Zhang et al. (2026) Interpretable deep learning method integrating spatial self-attention for generating bias-corrected high-resolution GFS precipitation forecasts

Identification

• Journal: Atmospheric Research
• Year: 2026
• Date: 2026-02-03
• Authors: Yufan Zhang, Shufeng Lai, Chongxun Mo, Tao Feng, Changhao Jiang, Na Li
• DOI: 10.1016/j.atmosres.2026.108832

Research Groups

• College of Architecture and Civil Engineering, Guangxi University, Nanning 530004, China
• Key Laboratory of Disaster Prevention and Structural Safety of Ministry of Education, Guangxi University, Nanning 530004, China
• Guangxi Key Laboratory of Disaster Prevention and Engineering Safety, Guangxi University, Nanning 530004, China
• Guangxi Provincial Engineering Research Center of Water Security and Intelligent Control for Karst Region, Guangxi University, Nanning 530004, China

Short Summary

This study introduces DualTransBU-Net-P, an explainable deep learning framework integrating spatial self-attention for end-to-end downscaling and bias correction of GFS precipitation forecasts, significantly enhancing accuracy and resolution while providing interpretability into its decision-making processes.

Objective

• To develop an explainable deep learning framework (DualTransBU-Net-P) that integrates spatial self-attention for joint downscaling and bias correction of Global Forecast System (GFS) precipitation forecast data to enhance regional precipitation forecasting accuracy and resolution.

Study Configuration

• Spatial Scale: Precipitation forecast resolution improved from 0.25° to 0.025°.
• Temporal Scale: Daily precipitation forecasts, analysis of heavy-precipitation day samples.

Methodology and Data

• Models used: DualTransBU-Net-P (explainable deep learning framework), DualTransBU-Net (core downscaling-bias correction model), SHAP (Shapley Additive Explanations) for interpretability. Base data from Global Forecast System (GFS) model.
• Data sources: GFS precipitation forecast data, multi-source data.

Main Results

• The proposed architecture significantly enhanced GFS precipitation forecast accuracy, improving resolution from 0.25° to 0.025°.
• The root mean square error (RMSE) of the test set was reduced by 4.6% to 18.7%.
• The fair threat score (ETS) was improved by an average of 43.9%.
• Among 437 heavy-precipitation day samples, RMSE decreased for 412 samples (94.3%).
• The ETS under the extreme precipitation threshold (>10 mm d⁻¹) increased by 50.6% to 63.1%.
• The model's performance remained high during seasonal analysis, demonstrating strong seasonal generalization.
• Interpretability analysis revealed distinct decision-making mechanisms of the deep learning model during heavy precipitation under typhoon and non-typhoon conditions, controlled by different underlying physical factors.

Contributions

• Proposes an interpretable deep learning framework (DualTransBU-Net-P) for end-to-end joint downscaling and bias correction of GFS precipitation forecasts.
• Integrates spatial self-attention and SHAP for enhanced accuracy, resolution, and interpretability in refined precipitation forecasting.
• Demonstrates significant improvements in forecast accuracy, particularly for extreme precipitation events, and strong seasonal generalization.
• Provides novel insights into the decision-making mechanisms of deep learning models under varying meteorological conditions (typhoon vs. non-typhoon).

Funding

Not specified in the provided text.

Citation

@article{Zhang2026Interpretable,
  author = {Zhang, Yufan and Lai, Shufeng and Mo, Chongxun and Feng, Tao and Jiang, Changhao and Li, Na},
  title = {Interpretable deep learning method integrating spatial self-attention for generating bias-corrected high-resolution GFS precipitation forecasts},
  journal = {Atmospheric Research},
  year = {2026},
  doi = {10.1016/j.atmosres.2026.108832},
  url = {https://doi.org/10.1016/j.atmosres.2026.108832}
}