Abstract: Based on hindcast experiment results from 1991 to 2020 using the dynamical climate prediction system of the Institute of Atmospheric Physics, Chinese Academy of Sciences (IAP-DCPv3.5), this study first evaluates the system's prediction skill for summer precipitation anomalies in the Lancang-Mekong River Basin (LMRB), as well as the impact of the Empirical Orthogonal Function (EOF) based correction method on seasonal prediction skill. Subsequently, by incorporating key atmospheric circulation features that influence precipitation in LMRB, a bias correction method based on the U-Net model is proposed, with the feature extraction technique for atmospheric circulations within the model being effectively optimized. Verifications with the observation during 2021–2024 demonstrate that the U-Net based correction method can substantially improve the model's seasonal rainfall prediction skill. For summer mean precipitation predictions initialized in June, May, and March, the mean absolute error is reduced by over 50% compared to the original predictions. Concurrently, it further improves the spatial distribution characteristics of the predicted summer precipitation anomalies by IAP-DCPv3.5 and overcomes the instability inherent in the traditional EOF based bias correction method. For the prediction skill of summer precipitation anomalies initialized in June, May, and March, the average spatial Pattern Correlation Coefficient (PCC) between the observation and the original predictions are merely 0.09, −0.01, and 0.09, respectively. After applying the U-Net correction method, the corresponding PCCs increase to 0.31, 0.21, and 0.19 respectively. In contrast, when using the EOF method, the PCCs for predictions initialized in May and June even decrease to −0.02 and 0.02, respectively. Verification of the prediction for typical basin-wide drought in 2021 reveals that, the U-Net method can successfully predict the observed magnitudes and spatial distribution patterns of the summer rainfall anomalies in the LMRB, with the PCC prediction skill initialized in March reaching as high as 0.50. This study demonstrates that the integration of deep learning technique with numerical model predictions is an effective approach for improving seasonal climate prediction capabilities.