We use the mobile sampler XHAQSN-508 from Hebei Sailhero Environmental Protection High-tech Co., Ltd. (Hebei, China) to measure the air quality in the Nanjing urban area. The instrument is equipped with internal gas sensors for CO (model XH-CO-50-7), NO2 (XH-NO2-5AOF-7), and O3 (XH-O3-1-7) (dimensions: 290 × 81 × 55 mm; weight: 1.0 kg) as well as two small in-built sensors for temperature and relative humidity and is fixed in the top lamp support pole (∼ 1.5 m above ground) of two Nanjing taxis (Fig. 1). Two taxis fueled with electricity and liquefied natural gas (one each) are selected to reduce the impact of emissions from the sampling vehicles themselves. All three sensors are electrochemical, which based on a chemical reaction between gases in the air and the electrode in a liquid inside a sensor that can detect gaseous pollutants at levels as low as parts per billion (Maag et al., 2018). Sensors are continuously powered by an external DC 12 V power supply provided by a taxi battery. The sample is refreshed by pumping air to the sensors. There is an air inlet at the bottom of the instrument, which is also checked periodically to avoid blockage. Because it is fixed in the taxi top lamp, it can reduce the impact of different wind direction airflow. This device integrates components for data integration, processing, and transmission and provides data management, quality control, and visualization functions. Pollutant concentration data are generated by different voltage values sensed by gas sensors, which are automatically uploaded to a database in the cloud via the 4G telecommunications network. We continuously measured the concentration of CO, NO2, and O3 in the street canyon in the urban area of Nanjing (with the center located at 32.07∘ N and 118.72∘ E) for a whole year (1 October 2019–30 September 2020). An instantaneous measurement of CO, NO2, and O3 concentrations is configured to continuous measurements at a frequency of once per 10 s sampling interval, and their limits of detection are 0.01 µmolmol-1, 0.1 nmol mol-1, and 0.1 nmol mol-1, respectively. The sampling routes were relatively random during taxi operations and were mainly on the arterial roads. A GPS device (u-blox, Switzerland) is utilized to record the spatial coordinates, and the spatial offsets are corrected by ArcGIS 10.2 software. Generally, the sampling campaign is conducted on both weekdays and weekends from 06:00 to 22:00 local time (LT). Occasionally the taxi drivers work for the night shift, and the instruments are run from 22:00 to 06:00 LT. The collected data cover 373 km2 with a population of 6 million (Fig. 1).

Different from traditional instruments, low-cost sensors have some limitations, such as nonlinear response, signal drift, environmental dependencies, low selectivity, and cross-sensitivity, so it is important that calibration procedures are applied with respect to these limitations (Maag et al., 2018; Lösch et al., 2008). For example, environmental conditions are known to cause nonlinear behavior of sensors (Popoola et al., 2016). Due to aging and impurity effects over a long time, low-cost sensors are prone to signal drift and low sensitivity (Kizel et al., 2018). In addition, cross-sensitivities differ largely according to the ambient temperature and level of gas the sensor is being exposed to (Lösch et al., 2008). So, multi-parameter joint calibration training is utilized to reduce the interference issue between air pollutants in our research, including air pollutant concentrations, temperature, and relative humidity. The sensors are usually trained with co-located data collected by reference methods before being deployed to actual measuring campaigns (Kaivonen and Ngai, 2020; Chatzidiakou et al., 2019; Bossche et al., 2015).

The XHAQSN-508 is calibrated every month starting from September 2019. The instrument is placed at the outdoor Station for Observing Regional Processes of the Earth System (SORPES) in the Xianlin Campus of Nanjing University (https://as.nju.edu.cn/as_en/obsplatform/list.htm, last access: 22 May 2021) for at least seven days before the taxi began sampling. The collected data are calibrated against standard instruments (Thermo Fisher Scientific 48i, 42i, and 49i, USA, for CO, NO2, and O3, respectively). The instrument precision is ±2 ppbv for O3 and ±1 % and ±4 % for CO and NO2, respectively, which have been used in many other studies and found to perform well for long-term runs (Ding et al., 2013; Herrmann et al., 2013). One drawback of our study is that the air pollutant concentrations observed at SORPES are lower than those observed in a road environment, which might cause issues for the calibration process. Comparing different calibration models, we found that a machine learning algorithm can improve sensor–monitor agreement with reference monitors, and many previous studies have used this method (Qin et al., 2020; Esposito et al., 2018; Vito et al., 2018). A supervised machine learning methodology based on the gradient boost decision tree (GBDT) is used for data calibration (Johnson et al., 2018). GBRT, an ensemble learning method, is a decision-tree-based regression model that implements boosting to improve model performance using both parameter selection and k-fold cross-validation. GBRT needs to be trained by the dataset with target labels (Yang et al., 2017). It takes input variables including raw signals of sensors, air pollutant concentrations (CO, NO2, and O3), temperature, and relative humidity. The stationary instrument data are taken as training targets. The parameters of the machine learning model are adjusted continuously based on a gradient descent algorithm. The R2 of the calibration results is generally high (> 0.90) for all the three air pollutants (e.g., Fig. 2a).

The success of supervised model training with target labels (i.e., co-located with SORPES, Fig. 2a) does not guarantee its predicting power for conditions without labels (i.e., on roads or co-located with SORPES but not feeding the station data to the algorithm; Fig. 2b). We use a calibration–validation methodology to evaluate the performance of the calibrated sensors (Chatzidiakou et al., 2019). This method includes two phases: first, the sampler was calibrated against the SORPES station for 10 d (1–10 June 2020), and the sensor data were used for sensor algorithm training as described above (Fig. 2a); second, we continued to place the sampler in the station (11–17 June 2020). However, the sensor data are not used for calibration but directly fed in the algorithm trained in the first phase. The results are compared with the station data (i.e., validation phase; Fig. 2b). We find that the sensor data agree well with standard instrumentation in the second phase. The sensor-retrieved CO, NO2, and O3 concentrations are 0.58 ± 0.12 mg m-3, 8.40 ± 4.30 µgm-3, 27.3 ± 16.5 µgm-3 respectively, not significantly different from those measured by standard instruments (0.50 ± 0.10 mg m-3, 10.5 ± 6.31 µgm-3, and 32.4 ± 20.2 µgm-3) (α=0.05, ANOVA analysis). The R2 values generally remain high (0.82–0.97) for different air pollutants (CO and O3) except for NO2 (R2=0.67). The lower R2 value for NO2 may be associated with the higher humidity during the validation period (13–16 June 2020). As NO2 is water dissolvable, high relative humidity may lead to a low bias for sensors (Wei et al., 2018). To improve performance of the NO2 model, temperature and relative humidity have also been involved in the training algorithm. However, the interaction between O3 and NO2 may influence the detection accuracy of these two chemicals, especially for NO2 (Ivanovskaya et al., 2001). The accuracy of the sensor generally decreases with time (a.k.a. aging) due to the evaporation of the electrolyte (Ribet et al., 2018). However, we find no significant decrease in the R2 values for the three pollutants during our campaign. It seems that the machine-learning algorithm could successfully compensate the aging of the sensors. Field calibration of low-cost sensors is still a challenging task, as it is greatly affected by atmospheric composition and meteorological conditions (Spinelle et al., 2017; Castell et al., 2017). Our results have high R2 values compared to previous studies, indicating relatively high accuracy (e.g., Castell et al., 2017). The results from the two sensors also agree with each other reasonably well, with R2 values ranging 0.97–0.99 for a linear regression. Their data are thus combined in the following analysis to achieve maximum data coverage. Overall, the sensor results have substantial uncertainty compared to reference methods. We thus focus on the relative temporal and spatial distributions rather than the absolute concentrations.

As the mobile monitoring platform samples along the trajectories of carrying vehicles, we need to sacrifice either the temporal information to calculate the spatial distribution of air pollutants, or the spatial information to temporal variations. Similar approaches have also been adopted by previous studies (Bossche et al., 2015; Apte et al., 2017; Farrell et al., 2015). To generate the spatial distribution of air pollutants at high spatial resolution, we divide the research area into grids with 50 m × 50 m resolution and calculate the mean values of the samples collected in each grid. The driving condition is highly variable and the taxi can travel more than 50 m in 10 s if the vehicle speed is over 18 km/h. However, given the complexity of the driving conditions, we ignore the vehicle trajectory in the past 10 s but assign the measured values to the location of the vehicle at the time of data uploading. Then, combined with GIS technology, we calculate the average of all the data points over one year that fall in the same grid. One drawback of our study is the impact of spike concentrations on sensor performance. The sensors keep reporting high concentrations in an approximate 1 min period after exposure to large environmental concentration spikes. This effect would reduce the effective resolution of our gridded concentration map. Similarly, we calculate the hourly average concentrations by considering only the data sampled in the same hour of different days. The GPS signal is missing when the taxis pass through the nine underground tunnels in Nanjing (e.g., Xuanwu lake tunnel and Jiuhuashan tunnel in the city center; Fig. 3). We assume the taxies travel at a constant speed and the sampling points are uniformly allocated along the tunnels. We use the ArcGIS 10.2 software for data processing. To calculate the air pollutant concentrations (CO, NO2, and O3) of different road types and the contribution of traffic emissions to them, we divide the urban roads in Nanjing area into six types, including highways, arterial roads, secondary roads, branch roads, residential streets, and tunnels (https://wiki.openstreetmap.org/wiki/Key:highway, last access: 21 January 2021). The roads and land use data of Nanjing shown in Fig. 3 are based on OpenStreetMap (OpenStreetMap contributors, 2020).

The mobile platform keeps sampling in the urban road network which carries a strong signal from traffic sources. By contrast, stationary stations are often located far away from major roads to represent a regional background air pollution level (Hilker et al., 2019). Seven state-operated air quality observation stations in Nanjing are selected in our research, including Maigaoqiao, Caochangmen, Shanxi Road, Zhonghuamen, Ruijin Road, Xuanwu Lake, and the Olympic Sports Center (Zhao et al., 2015; Zou et al., 2017), which are far away from major roads and large point sources, so they are usually used as regional backgrounds in different functional areas (Zou et al., 2017; An et al., 2015). For example, Zou et al. (2017) chose the Olympic Center station (G in Fig. 1) to get the background characteristics of CO and NO2 in Nanjing. Therefore, the normalized contribution from traffic-related emissions can be obtained by differencing the mobile measurements and the stationary ones to minimize the influence of daily meteorological variations on the urban air quality, following Bossche et al. (2015): 1APtraffic,ij=(APij-APmin⁡)/APij, where APtraffic,ij represents the air pollutant concentration contributed by traffic emissions for the ith pollutant at time j (%); APij is the sensor-measured concentration of air pollutants; and APmin⁡ means the ambient background concentration, which is calculated as the minimum of the measurements from all the stations in Nanjing in the national air quality network without major sources within a direct vicinity of 50 m (https://quotsoft.net/air/, last access: 1 November 2020, Fig. 1). We refer to this method as the “background site” (BS).

We also adopt a method similar to Apte et al. (2017) for traffic source attribution. This method includes a peak detection algorithm to calculate the contribution of local traffic emission sources to on-road pollutant concentrations. We calculate the mean and minimum of air pollutant concentrations in each grid as the “peak” and “baseline”, respectively. The difference between the two is considered as the contribution from traffic sources. We refer to this method as “peak detection” (PD). MATLAB R2019a is used for such data calculation.

There is a trade-off between the resolution of an air pollutant concentration map and its reproducibility; i.e., high-resolution maps are subject to large randomness due to the limited number of samples in each grid. We investigate the consistency of spatial patterns of different resolution (10–100 m). We calculated the standard error of the means of samples in each grid (SEM) and then averaged the SEM over all grid cells: 2SEM=σ/n‾, where σ and n are the standard deviation and number of samples in each grid, respectively. We find the calculated SEM first decays rapidly with the grid spacing but tends to be in a regime of linear decay after a resolution of approximately 50 m for all the three air pollutants (Fig. 4). Therefore, we choose a resolution of 50 m, which is consistent with previous studies (Bossche et al., 2015; Apte et al., 2017). For example, Bossche et al. (2015) used a spatial resolution of 20–50 m to map urban air quality and identify hotspots. Apte et al. (2017) found that reproducible results (with high precision and low bias) of NO, NO2, and black carbon can be generated by at least 10–25 repetitions in a specific area with 30 m median spatial aggregation.

A total of 1.32 million pieces of data were obtained during the observation period, which covers 66.4 % of the major roads in Nanjing in the sampling domain with a large repeat-visit frequency (median repetition = 61 (14 and 264 as the lower and upper quartiles, respectively, the same hereinafter)) (Fig. 5a). The type of road with the most visits is the Neihuan lines (258 (116, 526)), followed by the arterial roads (125 (35, 393)), secondary roads (151 (24, 442)), and highways (34 (12, 115)). The residential streets (22 (6, 100)) have the fewest visits.

Apart from the areas without roads, such as the Yangtze River, Xuanwu Lake, and Purple Mountain, the data cover 43.5 % of the 50 m grids in the research area with the two taxis contributing 36.8 % and 37.2 %. As shown in Fig. 5b, the median number of repeated frequency in each grid is 66 (18, 286), with the highest value of 15 449 in Nanjing South Railway Station and the lowest in some residential roads (1). The repeated frequencies in each 50 m grid along the arterial roads and Neihuan line are higher than other types of roads, i.e., Zhongyang road, Huju road, Neihuandong, and Neihuanxi lines (Fig. 5b). Our repeated frequency is generally higher than previous research on mobile monitoring of urban air pollution (Peters et al., 2013; Poppel et al., 2013; Bossche et al., 2015; Apte at al., 2017), which can lower the uncertainty of our results. By comparing the time series of the air pollutant concentrations with that from nearby state-operated air quality observation stations (A′ and E′, with repeated frequencies > 500), we find that the results are consistent (Fig. S1 in the Supplement), which shows the stability and reliability of our data.

Figures 6 and S2 show the coefficients of variation (CV ≡ standard deviation / mean × 100 %) for different air pollutants in each grid. For one thing, this matrix quantifies the sensing power of mobile monitoring, i.e., more data points reduce the uncertainty of observations. For another, it reflects the inherent variability of pollutants caused by factors such as meteorological conditions and hotspot emission sources. We find that the CV values are lower than 100 % on the main roads, including highways and arterial roads, but higher than 100 % on some tunnels, residential streets, and Nanjing railway station. As discussed above, the road network coverage is much higher over the main roads than smaller roads. This indicates that increasing the sampling numbers within secondary and residential roads is the most useful way to reduce the uncertainty of mobile observation. It is also interesting to note that a single taxi has a data coverage of ∼ 37 % but the second one only increases it by ∼ 6.5 % to 43.5 %, which implies that the marginal increase in spatial coverage decreases substantially with an increasing number of sensors. This is indeed one limitation of our monitoring platform, and a much larger fleet size or different sampling platforms (e.g., bikes) may be needed to reduce the uncertainty over these smaller roads.

Although the spatial patterns of CV are similar for different air pollutants, we find generally higher CV for O3 (67.3 %) and NO2 (59.5 %) than CO (51.6 %). This is associated with the spatial and temporal variability of different air pollutants, which are influenced by their lifetimes in the atmosphere. Lifetime (or residence time) is the average time for a chemical compound that is transported in the atmosphere before it is deposited or consumed by chemical reactions. It is associated with its spatial scale of variability. The longer the lifetime, the more uniformly the concentrations are distributed. The chemical properties of CO are the most stable in the environment (τ=1–2 months), and its spatial concentration difference is more affected by the sampling time and the number of samples. The lifetime of NOx is shorter (τ=2–11 h, Romer et al., 2016), so the measured concentrations are more influenced by local or “hotspot” emissions and meteorological factors. O3 has the shortest lifetime (τ=∼ 1 h in urban atmosphere, McClurkin et al., 2013) among the three pollutants. The level of ozone is affected by its precursors (NOx and VOCs), which both have large variability (Sharma et al., 2016). The complex chemical reactions also increase its spatial heterogeneity.

Figure 8 shows the temporal variation of the three air pollutant concentrations during the observation campaign, with the hourly mean concentrations over the research domain shown in Fig. 9 (the corresponding spatial distributions are shown in Figs. S4–6). The difference of the hourly variation of the mean sample of different types of roads over a year is small (Fig. S7), so the data in Fig. 9 are not filtered in anyway, but for each hour a similar mix of road types is sampled. We find that the median concentrations of CO and NO2 in rush hours (07:00–09:00 and 17:00–19:00 LT) are increased by 26.4 % and 27.3 % compared to non-rush hours, respectively. The hourly mean concentrations of CO and NO2 show a double-peak pattern with higher concentrations in rush hours (Fig. 9a), reflecting the contribution of traffic-related emissions (Tan et al., 2009), which we will elaborate in the next section. The observed O3 concentrations show a unimodal diurnal pattern with a peak at ∼ 14:00 LT as a result of photochemical formation. At night, O3 concentrations are maintained at a low level due to a lack of solar radiation and the NOx-titration effect (Xie et al., 2016; Li et al., 2013). These patterns generally agree with the measurements at stationary monitoring stations (Fig. S3).

No significant differences are observed for the median concentrations and spatial distribution of three air pollutants between weekdays and weekends (α=0.05; Figs. 8b and S4), even though the morning peaks for CO are slightly higher during weekdays (Fig. 9b), which is consistent with An et al. (2015). Wang et al. (2014) found that NOx displays a weekly cycle in the Beijing–Tianjin–Hebei metropolitan area, with higher levels on weekdays than weekends. Qin et al. (2004) observed a significant weekend effect in southern California, showing that during the morning traffic rush hour, the concentrations of CO and NO2 at weekends were about 18 % and 37 % lower than on weekdays. The difference between our study and other cities lies in the difference of fleet fuel structure, and the different weekly routine of human activities and the taxi driving trajectories (Xie et al., 2016).

The median concentrations of CO and NO2 during holidays are comparable to those during non-holidays but are 18.3 % lower for O3 (Fig. 8c). In addition, compared with the spatial distribution of O3 concentration during holidays, we find that the concentrations of O3 in Xinjiekou and its surrounding areas, where many shopping malls are located, are higher during non-holidays (Fig. S6). This may be related to the higher NO2 concentrations in this area during holidays (24.8 ± 10.2 µgm-3) than non-holidays (20.6 ± 4.82 µgm-3). The hourly concentrations show no significant difference between holidays and non-holidays (Fig. 9c). The holidays include the periods of National Day (1–7 October), the Spring Festival (24–31 February), Qingming Festival (4–6 April), international labor day (1–5 May), and the Dragon Boat Festival (25–27 June). The “holiday effect” has been observed extensively for urban and regional air quality. For example, Xu et al. (2017) found that VOC tracers were significantly enhanced during the National Day holiday (from 1–10 October 2014) in the Yangtze River Delta (YRD) region, indicating that the “holiday effect” had a strong influence on the distribution and chemical reactivity of VOCs in the atmosphere. The reason why this effect is not observed in our study may be related to the relatively smaller sample size during holidays. The sample size for holidays account for only 11.3 % of those for the non-holidays.

Figure 10a and b show the calculated contributions by traffic-related emission sources to the observed concentration of CO (referred to as contributions hereinafter). We find that the mean contribution calculated by the BS method (42.6 ± 11.5 %) is generally consistent with that obtained from the PD algorithm (43.9 ± 27.0 %). Their spatial patterns are also similar (Fig. 10a vs. b). Although our data coverage is much larger than that of the Apte et al. (2017) study, we find that the reference method is still applicable in our research area. The contributions in highways, near tunnel entrances and exits (e.g., Jiuhuashan and Xuanwuhu tunnel), at the railway station (Nanjing south station), and on arterial roads (44 %–59 %) calculated using both methods are higher than on secondary roads and residential streets and lowest on branch roads (29 %–39 %) (Table 3), which is consistent with the trend in traffic volumes. The patterns for NO2 are quite similar to CO (Fig. S8c and d, Table 1), but the mean contribution to NO2 calculated using the BS method (26.3 ± 14.7 %) is lower than that obtained from the PD algorithm (40.2 ± 29.9 %). This difference is associated with the relatively higher uncertainty for NO2 measurements by sensors (Sect. 2.2), while the results of the PD method seem unaffected as the sensor bias is canceled out when calculating the difference between “peak” and “baseline” (Sect. 2.4).

The bottom-up emission inventory indicates that on-road transportation contributed ∼ 11 % of total CO emissions from Nanjing in 2012 (Zhao et al., 2015). Considering the number of cars has increased by ∼ 80 % and the total CO emissions remained relatively stable (BSNM, 2019), the contribution of traffic sources in recent years is expected to be ∼ 20 %. These values are much lower than what we calculated based on mobile monitoring data because of the lower spatial resolution of these regional inventories (e.g., 0.05∘ × 0.05∘) (Zheng et al., 2014). They are unable to distinguish the emission characteristics of air pollutant within a street level (tens of meters), which leads to their underestimation of traffic-related emissions in the road micro-environment.

Figures 8d and 9d show the variation of air pollutant concentrations at different stages of the COVID-19 pandemic. The spatial distribution of concentrations and traffic contributions are also depicted in Figs. 11–12 and S9–S10. We divide the data into three stages: pre-COVID (P1, 1 October 2019–23 January 2020), COVID lockdown (P2, 24–31 January 2020 and 17–24 February 2020), and post-COVID (P3, 1 March–30 September 2020). We find the median concentrations of CO and NO2 were the lowest in P2 (Fig. 9d). For example, the CO and NO2 concentrations decreased by 44.9 % and 41.7 % from P1 to P2, respectively (Figs. 11 and S8). This pattern agrees well with the air quality station data over eastern China (Huang et al., 2021). We focus on the traffic sector as it is the most sensitive to lockdown measures, while other sectors, including power, industrial, and residential sectors, remain relatively unchanged (Guevara et al., 2021). We find that from P1 to P2, the average traffic source contributions of CO and NO2 using the BS method decreased by 59.9 % and 51.8 %, respectively (Figs. 12 and S9). This is consistent with the transportation index data, which shows a 70 % reduction in eastern Chinese cities during lockdown (Huang et al., 2021).

The observed CO and NO2 concentrations recovered to a level similar to P1 during P3. The traffic-related source contributions were increased by 120 % and 131 % from P2 to P3 for CO and NO2 (Figs. 11 and S9). Due to the limited data size and spatial coverage (only on some arterial roads and highways) during P2, the calculated contribution of traffic emissions to air pollutants may be not directly comparable to those shown in Fig. 9. But the changes in the contribution match well with the changes in traffic volume and human activities (Bao and Zhang, 2020). Our results also agree with top-down emission estimates from remote sensing data (Zhang et al., 2020), which showed the total NO2 emissions decreased by 31 %–44 % from P1 to P2 but increased 67 %–85 % from P2 to P3.

The observed ozone concentrations show a different trend from other pollutants in the three stages. We find a pattern of P1 < P2 < P3 for O3 median concentrations (Fig. 8d). The ozone concentration increased by 35.7 % from P1 to P2, and 48.7 % from P2 to P3 (Fig. S9). While the contribution of traffic emissions to ozone first decreased by 32.5 % from P1 to P2 and then increased by 39.3 % in P2 to P3 (Fig. S10). This is firstly associated with less titration of NOx during P2 as discussed earlier. In addition, the increased temperature and solar insolation in P2 and P3 also favor the photochemical formation of O3 compared to P1 (Xie et al., 2016; Fu et al., 2015; Reddy et al., 2010).