---
格式版本: 2
标题: "Earth observation embeddings are effective sub-grid descriptors for probabilistic weather downscaling"
原文链接: "https://arxiv.org/abs/2608.12271"
发布日期: "2026-08-12"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "**\\[v1\\]** Wed, 12 Aug 2026 17:10:42 UTC (10,246 KB)"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 18
发布时间严格候选数量: 6
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-08-13T18:31:36+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-08-13T18:24:32+08:00"
入库时间: "2026-08-13T10:31:36.190Z"
来源平台: "arXiv 学术论文搜索"
搜索渠道: "source_template"
搜索词: "https://arxiv.org/search/?query=AI&searchtype=all"
匹配关键词:
  - "AI"
相关厂家:
  []
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 5
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "该论文是关于天气降尺度的机器学习研究，与超节点/AI Rack/机柜级AI基础设施完全无关，无任何相关技术、商业或部署信息。"
AI质检模型: "ali-deepseek-v4-flash"
AI质检时间: "2026-08-13T18:32:04+08:00"
AI主题相关性: 0
AI来源权威性: 3
AI新颖性: 0
AI技术细节: 0
AI商业部署信号: 0
AI完整性: 2
AI摘要: "该研究提出用地球观测基础模型TESSERA生成的10米分辨率嵌入，作为概率天气降尺度的亚网格地表描述符，替代传统手工地形特征。"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-07T03:40:12.866Z"
采集批次: "2026年8月13日18点24分27秒"
采集批次ID: "20260813-182427-1564e572"
去重键: "https://arxiv.org/abs/2608.12271"
---

## Computer Science > Machine Learning

## Title:Earth observation embeddings are effective sub-grid descriptors for probabilistic weather downscaling

Authors:[Pedro Sousa](https://arxiv.org/search/cs?searchtype=author&query=Sousa,+P) (1), [Will Tebbutt](https://arxiv.org/search/cs?searchtype=author&query=Tebbutt,+W) (2), [Sadiq Jaffer](https://arxiv.org/search/cs?searchtype=author&query=Jaffer,+S) (1), [Robin Young](https://arxiv.org/search/cs?searchtype=author&query=Young,+R) (1), [Anil Madhavapeddy](https://arxiv.org/search/cs?searchtype=author&query=Madhavapeddy,+A) (1), [Richard E. Turner](https://arxiv.org/search/cs?searchtype=author&query=Turner,+R+E) (2) ((1) Department of Computer Science, University of Cambridge, (2) Department of Engineering, University of Cambridge)

[View PDF](https://arxiv.org/pdf/2608.12271) [HTML (experimental)](https://arxiv.org/html/2608.12271v1)

> Abstract:Global weather reanalyses and forecasts resolve the evolving atmospheric state on coarse grids, but site-specific applications require predictions at arbitrary locations where near-surface conditions also depend on unresolved terrain and land-surface properties. Existing probabilistic downscalers address this gap using hand-crafted topographic descriptors. We ask instead whether Earth observation foundation models can provide transferable sub-grid surface representations for probabilistic weather downscaling.  
> We augment a convolutional conditional neural process that downscales coarse ERA5 reanalysis fields at ~25 km resolution with a learned local surface descriptor, obtained by compressing a patch of TESSERA embeddings at 10 m resolution. Although these embeddings summarise surface conditions over annual timescales, they improve downscaling of instantaneous 2 m temperature and 10 m wind speed by encoding persistent surface properties that capture a location's departure from the coarse-grid atmospheric state. Across five climatically diverse regions, the embedding improves point and probabilistic skill at stations held out in both space and time, overall improving CRPS skill by 11.5% for 2 m temperature and 6.2% for 10 m wind speed. We further analyse how its contribution differs by variable, finding that topography explains more of temperature's sub-grid structure, while TESSERA provides additional surface information for wind speed.  
> These improvements persist when the coarse input is changed from ERA5 to forecasts from the Aurora AI forecasting model, and when predicting at newly deployed stations with no regional history. To our knowledge, this is the first evidence that long-timescale Earth-observation embeddings can support short-timescale weather downscaling where sub-grid departures are systematically structured by persistent surface properties.

| Comments: |  |
| --- | --- |
| Subjects: | Machine Learning (cs.LG); Atmospheric and Oceanic Physics (physics.ao-ph) |
| Cite as: | [arXiv:2608.12271](https://arxiv.org/abs/2608.12271) \[cs.LG\] |
|  | (or [arXiv:2608.12271v1](https://arxiv.org/abs/2608.12271v1) \[cs.LG\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2608.12271](https://doi.org/10.48550/arXiv.2608.12271) |

## Submission history

From: Pedro Miguel Marques Sousa \[[view email](https://arxiv.org/show-email/4ca1429f/2608.12271)\]  
**\[v1\]** Wed, 12 Aug 2026 17:10:42 UTC (10,246 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2608.12271) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))