Merging precipitation scheme design for improving the accuracy of regional precipitation products by machine learning and geographical deviation correction

Abstract

Global satellite precipitation products (SPPs) effectively obtain spatial precipitation information but frequently fail to meet application requirements in small-scale areas due to low accuracy. This study aims to design a merging precipitation scheme based on remote sensing and surface parameters to address the inaccuracy of regional precipitation. The scheme data are based on daily IMERG-FR, CHIRPS, and PDIR-Now satellite precipitation products from 2011 to 2018, combined with WorldClim model climate data and 39 rain gauge observations. The scheme generates merged precipitation with high-accuracy and high-resolution (0.04°) through the streamlined operation of multiple methods in the Songhua River basin in northeast China. First, the stacking algorithm was employed to perform preliminary merging of SPPs and decrease data noise errors (mean absolute error and standard deviation were reduced by 13.80% and 14.44%, respectively). Second, the correlation of merged and observed precipitation was improved by 1.18%–4.46% in different seasons after geographically weighted regression. Finally, the merged data were subjected to local intensity scaling, which reduced the precipitation error by an average of 3.14%. When compared to the original SPPs, the final merged precipitation (FMP) improved the correlation (increased by 0.19) and reduced the errors (root mean square error and relative error decreased by 1.71 mm and 0.08 mm, respectively). FMP performed well under precipitation event cases, with an average difference of 0.24 mm from observed precipitation. The study realized the merging analysis of remote sensing precipitation data with varying precision and spatial resolution. Furthermore, a systematic merging precipitation scheme coupled with multiple algorithms of machine learning, geographic regression, and mathematical statistics was formed. This study provides a reference for merging precipitation at a regional scale, which can be applied to other study areas as a process-oriented scheme.