Residents in the Northeastern United States closely monitored air quality warnings this summer as wildfire smoke cast an orange haze across the skies. Airborne pollutants, including minute particles known as fine particulate matter (PM 2.5), raised concerns due to their potential health risks when inhaled, particularly for individuals with pre-existing heart and lung conditions.
To evaluate PM 2.5 exposure and assist public health authorities in formulating effective strategies, a research team led by Penn State has developed enhanced models utilizing artificial intelligence and mobility data.
“Our findings demonstrate that integrating artificial intelligence and mobility data into air quality models enhances their accuracy. This aids decision-makers and public health officials in identifying areas requiring additional monitoring or safety alerts due to poor air quality or a combination of poor air quality and high pedestrian activity,” explained Manzhu Yu, an assistant professor of geography at Penn State and the study’s lead author.
Published in the journal Frontiers in Environmental Science, the study focused on PM 2.5 measurements in eight major metropolitan areas in the continental United States. Air quality data originated from Environmental Protection Agency (EPA) monitoring stations and affordable sensors typically acquired and distributed by local community groups. The researchers used this data to calculate hourly averages of PM 2.5 in each region.
The scientists incorporated the air quality data into a land use regression model, considering local geographical factors like aerosol optical depth measured by satellites, proximity to roads or streams, elevation, vegetation, and meteorological conditions such as humidity and wind speed. Unlike previous models that followed a linear approach, assigning fixed importance to each geographic factor, Yu and her colleagues opted for a nonlinear approach. This approach, incorporating automated machine learning, considers dynamic factors that change hourly or seasonally and may have intricate interactions affecting air quality.
The automated machine learning technique employed an ensemble method, allowing the machine to run and combine multiple models to identify the most effective model for each region. The researchers analyzed anonymized cell phone mobility data to identify areas with both poor air quality and high visitor numbers.
The study revealed that the automated machine learning method, coupled with data from low-cost sensors and EPA monitoring stations, enhanced the accuracy of air pollution exposure models by an average of 17.5%. This improvement offered greater spatial variation compared to relying solely on regulatory monitors.
Yu attributed the increased accuracy to the method’s ability to better consider dynamic variables such as aerosol optical depth and meteorological factors, which consistently emerged as crucial across all study regions. The inclusion of mobility data enabled the team to identify potential hotspots within regions and times of the day and year when large populations might encounter high PM 2.5 levels.
While some areas may consistently exhibit elevated air pollution levels, like those near industrial facilities and major transportation hubs, this information alone isn’t sufficient to prioritize places for additional monitoring or health alerts, Yu noted. The mobility-based exposure maps generated by the study provide public health officials and decision-makers with insights into hotspots featuring both unhealthy air quality and high visitor traffic. This information can be utilized to send alerts to individuals’ mobile phones when entering areas with notably high PM 2.5 levels, reducing their exposure to unhealthy air quality.