Singh et al. (2026) Comprehensive inter-comparison of generative AI models for super-resolution precipitation downscaling across hydroclimatic regimes

Identification

• Journal: Geoscientific model development
• Year: 2026
• Date: 2026-08-17
• Authors: Shivam H Singh, Simon Michael Papalexiou, Hebatallah Abdelmoaty, Tom Hartvigsen, Antonios Mamalakis
• DOI: 10.5194/gmd-19-7545-2026

Research Groups

• Environmental Institute, University of Virginia
• Department of Environmental Sciences, University of Virginia
• Institute for Global Water Security, Hamburg University of Technology
• Faculty of Environmental Sciences, Czech University of Life Sciences Prague
• Department of Civil Engineering, Schulich School of Engineering, University of Calgary
• Irrigation and Hydraulics Department, Faculty of Engineering, Cairo University
• School of Data Science, University of Virginia

Short Summary

This study presents a comprehensive intercomparison of three deep-learning architectures for precipitation super-resolution downscaling: U-NET, WGAN, and DDPM. The models are evaluated using a perfect-model experimental design based on ERA5-Land precipitation fields over distinct hydroclimatic regimes in the United States.

Objective

• To compare the performance of deterministic convolutional U-NET, conditional Wasserstein GAN (WGAN), and denoising diffusion probabilistic model (DDPM) for precipitation super-resolution downscaling.
• To assess their ability to reproduce precipitation occurrence, spatial structure, temporal consistency, precipitation distributions, extreme-event behavior, and uncertainty characteristics.

Study Configuration

• Spatial Scale: 128 × 128 grid domains (∼1150 km × 1150 km at mid-latitudes) representing major US hydroclimatic regimes.
• Temporal Scale: Daily precipitation fields from the ERA5-Land reanalysis dataset (1980–2014).

Methodology and Data

• Models used:
  • Deterministic convolutional U-NET as a baseline
  • Conditional Wasserstein GAN (WGAN)
  • Denoising diffusion probabilistic model (DDPM)
• Data sources: ERA5-Land reanalysis dataset

Main Results

• The study found that the DDPM outperforms both the U-NET and WGAN in terms of distributional consistency, spatial structure, and storm morphology.
• The DDPM also exhibits better performance in representing extreme events and uncertainty characteristics.
• The WGAN shows improved representation of fine-scale variability and extreme intensities compared to the U-NET.

Contributions

• This study provides a comprehensive comparison of three deep-learning architectures for precipitation super-resolution downscaling.
• It highlights the strengths and limitations of each model, providing insights into their relative performance and uncertainty representation.
• The results have implications for hydrologic and climate-risk applications requiring reliable representation of precipitation extremes, spatial organization, temporal persistence, and predictive uncertainty.

Funding

• This research was funded by the University of Virginia's Environmental Institute and the School of Data Science.

Citation

@article{Singh2026Comprehensive,
  author = {Singh, Shivam H and Papalexiou, Simon Michael and Abdelmoaty, Hebatallah and Hartvigsen, Tom and Mamalakis, Antonios},
  title = {Comprehensive inter-comparison of generative AI models for super-resolution precipitation downscaling across hydroclimatic regimes},
  journal = {Geoscientific model development},
  year = {2026},
  doi = {10.5194/gmd-19-7545-2026},
  url = {https://doi.org/10.5194/gmd-19-7545-2026}
}