Li et al. (2026) A tailored deep learning method to improve spatial rainfall downscaling

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-09
• Authors: Tinghui Li, Shuiqing Yin, Yuanyuan Xiao, Nadav Peleg
• DOI: 10.1016/j.jhydrol.2026.135272

Research Groups

• State Key Laboratory of Earth Surface Processes and Hazards Risk Governance (ESPHR), Faculty of Geographical Science, Beijing Normal University, China
• Institute of Earth Surface Dynamics, University of Lausanne, Switzerland
• Expertise Center for Climate Extremes, University of Lausanne, Switzerland

Short Summary

This study developed a tailored deep learning model, RM-ResNet, incorporating a spatial correction algorithm to downscale satellite rainfall data from 8 km to 1 km resolution. The method successfully improved the representation of rainfall spatial patterns, including extreme events and storm centers, demonstrating consistency with observations.

Objective

• To develop a deep learning method that accurately downscales rainfall, specifically addressing the challenge of correctly representing heavy rainfall intensities and spatial structures, especially when high-resolution training data are limited and agreement with low-resolution target data is low.

Study Configuration

• Spatial Scale: Downscaling from 8 km to 1 km resolution. Case study area: Beijing.
• Temporal Scale: Hourly and daily rainfall data, with training data available since 2015.

Methodology and Data

• Models used: RM-ResNet (a tailored deep learning model), coupled with a spatial correction algorithm for training data.
• Data sources:
  • High-resolution observations (1 km, hourly) for Beijing (training data).
  • CMORPH (Climate Prediction Center MORPHing technique) satellite estimates (8 km, hourly) for downscaling.

Main Results

• The RM-ResNet model, using spatially corrected training data, successfully downscaled hourly CMORPH estimates from 8 km to 1 km resolution.
• Both hourly and daily extreme rainfall events in the downscaled data were consistent with observations.
• The downscaled spatial structures of rainfall were well captured.
• Missing storm centers in the original CMORPH data were more realistically represented after downscaling, remaining within a reasonable physical range.

Contributions

• Introduction of a spatial correction algorithm for training data to address discrepancies between high-resolution training and low-resolution target data.
• Development of RM-ResNet, a tailored deep learning model capable of simultaneously learning rainfall occurrence and magnitude for improved downscaling.
• Demonstrated effective downscaling of rainfall, including extreme intensities and complex spatial patterns (e.g., storm centers), using lightweight deep learning networks.
• Provides a strategy for rainfall downscaling applicable in regions with limited high-resolution training data.

Funding

[No funding information was provided in the excerpt.]

Citation

@article{Li2026tailored,
  author = {Li, Tinghui and Yin, Shuiqing and Xiao, Yuanyuan and Peleg, Nadav},
  title = {A tailored deep learning method to improve spatial rainfall downscaling},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135272},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135272}
}