# Interpretable downscaling of local weather predictions

*2026-08-01 — idea*


Weather reanalyses and forecasts are usually plotted on grids of ~25km, but the
people's decisions about weather-related actions happen much more locally\!
Within a single coarse grid cell the land surface isn't uniform, so the weather
at a given spot can depart from the cell average quite significantly in ways
that terrain and land cover largely determine.

We've shown that a general purpose Earth observation embedding ([Tessera](https://anil.recoil.org/projects/tessera) in our case) closes some of
that gap and enables much more accurate local forecasting. In [Earth observation embeddings are effective sub-grid descriptors for probabilistic weather downscaling](https://anil.recoil.org/papers/2026-weather-downscaling) we compressed a patch of
embeddings at 10m into a per-location latent and gave it to
a convolutional conditional neural process downscaling ERA5.

Across five climatically diverse regions and 9000+ stations held out in both space and
time, this improved CRPS skill by 11.5% for 2m temperature and 6.2% for 10m
wind speed. We have ablations and hypotheses about why this happens. Topography explains more of
the sub-grid structure of temperature prediction variability, while the embedding itself carries surface
information that matters more for wind (e.g. is there a canyon?). I've jotted down
[an explainer](https://anil.recoil.org/notes/weather-downscaling-tessera) with more of our thoughts about the paper results.

So we know using satellite embeddings helps forecastign, and not much about _why_ they do. When it
comes to forecasting, it would be ideal if we know more about the causative factors behind
particular temperature variations, in case it helps with climate adaptation efforts.
This project aims to help figure out the "why" behind some of this data-driven prediction magic.
You would first investigate and pick one of these; the difficulty level here is fairly high.

## Which parts of the embedding carry the downscaling skill?

*Hypothesis:* The weather prediction skill rests on only a few interpretable surface properties, and
the rest of the Tessera embedding is inert for a given task like temperature or wind prediction.

The interpretability experiments in our paper looked only at the VAE-encoded
patch latent, and not at the original embedding features it was compressed from. We
therefore know a little about the summary and nothing about what survived the
compression.

An interprebility experiment would be to attribute skill back to the raw
per-pixel features, and to ask whether the embedding dimensions that are significant correspond to
anything that a person would recognise. These might include canopy structure,
built roughness, water or slope aspects.  This may work particularly well with
[Tessera v2](https://anil.recoil.org) which has Matryoshka embeddings whose first dimensions
pack in more information than later ones.

## Where should the next weather station go?

*Hypothesis:* The embedding space knows which places a network already represents, so it should also know where a _new_ station would provide the most prediction gain for a larger region.

This builds on Rich Turner's [environmental sensor placement work](https://arxiv.org/pdf/2211.10381) about figuring out how to get the best coverage of a large (and often remote) region.
Consider the Norwegian network deployment experiment in the paper. If a given station teaches the downscaler about the surfaces "near it" in embedding space, then station coverage is dependent on which representative surfaces we have instrumented, rather than of the overall surface geography.

What if we instead make this an objective; i.e., *"where do I put X stations to guarantee Y level of downscaling skill over this region?"*. We could then test how close a greedy or submodular selection over the embedding space gets to this goal. We're trying to make all stations "reachable" within the embedding space based on which representative surfaces we deploy at, thus maximising the overall surface coverage.

## Who this suits

This is a good project for an MPhil student who wants to get into data-driven
earth systems modelling. You would work alongside [Pedro Sousa](https://www.linkedin.com/in/pedro-marques-sousa) who is doing a PhD
in this space, and many of the models and data are already available from his
experiments.  Comfort and interest with probabilistic modelling will help a
lot, but you can pick this up as you go along too.
Status: Available
Level: MPhil
Year: 2026
Project: TESSERA, a pixelwise geospatial foundation model
Supervisors: Pedro Sousa, Sadiq Jaffer, Anil Madhavapeddy

## Related

- [Improving local weather forecasts using Tessera embeddings](https://anil.recoil.org/notes/weather-downscaling-tessera) (note, 2026-08-13)
- [Earth observation embeddings are effective sub-grid descriptors for probabilistic weather downscaling](https://anil.recoil.org/papers/2026-weather-downscaling) (paper, 2026-08-01)
- [TESSERA, a pixelwise geospatial foundation model](https://anil.recoil.org/projects/tessera) (project, 2025-01-01)

---
Canonical: https://anil.recoil.org/ideas/tessera-interpretable-downscaling
Type: idea
Tags: tessera, climate, ai, spatial