BaguanHR: Scaling Data for High-Resolution Weather Forecasting
我参与的论文 Pushing the Limits of High-Resolution Weather Forecasting Through Data Scaling 已发表于 ECCV 2026。我们提出 BaguanHR，通过逐变量超分辨率扩充训练数据，支持全球 0.1° 高分辨率天气预报。
论文链接：Springer Nature Link
预印本：arXiv:2608.14652
发表信息
| 项目 | 信息 |
|---|---|
| 论文题目 | Pushing the Limits of High-Resolution Weather Forecasting Through Data Scaling |
| 作者 | Yang Zhao, Peisong Niu, Tian Zhou, Ziqing Ma, Guanlong Ma, Rong Jin, Huiling Yuan, Liang Sun |
| 会议 | 19th European Conference on Computer Vision（ECCV 2026） |
| 论文集 | Computer Vision – ECCV 2026, Proceedings, Part LIX |
| 丛书 | Lecture Notes in Computer Science（LNCS），第 17059 卷 |
| 页码 | 135–152 |
| 出版社 | Springer, Cham |
| 正式上线日期 | 2026 年 10 月 5 日 |
| 电子版 ISBN | 978-3-032-37086-0 |
| DOI | 10.1007/978-3-032-37086-0_8 |
Yang Zhao、Peisong Niu 与 Tian Zhou 为共同第一作者；Huiling Yuan 与 Liang Sun 为通讯作者。
论文摘要
全球 0.1° 天气预报面临高分辨率训练数据不足的问题，而长期 ERA5 再分析资料主要提供 0.25° 数据。BaguanHR 使用逐变量超分辨率重建，将这些历史资料转为 0.1° 合成数据，再结合真实高分辨率资料训练预报模型。
实验中，BaguanHR 在 72 小时内超过 85% 的预报时效上优于所评估的机器学习方法及 IFS-HRES。数据规模扩展带来幂律式性能改善：论文报告 72 小时和 120 小时预报的 RMSE 分别降低 4.6% 和 4.9%。
核心贡献
- 扩充高分辨率数据：以逐变量超分辨率利用长期低分辨率再分析资料。
- 验证训练策略：实验与理论分析支持合成数据和真实数据结合的方案。
- 揭示数据规模效应：量化增加训练数据对不同预报时效的收益。
方法细节见论文全文。
Our paper, Pushing the Limits of High-Resolution Weather Forecasting Through Data Scaling, has been published at ECCV 2026. BaguanHR uses variable-wise super-resolution to expand training data for global 0.1° weather forecasting.
Paper: Springer Nature Link
Preprint: arXiv:2608.14652
Publication Information
| Item | Details |
|---|---|
| Title | Pushing the Limits of High-Resolution Weather Forecasting Through Data Scaling |
| Authors | Yang Zhao, Peisong Niu, Tian Zhou, Ziqing Ma, Guanlong Ma, Rong Jin, Huiling Yuan, Liang Sun |
| Conference | 19th European Conference on Computer Vision (ECCV 2026) |
| Proceedings | Computer Vision – ECCV 2026, Proceedings, Part LIX |
| Series | Lecture Notes in Computer Science (LNCS), volume 17059 |
| Pages | 135–152 |
| Publisher | Springer, Cham |
| First online | October 5, 2026 |
| Online ISBN | 978-3-032-37086-0 |
| DOI | 10.1007/978-3-032-37086-0_8 |
Yang Zhao, Peisong Niu, and Tian Zhou contributed equally. Huiling Yuan and Liang Sun are the corresponding authors.
Abstract
High-resolution training data is scarce. BaguanHR reconstructs historical 0.25° ERA5 fields at 0.1° and combines synthetic and real data to train a forecasting model.
It outperforms evaluated ML baselines and IFS-HRES at over 85% of lead times within 72 hours. Data scaling follows a power-law trend, with reported RMSE reductions of 4.6% at 72 hours and 4.9% at 120 hours.
Main Contributions
- Data expansion: variable-wise super-resolution makes historical reanalysis useful for high-resolution training.
- Training strategy: experiments and theory support combining synthetic and real data.
- Scaling analysis: quantifies benefits across forecast horizons.
See the full paper for methodology details.