Researchers have built a transformer-based model that turns coarse air quality forecasts into neighborhood-level pollution maps accurate to roughly one kilometer.
The system takes atmospheric composition data from the Copernicus Atmosphere Monitoring Service, then layers in satellite aerosol readings, land cover, elevation, wind fields, and human activity indicators to produce PM2.5 estimates at 40 times the original spatial resolution. The hard part was supervision: ground-level air quality sensors are sparse and scattered, so the team developed a propagation strategy that blends interpolated readings from the OpenAQ network using spatial Gaussian smoothing rather than requiring dense, temporally aligned observations. The result corrects known biases in the underlying CAMS data while recovering fine-grained pollution structure the source model smooths over.
At stake is the difference between knowing a city has bad air and knowing which neighborhoods bear the worst of it — information that matters for public health policy, urban planning, and environmental justice. Most operational air quality models operate at resolutions of tens of kilometers, which obscures the hyperlocal variation that tracks closely with where people actually live and work.
The approach sidesteps a design choice common in similar work — leaning on time-series modeling — which limits data requirements and could make deployment more practical; whether it holds up outside Europe remains the obvious next test.