# From satellites to the streets: Earth embeddings for local environmental decisions *2026-10-01 — idea* Status: Available Level: PhD Year: 2026 Project: TESSERA, a pixelwise geospatial foundation model Supervisors: David Topping, Sadiq Jaffer, Anil Madhavapeddy Nitrogen dioxide levels beside a busy road can be much higher than in a park a few hundred metres away. Despite this, the [national forecasts](https://uk-air.defra.gov.uk/forecasting/) used to warn the public use grid squares several _kilometres_ across. Long-term exposure to air pollution is linked to the [equivalent of around 30,000 deaths a year](https://www.gov.uk/government/publications/air-pollution-applying-all-our-health/air-pollution-applying-all-our-health) in the UK. Much of the decisions regarding such health issues are local ones; where to place an air monitor, when to warn people with asthma, or whether a low-emission zone has worked as intended. AI models trained on satellite and weather data should now be able to predict conditions street by street and show how far each prediction can be trusted. This project builds on two recent results. At Manchester, [David Topping](https://research.manchester.ac.uk/en/persons/david.topping) and his group have retrained generative AI models first built for weather on a year of hourly chemistry model output. They give UK-wide pollution maps with grid cells of 2–3 square kilometres and [can run on a desktop AI computer](https://blogs.nvidia.com/blog/uk-air-pollution-research-earth-2/). At Cambridge, [Tessera](https://anil.recoil.org/projects/tessera) condenses a year of Sentinel-1 and Sentinel-2 satellite data into a 128-number description of every 10 m pixel of the Earth's land surface. Our recent work on [weather downscaling](https://anil.recoil.org/papers/2026-weather-downscaling) showed that these embeddings make probabilistic weather downscaling more accurate at stations and times the model has never seen, with temperature gains generally growing as stations become sparser (see the [explainer](https://anil.recoil.org/notes/weather-downscaling-tessera)). This project will use Earth embeddings to add local detail and reliable estimates of uncertainty to environmental models, starting with air pollution in the UK. A first step is to add embedding-based descriptors to probabilistic and generative pollution models, trained on data from regulatory monitors and low-cost sensor networks, and to find out which pollutants benefit and why. From there, the work can go in several directions depending on the student's interests. On the methods side, this could mean models that give joint uncertainty across a whole city, or that use embeddings from different years to track change. On the decision side, the same models can be used to choose where a new sensor would reduce uncertainty most, or to judge when a forecast is confident enough to trigger a health alert. They can also be moved to a new city with few monitors. Embeddings can also find comparable places to test whether interventions, such as low-emission zones or new green space, changed local air quality. Beyond pollution, the same approach applies to other problems shaped by the land surface, such as urban heat, pollen and wildfire risk. One of our key goals is to make the results usable by people who are not modelling experts, such as council officers and clinicians. We plan to build tools that explain each prediction and how far it can be trusted. In the longer term we aim for an open workflow that any city can use to build its own detailed environmental models from public satellite data and a small number of local measurements. ## References - P. Sousa et al., "[Earth observation embeddings are effective sub-grid descriptors for probabilistic weather downscaling](https://anil.recoil.org/papers/2026-weather-downscaling)", [arXiv:2608.12271](https://arxiv.org/abs/2608.12271), 2026. - Z. Feng et al., "[TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis](https://anil.recoil.org/papers/2025-tessera)", [Proc. IEEE/CVF CVPR](https://openaccess.thecvf.com/content/CVPR2026/html/Feng_TESSERA_Temporal_Embeddings_of_Surface_Spectra_for_Earth_Representation_and_CVPR_2026_paper.html), pp. 34818–34831, 2026. - I. Salian, "[University of Manchester Uses NVIDIA Earth-2 to Forecast Air Pollution Across the UK](https://blogs.nvidia.com/blog/uk-air-pollution-research-earth-2/)", NVIDIA Blog, September 2026. ## Related - [Improving local weather forecasts using Tessera embeddings](https://anil.recoil.org/notes/weather-downscaling-tessera) (note, 2026-08-13) New paper showing that a frozen TESSERA embedding improves probabilistic weather downscaling, improving local temperature and wind predictions at never-observed sites. - [Earth observation embeddings are effective sub-grid descriptors for probabilistic weather downscaling](https://anil.recoil.org/papers/2026-weather-downscaling) (paper, 2026-08-01) Pedro Sousa, Will Tebbutt et al. - [TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis](https://anil.recoil.org/papers/2025-tessera) (paper, 2026-06-01) Zhengpeng Feng, Clement Atzberger et al. — Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) - [TESSERA, a pixelwise geospatial foundation model](https://anil.recoil.org/projects/tessera) (project, 2025-01-01) --- Canonical: https://anil.recoil.org/ideas/tessera-street-air-quality Type: idea License: CC BY 4.0 Tags: tessera, climate, ai, spatial