Comparison of methods for resolving the contributions of local emissions to measured concentrations

Abstract

To accurately study the characteristics of an air pollution emitter, it is necessary to isolate the contribution of that emitter to total measured pollution concentrations. A variety of published methods exist to complete this task, like placing measurements upwind the emitter, employing a distant background measurement station, or algorithmic methods that extract a background from the time series of measured concentrations (e.g. wavelet decomposition). In this study, we measured nitrogen oxides (NOx), carbon monoxide (CO), carbon dioxide (CO2), and ne particulate matter (PM2.5) at four sites spanning Toronto, Ontario, Canada. We rst characterized the spatial variability of background concentrations across the city and then tested the accuracy of seven different algorithmic methods of estimating true measured upwind-of-emitter backgrounds near Toronto’s Highway 401 by using the data collected at a downwind site. These methods included time-series and regression methods, including machine learning (XGBoost). We observed background concentrations had notable spatial variability, except for PM2.5. When predicting backgrounds upwind the highway, we found a distant measurement station provided an accurate background only during some times of day and was least accurate during rush hours. When testing algorithmic predictions of upwind-of-highway backgrounds, we found that regression models surpassed the performance of time-series methods, with best predictions having R2 exceeding 0.8 for all four pollutants. Despite the better performance of regression models, time-series methods still provided reasonable estimates. We also found that emitter-specic covariates (e.g. trafc counts, on-site dispersion modelling) did not play an important role in regressions, suggesting backgrounds can be well characterized by time of day, meteorology, and distant measurement stations. Based on our results, we provide ranked recommendations for choosing background estimation methods. We suggest future air pollution research characterizing individual emitters includes careful consideration of how background concentrations are estimated.