Constrained by atmospheric internal variability and complexity of precipitation processes, current numerical models demonstrate limited skill for subseasonal precipitation. This study develops a U-Net-based correction model for subseasonal rainfall in summer over China, incorporating precipitation, atmospheric circulation variables, and topographic features from the regional Climate–Weather Research and Forecasting (CWRF) model. Results indicate that the U-Net model systematically corrects biases over 6 mm day^-1 on average, extending prediction skill from the lead time of 2 pentads (10–14 days) to 5 pentads (25–29 days). The U-Net model provides substantial forecast improvement in Central China, South China, and Southwest China while maintaining the high prediction skills of the CWRF model over North China. Interpretability analysis further demonstrates topography, 200-hPa zonal wind and 500-hPa geopotential height (500-hPa temperature) collectively contribute over 45%, with underlying mechanism also discussed. This study confirms the value of a hybrid dynamical–statistical approach in enhancing regional model prediction performance, providing a novel methodology for subseasonal precipitation prediction.