AI pinpoints local pollution hotspots via satellite images
Image credit: Dmitrii Melnikov/Dreamstime
Researchers in the US have developed a method that uses machine learning, satellite imagery, and weather data that can autonomously identify hotspots of heavy air pollution, city block by city block.
According to its developers from Duke University, the technique could help in finding and mitigating sources of hazardous aerosols, studying the effects of air pollution on human health, and making better informed, socially just public policy decisions.
“Before now, researchers trying to measure the distribution of air pollutants throughout a city would either try to use the limited number of existing monitors or drive sensors around a city in vehicles,” said Mike Bergin, professor of civil and environmental engineering at Duke. “But setting up sensor networks is time-consuming and costly, and the only thing that driving a sensor around really tells you is that roads are big sources of pollutants. Being able to find local hotspots of air pollution using satellite images is hugely helpful.”
Bergin and his colleagues focused their research on specific air pollutants – tiny airborne particles called PM2.5. These are particles that have a diameter of less than 2.5 micrometres, about 3 per cent of the diameter of a human hair. They are known to affect human health because of their ability to travel deep into the lungs.
Research led by the Institute for Health Metrics and Evaluation (IHME), called the ‘Global Burden of Disease Study’ (GBD), ranked PM2.5 fifth on its list of mortality risk factors in 2015. The study showed that PM2.5 was responsible in one year for about 4.2 million deaths and 103.1 million years of life lost or lived with disability.
Meanwhile, a recent study from Harvard University found a link between areas with higher PM2.5 levels and higher death rates because of Covid-19. But the Harvard researchers stressed they could only access PM2.5 data on a county-by-county level within the US. While a valuable starting point, county-level pollution statistics can’t drill down to a neighbourhood next to a coal-fired power plant versus one next to a park that is 30 miles upwind, the researchers said. And most countries outside of the Western world don’t have that level of air quality monitoring.
“Ground stations are expensive to build and maintain, so even large cities aren’t likely to have more than a handful of them,” Bergin explained. “So while they might generally indicate the amount of PM2.5 in the air, they don’t come anywhere near giving a true distribution for the people living in different areas throughout that city.”
In previous work with doctoral student Tongshu Zheng and colleague David Carlson, assistant professor of civil and environmental engineering at Duke, the researchers showed that satellite imagery, weather data and machine learning could provide PM2.5 measurements on a small scale. And building off that work and focusing on Delhi, the team said they have now improved their methods and taught the algorithm to find hotspots and cool spots of air pollution with a resolution of 300m.
For the study, the researchers used a technique called residual learning. Here, the algorithm first estimates the levels of PM2.5 using weather data alone. It then measures the difference between these estimates and the actual levels of PM2.5 and teaches itself to use satellite images to improve its predictions.
The researchers then used an algorithm initially designed to adjust uneven illumination in an image to find areas of high and low levels of air pollution. The technique, called local contrast normalisation, looks for city-block-sized pixels that have higher or lower levels of PM2.5 than others in their vicinity.
“These hotspots are notoriously difficult to find in maps of PM levels because some days the air is just terrible across the entire city, and it is really difficult to tell if there are true differences between them or if there’s just a problem with the image contrast,” said Carlson. “It’s a big advantage to find a specific neighbourhood that stays higher or lower than everywhere else, because it can help us answer questions about health disparities and environmental fairness.”
While the exact methods the algorithm teaches itself can’t transfer from city to city, the researchers said the algorithm could easily teach itself alternative methods in different locations. The researchers also said the number of air quality sensors is only going to increase in coming years and believe their approach will only get better with time.
“I think we’ll be able to find built environments in these images that are related to the hot and cool spots, which can have a huge environmental justice component,” said Bergin. “The next step is to see how these hotspots are related to socioeconomic status and hospital admittance rates from long-term exposures.”
Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.