An analysis of roadside particulate matter pollution and population exposure over the Pearl River Delta region of China under clear-sky condition using new ultra-high-resolution PM_2.5 satellite-retrieval algorithms

The Landsat-8 OLI/TIRS C1 Level-1 data were used for the AOD retrieval. Detailed parameters can be found in table S1. Three granules with a similar scene time inside the PRD region were chosen (figure 1). Any day during 2014–2022 with cloud coverage less than 30% in either of the first two Landsat-8 OLI granules (marked in blue and pink) were selected. As a result, 226 Landsat-8 data in the total of 77 d were selected.

Zoom In Zoom Out Reset image size

Auxiliary data related to the AOD-PM_2.5 relationship model used in this study consisted of the 30 m Shuttle Radar Topography Mission 1 arc-second global elevation data, the 30 m Global land cover dataset (Globelland30) (Chen and Chen 2018), and the yearly 100 m WorldPop population distribution data (hub.worldpop.org). The meteorological data were obtained from ECMWF (https://cds.climate.copernicus.eu/) ERA5-Land hourly data with the continuous real-time gridded data measured at spatial resolutions of 0.1°, including 10 m U-component of wind (m s⁻¹), 10 m V-component of wind (m s⁻¹), 2 m temperature (K), surface net solar radiation (J m⁻²), surface pressure (Pa), leaf area index with low vegetation (m² m⁻²), leaf area index with high vegetation (m² m⁻²), total evaporation (m.w.e meter of water equivalent) and total precipitation (m). The wind velocity (m s⁻¹) and direction (°) were also calculated using 10 m U and V components. The relative humidity (%) was calculated using the 2 m dewpoint temperature (K), 2 m temperature, and surface pressure data. To improve the accuracy and spatial resolution, we also combined the ECMWF grided data with hourly meteorological in-situ measurements from NOAA Integrated Surface Data lite data set (www.ncei.noaa.gov/pub/data/noaa/isd-lite/) including temperate, surface pressure, wind speed, and wind direction.

For AOD results validation, we used two sites with level 2.0 AOD AERONET data (Giles et al 2019) at 440 nm with a network of ground-based Sun photometers within ±1 h of the Landsat-8 scenes acquisition time. PM_2.5 results validation used PM_2.5 data at 62 stations obtained from China National Environmental Monitoring Centre (CNEMC) (http://106.37.208.233:20035/) measured by ambient air quality continuous automated monitoring system, at six stations from Macao Meteorological and Geophysical Bureau (SMG) (www.smg.gov.mo) measured with tapered element oscillating microbalance or beta ray attenuation method, and at 16 stations from Hong Kong Environmental Protection Department (HKEPD) (www.aqhi.gov.hk/) collected by continuous ambient air monitor.

It is noted that the PM_2.5 results validation included the roadside PM_2.5 measurements collected at two SMG roadside stations, nine near-road CNEMC stations (⩽10 m away from the edge of roads) and three HKEPD roadside stations.

The digital map obtained from Open Street Map (OSM) (www.openstreetmap.org) was utilized to abstract the PM_2.5 concentration through 15 types of roads, including cycleway, footway, living road, motorway, path, pedestrian road, primary road, residential road, secondary road, service road, steps, tertiary, track road, trunk road, and unclassified road. The definitions and examples for each road type are provided in table S2.

The AOD retrieval algorithm was developed by combining the DT method with the SVM model. We first applied the DT method on vegetation and water surface, and then used the surrounding AOD results in the urban area as training data to estimate the AOD values over urban land surfaces. SVM model was chosen because of its ability to explain nonlinear relationships with small samples.

The detailed processes included pre-processing, classification, and modelling. Firstly, the raw Landsat 8 data were pre-processed using radiation calibration, atmospheric correction and cloud detection algorithms (Qiu et al 2019). Secondly, we removed the areas covered by clouds and the coastal areas, and classified the remaining data into dense vegetation, water, and complex urban surface. The surface reflectance of vegetation and water surfaces was dominated by groups of endmembers corresponding to water, soil, and vegetation. The DT method was thus applied to these two types of surfaces for the AOD estimation.

Regarding the AOD retrieving algorithm for urban surfaces, machine learning was chosen because the urban land cover is complex and always contains bright areas, which could highly affect the accuracy of the DT method. To avoid such uncertainty, we used the average AOD of vegetation and water surfaces near the urban area within 3 × 3 pixels as the training data to fit the relationship between AOD and Landsat-8 reflectance. Deep blue (0.440 µm), green (0.560 µm), near-infrared (0.865 µm), and short-wave infrared (2.200 µm) band were used as parameters. Pixels with more than three vegetation and water pixels nearby were chosen as samples to estimate the relationship within urban surface areas using the SVM model. To choose representative samples and improve algorithm efficiency over Landsat-8 urban areas with a complex morphology structure, we used simple linear iterative clustering (SLIC) super-pixels (Achanta et al 2012) to efficiently group similar pixels to include as many types of urban areas as possible. 1000 SLIC Super-pixels were first built (figure S4), whereas 10 sample pixels were chosen within each Super-pixel. The model was built and validated by 10-fold cross-validation (10-fold CV) to avoid the over-fitting problem and then applied to all urban area pixels. More detailed information can be found in the Supporting Information (SI).

The RF model was applied to analyse the AOD-PM_2.5 relationship because it can predict both continuous and categorical features and determine the relative importance of each feature. All the Landsat-8 AOD results from 2014 to 2022 were used in this model. Overall, AOD results in 226 Landsat-8 images on 77 d in the study period were retrieved and used in our roadside PM_2.5 pollution analyses. For each day, two to three Landsat-8 granules were used, and the average AOD was calculated for the Landsat-8 scene-overlapping areas. Moreover, to further improve the model performance, auxiliary data related to the AOD-PM_2.5 relationship were also included in the model, such as land-use data, population data, and meteorological data. As shown in figure S5, all the parameters were analysed based on the relative importance of PM_2.5 and the collinear relationship between one another, according to the p-value of the F-test and the variance inflation factor method, respectively.

For PM_2.5 results validation, our retrieval PM_2.5 concentration was compared to ground-level measurements at 84 sites. Cubic spline interpolation was used with the coordinates of stations to match the interpolated values of our results. To avoid the common over-fitting problem, three kinds of cross-validation were conducted and cross-compared: (1) the 10-fold CV, in which 90% of the sample were randomly selected as training data, and the remaining 10% data were used for validation. This validation was repeated 10 times; (2) the cross-validation with a single station withheld (1-station left-out CV), in which one of the stations was taken turns to be withheld when building another model, and the modelled PM_2.5 results were then compared with the data of the withheld station; and (3) validation by the whole data set (whole validation), in which the model built based on all the PM_2.5 site data was applied to the Landsat-8 AOD retrievals to get the PM_2.5 ground level estimation.

Figure 2(a) shows the PM_2.5 of 10-18-2015 as an example, whereas the complete results during 2014–2022 are shown in figure S8. Before analysing roadside PM_2.5, we first intensively validated for our developed PM_2.5 retrieval algorithms in two parts: (1) AOD and (2) PM_2.5.

Zoom In Zoom Out Reset image size

The retrieved AOD were validated using AERONET AOD. Figure S6 (a) shows the validation of AOD against AERONET AOD within ±1 h of the Landsat-8 scenes acquisition time from two sites in Hong Kong. The overall R² and RMSE were 0.84 and 0.15 µg m⁻³, respectively, indicating that our Landsat-8 AOD retrieval model was reliable. We also compared our AOD results with MODIS MAIAC AOD (R² = 0.58; RMSE = 0.14 µg m⁻³), and the traditional DT method-retrieved AOD (R² = 0.30; RMSE = 0.19 µg m⁻³), as shown in figure S6 (b). In addition, figure S7 compares the spatial patterns between MODIS MAIAC AOD and Landsat-8-derived AOD products on 15-10-2014 and 05-11-2016 as an example. The results indicate that both results had a similar spatial distribution; nevertheless, the Landsat-8 AOD contained more detailed spatial information.

We also compared our results with the hourly PM_2.5 ground-level measurements at 84 stations in the matching time (figures 2(b)–(g)). On average, R² and RMSE were estimated to be 0.86 and 7.72 µg m⁻³, respectively. On the other hand, 10-fold CV shows R² = 0.75 and RMSE = 9.81 µg m⁻³, whereas 1-station left-out CV shows R² = 0.74 and RMSE = 9.94 µg m⁻³ (figure 3). In addition, we compared 1 km PM_2.5 product generated from MODIS MAIAC AOD products by (Wei et al 2019) (figures S9–S11). The results show R² and RMSE to be 0.72 and 10.0 µg m⁻³, respectively. We also compared the range of our PM_2.5 estimation in each PRD city with all the PM_2.5 ground-level measurements. We selected the stationary data from 2014 to 2022 under clear sky condition with less than 30% cloud coverage and no precipitation. Figure S9 shows that the range of our PM_2.5 results in all 11 PRD cities was consistent with the PM_2.5 stationary data. The results above overall indicate that our results had a comparable accuracy and provided more spatial distribution details.

Zoom In Zoom Out Reset image size

With the 30 m PM_2.5 results, we analysed PM_2.5 distribution in the 11 cities in PRD region at the road level. Our retrieved PM_2.5 concentration was first overlaid onto OSM digital road data to extract roadside PM_2.5 on 15 types of roads. The definitions and examples for each road type can be found in table S2. Figure 4 depicts the highly polluted roads in each city as a demonstration. Our results show high PM_2.5 values often occurred in heavy traffic road types such as motorway, trunk, and primary roads. Within busy expressway roads, hot spots were often found in the middle part of roads due to high-speed car movement (e.g. up to 107.60 µg m⁻³ on Motorway in Foshan, 103.99 µg m⁻³ on Motorway in Zhuhai). On the contrary, within inner city roads, high PM_2.5 were often located at street intersections because of congestions, acceleration, and stop-and-go activities (e.g. up to 100.09 µg m⁻³ on secondary road in Guangzhou, 90.71 µg m⁻³ on living road in Shenzhen).

Zoom In Zoom Out Reset image size

Considering the distribution difference within different parts of the roads, we applied the Monte Carlo method (Garcia et al 1987) to statistically evaluate the differences in roadside air pollution between different road types, cities, and dates. In each comparison, we randomly selected 1000 000 samples, and removed any insignificant results (p-value ⩾ 0.05). Comparing to averaging, our method can analyse the difference in PM_2.5 between roads on a more detailed level. The average PM_2.5 concentration for each road type, cities and dates can also be found in figure S15.

Figure 5(a) shows the difference in roadside PM_2.5 between each city. It notes that warmer colour in the vertical direction indicates higher PM_2.5 concentration whereas cooler colour means lower values. The average PM_2.5 concentration for each city and date were also provided in figure S16 in SI. Among all the cities in the PRD region, roads in traffic hub cities such as Foshan and Guangzhou were the most polluted, followed by surrounding cities such as Dongguan, Zhongshan, and Jiangmen. Meanwhile, Hong Kong's roads showed the lowest PM_2.5 concentration of 31.71 µg m⁻³ on average, almost 29.70% higher than the average value in Foshan. Seaside cities such as Shenzhen, Zhuhai and Macao were also relatively clean, with average PM_2.5 of 34.54 µg m⁻³, 35.60 µg m⁻³, and 36.48 µg m⁻³, respectively. We also calculated the difference in PM_2.5 in each road type among each city. The difference in PM_2.5 between each city tended to be more obvious on roads for motorized vehicles, such as motorway, primary road, secondary road, and trunk road, and less noticeable on roads for non-motorized vehicles and pedestrians such as cycleway, footway, and track roads. This indicates that the PM_2.5 concentration level on non-motorized roads in most PRD cities were not dominated by traffic air pollution. Instead, they were likely to be dominated by background PM_2.5 level.

Zoom In Zoom Out Reset image size

Figure 5(b) displays the differences in PM_2.5 among each road type in the PRD region. In general, warmer colours on the vertical axis often appeared on road types related to motorized vehicles. Cooler colours with negative values often appeared on roads related to non-motorized vehicles and pedestrians, such as cycleway, footway, path, pedestrian, and steps. Average PM_2.5 higher than 37.00 µg m⁻³ only occurred on motorized road types such as trunk, tertiary, motorway, primary, secondary, residential road, and living road. The highest PM_2.5 was observed on the trunk roads (38.98 µg m⁻³ on average), whereas the lowest PM_2.5 was observed in road type of steps. The PM_2.5 level of trunk roads was 18% higher than that of the road type of steps, which were typically on hiking trails and more far away from main roads compared to cycleway and pedestrian roads.

We also analysed the differences in PM_2.5 among road types in each city (figure S17). In most cities in the region, the highest PM_2.5 appeared in primary roads, trunk, and motorways, which were usually major divided highways with 4–8 lanes. PM_2.5 on these roads were 9%–22% higher than that on the cleanest non-motorized roads. The highest PM_2.5 were located on motorway in Foshan (max: 107.60 µg m⁻³; mean: 48.03 µg m⁻³). In Hong Kong, tracks roads had the highest PM_2.5 (max: 85.10 µg m⁻³; mean: 33.54 µg m⁻³), which were 12% higher than the average PM_2.5 on step roads. Track roads were usually routes that were separated from traffic, but several hot spots (figure S18) were found in track roads contributing to high air pollution, such as the one inside the WENT (West New Territories) landfill, and the road outside of an asphalt company, which has already been convicted and fined by the government for violating the Air Pollution Control Ordinance license conditions.

Figure 5(c) depicts the differences in PM_2.5 between holidays and workdays in each road type and city according to different local laws in mainland China, Hong Kong, and Macao. According to the figure, Hong Kong was the only city with higher PM_2.5 concentration (1.93–4.50 µg m⁻³) during workdays, whereas PM_2.5 level remained the same between holidays and workdays in Shenzhen, Zhuhai, and Macao. In the rest of PRD cities, PM_2.5 concentration in holidays were 8%–17% higher than that in workdays. To prove our findings, we also calculated the holiday-workday differences in PM_2.5 using the roadside PM_2.5 measurements with the same time periods of our retrieval PM_2.5 data. The roadside data measurements show comparable results with the satellite retrieved data showing larger differences in PM_2.5 between holiday and workday. The holiday-workday pattern could be because tourists in Hong Kong may prefer to use the advanced public transfer system inside the city, whereas people in mainland China tended to drive between cities on holidays; the other reason could be Hong Kong cross-border trucking with mainland China on workdays was busier than on holidays. To further understand the traffic pattern, we evaluated the traffic volume data in Hong Kong and Shenzhen (figure S21). It shows that the traffic volume in Hong Kong was almost 30% higher on holidays then on workdays, whereas Shenzhen shows negligible difference in its traffic volume between holidays and workdays. This is consistent with our results. It is noted that our analysis focused on clear-sky condition, whereas traffic volume may reduce during severe weather conditions (Miao et al 2019, Akter et al 2020).

Moreover, we calculate the roadside the population-weighted PM_2.5 exposure using the formula ∑(P_i * C_i )/∑P_i, where P_i is the population and C_i is the PM_2.5 concentration. We analysed the population-weighted PM_2.5 as a function of distance from roads starting from 0 m to 1500 m in the PRD region and in each city. This analysis only considered the roads for motorized vehicles with high PM_2.5 including motorway, primary, residential, secondary, service, tertiary, track, trunk and unclassified roads. We calculated the distance between each pixel within boundary of each city and the nearest road (figure S19), and then calculated the average population-weighted PM_2.5 within each 30 m distance ring from roads.

Figure 6 shows that the population-weighted PM_2.5 tended to decrease with the increasing distance from roads in the whole PRD region and in each city. Among all the PRD cities, Dongguan, Foshan, Zhongshan and Jiangmen had relatively high population-weighted PM_2.5 near roads. Guangzhou had a comparable level of population-weighted PM_2.5 within 200 m from roads, and then the population-weighted PM_2.5 declined sharply and became the second last after 1 km from roads. Similar to Guangzhou, Shenzhen and Hong Kong show the similar sharp reduction in population-weighted PM_2.5 after several meters from roads. This characteristic clearly reflected the critical impacts of the traffic emissions to the air pollution near roads in the cities. On the other hand, Zhuhai had relative marginal changes along distances from roads due to its relatively less traffic. The stable level of population-weighted PM_2.5 reflected the minor contribution of traffic emissions to the air quality in Zhuhai.

Zoom In Zoom Out Reset image size

Our results also show that the population-weighted PM_2.5 profiles of the PRD region and most of the PRD cities (8 out of 11) had an increase within 100 m from roads. The slight increases in population-weighted PM_2.5 indicate that the traffic pollution tended to accumulate within a certain distance, and there might be a heightened population-weighted PM_2.5 zone near the roads. Overall, the population-weighted PM_2.5 was 2%–22% higher at roadsides than at distances more than 1500 m away from roads in the PRD region.

This study provided a comprehensive analysis of roadside PM_2.5 distribution and the resultant population exposure. It should be noted that evaluating the contribution of traffic-related emissions to roadside PM_2.5 is critical. Future studies are recommended to conduct a source apportionment study to assess the contribution of vehicular emissions and the background PM_2.5 to the roadside PM_2.5 in various cities. Additionally, we plan to enhance the temporal resolution of our results by combining multiple satellite sources. These proposed improvements are anticipated to further enhance our understanding of roadside PM_2.5 distribution.