Getting your Trinity Audio player ready...
|
This summer, Northeastern Americans faced heightened air quality concerns due to thick wildfire smoke containing health-risky fine particulate matter (PM 2.5). A Penn State-led team used AI and mobility data to develop improved PM 2.5 exposure models for public health strategies.
In light of this, the researchers analysed PM 2.5 measurements in eight US metropolitan areas, using data from EPA monitoring stations and local community-distributed low-cost sensors to calculate hourly PM 2.5 averages. They then fed this data into a land use regression model that considered geographical factors like aerosol optical depth, proximity to roads or waterways, elevation, vegetation, and meteorological conditions.
These factors are examined to understand their impact on air quality. It’s noteworthy that prior models took a linear approach to assess air pollution, assigning a fixed level of importance to each geographical factor and its influence on air quality. However, as Yu clarified, certain factors like vegetation and meteorological conditions exhibit hourly or seasonal variations. They may interact with other elements that impact air quality, rendering a linear approach inadequate.
Yu and her research team employed an innovative nonlinear strategy to tackle the intricate and ever-changing factors that influence air pollution exposure. They integrated automated machine learning, a type of artificial intelligence capable of independently managing labour-intensive tasks like data preparation, parameter selection, and model deployment, into their land use regression model.
This approach leveraged an ensemble method, enabling the machine to run and consolidate multiple models to determine the most effective one for each region. Additionally, they analysed anonymised cell phone mobility data to identify areas with poor air quality and high visitor volumes.
The process began with collecting diverse datasets, including information from low-cost sensors, EPA monitoring stations, and anonymised cell phone mobility data. These datasets provided essential insights into air quality, weather conditions, and human movement patterns.
AutoML then took charge of model selection, identifying the most suitable algorithms for predicting air pollution levels. Feature selection followed, with the system pinpointing critical variables like aerosol optical depth and meteorological factors. Machine learning models were trained on the data, learning the intricate relationships between input features and air pollution levels.
Moreover, integrating mobility data into the models allowed for identifying high PM 2.5 exposure areas and creating an alert system to notify individuals about unhealthy air quality, empowering them to take necessary precautions. This comprehensive approach demonstrates how machine learning can substantially enhance the accuracy and practicality of air quality models, offering valuable insights for public health strategies.
Yu elaborated, “Numerous regions may consistently grapple with elevated air pollution levels, particularly those near factories and major transportation hubs. Yet, this information alone does not suffice to prioritise locations requiring additional monitoring or health alerts.”
Their exposure maps, grounded in mobility data, provide a more comprehensive perspective by spotlighting areas characterised by both subpar air quality and significant visitor traffic. This dataset can then be harnessed to transmit alerts to individuals’ mobile devices as they enter regions with exceptionally high PM 2.5 levels, allowing them to curtail their exposure to unhealthy air quality and ultimately bolstering public health efforts.