A fuel-based method for updating mobile source emissions during the COVID-19 pandemic

We follow the methodology developed by McDonald et al (2014) to update the spatial mapping of the fuel-based inventory for vehicle emissions (FIVE). The most recent national traffic count data are available at a roadway-link level from the highway performance monitoring system (HPMS) for the year 2018 (FHWA 2020a), which spatially allocate ∼82% and ∼71% of LD and HD traffic, respectively, when compared with state-level traffic statistics (FHWA 2020b). In the US, gasoline is consumed primarily by LD vehicles and diesel by HD vehicles. Therefore, we use LD traffic data to downscale annual state-level taxable gasoline sales (FHWA 2020b), and HD traffic for diesel sales. Untaxed diesel consumption by buses is also included (Davis and Boundy 2020). The 2018 fuel use maps are then projected to 2019 (BAU fuel is assumed unchanged from 2019 to 2020) using the state-level taxable fuel sales records of gasoline and diesel. For simplicity, the balance of traffic unaccounted by HPMS traffic count data is spatially apportioned using population as a surrogate (EPA 2020). While distributing on-road emissions by population density is known to introduce bias (Gately et al 2015), this is diminished by the large portion of fuel consumption already distributed at the roadway link level. The result is a gridded 4 km × 4 km contiguous US emissions map of on-road fuel consumption, separately apportioned between LD and HD traffic.

Off-road gasoline engines consist of small two- and four-stroke gasoline engines (e.g. lawn equipment, generators, forklifts, recreational vehicles, watercraft), consuming ∼6% of total gasoline (FHWA 2020b). Though this consumption is relatively small, it can be a significant source of CO and VOCs due to high emission factors relative to on-road transportation (Gordon et al 2013). Off-road diesel engines are used in heavy equipment (e.g. agricultural, construction, other industrial/commercial engines), and are important sources of NO_x and PM (Dallmann and Harley 2010). Off-road diesel fuel use is estimated following Kean et al (2000), utilizing end-use surveys reported by the Energy Information Administration (EIA) (EIA 2019). Off-road fuel use is mapped utilizing spatial and temporal surrogates from the National Emissions Inventory (NEI) 2011 (EPA 2014) for two- and four-stroke gasoline equipment, small watercraft, and agricultural and non-agricultural diesel engines (McDonald et al 2018a). While there are more recent versions of the NEI, spatial and temporal surrogates are not expected to change significantly until the US 2020 Census is complete.

2.3.1. Monthly scaling factors of gasoline and diesel fuel sales

2.3.2. Traffic monitoring data

Based on previous studies described below, we compile fuel-based mobile source emission factors (in g kg⁻¹ fuel) for CO, NO_x , VOCs, NH₃, SO₂, and PM_2.5 for the 2020 BAU year (table S4). A brief description of the methodology is provided here, and more detail can be found in appendix S3, appendix S4 and table S5.

On-road fuel-based emission factors used in FIVE for CO, NO_x , and VOCs have been described by several studies (McDonald et al 2012, 2013, 2018a, 2018b, Hassler et al 2016). Emission factors are adjusted to account for the effects of cold-starting engines based on the EPA MOVES2014a model (EPA 2015), as well as evaporative emissions of gasoline (see appendix S4 and table S5 for details). Primary PM_2.5 emission factors from gasoline and diesel exhaust have been described by McDonald et al (2015). Trends in LD PM_2.5 exhaust emissions scale with CO (McDonald et al 2015), and we update HD PM_2.5 exhaust emission factors with more recent measurements of HD trucks (figure S2). We also include brake wear and tire wear emissions from the EPA MOVES2014a model (EPA 2015). Cao et al (2021) presents trends in tailpipe NH₃ emission factors. Lastly, the SO₂ emission factor is estimated based on a sulphur content of 10 ppm and 20 ppm (by weight) for gasoline and diesel fuel, respectively.

For off-road engines, we use emission factors from the EPA NONROAD model (EPA 2010), except where otherwise noted in table S4 for off-road gasoline engines. Off-road gasoline CO, NO_x , VOC, and PM_2.5 emission factors have been described previously for FIVE (McDonald et al 2015, 2018a, 2018b). Emission factors of SO₂ are assumed to be the same for on-road gasoline and diesel engines. Since 2014, all non-road engines are required to utilize ultra-low sulphur diesel (EPA 2017).

In figure 1, we compare Apple's COVID-19 Mobility Trends Report (Apple 2020) and Google's COVID-19 Community Mobility Report (Google 2020) with highway traffic monitoring data across four cities (figure 1: (a) New York, (b) San Francisco, (c) Atlanta and (d) Los Angeles). For Google's mobility report, we average the Workplace Mobility and the Retail and Recreation Mobility metrics. For Apple's mobility report, we use the driving mobility index. We note that neither mobility report attempts to directly quantify changes to on-road transportation. Google's report is derived from mobile phone location activity at marked locations while Apple's report is derived from Apple Maps personal vehicle routing requests. Additionally, the Apple and Google mobility datasets do not explicitly capture HD traffic trends, which are important to on-road trends due to differences in emission factors by engine type (table S4).

Zoom In Zoom Out Reset image size

Considering the mobility datasets alone, we find they perform similarly in capturing the beginning of lockdowns (figure 1). When comparing timing of decreases in traffic monitoring and mobility to timing of public health measures, traffic decreases generally precede stay-at-home orders (table S6) by several weeks. Apart from Atlanta, reopening measures do not generally coincide with a significant traffic increase. Traffic counters in each city (black lines in figure 1) show that by July, traffic has nearly recovered from the minimum in travel seen in April. In June and beyond, Google's datasets have not recovered to prior levels, whereas Apple's dataset shows a stronger recovery, and even exceeds 2019 traffic levels in some cities.

Across the cities considered, neither mobility dataset performs well at capturing trends in 2020 traffic. Considering the month of April, normalized mean bias relative to traffic counters of Google mobility data is −29% and −27% for Apple (figure 1). Considering the COVID-19 timeframe from March to December, biases are −28% for Google (R = 0.72) and +5% for Apple (R = 0.80). While Apple mobility data exhibits little overall bias, the direction and magnitude of bias varies largely by month and location (figure 1). For Google, the bias is more consistent throughout the year but varies by location.

While the mobility datasets evaluated here may not precisely capture the magnitudes of traffic trends, their value for rapidly understanding the onset of the pandemic and general traffic trends should not be underestimated. Mobility datasets are daily estimates, available within a few days at a county-level spatial resolution. With proper calibration of the relationship between mobility data and traditional traffic counting and fuel sales data used in bottom-up emission inventories (Gurney et al 2012, Gately et al 2013, McDonald et al 2014), these datasets could be improved for future research and air quality forecasting efforts. As an alternative to using mobility data to estimate COVID-19 emission scaling factors, and to facilitate future use of mobility datasets, we next explore scalings based on monthly fuel sales data.

Figure 2 shows changes in US gasoline (panel a) and diesel (panel b) fuel sales, as well as a map of 2020 BAU on-road CO₂ emissions with reductions due to COVID-19 lockdowns by PADD district overlaid (panel c). Fuel sales data are de-seasonalized by representing each month as a fraction relative to the same month in 2019. Under BAU conditions, on-road traffic peaks in summer and is lower in winter (McDonald et al 2014), so a de-seasonalized trend better portrays how COVID-19 lockdowns impact transportation. The coloured uncertainty bands reflect regional variability at a PADD level. Although fuel sales trends are similar across PADDs (figures 2(a) and (b)), variations in the fraction of total on-road fuel consumed as diesel result in on-road CO₂ reductions (figure 2(c)) that are 35% larger in the most affected PADD than the least affected PADD (PADD scaling factors included in table S7). The PADD district to state-level downscaling we apply to on-road LD activity (appendix S1) suggests even greater variability in fuel sales when examined by state rather than by region (green error bars shown in figure 2(a), table S8).

Zoom In Zoom Out Reset image size

Overall, we estimate US states experienced 25%–51% decreases in gasoline sales in April 2020, relative to April 2019. Traffic and fuel sales both decrease in March and reach a minimum in April as lockdown measures go into effect (figure 2(a)). Public health policies surrounding the reopening of states have been staggered and vary by state, but in both traffic and fuel sales we observe that most of the recovery occurs in May and June. Traffic recovery then plateaus and remains below normal throughout July and August. By July, we estimate gasoline sales decreased 1%–15% across states, relative to 2019 levels.

We also compare gasoline fuel sales trends with LD traffic trends from publicly available embedded traffic counters at urban and national scales. The urban locations included are the same as in figure 1. Generally, we observe that both urban only and national LD traffic on interstates follow a similar trend throughout 2020, which also follow the gasoline fuel sales trend line (figure 2(a)). In contrast, diesel sales follow a different monthly pattern throughout 2020 (figure 2(b)). In April 2020, diesel sales decreased by 4%–17% relative to 2019. Urban HD traffic has a greater decrease in April 2020 (−30%) than national interstate HD traffic (−16%) by a factor of ∼2, a difference not apparent in LD traffic. This suggests that intra-city HD truck traffic was impacted greater than interstate goods movement. We provide COVID-19 inventory scaling factors for LD/HD traffic separated by urban and rural regions in table S8 (see appendix S1 for methodology).

State-level on-road gasoline fuel scalings are plotted for the four cities in figure 1. The April fuel scalings we develop exhibit less normalized mean bias relative to monthly averaged traffic counting data (+12%) than monthly averaged mobility data (−29%, Google; −27%, Apple). Between March and December, bias in our fuel scalings is +8%, which outperforms Google mobility (−28%). Bias in Apple's dataset is smaller (+5%) due to offsetting directions in the bias throughout the year across locations. Over the same timeframe, our fuel scalings have lower normalized mean absolute error (NMAE = 0.09) compared to either mobility dataset (NMAE = 0.28, Google; NMAE = 0.17, Apple). Fuel scalings also correlate with traffic counters (R = 0.90) better than for mobility data (R = 0.72, Google; R = 0.80, Apple). While a state-level fuel-based method has reasonable bias and correlation with respect to traffic counters, a positive bias in each city suggests our fuel-based scaling may slightly overestimate urban traffic throughout 2020. Further research is needed to separate urban and city-level variations in traffic from state-level variations, such as top-down emissions estimate techniques using high-spatial resolution satellite data (e.g. TROPOMI).

Traffic volume reductions in 2020 likely impacted speed patterns, which affect vehicle fuel consumption rates. We investigate this by looking at highway speeds in several locations. In Los Angeles, the average highway speed increased +17% (84 km h⁻¹ to 98 km h⁻¹) in April 2020 versus April 2019. Similar trends are found in San Francisco (+18%, 87 km h⁻¹ to 103 km h⁻¹) and Atlanta (+21%, 89 km h⁻¹ to 108 km h⁻¹). Throughout the rest of the 2020, highway speed increases averaged +10%, +16% and +9% for Los Angeles, San Francisco, and Atlanta, respectively. While a reduction in stop-and-go congestion improves fuel economy, high speed driving exceeding 90 km h⁻¹ also degrades fuel economy and could offset fuel economy gains from reduced congestion (Davis and Boundy 2021). While spatially resolved speed information could help improve our bottom-up emissions inventory at finer spatial scales, at coarser resolutions (e.g. state-level), fuel sales data reflect the combined effect of changes in the amount of driving and in fuel economy. For co-emitted air pollutants, fuel-based emission factors are much less sensitive to changes in speed and engine loading in comparison to activity-based emission factors (i.e. g km⁻¹) (Bishop and Stedman 2008, McDonald et al 2013, McDonald et al 2018a). This implies that reductions in fuel use also translate to similar reductions in co-emitted air pollutants.

When examining changes to weekday passenger diurnal traffic patterns throughout 2020 versus 2017–19, we find insignificant changes during morning (5–9 AM) and evening (2–6 PM) rush hours, by +1% and +6%, respectively (figure S3). Daytime truck traffic (6 AM—6 PM) in April also changes insignificantly by +4% (figure S4). For LD and HD vehicles, weekday traffic saw minor increases while weekend traffic saw minor decreases (figure S5). In the months following April, the magnitude of these deviations decreases. Traffic monitoring data aggregated at a state level, typically has 95% confidence bounds of ±15%–19% (FHWA 2016), meaning that the changes in day-of-week and diurnal traffic patterns due to COVID-19 are likely not statistically significant relative to uncertainties in traffic monitoring datasets. In general, we find that while the overall level of traffic changed due to the COVID-19 pandemic, the timing of traffic was not significantly altered.

Figure 3 presents a map of COVID-19 changes to mobile source NO_x emissions (panel a) and their impact on total anthropogenic NO_x emissions (panel b) for the month of April. Non-mobile emissions (e.g. industry and power plants) are from the NEI 2017 (EPA 2020) and are not adjusted, thus our estimates of how total anthropogenic emissions have changed due to COVID-19 should be treated as a lower bound estimate as other emission sectors have likely been impacted by the pandemic. The fuel sales data used in estimating fuel scaling factors include both on-road and off-road mobile source engines, and we apply the same fuel scaling factors to both types of engines at a PADD level. Ambient NO_x /CO ratios in 2019 and 2020 are assessed by month (see figure S6 and appendix S5), a diagnostic previously used to evaluate mobile source emission inventories (Parrish et al 2006, McDonald et al 2013, Hassler et al 2016). For Los Angeles and New York City, ambient NO_x /CO stayed remarkably similar in 2019 and 2020, suggesting that the mix of gasoline and diesel fuel use has not changed in urban areas during the COVID-19 pandemic though overall activity levels are down.

Zoom In Zoom Out Reset image size

At a state level, reductions to April NO_x emissions vary between 6% and 39% (figure 3(a)) for mobile source emissions, which are expected to lower total anthropogenic emissions by 1%–26% (figure 3(b)). Summed over the contiguous US during the month of April, mobile source NO_x emissions decreased 20%–25% (figure S7(a)), which lowers total anthropogenic NO_x emissions by 9%–12% (figure 4(c)). The recovery in emissions reductions is steady throughout May and June but slows in July and beyond, where emissions remain below normal (figure S8). In July, NO_x reductions are limited to 6%–7% (mobile sources) and 3%–4% (total), respectively (figures S7(b) and 4(d)). Overall, we find that 2020 NO_x reductions are greatest for East and West Coast states, where transportation appears to be more impacted by public health measures to combat the spread of COVID-19. Additionally, in many coastal states, diesel fuel makes up a smaller fraction of on-road fuel consumption than in central states. For example, in Wyoming 52% by volume of on-road fuel consumption is diesel whereas only 17% is diesel in California (FHWA 2020b). Diesel consumption was not as heavily impacted by COVID-19 lockdowns as gasoline (figure 2), which helps to explain the greater relative decrease in mobile source NO_x emissions in California versus Wyoming.

Zoom In Zoom Out Reset image size

Lastly, we assess the impact COVID-19 lockdowns had on emissions in urban areas. In figure 4, we present a sectoral breakdown of mobile source related total emissions reductions in urban areas (figures 4(a) and (b)) and across the contiguous US (figures 4(c) and (d)) in April and July. Urban is defined by the US Census Bureau, as >1000 people mi⁻² (386 people km⁻²). In both urban areas and across the contiguous US, COVID-19 related reductions to highly impacted pollutants such as CO, NO_x , and NH₃, are generally dominated by on-road and off-road gasoline, due to greater reductions in gasoline sales than diesel sales. The impacts on emissions we estimate are also relatively larger in urban areas than across the whole US (figure 4(c)) or in rural areas. We estimate that mobile source reductions lower total urban CO emissions by 25%–35% and total urban NO_x emissions by 18%–22% in April (figure 4(a)), which are larger than the 20%–34% and 9–12%, respectively, reductions seen nationally (figure 4(c)). By July, total emissions reductions from mobile sources, as well as the enhanced reductions to urban emissions that we estimate, have greatly diminished (figures 4(b) and (d)).

Keller et al (2020) estimates a ∼30% decrease to NO_x emissions in the US in April. The NO₂ monitors used in their analysis are located predominantly in urban areas, meaning their estimate may better reflect urban reductions than total US reductions. Adjusting mobile sources alone (67% of urban NO_x emissions), we estimate an 18%–22% decrease to urban NO_x emissions, similar to Keller et al. If we accounted for COVID-19 impacts for other sources of NO_x (e.g. industry), then it is possible that the inventory-observation agreement could be closer. Our urban reduction estimate (figures 4(a) and (b)) is also comparable to the decreases in tropospheric column NO₂ observed by TROPOMI over US cities, when meteorology and seasonality are accounted for (Goldberg et al 2020). Specifically, we estimate reductions of 17%, 30%, and 35%, for New York City, Atlanta, and Los Angeles, respectively, compared to the mean decreases reported by Goldberg et al (2020) of 20%, 27%, and 33%.

Finally, our estimate of maximum change to on-road CO₂ emissions (−31%, figure S7(c)) is notably less than estimates by Doumbia et al (2021), Forster et al (2020), and Liu et al (2020) (−40%, −45% and −50% respectively), which is likely due to using the fuel-based, rather than mobility-based, method presented here. While these studies do not separate LD and HD traffic trends, their on-road CO₂ emission scalings can be applied to on-road NO_x emissions, resulting in a mean mobility-based estimate of US on-road NO_x reductions of 45% in April. These mobility-based reduction estimates are twice as large for US on-road NO_x emissions in April compared to our fuel-based method (22%, figure S7(c)), which demonstrates both the importance of separating LD and HD trends and the need for calibrating the relationship between mobility datasets and traffic when scaling emissions.