Fusion and Evaluation of Multi-Source Satellite Remote Sensing Precipitation Products Based on Transformer Machine Learning

Satellite precipitation products offer great potential for acquiring reliable precipitation data in data-sparse areas, yet they have inherent uncertainties and errors as indirect observations. This study evaluated the accuracy of multi-source satellite precipitation products from daily and precipitation magnitude perspectives and discussed the spatiotemporal variation in their inversion errors. Based on ground rainfall observations, satellite products, and environmental factors, a Transformer-based multi-source precipitation fusion method was proposed, with its effectiveness preliminarily analyzed for daily precipitation in the Jingle River Basin. The main conclusions are as follows: (1) Compared with the observed precipitation data, the GSMaP_Gauge satellite remote sensing precipitation product showed the closest agreement with the observations, ranking first in all indicators except the Probability of Detection (POD). The MSWEP satellite remote sensing precipitation product followed in performance, while the CHIRPS satellite product performed the poorest. Satellite products showed distinct error characteristics across seasons and rainfall intensities, as well as general overestimation of light rain frequency and insufficient heavy rain capture; however, these products also showed better detection capability in flood seasons. Error spatial distribution was consistent with topography, vegetation coverage, and temperature. (2) Verification demonstrated that the Transformer fusion algorithm effectively reduced relative bias and improved correlation with ground data. The scheme which incorporated environmental factors outperformed the other, which only considered precipitation characteristics, achieving higher estimation accuracy and fusion stability.