Ryu et al. (2026) Increasing resolution and accuracy in sub-seasonal forecasting through 3D U-Net: the western US

Identification

• Journal: Geoscientific model development
• Year: 2026
• Date: 2026-01-05
• Authors: Jihun Ryu, Hisu Kim, Shih-Yu (Simon) Wang, Jin‐Ho Yoon
• DOI: 10.5194/gmd-19-27-2026

Research Groups

• School of Environment and Energy Engineering, Gwangju Institute of Science and Technology, Gwangju, South Korea
• Department of Plants, Soils and Climate, Utah State University, Logan, UT, USA
• School of Electrical Engineering and Computer Science, Gwangju Institute of Science and Technology, Gwangju, South Korea
• Department of Agronomy, Kasetsart University, Bangkok, Thailand

Short Summary

This study investigates using a 3D U-Net architecture for post-processing sub-seasonal forecasts in the western U.S. to enhance predictability and spatial resolution. The model significantly improves temperature predictability and spatial resolution, outperforming traditional Numerical Weather Prediction (NWP) models, with the most efficient configuration utilizing the ensemble mean and only target variables.

Objective

• To investigate the use of 3D U-Net architecture for post-processing sub-seasonal forecasts to enhance both predictability and spatial resolution, focusing on the western U.S.
• To identify the role played by ensemble members and additional variables in enhancing predictability and to determine if neural network-based downscaling leads to meaningful improvements at smaller scales, such as the county level.

Study Configuration

• Spatial Scale: Western United States, with analysis downscaled to county-level regions (e.g., San Francisco, Orange County, Great Salt Lake area, Seattle, eastern Washington wheat farming region). Input data resolutions were 1.5° × 1.5° (approximately 120 km × 120 km) for ECMWF and 0.042° × 0.042° (approximately 4 km) for PRISM, interpolated to a 0.25° × 0.25° grid for model input.
• Temporal Scale: Sub-seasonal forecasts with lead times up to 32 days, generating daily forecasts. Data covered January 2015 to December 2023 for training/validation/testing.

Methodology and Data

• Models used: 3D U-Net architecture for post-processing. The model was trained using the Adam optimizer, mean squared error (MSE), and spatial pattern correlation as the loss function.
• Data sources:
  • Input: European Centre for Medium-Range Weather Forecasts (ECMWF) real-time perturbed forecasts (1.5° × 1.5° resolution, 50 ensemble members, 32 days lead time, daily/6-hourly frequency, CY40R1 to CY48R1 from January 2015 to December 2023). Variables included 2 m temperature, precipitation, total column water, mean sea level pressure, 10 m u/v wind, elevation, geopotential height, and u/v wind at 850, 500, and 200 hPa levels.
  • Target/Validation: Parameter-elevation Regressions on Independent Slopes Model (PRISM) dataset (0.042° × 0.042° resolution, daily, January 2015 to January 2024) for temperature, precipitation, and elevation.
  • Complementary: ERA5 data for additional experiments.

Main Results

• The 3D U-Net model consistently outperformed raw NWP forecasts across multiple evaluation metrics (pattern correlation, RMSE, Epre) for both temperature and precipitation, except for the Epre metric in precipitation where results were mixed.
• Temperature predictability was significantly improved, with the 3D U-Net model enhancing spatial resolution and reducing overall forecast errors.
• The pattern correlation coefficient for temperature and precipitation improved by 0.12 and 0.19, respectively, over a 32-day lead time compared to NWP.
• Root Mean Square Error (RMSE) was reduced by approximately 31% for temperature and 22% for precipitation.
• The most efficient model configuration was found to be using the ensemble mean (E50M) and only the target variables (V1), demonstrating comparable performance to using the full set of 50 ensemble members or additional meteorological variables.
• Challenges remain in accurately forecasting extreme precipitation events, with the model tending to underestimate precipitation in coastal and mountainous regions, despite improved spatial detail.
• At the county level, 3D U-Net models generally showed improved or comparable performance for temperature forecasts, but precipitation forecast improvements were less consistent and varied significantly by region.

Contributions

• This study introduces and validates a 3D U-Net architecture for post-processing sub-seasonal forecasts, demonstrating significant enhancements in both predictability and spatial resolution for temperature and precipitation in the complex terrain of the western U.S.
• It extends previous findings from short-range forecasting to the sub-seasonal prediction regime, showing that the ensemble mean and target-only variables are sufficient for optimal performance, leading to a more computationally efficient approach.
• The research provides a framework for daily, high-resolution sub-seasonal post-processing that is operationally feasible on commodity GPUs, offering practical value for regional decision-making in water, fire, and agricultural management.
• It systematically investigates the role of ensemble members and additional input variables in deep learning-based sub-seasonal forecasting, challenging the conventional wisdom that more input data invariably leads to better predictions in this context.

Funding

• National Research Foundation of Korea (RS-2025-02363044)
• Korean Meteorological Agency (KMI2018-07010)

Citation

@article{Ryu2026Increasing,
  author = {Ryu, Jihun and Kim, Hisu and Wang, Shih-Yu (Simon) and Yoon, Jin‐Ho},
  title = {Increasing resolution and accuracy in sub-seasonal forecasting through 3D U-Net: the western US},
  journal = {Geoscientific model development},
  year = {2026},
  doi = {10.5194/gmd-19-27-2026},
  url = {https://doi.org/10.5194/gmd-19-27-2026}
}