Traffic-related air pollution backcasting using convolutional neural network and long short-term memory approach

Abstract

Air pollution backcasting, especially nitrogen dioxide (NO2), is crucial in epidemiological studies, thus enabling the reconstruction of historical exposure levels for assessing long-term health effects. Changes in NO2 concentrations in urban areas are typically influenced by vehicle composition, technology, and traffic volumes. However, the observed NO2 levels at a monitoring site also reflect contributions from other sources, such as industrial and regional backgrounds. This study proposes a model that captures the spatial variability of NO2 concentrations, incorporating temporal trends through traffic-related predictors like nitrogen oxides (NOx) emissions and Annual Average Daily Traffic (AADT). Our approach integrates a Convolutional Neural Network (CNN) for spatial variation and Long Short-Term Memory (LSTM) for long-term temporal dynamics, yielding optimal spatiotemporal predictions for NO2 levels across the City of Toronto, Canada. The model, trained with NO2 measurements collected via the Urban Scanner mobile platform in 2020 and 2021, utilizes a Traffic Emission Prediction scheme (TEPs) to develop NOx and AADT inventories, serving as input to the LSTM model. Our proposed approach successfully estimates traffic-related NO2 levels across Toronto from 2006 to 2020. By intersecting the backcasted levels with census data, we noted that despite an overall decrease in NO2 levels between 2006 and 2020, disparities in exposure grew as more marginalized communities faced environmental injustice.