A two‑stage deep learning framework for multi‑source precipitation fusion to optimize seasonal, extreme and precipitation event estimation

Study region The study is conducted over mainland China (18.0–53.5°N, 73.5–135.0°E), with a focus on the nine major river basins across China. Study focus A two-stage deep learning multi-source precipitation fusion framework (TDL-MSP) is developed to mitigate systematic biases in existing products and overcome limitations of conventional single-stage regression merging. The framework integrates six widely used precipitation datasets and four near-surface meteorological variables from CMFD 2.0, generating daily 0.25° gridded precipitation across China for 2001–2022. It adopts a classification-regression architecture: UNet identifies precipitation events in the first stage, and SegNeXt performs quantitative rainfall regression in the second. Model performance is evaluated against original products using the CHM_PRE ground dataset as reference. New hydrological insights for the region The TDL‑MSP framework outperforms all original multi‑source precipitation products across China. The optimal SegNeXt_11x model achieves a Pearson correlation coefficient (CC) of 0.817, Kling‑Gupta efficiency (KGE) of 0.608, and root mean square error (RMSE) of 2.364 mm/day. Specifically, its CC is 0.267 higher than ERA5, KGE is 0.388 higher than MSWEP, and RMSE is reduced by up to 1.538 mm/day. Overall, the UNet‑SegNeXt framework improves seasonal and extreme precipitation estimation, exhibiting promising prospects for next‑generation high‑resolution gridded precipitation datasets.