Bias adjustment of satellite rainfall data through Gaussian process regression (GPR) based on rain intensity classification in the Greater Bay Area, China.

Estimating precipitation over large spatial areas remains a challenging problem for hydrologists. Satellite-based remote sensing rainfall products have the advantage of large-scale synchronous coverage, but the reliability of their inversion still needs to be improved. Correcting the bias of satellite-based precipitation estimates (SPEs) is a major challenge in applications such as environmental modeling, hydrology, and water resource management. In this paper, a new bias correction method—a Gaussian process regression (GPR) model method based on rain intensity classification—is proposed to improve the accuracy of a satellite precipitation product—the Integrated Multi-satellite Retrievals for GPM (IMERG) Final Run (FR) product—at the daily scale in the Guangdong-Hong Kong-Macao Greater Bay Area. By comparing the effects of the proposed method with those of other existing classical correction methods, namely, quantile mapping (QM), the support vector machine (SVM) approach and direct GPR, it is found that all four methods improve the accuracy of the FR product to varying degrees. GPR based on rain intensity classification has the best effect of FR product improvement, with the CORR increasing from 0.55 to 0.59, the RMSE ranging from 14.37 to 12.79 mm/day, and the BIAS ranging from − 0.14 to 0.03 during the validation period. GPR without rain intensity classification also yields good results, with the QM and SVM methods being the least effective. [ABSTRACT FROM AUTHOR]