Pedro Sousa began his PhD in January 2026 working on what a geospatial foundation model can contribute to near-surface weather prediction. This is joint work with the Aardvark group in the Department of Engineering, with Richard Turner and Will Tebbutt, whose end-to-end data-driven weather prediction system supplies the downscaling machinery that we extend.
Weather reanalyses and forecasts are currently produced on grids of ~25km, but most downstream decisions need the near-surface state at a specific point. Within a single grid cell the land surface is far from uniform and so the resulting sub-grid variation is large and structured rather than incidental. For example, urban heat islands vary by 1–3 °C, cold-air drainage can leave valley floors 4–8 °C colder than slopes nearby, and wind accelerates through terrain gaps while slowing over forest canopy and built roughness.
This project explores how a pre-trained Earth observation embedding is a valid complementary sub-grid descriptor to improve the accuracy of weather downscaling. A single TESSERA pixel describes only a single 10m spot, so we will need to use more sophisticated downstream heads (e.g. a variational autoencoder) to compress a patch of per-pixel embeddings into a compact per-location latent.
The goal is to use this resulting embedding to improve both point skills (MAE, RMSE) and also calibrated probabilistic skill (NLL, CRPS) over ERA5 interpolation, persistence, and a matched-capacity model that is used without the TESSERA embedding, and to work across climatically diverse regions.
The motivating application for this project is that decisions about energy, agriculture and hazard are all taken at local points rather than on large grids. Later work in the PhD is likely to extend to joint multi-variable modelling, precipitation, and other embeddings for comparison.
1 Progress
August 2026: the first preprint from this project is now out: "Earth observation embeddings are effective sub-grid descriptors for probabilistic weather downscaling" (also on arXiv).
Across five climatically diverse regions and over 9,000 stations, adding a frozen Tessera embedding improved probabilistic skill by ~11.5% for 2m temperature and ~6.2% for 10m wind speed, and the advantage holds when downscaling Aurora forecasts out to 72 hours ahead. I wrote up an explainer of the results as well. Precipitation and time-indexed embeddings are next.
2 See also
Pedro's earlier work with us on Reverse emulating agent-based models for policy simulation for his Masters, and the broader TESSERA programme this builds on.