Cornell team develops computationally efficient machine learning models for hyperlocal models of PM2.5 concentrations
Cornell engineers have developed machine learning models to simplify and reinforce models to calculate the fine particulate matter (PM2.5) contained in urban air pollution. Described in a paper in the journal Transportation Research Part D: Transport and Environment, the modeling approach has low data requirements and is computationally efficient.
Previous methods to gauge air pollution were cumbersome and reliant on extraordinary amounts of data points.
Older models to calculate particulate matter were computationally and mechanically consuming and complex. But if you develop an easily accessible data model, with the help of artificial intelligence filling in some of the blanks, you can have an accurate model at a local scale.—Senior author Oliver Gao, the Howard Simpson Professor of Civil and Environmental Engineering in the College of Engineering
In this work, the group developed four machine learning models for traffic-related particulate matter concentrations in data gathered in New York City’s five boroughs, which have a combined population of 8.2 million people and a daily-vehicle miles traveled of 55 million miles.
The equations use few inputs such as traffic data, topology and meteorology in an AI algorithm to learn simulations for a wide range of traffic-related, air-pollution concentration scenarios.
The ML model’s overall performance in predicting air pollution concentrations at receptors outperforms previous methods. Our best model, Convolutional Long Short-Term Memory (ConvLSTM), has a Mean Relative Error (MRE) of 38.9%, which is lower than the 47.5% MRE for the single hidden layer model, the 63.2% MRE for the Convolutional Neural Network model, and the 41.5% MRE for ConvLSTM with time-series data. Memory cells help the ConvLSTM model predict a large number of spatially correlated observations.—Desai et al.
Instead of focusing on stationary locations, the method provides a high-resolution estimation of the city street pollution surface. Higher resolution can help transportation and epidemiology studies assess health, environmental justice and air quality impacts.
Funding for this research came from the US Department of Transportation’s University Transportation Centers Program and Cornell Atkinson.
Salil Desai, Mohammad Tayarani, H. Oliver Gao (2022) “Developing Machine learning models for hyperlocal traffic related particulate matter concentration mapping,” Transportation Research Part D: Transport and Environment, Volume 113, 103505, doi: 10.1016/j.trd.2022.103505