVS components in influencing the forecasts, if any.