A hybrid model approach for estimating health burden from NO₂ in megacities in China: a case study in Guangzhou

The city of Guangzhou is located on the north side of the Pearl River Delta (figure 1). It has a total area of 7434 km² and a population of 14 million, divided into six districts (Conghua, Guangzhou city centre, Huaxian, Nansha, Panyu, and Zengcheng) (table 1). Concentrations of NO₂ are measured at 11 sites (figure 1). The workflow for this study is shown in figure 2. The ADMS-Urban dispersion model v4.1 (CERC 2017) was used to derive NO₂ concentrations at 83 additional strategically-selected receptors locations across Guangzhou, which were then used to develop an LUR model for NO₂ concentrations across the whole city.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The ADMS-Urban dispersion model v4.1 (Carruthers et al 1994, CERC 2017) is a widely-used (McHugh et al 1997, Carruthers et al 2000, Di Sabatino et al 2007, Righi et al 2009) quasi-Gaussian model for simulating dispersion in the atmosphere of continuous releases from a range of explicit sources in an urban area. The NO_x–NO₂ chemistry is simulated using the 8-reaction set of Venkatram et al (1994) that includes reactions with ozone and hydrocarbons. This General Reaction Set has been thoroughly evaluated against measurements (see e.g. Tonnesen and Jeffries 1994, Venkatram et al 1994, Chaney et al 2011, Carruthers et al 2017). Surface deposition is included in the model. The sources of data used for the ADMS-urban dispersion modelling of NO₂ concentrations at monitor locations and at the virtual sites in this study are given in supplementary information (SI) table S2. Gridded emissions of NO_x and VOCs for the year 2016 were obtained from the Multi-resolution Emission Inventory for China (Zheng et al 2018a) and downscaled to 4 km × 4 km horizon resolution using various spatial proxies (Geng et al 2017, Zheng et al 2018b). Shipping emissions were taken from MarcoPolo-Panda for the year 2014 (MarcoPolo-Panda 2017). A map of the total NO_x emissions is shown in SI figure S1. Motorway, trunk, primary, secondary, and tertiary roads, as defined by OpenStreetMap, were explicitly modelled as road sources (OpenStreetMap 2018). Emissions from roads were calculated by assigning total on-road traffic emissions to the explicitly modelled roads according to road type and total length of each road, as described in SI section A.

Wind speed and direction at 10 m above the ground, and cloud cover were obtained from ECMWF-ERA5 (ECMWF 2019). Background concentrations of NO₂, NO_x, and O₃ were obtained from the Copernicus Atmosphere Monitoring Service (CAMS) (ECMWF 2019). As the prevailing wind direction in Guangzhou is from the northeast (SI figure S2), the background site for pollutant concentrations for the model was chosen to be a rural location to the northeast of the model domain (SI figure S3). The model was evaluated using the NO₂ concentrations at the 11 monitoring sites shown in figure 1.

As the main purpose of modelling NO₂ concentrations is to estimate NO₂ where people live, the virtual sites were primarily selected to be proportionally located across the residential areas of Guangzhou. The six districts have a large range in population (table 1) so site selection was first stratified across the six districts so as to ensure that sites in the virtual network were located across the full domain area. Centroids of each cell of population data (100 m × 100 m) (Worldpop 2015) were extracted as potential virtual receptor sites. Three types of virtual sites were defined for each district so as also to span the range in anticipated NO₂ concentration:

• Background site: centroid of the population cell with greatest distance to the nearest road.
• Roadside sites: centroids of the population cells with a distance to nearest road of no more than 5 m.
• Residential sites: centroids of the population cells with the highest population density.

In each district, residential sites were selected first, with number of residential sties proportional to the relative proportion of the total Guangzhou population in that district (the specific number is given in table 1). One residential site in two districts also fitted the definition of roadside site (see table 1), but these were retained under the category of residential. All other sites that satisfied the definition of roadside site (17 across the six districts, table 1) were selected as receptor locations within the roadside site category. A background site was selected in each district. An additional overarching criterion was a minimum distance of 300 m between any pair of sites in each district to ensure that all sites were distributed across the range of localities in the study area. The locations of the 83 sites are shown in figure 1. ADMS-Urban was used to simulate annual-mean NO₂ concentrations at these locations at a height of 1.5 m.

Sources of data used to develop the LUR model are given in SI table S2. Potential predictor variables (SI table S5) for the LUR model were chosen based on expectation that emission sources, dispersion, and physical geography may contribute to NO₂ concentration variation in urban areas (ESCAPE 2008). The inclusion of different buffer sizes allows for potential different influences of a predictor variables over different distances from the receptor (Su et al 2009, Beelen et al 2013).

Road lengths in a buffer had to be used as surrogates for traffic flow and fleet composition, as neither of these data are publicly available for Guangzhou. The variables of number of people and artificial (i.e. non-natural) area within a buffer do not directly affect NO₂ concentration, but they are indirectly related to road transport emissions and to domestic, industrial and electrical generation emissions which contribute to total NO_x emission in Guangzhou (SI figure S4). Distance to nearest port and its derivatives were included as non-buffer variables to account for the impacts of shipping, Guangzhou is close to the South China Sea, and is a major port, with ship activities significantly contributing to total NO_x emissions in the southern part of the model domain (Liu et al 2016, Johansson et al 2017) (SI figure S4). A maximum distance of 35 000 m was specified for the distance to port variable as the dispersion modelling at the virtual sites indicated that NO₂ concentration decreased to urban background level at this distance from nearest port as shown in SI figure S5.

Green, i.e. not built-up, areas are assumed to mitigate NO₂ concentration since they are not sources of NO₂ and also act as areas over which NO₂ concentrations disperse and dilute (and deposit), hence the a priori direction of effect of the green area variable is negative (Jim and Chen 2008). Coordinates and altitude were included to reflect physical geography influence and wind speed and direction proxies.

A stepwise multiple regression approach was used to select the potential predictor variables to maximise the adjusted percentage explained variance (R²) and minimise the root mean square error (RMSE) (ESCAPE 2008). First, all predictor variables were individually regressed against the dependent variable data (annual-mean NO₂ concentrations modelled by ADMS-urban at the 83 receptor sites). The predictor variable which explained the most NO₂ concentration variation (highest R²) and with a coefficient in the a priori-defined direction (SI table S5) formed the initial model. The univariate linear regression process was repeated with the remaining variables. A variable with the highest increase in adjusted R² was added into the initial model if it met the following criteria: (1) the increase in adjusted R² was greater than 0.01; (2) the coefficients of this variable and the variables already in the model were in accord with the a priori-defined direction. This process was repeated until no further variable satisfied the criteria. In the final step, variables with p-value greater than 0.1 were removed from the model starting from the variable with highest p-value.

Tests were then performed to check multicollinearity and influential observations. Multicollinearity in the variables was checked using Variance Inflation Factor (VIF). Variables with VIF values greater than three were removed from the model. Extreme values or many zero values in a variable data can skew the final model and this can be indicated by Cook's distance >1. In this study, no observation was removed. The model was validated using both the measurement and virtual site concentration data.

Calculation of total premature deaths from concentrations of NO₂ followed the methodology described by Walton et al (2015). We used the association with all-cause mortality of 2.45% (95% CI: 2.34%, 2.58%) per 10 μg m⁻³ elevation of NO₂ reported by Zhang et al (2011). This health risk coefficient was used since it was derived from data in Shenyang, a province in China. The number of deaths in 2017 in Guangzhou was 60 900 (Guangzhou Statistics Bureau 2018). The population data (100 m × 100 m) was resampled to the resolution of the concentration map (25 m × 25 m). The population-weighted average concentration (E) for NO₂ across the whole of Guangzhou was calculated as follows

$$\begin{eqnarray}&&E=\displaystyle \frac{1}{Pop}\displaystyle \sum _{i}{C}_{i}Po{p}_{i}\end{eqnarray} \tag{ 1 }$$

where ${C}_{i}$ and $Po{p}_{i}$ are the concentration and the number of people in each cell i of the concentration map.

The attributable deaths from exposure to ambient NO₂ in Guangzhou was calculated by multiplying the attributable fraction by number of all-cause deaths (equations (2)–(4)), where RR refers to relative risk

$$\begin{eqnarray}&&RR={1.0245}^{\left(E/10\right)},\end{eqnarray} \tag{ 2 }$$

$$\begin{eqnarray}&&{\rm{AF}}=({RR}-1)/{RR},\end{eqnarray} \tag{ 3 }$$

$$\begin{eqnarray}&&{\rm{Attributable}}\,{\rm{death}}={\rm{the}}\,{\rm{number}}\,{\rm{of}}\,{\rm{deaths}}\times {\rm{AF}}.\end{eqnarray} \tag{ 4 }$$

The ADMS-Urban model showed good statistical performance against the annual-mean NO₂ concentrations at the 11 monitoring sites in Guangzhou (figure 3(A)). The model explains 72% of the spatial variation in the measured NO₂, with an RMSE of 17.7 μg m⁻³, albeit with a tendency to overpredict NO₂ concentrations (MB = 11.2 μg m⁻³, NMB = 0.22). A likely explanation for the general overprediction is that the precise height of each monitoring site is unknown. The only information provided to us is that they range in elevation from 2 to 20 m, which is consistent with anecdotal information that pollutant monitoring sites in Chinese cities are often situated on tops of buildings, and thus at higher elevation (where NO₂ concentrations would be lower) than the receptor height used in the model. The greatest model overprediction at site 1354A (figure 3(A)) is likely explained by the fact that this site is located in a park in the city centre and is surrounded by dense vegetation whose effects on NO₂ concentration (through dispersion, deposition, and temperature reduction) are not sufficiently captured by the model. As the distance to nearest road for all 11 monitoring sites is greater than 5 m, the observed data can not verify the model for roadside locations where highest NO₂ concentrations are anticipated, but at the other end of the concentration range the model predicted well the background site 1355A which is located in an area of natural land cover (Maofengshan). It is also important to recognise that the observational data have uncertainty. Measurement errors have been reported previously in China but more recent literature indicates that data is more reliable since 2013 (Yuyu et al 2012, Ghanem and Zhang 2014, Stoerk 2016). The monitoring specification HJ654-2013 (Ministry of Environmental Protection 2018b) also specifies detailed calibration procedures for each instrument such as zero and scale noises, error of indication, zero and span drifts tests. However, overestimation of NO₂ concentrations using instruments based on the chemiluminescence method (TEI Model 42i from Thermo Fisher Scientific Inc., USA) ranged from 6% to 280% depending on the composition of the oxidation products of NO_x has been reported (Xu et al 2013). It has not been feasible to robustly quantify uncertainty for the NO₂ measurements in Guangzhou within the scope of this study, as documentation of the processes applied for data quality control and assurance (QA/QC) and the operating principles at each monitoring site is not publicly available.

Zoom In Zoom Out Reset image size

The final LUR model contained three variables (table 2): pop_5000 (the number of people within a 5000 m buffer) sq_dis_port (the squared distance to the nearest port), and all_len_25 (the length of all type of roads within a 25 m buffer). The three variables respectively represent the influence of people (indirectly related to residential, industrial, and traffic emission, and inversely related to green space), shipping emissions, and traffic emissions. Table 3 summarizes the results of hybrid LUR model evaluation against both the 83 virtual sites and the 11 monitoring sites and figure 3(B) shows the scatter plot for the LUR model predictions at the 11 monitoring sites.

Table 2.  The final LUR model variables and their coefficients and standard errors.

Variable	Definition	Coefficient	Standard error
Intercept	/	$33.09$	1.77
pop_5000	Number of people within a 5000 m buffer	$2.875\times {10}^{-5}$	$1.087\times {10}^{-6}$
$sq\_dis\_port$	Squared distance to the nearest port	$-9.543\times {10}^{-9}$	$1.832\times {10}^{-9}$
all_len_25	Length of all type of roads within a 25 m buffer	$6.362\times {10}^{-2}$	$1.261\times {10}^{-2}$

Against the 83 virtual sites, LOOCV evaluation shows the hybrid LUR model explains 96% of the spatial variation of NO₂ concentration with an RMSE of 5.64 μg m⁻³. The NO₂ concentrations at the monitoring site provide an additional evaluation of the hybrid LUR model independent of the NO₂ data used to construct the model (table 3 and figure 3(B)). For this plot, R² = 0.63 and MB = 9.08 μg m⁻³, which corresponds to a NMB of 18.2%. These statistical uncertainties seem reasonable for intra-urban modelling/mapping of a domain of this size and with the paucity of detailed emissions data to support modelling. Figure 3(B) shows that the hybrid model slightly overestimates NO₂ concentrations, which as discussed earlier is likely the consequence of the model predicting NO₂ concentration at 1.5 m height, which is lower than the height of most monitoring sites (2–20 m). Closer inspection of figure 3(B) shows that the model generally over-predicts more at locations with higher NO₂ concentrations which can be rationalised because these are the locations most affected by local traffic emissions for which there is insufficiently detailed input data. As noted above, it is also important to recognise that there is uncertainty in the observed values.

Figure 4 shows the spatial distribution of annual-mean NO₂ concentration across Guangzhou in 2017. Modelled concentrations vary between 21.5 μg m⁻³ in the most rural areas in the north of the domain to 99.7 μg m⁻³ in the most polluted district of Guangzhou city centre. The map indicates that, as expected, concentrations of NO₂ are highest on road links, particularly those with heavy traffic as shown on the expanded map of the city centre. The model NO₂ concentration show high values at the boundary of Nansha, which is likely due to shipping (SI figure S4). The population-weighted annual-mean concentration of NO₂ across Guangzhou is 52.5 μg m⁻³. Even this population-weighted concentration exceeds the Chinese air quality standard and WHO guideline of 40 μg m⁻³ (WHO 2006, Ministry of Environmental Protection 2012), but figure 4 indicates that, in large areas of Guangzhou, concentrations of NO₂ are substantially in excess of this value.

Zoom In Zoom Out Reset image size

The modelled number of deaths attributable to long-term exposure to NO₂ in Guangzhou in 2017 was 7270 (6960–7620; range based only on the 95% CI for the health risk coefficient). For comparison, the total number of deaths from road-traffic accidents in Guangzhou was 847 in 2015 (Guangzhou Municipal Public Security Bureau 2016). Approximately 60% of people in Guangzhou experienced NO₂ concentration above the China air quality standard/WHO guideline of 40 μg m⁻³. The air quality plan (2016–2025) published by the Guangzhou government to combat air pollution problems sets the target of meeting the NO₂ air quality standard of 40 μg m⁻³ by 2020 (Guangzhou Government 2017). Our model predicts that if all areas with NO₂ concentrations greater than 40 μg m⁻³ were reduced to the air quality standard, the number of lives saved would be 1900 (1820–1980) compared to the 2017 estimate.