Enhancing precipitation merging accuracy in China with machine learning and rain-snow classification

AbstractRising global temperatures have increased the complexity of precipitation patterns and phase transitions, posing significant challenges for accurate satellite precipitation estimation. Although machine learning has demonstrated strong capability in modeling complex precipitation processes, the influence of precipitation phase has been largely overlooked in precipitation merging studies. This study explicitly integrates precipitation phase information into satellite precipitation merging over China using Random Forest (RF) and Artificial Neural Networks (ANN), with Geographically Weighted Regression (GWR) serving as a benchmark. Three training strategies are evaluated, including phase independent training, phase aware combined training, and phase specific independent training, to evaluate the effectiveness of phase informed machine learning. The results indicate that: (1) machine learning markedly improves the accuracy of snowfall data merging, with RF consistently outperforming ANN, whereas ANN exhibits a systematic tendency to overestimate precipitation across all phases. (2) Both Merge_p1 and Merge_p2 further enhance merging performance, with Merge_p2 being particularly effective in alleviating ANN-related overestimation, reducing the snowfall RMSE from 2.30 to 1.93 mm day–1. (3) Precipitation merging accuracy declines as station density decreases, with model sensitivity following the order ANN > GWR > RF, with higher sensitivity observed for rainfall than snowfall. These findings contribute to improving precipitation estimation by incorporating phase differentiation and provide a robust methodological foundation for identifying extreme climate events.