U-net-based spatial synergistic correction of gridded summer precipitation forecasts for CMA-MESO in central china
He, G., Wu, W., Zhang, Z., Ma, Z., Sun, J., et al. (2026). U-net-based spatial synergistic correction of gridded summer precipitation forecasts for CMA-MESO in central china. Atmospheric Research, doi:https://doi.org/10.1016/j.atmosres.2026.109217
| Title | U-net-based spatial synergistic correction of gridded summer precipitation forecasts for CMA-MESO in central china |
|---|---|
| Genre | Article |
| Author(s) | Guangxin He, Wei Wu, Z. Zhang, Z. Ma, Juanzhen Sun, H. Chen, J. Luo |
| Abstract | This study addresses the systematic biases in the quantitative precipitation forecasts produced by the China Meteorological Administration's regional high-resolution model CMA-MESO over Central China and proposes a spatial synergistic correction method based on the U-Net convolutional neural network. By constructing multi-source meteorological features into multi-channel two-dimensional matrices and designing a hybrid loss function that integrates weighted mean squared error with a differentiable TS score the model achieves an effective balance between numerical accuracy and spatial pattern fidelity. Experimental results indicate that compared with the original forecasts the U-Net model reduces the root mean square error by approximately 23% doubles or more the correlation coefficient and outperforms traditional loss functions under different precipitation thresholds and forecast lead times showing higher hit rates and lower false alarm rates. In addition with the help of Grad-CAM interpretability analysis the model demonstrates automatic focusing capability on key physical features such as composite reflectivity and dynamic fields and its attention distribution exhibits a transition mechanism from dynamically driven to thermodynamically regulated features as the forecast lead time increases thereby validating the consistency between the decision-making process and atmospheric physical laws. |
| Publication Title | Atmospheric Research |
| Publication Date | Jul 1, 2026 |
| Publisher's Version of Record | https://doi.org/10.1016/j.atmosres.2026.109217 |
| OpenSky Citable URL | https://n2t.net/ark:/85065/d7js9w1n |
| OpenSky Listing | View on OpenSky |
| RAL Affiliations | HAP |