Researchers from Duke University used machine learning, satellite imaging and weather data to autonomously find hotspots for heavy air pollution, down to the exact city block. This new system could help researchers and policymakers find and mitigate sources of hazardous aerosols, study the effects of air pollution and make informed policy decisions.
In the past, gathering air pollution data was limited by the number of monitors available, and researchers had to drive sensors around the city in cars to detect air pollution. This procedure is unnecessarily time-consuming, expensive and limited because driving around only gathers data on roads that are major pollution sources.
Hot spots are typically hard to find because the air quality of an entire city varies day by day. Finding an exact neighborhood with higher levels is important because it helps answer questions about health disparities and environmental fairness.
A new AI algorithm picked out these city-block-sized satellite images as local hotspots (top) and cool spots (bottom) for air pollution in Beijing. Source: Tongshu Zheng, Duke University
During this study, the team specifically surveyed particulate matter with diameters less than 2.5 micrometers (PM2.5). Areas with higher PM2.5 levels were associated with higher COVID-19 death rates, according to a Harvard study. Currently, the only data available on PM2.5 is on a county by county basis, which is not specific enough to find problem areas.
In a previous study, satellite imagery, weather data and machine learning could provide PM2.5 measurements on a small scale. For their study, the team improved these methods and taught the algorithm to automatically find hot and cool spots of air pollution with a resolution of 300 m.
The algorithm can estimate levels of PM2.5 with weather data and measures the difference between estimates and actual PM2.5 levels. The algorithm teaches itself to use satellite images to make better predictions and searches for city block-sized pixels with higher or lower levels of PM2.5.
Eventually, the algorithm could teach itself new methods in different locations. It can evolve with the cities and environments it is analyzing. The team’s next step is to see how hotspots are related to socioeconomic status and hospital admittance rates from long-term exposure.
This study was published in Remote Sensing.
