Near-Real-Time Error Correction of Satellite Precipitation Products Using a U-Net-Based Deep Learning Approach

This study investigates the application of a U-Net deep learning model for real-time error correction of satellite precipitation products, using IMERG-Late data over mainland China during 2015–2017. The performance of U-Net is systematically compared with five widely used machine learning approaches, including Decision Tree Regressor, Gradient Boosting Decision Tree, Multilayer Perceptron, Random Forest, and Extreme Gradient Boosting. The results indicate that U-Net consistently improves overall precipitation accuracy and is particularly effective in reducing false precipitation events, outperforming the comparative methods across most evaluation metrics. However, a relatively lower probability of detection (POD) is observed for U-Net, which is primarily associated with over-correction effects that suppress weak or marginal precipitation signals. Despite this limitation, the overall performance demonstrates that the U-Net-based approach provides a robust and reliable solution for real-time error correction of satellite precipitation products, with clear potential for hydrological and meteorological applications.