A Hierarchical Bayesian Approach for Bias Correction of NEXRAD Dual-Polarization Rainfall Estimates: Case Study on Hurricane Irma in Florida

Abstract

It still remains a challenge to obtain accurate rainfall estimates for weather radar on extreme events. In this letter, a hierarchical Bayesian approach is proposed for bias correction of the weather service surveillance radar (WSR-88D) dual-polarization (Dual-Pol) rainfall estimates in a hurricane. Based on various assumptions of the probabilistic likelihood function of hourly rainfall and spatial dependences between residual errors, four specific models are developed and intercompared to quantify the bias of the existing WSR-88D Dual-Pol rainfall product. The case of Hurricane Irma in Melbourne, Florida, is used to demonstrate the proposed technique. It is found that Bayesian analysis is capable of improving the Dual-Pol rainfall estimates and quantifying the associated predictive uncertainty: both root-mean-square error and normalized mean absolute error are declined by at least 30% after the Bayesian correction at the validation sites. Additionally, the consideration of the spatial Gaussian process in the spatial error model achieves a better skill score compared to the model without the spatial Gaussian process, and the usage of spatial Gaussian process has a positive impact on the Bayesian rainfall estimates at short distance from ground reference sites in regions of interest.