Climate

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AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content.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

By using Earth observation embeddings to describe local surface properties, a neural network can accurately predict local weather conditions, like temperature and wind, even at places it has never seen before.

In depth

The paper introduces a novel approach to probabilistic weather downscaling by augmenting a convolutional conditional neural process (ConvCNP) with a learned local surface descriptor. This descriptor, derived from compressing Tessera Earth observation embeddings, captures persistent sub-grid properties like land cover and terrain, significantly improving predictions of 2m temperature and 10m wind speed at previously unobserved locations and times.

Key Takeaways

1
The study demonstrates that Earth observation foundation models can provide transferable sub-grid surface representations, enhancing probabilistic weather downscaling accuracy.
2
Integrating a compressed Tessera embedding into a ConvCNP consistently improves point and probabilistic skill for 2m temperature and 10m wind speed across diverse climatic regions.
3
The proposed method exhibits robustness to forecast fields and significantly improves sample efficiency, particularly for wind speed, by providing valuable surface information in data-scarce environments.

Conceptual Flow

HIGH LEVEL

Methodology: Enhancing Local Weather Prediction

The paper's method combines big-picture weather data with a special 'surface code' to make super-accurate local weather predictions.

Big Weather Grid

Local Surface Code

Combine & Learn

Accurate Local Forecast

Results: Better Predictions, Especially for Wind

Adding the 'surface code' makes local weather predictions much better, especially for wind, even in places with little past data.

Old Local Forecast

New Local Forecast

Compare Accuracy

Much Better Wind

Better Temperature

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.