Li et al. (2026) Domain-knowledge-guided deep learning for probabilistic precipitation downscaling over southeastern Tibetan Plateau

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-08-17
• Authors: Dayang Li, Baoxiang Pan, Yuan Liu, Ashish Sharma, Yan Zhou, Yi Yao, Jun Yin, Long Yang
• DOI: 10.1016/j.jhydrol.2026.136242

Research Groups

• College of Civil Engineering, Yancheng Institute of Technology
• Key Laboratory of Hydrometeorological Disaster Mechanism and Warning of Ministry of Water Resources, Nanjing University of Information Science & Technology
• School of Geography and Ocean Science, Nanjing University
• Frontiers Science Centre for Critical Earth Material Cycling, Nanjing University
• Institute of Atmospheric Physics, Chinese Academy of Sciences
• Department of Civil and Environmental Engineering, University of Wisconsin-Madison
• School of Civil and Environmental Engineering, University of New South Wales
• The National Key Laboratory of Water Disaster Prevention, Nanjing Hydraulic Research Institute

Short Summary

This study proposes a deep learning-based downscaling method, P-HPDNet, to predict probabilistic precipitation over the southeastern Tibetan Plateau. P-HPDNet incorporates domain knowledge and outperforms baseline models in capturing spatial and temporal dependencies.

Objective

• Investigate the feasibility of using deep learning for probabilistic precipitation downscaling over complex terrains

Study Configuration

• Spatial Scale: 9-km and 1-km resolutions over southeastern Tibetan Plateau
• Temporal Scale: Daily precipitation data from a ten-year period

Methodology and Data

• Models used: P-HPDNet, a deep learning-based downscaling network with convolutional layers, long short-term memory network layers, and spatial attention mechanisms
• Data sources: Weather Research and Forecasting model simulations, satellite-based precipitation product

Main Results

• P-HPDNet reduces the continuous ranked probability score by 28% for all precipitation values, with an even larger reduction of 55% for extreme precipitation
• Significant improvements in capturing both temporal and spatial dependencies of precipitation when applied to downscale a satellite-based precipitation product without retraining

Contributions

• This study highlights the importance of incorporating domain knowledge of precipitation characteristics into deep learning models, offering an efficient downscaling approach for regions with complex terrains.

Funding

• This research was supported by the National Key Research and Development Program of China (Grant No. 2021YFA1501900), the Natural Science Foundation of Jiangsu Province (Grant No. BK20210411)

Citation

@article{Li2026Domainknowledgeguided,
  author = {Li, Dayang and Pan, Baoxiang and Liu, Yuan and Sharma, Ashish and Zhou, Yan and Yao, Yi and Yin, Jun and Yang, Long},
  title = {Domain-knowledge-guided deep learning for probabilistic precipitation downscaling over southeastern Tibetan Plateau},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136242},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136242}
}