Environment

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

Pedro Sousa, Will Tebbutt, Sadiq Jaffer, Robin Young, Anil Madhavapeddy, Richard E. Turner

Featured August 19, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Using satellite-derived surface maps helps predict local weather like temperature and wind much better than before, especially in places without many weather stations, by understanding how the ground itself shapes the weather.

In depth

The paper introduces a novel approach to improve local weather predictions by augmenting a convolutional conditional neural process with a learned local surface descriptor. This descriptor is derived from compressing high-resolution Earth observation (EO) foundation model embeddings (Tessera) using a Variational Autoencoder. This allows the downscaling model to capture complex, persistent sub-grid surface properties beyond traditional topographic features, leading to more accurate and robust probabilistic forecasts, especially for 10m wind speed and in data-scarce regions.

Key Takeaways

1
The authors demonstrate that Tessera embeddings, a learned representation of surface properties from satellite imagery, significantly enhance the accuracy of probabilistic weather downscaling for 2m temperature and 10m wind speed.
2
The proposed method shows robustness when applied to medium-range AI weather forecasts (Aurora) and dramatically improves sample efficiency for new station deployments, outperforming baselines even with no local historical data.
3
The Tessera descriptor provides unique, transferable land-surface information that is particularly crucial for accurately downscaling 10m wind speed, which is less effectively captured by topography alone.

Conceptual Flow

HIGH LEVEL

Methodology: Combining Global Weather with Local Ground Details

The system takes big-picture weather data and combines it with detailed ground information from satellites to guess local weather.

Big Weather Map

Detailed Ground Info

Combine & Learn

Local Weather Guess

Results: More Accurate Local Weather Predictions

Adding detailed ground info made local weather predictions much more accurate, especially for wind and in areas with few sensors.

Old Weather Guess

New Weather Guess

Compare Accuracy

Much Better Predictions

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.