University of Pretoria: Dangerous air pollution detected with machine learning
Researchers found high levels of air pollution in areas of Mpumalanga and Gauteng. UP’s Prof Rebecca Garland helped create an artificial intelligence tool to monitor and predict daily air pollution against national air quality standards.
She was part of an international research team that fed data into a smart computer model to predict PM2.5 levels where there are blind spots in the available data. PM2.5 refers to particles less than 2.5 micrometres in size emitted by polluting industries like coal-fired power plants and many other sources in the atmosphere.
These particles can cause severe damage to the lungs. According to the World Health Organisation (WHO), PM2.5 particles have been linked to 4.2 million premature deaths worldwide. Since so many different sources emit PM2.5 where people live, it is crucial to make sure PM2.5 levels don’t exceed healthy levels.
The researchers used satellite, weather, and land use data to develop a highly accurate machine learning model that estimates daily PM2.5 levels in Gauteng and Mpumalanga. The team used their advanced machine learning model to fill the gaps left by the sparse ground monitors.
“When you want to know about pollution and its impacts on health, you want to understand how this pollution is distributed in space,” says Garland, who is based in UP’s Department of Geography, Geo-informatics and Meteorology. She brought her atmospheric science experience to the international team of researchers at Emory University in the USA.
“One way of doing this is a machine learning approach to integrate multiple data streams to map out pockets of air pollution.”
There are 130 ground stations that monitor air quality all over South Africa, but these are not enough to get an accurate picture of daily PM2.5 levels. Most of these stations also do not have adequate PM2.5 data, creating even larger areas where air quality isn’t monitored properly.
Their model successfully mirrored the real world’s seasonal trends and used them to predict daily PM2.5 levels. The model identified that many areas across Gauteng and neighbouring provinces have high levels of PM2.5, with high levels seen in some low-income settlements.
“There are PM2.5 sources within low-income settlements and we see high particulate matter levels in winter in the highveld region,” she says. “So, part of this increase is the use of solid fuels for cooking and for heating, which of course, gets higher in the winter.”
The model showed reduced PM2.5 levels in Tshwane from 2016 to 2018, but worryingly, they have stayed the same in Johannesburg and surrounding areas.
This research, while specific to Gauteng, has opened the door for governments all over the world to use this method to track PM2.5 for their own areas. In addition, this method can help create and adequately monitor policies that manage the risk associated with high PM2.5 levels.