Bias correction of seasonal precipitation prediction in Southern China: Pointwise and spatial corrections using deep learning methods

Southern China, characterized by complex and highly variable rainfall patterns, serves as the geographical focus of this study. The analysis covers 17 major cities and a broader spatial domain extending from 102°E to 121°E and 16°N to 35°N. This region encompassing all provinces south of the Qinling–Huaihe Line, including the middle and lower Yangtze River basin, southeastern coastal areas, and southwestern inland regions. This study aims to improve seasonal precipitation forecasts from the Climate Forecast System Version 2 (CFSv2) through deep learning-based bias correction. We explore two complementary strategies—pointwise and spatial correction—by integrating data-driven architectures with customized input configurations and training objectives, aiming to enhance predictive performance under complex regional precipitation regimes. The pointwise correction using LSTM significantly improves correlation coefficients in cities such as Shanghai, Nanjing, Hangzhou, and Hefei—from 0.24 to 0.52 on average—and reduces RMSE by up to 43.2 % in Nanning. All southwestern cities show RMSE reductions exceeding 24 %. For spatial correction, the U-Net model enhanced with the Extreme Value Penalty (EVP) loss function increases one-month lead anomaly correlation from 0.10 to 0.24 and reduces RMSE by 37.4 %. These findings underscore the potential of deep learning to enhance seasonal precipitation prediction in regions characterized by complex hydrometeorological conditions. • LSTM and U-Net show consistent spatial bias correction patterns in southern China. • Ensemble-mean input balances prediction accuracy with training efficiency. • Extreme penalty loss improves TCC/ACC and reduces RMSE in seasonal precipitation.