Physics-Guided AI Framework to Predict and Correct Radar QPE Errors

A Physics-guided Artificial Intelligence (PAI) framework consisting of a weather regime classifier and a suite of precipitation error prediction models applied to Radar Quantitative Precipitation Estimates (QPE) using machine learning is presented. The weather regime classifier is based on Empirical Orthogonal Function (EOF) analysis of Residual Errors from standard products and from the previous application of hydrologic Inverse Rainfall Correction (IRC, Liao and Barros, RSE, 2022) to storms producing large floods in headwater basins. The precipitation error model is a Multilayer Perceptron (MLP) that predicts the space-time evolution of precipitation errors during a storm event conditional on season, antecedent hydrometeorological conditions, that is initial conditions, and local and basin-scale precipitation characteristics. The PAI framework is demonstrated for warm and cold season precipitation in the Southern Appalachian Mountains (SAM) for 50 largest flood-producing storms during 2008-2018, achieving large improvements on Nash-Sutcliffe Efficiency (NSE) across most events with an average improvement of 0.75 for warm season events, and 0.42 for cold season events. Corrected QPEs demonstrate excellent skill against high elevation gauge locations. The probability distribution function (PDF) of the predicted precipitation errors varies significantly with season, and the spatial distribution of errors for the same precipitation regime varies from basin to basin depending on landform.