Comparison and accuracy enhancement of latest GPM IMERG and ERA5-Land precipitation products using machine learning methods over typical data-scarce high mountains in China

The correction of gridded precipitation products (GPPs) is a critical approach for obtaining high-quality precipitation data in data-scarce regions. However, a comprehensive understanding of the accuracy of corrected reanalysis and satellite precipitation products remains lacking in such regions, posing a challenge for selecting a higher-quality corrected GPP. In this study, two representative and newly released GPPs, including the Integrated Multi-Satellite Retrievals for Global Precipitation Measurement (GPM IMERG) V07_FR (GPMv07) and the European Centre for Medium-Range Weather Forecasts Reanalysis V5 for Land (ERA5-Land), were evaluated and corrected using three machine learning (ML) methods in the Tianshan Mountains in China. The potential advantages of the optimally corrected GPP were further identified. The results showed that both products exhibited significant errors, as reflected by the fact that the proportion of stations with absolute percent bias (PB) exceeding 20% was above 64.0%. In addition, the accuracy of GPMv07 and ERA5-Land exhibited altitude dependence. The method using random forest for correcting GPMv07 (RF-G) outperformed other bias-correction methods in reproducing monthly precipitation. Compared with the original GPMv07, the RF-G method improved the mean Kling–Gupta efficiency (KGE) and the proportion of stations with absolute PB no greater than 20% by 0.57 and 67.7%, respectively. Furthermore, the monthly precipitation data corrected by the RF-G method was superior to several state-of-the-art merged precipitation products, including CMFD, AIMERG, AERA5-Asia, and GMCP, at the monthly scale. These findings highlight the value of correcting the newly released GPMv07 and can provide support for water resources management in data-scarce mountainous areas.