Supplement of High-resolution mapping of vehicle emissions of atmospheric pollutants based on large-scale , real-world traffic datasets

Supplement of High-resolution mapping of vehicle emissions of atmospheric pollutants based on large-scale, real-world traffic datasets Daoyuan Yang 1, , Shaojun Zhang 2, , Tianlin Niu 1, , Yunjie Wang , Honglei Xu , K. Max Zhang , Ye Wu 1, 4, * 5 School of Environment, State Key Joint Laboratory of Environment Simulation and Pollution Control, Tsinghua University, Beijing 100084, P R China. Sibley School of Mechanical and Aerospace Engineering, Cornell University, Ithaca, NY 14853, U.S.A. Ricardo Energy & Environment, Beijing 100028, P R China. State Environmental Protection Key Laboratory of Sources and Control of Air Pollution Complex, Beijing 100084, P R China. 10 Transport Planning and Research Institute, Ministry of Transport, Beijing 100028, PR China * Correspondence to: Y. Wu (ywu@tsinghua.edu.cn)


Section S1. Supplementary figures and tables Supplementary Figures
Note: The maps cover the entire area within the Sixth Ring Road, but most of the available data were obtained from the area within the Fifth Ring Road (i.e., the urban area). Note: The red markers indicate the inter-city highway traffic monitoring sites, which generate detailed observations for 25 total volume, fleet mix and road speed and report to the Ministry of Transport (MOT) of China. The yellow markers represent the sites with traffic video records, which were conducted by researchers from Tsinghua University.  Note: The vehicle inspection dataset includes vehicle mileage records, and we obtained the processed data from local environmental protection authorities to represent the fleet-average annual vehicle kilometer travelled (VKT) (Zhang et al., 2014). The traffic volume informed annual VKT is estimated according to Eq. (S1) .
where ̅̅̅̅̅̅ , is fleet-average annual traffic VKT for vehicle category c, km; ̅̅̅̅ ,ℎ, is the average traffic volume for vehicle category c, hour h and road link l, veh, which is weighted by various traffic scenarios (e.g., weekday, weekend); Pc is total population for category c that are issued with Beijing license plates, veh.    according to the registration place; d The GVW values of non-local HDTs are mostly above 12000 kg, which use the emission factors of HDT3 (GVW>12000 kg) estimated by Wu et al. (2012) (Wu et al., 2012). By contrast, emission factor for local HDTs are weighted by HDT2 and HDT3 according to their registration number and annual VKT (Zhang et al., 2014). Non-local HDTs have been only allowed to drive within the Fifth Ring Road from 0:00 to 6:00 (GMT+8), and the restrictive boundary has been extended to the Sixth Ring Road since April 2014.
HDTs were only allowed to drive within the Sixth Ring Road from 0:00 to 6:00 (GMT+8) Non-local HDTs were not allowed to drive within the Sixth Ring Road all day around. Odd-even driving restrictions were implemented from 3:00 to 0:00 (GMT+8) for all the vehicles in the entire city, and especial public and municipal fleets were exempted. Section S2. Generating high-resolution traffic profiles

Calculating hourly road speeds based on real-time traffic congestion index data
According to the official guideline of congestion index (Beijing Municipal Administration of Quality and Technology Supervision and BTI, 2011), the index ranges from 0 to 5 representing more serious levels of congestion. Table S3 summarizes the relationship between color indicator, congestion index range and speed interval by road type. For example, grey indicate a congestion index below 3, representing average speed greater than 20 km/h for arterial roads. We proposed quadratic functions to express the relationship between congestion index and road speed and used breakpoints in Table S3 as known data for function fitting (see Fig. S11). For example, the data pairs (congestion index, road speed) used for determining the function for arterial roads were (1, 45), (2,35), (3, 25) and (4, 15). Red 4-5 <20 <15 <10 Figure S11. The relationships between the speed and the congestion index of different road types.
The original congestion map was updated every 5 mins. However, the color indicator embodies a large uncertainty in estimating road speed for that short episode (e.g., grey color indicates average speed greater than 20 km h -1 for arterial roads). To reduce the uncertainty, we used the midpoint of congestion index range to digitalize the color indicator (e.g., 1.5 for grey, 3.5 for yellow and 4.5 for red) and aggregated 12 color indicators (i.e., proxy instantaneous congestion indexes) into an hourly average index for all the road segments included in the congestion map. We applied the relationship between congestion index and road speed to derive hourly speeds in the research domain through the daytime (6:00 to 23:00 GMT+8). Fig. S12 provided an example of the entire process regarding speed estimation. The official website also reported average speeds (excluding minor roads) in various zones of Beijing during rush hours. As Fig. S13 indicates, the estimated average speeds in this study were close to the corresponding speeds reported by the official website on the annual level. As yellow and red indicators were rarely reported during nighttime (23:00 to 6:00 GMT+8), using the relationship functions in Fig. S11 would lead to underestimated road speeds. We referred to real-world nighttime vehicle speed data to calibrate our model, which were obtained from another project regarding large-scale GPS trajectory data collection from personal vehicles in Beijing (He et al., 2016).

Localized traffic density functions
Local traffic density functions for expressways and arterial roads were developed based on annual average hourly traffic volume and speed data. 45 expressways and 12 arterial roads (hourly average single lane volumes ranged between 800 to 1500 veh h -1 ) were used in this study, as road speed data were not available from other roads. We compared the fitting performance between two common traffic density models, i.e., Underwood and Greenshields functions. As where TV0 h,l and V0 h,l are annual-average hourly traffic volume and speed for hour h and road link l ,unit in veh h -1 and km h -1 , respectively; TVh,l is estimated traffic volume in response to traffic congestion index informed speed Vh,l unit in veh h -1 and km h -1 , respectively. Notes: a normalized mean bias; b mean fractional error.

Estimating traffic activities for nonmonitored roads
Traffic monitoring data (e.g., congestion index, annual traffic data, and intercity highway traffic monitoring) were not available for mostly minor roads and suburb/rural roads. We estimated traffic activities for these nonmonitored roads based on the average traffic data of monitored roads in the same corresponding traffic zone (see Fig. S15) and suggested volume/speed ratios between various road types (see Eq. S3).
where fn,i,j,k is the traffic volume of the road link n with road type k in the traffic zone i at hour j, unit in veh h -1 ; Fi,j is the average single-line traffic volume of monitored baseline roads in traffic zone i at hour j, unit in veh h -1 ; αk,i,j is the traffic volume with the road type k in proportion to the average volume of baseline expressways in the traffic zone i at hour j (See Table S5); mn is the number of lanes of the road number n. The suggested traffic volume ratios were reported by Beijing Jiaotong University, and summarized in Table S5.
where vn,i,j,k is the traffic speed of the road number n with road type k in the traffic zone i at hour j, unit in km h -1 ; Vi,j is the basic traffic speed of traffic zone i at hour j, unit in veh h -1 ; βk,i,j is the traffic speed with the road type k in proportion to the average speed of expressways in the traffic zone i at hour j. The speed ratios were derived based on trajectory data collected from nearly 500 personal vehicles in Beijing (He et al., 2016).

Part S3. Simulating vehicular NOX concentrations by using the RapidAir® model
Concentrations were modelled at 10m spatial resolution over the entire municipality of Beijing (16410 km 2 ). Hourly vehicular emissions of NOX from all road links were used as linear emission input data. The RapidAir® model spatially allocates emissions to line sources at the desired resolution (i.e., 10m). For the city-level simulations, we didn't include localized highresolution terrain and building profiles in the RapidAir modelling (i.e., default urban setting used). Meteorological profiles   For the fine-grained hotspot simulations (gridded spatial resolution at 1 m×1m), additional data are applied to improve the accuracy. First, the number of traffic lanes for each road has been specified, which further design the emission sources as multiple parallel lines (e.g., more than 10 lines for the major expressways in each hotspot domain, including the side road segments). Second, the profiles of building profiles are required to estimate the street canyon effect. The detailed methodology (based on the AEOLIUS model developed by the UK Meteorological Office) and evaluation of using geospatial surrogates to represent street canyon effects have been reported by Masey et al. (2018) (Masey et al., 2018). The buildings and height data were derived from OpenStreetMap and observation, traffic volume and speed data were from emission inventory. Figure S17 uses yellow area to represent the street canyon domains identified by the RapidAir®. Furthermore, the concentration gradients across major expressways (A) and arterial roads (B) in two hotspot regions are also summarized in Fig. S18.

Figure S18. Maps and cross-road NOX concentrations for two hotspot regions: (a) Guomao and (b) Xisanqi.
We compared the simulated NOX concentrations and the observed results from official air quality monitoring sites in Beijing (12 urban environment sites all located within the Sixth Ring Road and 5 urban traffic sites). We referred to site-specific cells in the RapidAir® results to derived the simulated vehicular NOX concentrations. However, only NO2 data rather than NOX (i.e., separate NO2 and NO concentrations) are reported to the public in China. To crossover the barrier, we referred to the approximate photostationary state and established a chemical equilibrium for NO2 between the NO2 photolysis and the O3 depletion (important reactions), which was applied to estimate site-specific NO concentrations (See Eq. S5). We constrain the comparison within the daytime time framework (6:00 to 18:00 GMT+8) in 2013. The nighttime simulations have not been included because of the complicated nighttime NOX chemistry.
where EM2 j,p is the allocated emissions of air pollutant p in cell j; Ep is total emissions of air pollutant p estimated according to M1; TP is the total population of the entire municipality of Beijing; Pj is the population in the grid cell j.
M3: Allocation based on road length and type. Eq. S8 illustrates the way to allocate vehicle emissions according to total standard road length in each cell.
where EM3 j,p is the allocated vehicle emissions of air pollutant p in cell j; ULk and RLk represent the total actual road length of the road type k in the urban (including the area between Fifth Ring Road and Sixth Ring Road) and suburb/rural areas (outside Sixth Ring Road), respectively; ULj,k and RLj,k represent the actual road length of the road type k in grid cell j, respectively; UWk and RWk are coeffients to estimate standard road length (e.g.,