Three-step Merging of Daily Multi-satellite Rainfall Estimates Based on Probability Density Function Matching and Dynamic Bayesian Model Averaging

High-resolution precipitation data is significant for hydrological, climatological, environmental, and agricultural research. Fusing satellite precipitation data from different sources is an effective way for obtaining high-quality precipitation estimates. Thus we presented a three-step strategy within a dynamic framework for merging satellite precipitation products, thereby enhancing the precision of rainfall estimates. Firstly, the bilinear interpolation was adopted to downscale the spatial resolution of coarse products. Then, systematic biases in the downscaled products were individually eliminated utilizing the probability density function (PDF) matching method. Finally, the corrected products were fused utilizing the dynamic Bayesian model averaging (DBMA) method, producing a final merged precipitation product with a daily 1-km scale. The three-step framework generated dynamic weights that varied spatiotemporally and was applied to merge three satellite rainfall estimates, namely the Climate Prediction Center Morphing Technique (CMORPH), Integrated Multi-satellite Retrievals for GPM (IMERG), and Global Satellite Mapping of Precipitation (GSMaP), during the flood season (April to October) in the Kuye River Basin of China from 2008 to 2012. A total of 36 (80%) ground precipitation gauges were randomly chosen for calibration, while the remaining 8 (20%) were allocated for validation. The findings revealed that the merged product significantly outperformed each of the original satellite products on five evaluation metrics (root mean square error (RMSE) = 6.47 mm and correlation coefficient (CC) = 0.65), and DBMA assigned higher weight (0.34) to corrected CMORPH. Moreover, the corrected CMORPH demonstrated higher skills in northern regions with average weights ranging from 0.327 to 0.355. It was found that the proposed three-step merging approach not just enhanced the spatial resolution, but also provided significantly improved precipitation distribution details. In addition, we defined five different magnitudes of precipitation, namely light rain, moderate rain, heavy rain, rainstorm, and heavy rainstorm, to investigate the performance of the proposed method. The merged product had smaller variation ranges of the RMSEs and MAEs (mean absolute errors) over these five precipitation magnitudes, indicating more stable precision. The findings provide a promising and easily implementable alternative for generating high-quality precipitation data.