A comprehensive study of the COVID-19 impact on PM2.5 levels over the contiguous United States: A deep learning approach

Abstract

We investigate the impact of the COVID-19 outbreak on PM_2.5 levels in eleven urban environments across the United States: Washington DC, New York, Boston, Chicago, Los Angeles, Houston, Dallas, Philadelphia, Detroit, Phoenix, and Seattle. We estimate daily PM_2.5 levels over the contiguous U.S. in March-May 2019 and 2020, and leveraging a deep convolutional neural network, we find a correlation coefficient, an index of agreement, a mean absolute bias, and a root mean square error of 0.90 (0.90), 0.95 (0.95), 1.34 (1.24) μg/m³, and 2.04 (1.87) μg/m³, respectively. Results from Google Community Mobility Reports and estimated PM_2.5 concentrations show a greater reduction of PM_2.5 in regions with larger decreases in human mobility and those in which individuals remain in their residential areas longer. The relationship between vehicular PM_2.5 (i.e., the ratio of vehicular PM_2.5 to other sources of PM_2.5) emissions and PM_2.5 reductions (R = 0.77) in various regions indicates that regions with higher emissions of vehicular PM_2.5 generally experience greater decreases in PM_2.5. While most of the urban environments ⸺ Washington DC, New York, Boston, Chicago, Los Angeles, Houston, Dallas, Philadelphia, Detroit, and Seattle ⸺ show a decrease in PM_2.5 levels by 21.1%, 20.7%, 18.5%, 8.05%, 3.29%, 3.63%, 6.71%, 4.82%, 13.5%, and 7.73%, respectively, between March-May of 2020 and 2019, Phoenix shows a 5.5% increase during the same period. Similar to their PM_2.5 reductions, Washington DC, New York, and Boston, compared to other cities, exhibit the highest reductions in human mobility and the highest vehicular PM_2.5 emissions, highlighting the great impact of human activity on PM_2.5 changes in eleven regions. Moreover, compared to changes in meteorological factors, changes in pollutant concentrations, including those of black carbon, organic carbon, SO₂, SO₄, and especially NO₂, appear to have had a significantly greater impact on PM_2.5 changes during the study period.

Keywords: COVID-19; Community multiscale air quality (CMAQ) model; Deep convolutional neural network; Google mobility reports; PM2.5 estimation; United States.

Graphical abstract

Fig. 1

Map of the CONUS (study areas). Pink dots represent the location of the eleven urban environments analyzed in this study. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Fig. 2

Scatterplots of the ten-fold cross-validation (10-CV) results showing the performance of the deep convolutional neural network (Deep-CNN) at estimating surface concentrations of PM_2.5 over the CONUS from March to May 2019 (a) and 2020 (b). The R, IOA, and the MAB refer to the correlation coefficient, index of agreement, and the mean absolute bias, respectively.

Fig. 3

The spatial distribution of mean estimated PM_2.5 levels over the CONUS from March to May 2019 (top) and 2020 (bottom). The circles show the mean observed PM_2.5 concentrations from the EPA stations during the study period, and the background maps represent the estimated PM_2.5 levels using deep learning (DL).

Fig. 4

Scatterplot showing the relationship between percentage changes of PM_2.5 between March–May of 2020 and 2019 and vehicular PM_2.5 (i.e., the ratio of vehicular PM_2.5 to other sources of PM_2.5) in the eleven urban environments over the CONUS, including Washington DC (DC), New York (NY), Boston (BO), Chicago (CH), Los Angeles (LA), Houston (HO), Philadelphia (PHI), Detroit (DE), Phoenix (PH), Dallas (DA), and Seattle (SE). The Upper CI and Lower CI refer to the upper and lower limits of the 95% confidence interval, respectively.