A new air-quality mapping project aims to bring pollution estimates down to a 100-meter grid and update them hour by hour. Developed in research by Andrea Porcheddu at the University of Eastern Finland in Kuopio, the system combines satellite observations with model data and readings from ground stations.
Air pollution can harm health and contributes to deaths around the world, but many places lack enough monitoring stations to show where pollution is worst. Existing stations are sparse and are mostly found in developed countries. Satellite observations offer a way to widen coverage, while ground measurements can provide information about conditions at the surface. Putting those sources together could give officials a clearer picture of pollution and help them decide where action is most needed.
Porcheddu's doctoral research developed a framework to estimate air quality by combining global model products with satellite data. Global models are available across large areas and often, but their maps have relatively coarse detail. Satellites can provide finer spatial information, though how often they collect data depends on the type of satellite. The research used polar-orbiting satellites, which pass over an area once a day, and geostationary satellites, which observe at intervals of 10 minutes.
The team first used an ensemble of multilayer perceptrons, a kind of neural network, to estimate ground-level PM2.5 when polar-orbiting satellites passed overhead. PM2.5 refers to very small airborne particles that can be breathed into the lungs. Researchers then used a 3D U-Net, another type of neural network, to combine those satellite-based maps with model information and ground-station readings. The resulting framework is designed to produce hourly PM2.5 maps on a grid with cells 100 meters across.
That level of detail could help reveal differences between nearby parts of a city that a broad regional estimate might miss. The work does not mean every location already has a new monitoring system in place. Rather, it offers a method for turning several incomplete sources of information into maps that may help policymakers understand urban pollution more closely. Satellite data could also help fill gaps in regions with few ground stations, although the framework still relies on data sources that vary in coverage and timing.
Porcheddu also took an initial step toward using geostationary satellites. The research combined observations from GOES-East and GOES-West across the contiguous United States, using a transformer, a neural-network design often used to find patterns across data. This work points toward a way to make use of satellites that observe frequently, alongside those that offer different kinds of coverage.
Porcheddu's dissertation, titled “From space to street level: Deep learning-enhanced air quality monitoring using satellite data,” is to be examined at the Faculty of Science, Forestry and Technology on the Kuopio campus. The next challenge is to develop and assess these methods further, so that frequent, fine-scale estimates can become a more useful part of how communities track polluted air and plan responses.