Wan et al. (2026) A self-supervised deep learning method for retrieving historically similar extreme rainfall patterns

Identification

• Journal: Environmental Modelling & Software
• Year: 2026
• Date: 2026-07-21
• Authors: Tianyu Wan, Chao Gao, Xiongpeng Tang, Xicheng Zhang, Xin Yin, 汤静静, Yue‐Ping Xu
• DOI: 10.1016/j.envsoft.2026.107116

Research Groups

• Guangdong-Hong Kong Joint Laboratory for Water Security, Beijing Normal University, Zhuhai, China
• Institute for Advanced Study, Beijing Normal University, Zhuhai, China
• College of Water Sciences, Beijing Normal University, Beijing, China
• Nanjing Hydraulic Research Institute, Nanjing, China
• Qingyuan Hydrology Branch of Hydrology Bureau of Guangdong Province, Qingyuan, China
• Institute of Hydrology and Water Resources, College of Civil Engineering and Architecture, Zhejiang University, Hangzhou, China

Short Summary

This study proposes a novel self-supervised deep learning method for retrieving historically similar extreme rainfall patterns at hourly, daily, and event scales. The method integrates a vision transformer with masked pretraining to learn rainfall representations and demonstrates accurate and efficient retrieval of diverse rainfall patterns.

Objective

• Investigate the feasibility of using a self-supervised deep learning approach for retrieving historically similar extreme rainfall patterns

Study Configuration

• Spatial Scale: Hourly, daily, and event scales
• Temporal Scale: Historical rainfall archives

Methodology and Data

• Models used: Vision transformer with masked pretraining
• Data sources: Satellite and observation data

Main Results

• Accurate retrieval of diverse rainfall patterns at hourly, daily, and event scales
• Mean query time of 21.35 ms over more than 10,000 rainfall maps
• Scalability to expanding rainfall archives and retrieval of dynamic analogues for localized extremes and rare large-scale events

Contributions

• Original contribution to the field of hydrometeorological modelling by proposing a novel self-supervised deep learning method for retrieving historically similar extreme rainfall patterns
• Extension of retrieval from static pattern matching to dynamic event tracking

Funding

• This research was funded by the Guangdong-Hong Kong Joint Laboratory for Water Security (Grant No. [not specified]) and the National Natural Science Foundation of China (Grant No. [not specified]).

Citation

@article{Wan2026selfsupervised,
  author = {Wan, Tianyu and Gao, Chao and Tang, Xiongpeng and Zhang, Xicheng and Yin, Xin and 汤静静 and Xu, Yue‐Ping},
  title = {A self-supervised deep learning method for retrieving historically similar extreme rainfall patterns},
  journal = {Environmental Modelling & Software},
  year = {2026},
  doi = {10.1016/j.envsoft.2026.107116},
  url = {https://doi.org/10.1016/j.envsoft.2026.107116}
}