Uneven Progress in Urban Air Quality: Evidence From the Greater Bay Area and Implications for Metropolitan Management
Abstract:
Improving urban air quality requires attention to both regional trends and persistent differences between cities. This study examines how these patterns inform urban environmental management across the Guangdong–Hong Kong–Macao Greater Bay Area. Official observations from 23 monitoring stations across 11 cities during 2019–2025 were analysed using city comparisons, a continuity-screened sample of 18 stations, and monthly regressions with seasonal controls and heteroskedasticity- and autocorrelation-consistent covariance estimates. Economic associations were examined using matched air quality and gross domestic product (GDP) per capita data for 2024. Across all stations, annual mean PM$_{10}$, PM$_{2.5}$ and NO$_2$ concentrations declined by 18.4%, 23.8% and 25.0%, respectively, with reductions recorded in every city. However, the between-city coefficient of variation for PM$_{10}$ increased from 0.126 to 0.153, and the correlation between the 2019 and 2025 city rankings was 0.835. Monthly models estimated an annual PM$_{2.5}$ decline of 5.50%, whereas the ozone trend was −0.12% per year, with a 95% confidence interval of −3.21% to 3.07%. Higher GDP per capita was associated with lower PM$_{2.5}$ in 2024, but the relationship varied across pollutants. The official dust-adjusted PM$_{10}$ mean for 2025 matched the reported 2024 value, removing the apparent rebound at the published precision. These findings show that regional improvement alone provides an incomplete assessment of urban air quality progress. Urban environmental management requires pollutant-specific monitoring, continued assessment of intercity disparities and coordinated evaluation of exceptional pollution episodes. The study provides a reproducible basis for identifying monitoring priorities across cities and distinguishing regional fluctuations from persistent local conditions.
1. Introduction
Metropolitan air quality management requires more than evidence that a regional average has fallen. Authorities also need to know whether improvement is shared across cities, which differences persist and whether a change reflects an exceptional episode. Global research links pollution to economic development and poverty, but shows considerable variation across places and development settings [1], [2], [3]. The health implications of combustion-related pollution provide a further reason to connect urban development decisions with environmental management [4], [5]. A regional monitoring assessment can inform that connection if it distinguishes the spatial distribution of concentrations from the causes of change and the distribution of human exposure.
International evidence illustrates why this distinction matters. Individual-level research in the Netherlands identifies socioeconomic and ethnic differences in exposure, whereas assessments of European cities distinguish contributions from local and more distant emissions [6], [7]. These studies use population, spatial or source information that a small network of fixed stations cannot reproduce. Policy research consequently calls for explicit assessment of distributional outcomes and for exposure studies to account for the communities they seek to serve [8], [9]. Differences between city station means are relevant to the allocation of monitoring and analytical resources, but they do not establish population exposure inequality or environmental injustice by themselves.
Urban planning research connects these concerns to the location of activity, transport systems and the design of the built environment. Recent work sets out planning priorities for respiratory health, examines how urban form and transport interact with environmental outcomes, and identifies data needs for studies across large numbers of cities [10], [11], [12]. Research on European cities also documents opportunities to coordinate climate and air quality strategies while recognising their different policy objectives [13]. For metropolitan managers, the practical issue is how to use available observations to identify the scale at which further diagnosis or intervention is needed. It is not sufficient to rank cities without considering monitoring coverage and pollutant-specific processes.
The institutional setting is equally important. Polycentric governance recognises that public authorities can act at several scales, with opportunities for coordination, experimentation and learning [14]. The institutional collective action framework explains why fragmented authority creates externalities and transaction costs that can obstruct joint decisions [15]. Applied to air quality, these perspectives suggest a division between shared tasks, such as regional episode analysis, and local tasks, such as site investigation and implementation. They do not imply that a common monitoring trend measures the effectiveness of cooperation. Evidence on local instruments reinforces the need for context: reviews of low-emission and charging zones report variation across interventions and outcomes, while a port-area study connects source assessment with the distribution of local pollution burdens [16], [17], [18].
The Guangdong–Hong Kong–Macao Greater Bay Area combines nine cities in Guangdong Province with Hong Kong and Macao within a coastal airshed and a heterogeneous urban economy. It therefore presents both an environmental management problem and a coordination problem across jurisdictions. Studies of China’s clean air actions document strong particulate reductions and a more complicated ozone response [19], [20]. Regional work also identifies consequences for health and household expenditure [21], [22]. The value of a contemporary monitoring assessment lies in showing whether recent progress has been shared across the constituent cities and whether the remaining pattern calls for the same response for every pollutant.
Research on cooperation in the Greater Bay Area provides an institutional context for that assessment. Empirical studies associate regional collaboration with air quality improvement, describe unequal intercity cooperation networks and examine how cooperation relates to regional inequality [23], [24], [25]. National evidence also investigates collaborative pollution governance through quasi-experimental designs [26]. These studies address institutional mechanisms or policy effects. A monitoring-based comparison answers a different question: what changes are visible in the shared environmental record, and what additional evidence is required before those changes can be used to judge local performance? Keeping these questions separate is essential when municipal responsibilities coexist with common atmospheric influences.
Economic explanations also depend on the empirical design. The environmental Kuznets curve offers a framework for studying nonlinear income–pollution relationships, but its interpretation is sensitive to the pollutant, scale and model specification [27], [28]. Recent work traces distinct spatial patterns of development and environmental pressure, examines environmental information disclosure and economic outcomes, and evaluates green development pathways for the Greater Bay Area [29], [30], [31]. These findings justify considering economic context without assuming that growth automatically reduces every pollutant. A matched single-year comparison can describe an income gradient; a longitudinal causal claim requires a different design and a harmonised economic panel.
Ozone further complicates the interpretation of urban air quality progress. Sensor observations and source apportionment studies in the Greater Bay Area identify local formation, regional transport and meteorological conditions as relevant to episodes [32], [33]. Research on joint particulate and ozone control, urbanisation effects, exposure metrics and the geographical scale of intervention shows why ozone requires its own indicators and diagnostic methods [34], [35], [36], [37]. An annual ozone mean and a daily maximum eight-hour percentile serve different purposes. Neither can be replaced by evidence that particulate concentrations have declined.
The reliability of the monitoring record places a further constraint on interpretation. Changes in station composition can affect network trends, and weather can obscure or accentuate changes associated with emissions [38], [39]. Research on sensor placement also shows why coverage matters for the communities and locations represented in an assessment [40]. Exceptional mineral-dust transport provides another source of variation, as documented for an episode affecting Central Europe [41]. These studies motivate explicit checks of station continuity, source flags and event reporting. They do not justify imputing unobserved local concentrations or treating a seasonal time trend as meteorological normalisation.
This study examines the Greater Bay Area from 2019 through 2025 using the official regional monitoring network. The research asks whether observed improvement was shared across cities, whether absolute and relative differences narrowed, and whether particulate matter and ozone imply the same management priorities. A 2024 economic cross section provides context, while monthly models and a screened station sample test whether the main interpretation depends on endpoints, temporary years or network composition. The contribution is an updated and reproducible assessment that connects three distinct dimensions of metropolitan monitoring: regional progress, persistent city differences and common episodic variation. The resulting management recommendations are tied to those observations, with source attribution and causal policy evaluation reserved for evidence that can support them.
2. Study Area and Data
The study covered Guangzhou, Shenzhen, Zhuhai, Foshan, Huizhou, Dongguan, Zhongshan, Jiangmen and Zhaoqing in Guangdong Province, together with Hong Kong and Macao. Air quality observations came from the official Guangdong Hong Kong Macao Pearl River Delta Regional Air Quality Monitoring Network annual reports [42], [43], [44], [45], [46], [47], [48]. The network contained 23 stations, distributed unevenly across the 11 cities. Guangzhou had five stations, Hong Kong four, Jiangmen and Huizhou three each, Foshan two, and the remaining cities one each. These observations describe conditions at the reporting sites. Their averages should not be interpreted as population-weighted exposure or complete municipal air quality inventories.
The primary period was January 2019 to December 2025. This starting point was chosen for comparability. The reports state that, from 2019, gaseous concentrations were reported at a reference temperature of 298.15 K and a reference pressure of 101.325 kPa, whereas particulate concentrations were measured under ambient temperature and pressure. Earlier station tables were therefore not appended directly to the primary panel. The 2019 starting point also places the analysis within the contemporary institutional development of the Greater Bay Area. It does not define a treatment date or establish a counterfactual for the effects of regional integration.
Monthly and annual station values were extracted for PM$_{10}$, PM$_{2.5}$, NO$_{2}$, SO$_{2}$ and O$_{3}$. The primary annual panel contained 805 pollutant–station–year observations, corresponding to 23 stations, seven years and five pollutants. The monthly source grid contained 9,660 possible observations. Missing entries and source flags were retained in the data files. Annual values were taken directly from the annual columns; they were not reconstructed by averaging rounded monthly values. All five pollutants were expressed in μg/m$^3$. Carbon monoxide was excluded from the principal comparisons because its unit and concentration scale differ and it was not needed to answer the research questions.
Station names and physical continuity were treated separately. Wanqingsha and Nansha-HKUST, and Liyuan and Tongxinling, were reconciled as name changes documented by the reports. A conservative sample excluded Modiesha and its replacement Nanshadawen, Xijiao and its replacement Shixia, Zhudong, Duanfen and Tap Mun. These exclusions covered replacement or relocation, interrupted operation, or historical completeness concerns. The resulting sample contained 18 station identities in all 11 cities, with 630 annual observations. It is described as the fixed-station sample; this designation refers to the documented continuity screen and does not establish that every instrument and surrounding land use remained unchanged.
The full annual sample retained flagged values for a transparent comparison with the published station tables. A separate sensitivity analysis removed every flagged annual observation. Monthly models and co-movement calculations used only observed, unflagged monthly values. For the 18-station monthly regional series, between 16 and 18 eligible station values were available in each pollutant–month. No missing concentrations were interpolated. The source report, PDF page, station, date, pollutant and flag are recorded in the accompanying data.
Economic indicators were taken from Appendix 5 of the Guangzhou Statistical Yearbook 2025, which reports comparable Greater Bay Area indicators for 2024 [49]. Gross domestic product (GDP) per capita was used to describe development levels across the 11 cities in a single year (Table 1). Total GDP and the preceding-year real GDP growth index were retained as contextual records, not explanatory variables in the trend models. The common currency values were used as published, without an additional exchange-rate conversion. This economic comparison was restricted to 2024, matching the air quality year. The 2025 pollution observations were not paired with 2024 GDP and described as a synchronous relationship.
| Records | Coverage | Role |
|---|---|---|
| Annual station concentrations | 23 sites $\times$ 7 years $\times$ 5 pollutants = 805 | Regional and city comparisons |
| Screened annual concentrations | 18 sites $\times$ 7 years $\times$ 5 pollutants = 630 | Station continuity sensitivity |
| Monthly source grid | 23 sites $\times$ 84 months $\times$ 5 pollutants = 9,660 possible cells | Seasonality and co-movement |
| Monthly model series | 84 months per pollutant; 16--18 eligible sites per month | Conditional temporal trends |
| Economic statistics | 11 cities in 2024 | Exploratory development gradient |
| Ozone daily eight-hour metric | 23 station percentiles in 2025 | Episode benchmark comparison |
Two types of air quality benchmark were kept distinct. The national Grade II annual limits reported by the monitoring network provide a common reference for interpreting PM$_{10}$ and PM$_{2.5}$. The World Health Organization guidelines provide a health-oriented reference, including annual values of 15 μg/m$^3$ for PM$_{10}$ and 5 μg/m$^3$ for PM$_{2.5}$ [50]. These common research benchmarks were distinguished from the legal objectives applied by each jurisdiction. Annual mean ozone was used for trend comparison only. The 2025 station-level 90th percentiles of daily maximum eight-hour ozone were examined separately against the 160 μg/m$^3$ benchmark used in the network report (Table 1).
3. Analytical Methods
The analysis used three complementary aggregation levels. A station mean gave every eligible station equal weight. A city station mean averaged the eligible sites within a city, after which a regional city mean gave each of the 11 cities equal weight. The fixed-station mean used only the 18 screened station identities. These summaries answer different questions: the station mean describes the sampled network, while equal city weighting prevents municipalities with more monitoring sites from receiving greater weight in comparisons of urban jurisdictions. None is an area-weighted or population-weighted exposure estimate.
where, $C_{sy}$ denotes the source-reported annual concentration at station $s$ in year $y$, and $n_{iy}$ is the number of eligible stations in city $i$. Here, $\bar{C}_{iy}$ represents the annual mean concentration across eligible stations in city $i$, and $\bar{C}_y$ denotes the equal-city-weighted regional mean across the 11 cities. Pollutant subscripts are suppressed for readability. Endpoint changes were calculated as $100(C_{2025}/C_{2019}-1)$. Theil–Sen slopes were calculated from annual regional series as an additional descriptive check [51]. With seven annual observations, these slopes were used to assess direction and sensitivity, without treating their conventional independent-observation intervals as reliable measures of temporal uncertainty.
Absolute differences were measured by the population standard deviation (SD) of the 11 city station means, using a divisor of 11. Relative differences were measured by the coefficient of variation (CV), defined as this SD divided by the city mean. The same cities were included throughout. A decline in absolute dispersion and an increase in relative dispersion can occur together when the overall mean falls faster than the spread. Reporting both measures avoids describing every reduction in the mean as convergence between cities.
Spearman correlations between the 2019 and 2025 city values described the persistence of city rankings. These correlations do not measure atmospheric transport. Nor does a higher concentration establish weaker municipal management: meteorology, local sources and the placement of monitoring sites can all affect the observed value. The city comparisons were therefore interpreted as indicators of where further investigation and management attention are warranted.
For each pollutant, monthly means were calculated from unflagged observations at the 18 screened station identities. Its logarithm was regressed on elapsed time in years, calendar-month indicators, separate indicators for 2020 and 2021, and an April 2025 indicator. January was the reference month. The pandemic-year indicators accommodated large temporary changes during those years without assigning them a causal interpretation. The April 2025 indicator accommodated an unusual month, documented as dust-affected in the source report. For the gaseous pollutants it served as the same calendar anomaly control, not as a measure of a dust mechanism.
where, $\bar{C}_t^{\mathrm{mon}}$ denotes the mean of the available unflagged observations from the 18 screened station identities in month $t$. The variable $\tau_t$ measures elapsed time in years from January 2019. The indicator $I_{mt}$ equals 1 when observation $t$ corresponds to calendar month $m$ and 0 otherwise. Similarly, $I_{2020,t}$, $I_{2021,t}$ and $I_{\mathrm{Apr2025},t}$ equal 1 for observations in 2020, 2021 and April 2025, respectively, and 0 otherwise. The coefficients $\delta_m$, $\gamma_{20}$, $\gamma_{21}$ and $\kappa$ represent the corresponding regression coefficients. The intercept is $\alpha$, the temporal trend coefficient is $\beta$, and $u_t$ is the regression error.
Each main model contained 84 monthly observations and 16 coefficients, including the intercept. The estimated annual percentage change was $100[\exp(\beta)-1]$. Newey–West heteroskedasticity- and autocorrelation-consistent (HAC) covariance estimates used Bartlett weights, a 12-month lag window and the finite-sample correction $n/(n-k)$, where $n$ denotes the number of monthly observations and $k$ denotes the number of estimated coefficients, including the intercept [52]. Approximate 95\% intervals used a Student $t$ critical value with $n-k$ degrees of freedom. Windows of six and 18 months were examined as sensitivity checks. The models were also refitted after omitting 2020 and 2021. In that check, covariance weights used actual calendar distances, preserving the gap between 2019 and 2022.
These models estimate a conditional temporal pattern in the monitored regional mean. They are not meteorological normalisation models and contain no weather, emissions inventory or policy-treatment variables. A coefficient cannot therefore be assigned to a particular regulation, energy substitution or urban development intervention. The April indicator likewise measures a conditional anomaly and does not isolate the fraction of concentration attributable to mineral dust.
Monthly co-movement was examined separately from the long-term trend. For each city and pollutant, the mean for each calendar month over 2019–2025 was subtracted from its monthly series. The resulting seasonal anomalies were differenced across consecutive months. Spearman correlations between city pairs were then calculated from these changes, requiring at least 80 paired observations. There were 55 possible city pairs; the number meeting the completeness rule was reported separately for each pollutant. Deseasonalisation and differencing reduce the contribution of recurring seasons and smooth common trends, but do not remove shared meteorological influences. The resulting correlations indicate synchronisation, not a direction of pollution transmission or a causal spillover coefficient.
The economic comparison calculated Spearman correlations between 2024 GDP per capita and pollutant-specific city station means in 2024. Sensitivity checks excluded Hong Kong and Macao and, separately, omitted one city at a time. These were exploratory comparisons of 11, nine and ten cities, respectively. The matched cross section described the contemporary income gradient. It was not used to estimate an environmental Kuznets curve or an income effect over time.
4. Results
The all-station annual mean of PM$_{10}$ declined from 42.35 μg/m$^3$ in 2019 to 34.57 μg/m$^3$ in 2025, a reduction of 18.4%. PM$_{2.5}$ fell from 25.35 to 19.30 μg/m$^3$, or 23.8%, and NO$_2$ from 29.57 to 22.17 μg/m$^3$, or 25.0%. SO$_2$ declined from 7.26 to 5.74 μg/m$^3$. Ozone changed little at the endpoints, from 60.35 to 60.22 μg/m$^3$. The screened sample produced the same substantive distinction between declining particulate and nitrogen dioxide concentrations and an ozone series without a clear sustained decline (Table 2).
Pollutant | 2019 Mean | 2025 Mean | All Sites Change (%) | 18 Sites Change (%) | Unflagged Change (%) |
|---|---|---|---|---|---|
PM$_{10}$ | 42.35 | 34.57 | −18.4 | −18.5 | −18.4 |
PM$_{2.5}$ | 25.35 | 19.30 | −23.8 | −25.4 | −22.6 |
NO$_2$ | 29.57 | 22.17 | −25.0 | −25.3 | −25.0 |
SO$_2$ | 7.26 | 5.74 | −21.0 | −22.3 | −21.0 |
O$_3$ | 60.35 | 60.22 | −0.2 | −1.3 | −0.2 |
The concentration indices for the 18 screened station identities, with 2019 set to 100, are shown in Figure 1. Indices use the source-reported annual values and are calculated separately for each pollutant.

The principal endpoint reductions remained similar after excluding stations affected by documented continuity concerns. PM$_{10}$ decreased by 18.5% in this sample, PM$_{2.5}$ by 25.4%, NO$_2$ by 25.3% and SO$_2$ by 22.3%. Removing all source-flagged annual observations also retained the direction of these changes. The annual Theil–Sen slopes were negative for these four pollutants under each sample definition. For ozone, however, the slope was close to zero and changed sign between sample definitions. Its endpoint reduction of 0.2% in the full sample is consequently insufficient evidence of a persistent downward trend.
All 11 cities recorded lower PM$_{10}$, PM$_{2.5}$ and NO$_2$ station means in 2025 than in 2019. The 2025 PM$_{2.5}$ city station means ranged from 14.75 μg/m$^3$ in Hong Kong to 24.00 μg/m$^3$ in Zhaoqing. Jiangmen and Foshan recorded 22.33 and 21.50 μg/m$^3$, respectively. For PM$_{10}$, Hong Kong recorded 23.00 μg/m$^3$, compared with 43.00 μg/m$^3$ in Foshan and 40.00 μg/m$^3$ in both Jiangmen and Zhaoqing. These are comparisons among the network’s sampled locations; they do not establish the full distribution of exposure within each city (Table 3 and Figure 2).
| City | PM$_{\mathbf{10}}$ | PM$_{\mathbf{2.5}}$ | NO$_{\mathbf{2}}$ | O$_{\mathbf{3}}$ |
|---|---|---|---|---|
| Dongguan | 37.00 | 21.00 | 22.00 | 66.00 |
| Foshan | 43.00 | 21.50 | 28.50 | 60.00 |
| Guangzhou | 37.60 | 20.40 | 23.40 | 60.80 |
| Hong Kong | 23.00 | 14.75 | 27.00 | 56.75 |
| Huizhou | 34.67 | 19.00 | 13.33 | 59.67 |
| Jiangmen | 40.00 | 22.33 | 19.67 | 58.00 |
| Macao | 32.00 | 16.00 | 22.00 | 66.00 |
| Shenzhen | 30.00 | 16.00 | 16.00 | 64.00 |
| Zhaoqing | 40.00 | 24.00 | 28.00 | 58.00 |
| Zhongshan | 34.00 | 20.00 | 21.00 | 61.00 |
| Zhuhai | 32.00 | 19.00 | 20.00 | 66.00 |

For PM$_{2.5}$, the between-city SD declined from 3.98 to 2.75 μg/m$^3$ and the coefficient of variation fell from 0.155 to 0.141. Both measures indicate a narrowing of differences over the period. The pattern for PM$_{10}$ was different. Its SD changed only slightly, from 5.41 to 5.33 μg/m$^3$, while its city mean declined from 43.04 to 34.84 μg/m$^3$. The coefficient of variation consequently increased from 0.126 to 0.153. The region achieved lower PM$_{10}$ concentrations without a corresponding proportional reduction in differences between cities.
City rankings also remained persistent. Spearman correlations between the 2019 and 2025 distributions were 0.835 for PM$_{10}$, 0.870 for PM$_{2.5}$ and 0.929 for NO$_2$. This persistence indicates that broad improvement was accompanied by limited reordering of relative positions. Ozone displayed a lower coefficient of variation in 2025 than in 2019, but this narrowing occurred around a regional concentration that showed little endpoint improvement. Convergence in the distribution of one pollutant is therefore not sufficient to establish an acceptable level of that pollutant (Table 4 and Figure 3).
| Pollutant | SD 2019 | SD 2025 | CV 2019 | CV 2025 | Rank $\boldsymbol{\rho}$ |
|---|---|---|---|---|---|
| PM$_{10}$ | 5.41 | 5.33 | 0.126 | 0.153 | 0.835 |
| PM$_{2.5}$ | 3.98 | 2.75 | 0.155 | 0.141 | 0.870 |
| NO$_2$ | 6.75 | 4.54 | 0.231 | 0.207 | 0.929 |
| SO$_2$ | 1.92 | 1.85 | 0.269 | 0.301 | 0.581 |
| O$_3$ | 4.81 | 3.31 | 0.077 | 0.054 | 0.751 |

The monthly models with calendar-month controls estimated annual changes of −4.42% for PM$_{10}$, −5.50% for PM$_{2.5}$, −4.64% for NO$_2$ and −4.53% for SO$_2$. The corresponding 95% HAC intervals were −6.83% to −1.94%, −7.96% to −2.98%, −6.45% to −2.80% and −5.71% to −3.34%. For ozone, the estimate was −0.12%, with an interval of −3.21% to 3.07%. The evidence therefore supports a conditional decline in the monitored monthly regional means for the first four pollutants, while providing no comparable evidence of an ozone decline (Table 5 and Figure 4).
Pollutant | Main Rate (%) | Main 95% Interval | Omit 2020–2021 Rate (%) | Omission 95% Interval |
|---|---|---|---|---|
PM$_{10}$ | −4.42 | −6.83 to −1.94 | −4.47 | −6.99 to −1.89 |
PM$_{2.5}$ | −5.50 | −7.96 to −2.98 | −5.55 | −8.09 to −2.93 |
NO$_2$ | −4.64 | −6.45 to −2.80 | −4.69 | −6.56 to −2.79 |
SO$_2$ | −4.53 | −5.71 to −3.34 | −4.57 | −5.76 to −3.36 |
O$_3$ | −0.12 | −3.21 to 3.07 | −0.18 | −3.14 to 2.88 |

Changing the covariance window from 12 months to six or 18 months retained the direction of the particulate, NO$_2$ and SO$_2$ estimates, with their intervals remaining below zero. Omitting 2020 and 2021 produced annual estimates of −4.47% for PM$_{10}$, −5.55% for PM$_{2.5}$, −4.69% for NO$_2$ and −4.57% for SO$_2$. The ozone estimate remained close to zero at −0.18%. Thus, the principal distinction was not dependent on retaining the pandemic years. The fitted annual rates differ from annualised endpoint changes because they use every monthly observation and account for seasonal and temporary calendar differences.
In the 2024 cross section, GDP per capita was negatively associated with PM$_{2.5}$, with a Spearman correlation of −0.853. The association remained negative when Hong Kong and Macao were excluded, at −0.746, and across the leave-one-city-out comparisons, which ranged from −0.926 to −0.804. The PM$_{10}$ association was weaker, at −0.461. SO$_2$ also showed a negative association, while the corresponding NO$_2$ correlation was close to zero and the ozone correlation was positive (Table 6 and Figure 5).
Pollutant | All 11 Cities $\boldsymbol{\rho}$ | Mainland 9 Cities $\boldsymbol{\rho}$ | Leave-One-Out Minimum | Leave-One-Out Maximum |
|---|---|---|---|---|
PM$_{10}$ | −0.461 | −0.477 | −0.620 | −0.341 |
PM$_{2.5}$ | −0.853 | −0.746 | −0.926 | −0.804 |
NO$_2$ | 0.091 | −0.268 | −0.116 | 0.286 |
SO$_2$ | −0.863 | −0.778 | −0.896 | −0.817 |
O$_3$ | 0.420 | 0.576 | 0.272 | 0.659 |

These results describe different contemporary distributions of development and pollutants. They do not show that a rise in a city’s income would cause the observed concentration difference. The mainland-only NO$_2$ correlation changed from the slightly positive full-sample value to a negative value, illustrating the sensitivity of a small cross section to its institutional and spatial coverage. The ozone pattern also rules out treating higher GDP per capita as a universal indicator of cleaner air. A policy argument based on economic upgrading consequently needs to specify the emissions and pollutant pathways through which improvement is expected.
The city-weighted monthly series showed an unusually high PM$_{10}$ value in April 2025 compared with the same month in 2024 (Figure 6). The official 2025 report identifies a sand-dust episode during that month and gives a network annual PM$_{10}$ concentration of 35 μg/m$^3$ before its specified event exclusion and 33 μg/m$^3$ afterwards [48]. The latter equals the reported network value for 2024 at the report’s integer precision. The apparent annual rebound therefore requires an episode-aware interpretation. These official network figures are reported as supplied; they are not substituted for the independently calculated city and station means.

The April 2025 indicator accommodated this unusual month in the trend models. The published monthly means do not permit reconstruction of station-specific annual values with the dust episode removed. Observed station concentrations were therefore retained, and the official event-adjusted regional statistic was presented separately. This distinction preserves the environmental record while avoiding an unsupported allocation of the episode to individual cities.
Ozone required a different interpretation. Although its annual mean did not rise substantially between 2019 and 2025, four stations had 2025 90th percentiles of daily maximum eight-hour ozone above 160 μg/m$^3$: Nanshadawen, Zhudong, Huijingcheng and Nanchengyuanling, at 164, 165, 176 and 164 μg/m$^3$, respectively. These observations indicate continuing episode-related concerns under the network’s common benchmark. They are not a statement of statutory non-compliance under each jurisdiction’s separate objectives. The annual ozone mean and the eight-hour percentile answer different management questions.
The median correlation among eligible city pairs was 0.871 for changes in deseasonalised PM$_{10}$ anomalies and 0.824 for PM$_{2.5}$. Ozone also showed strong co-movement, with a median of 0.872. The corresponding medians were 0.450 for NO$_2$ and 0.203 for SO$_2$. The completeness rule retained 54 pairs for each particulate indicator and SO$_2$, 55 for NO$_2$ and 45 for ozone. These differences are consistent with pollutants responding differently to regional and local conditions, although the correlation analysis cannot distinguish common meteorology from transport, synchronous activity changes or other shared influences. The PM$_{2.5}$ pairwise matrix is shown in Figure 7; all five matrices are provided in the accompanying results.

The co-movement results support managing episodes at an urban-region scale while retaining local diagnostics. They do not identify an upwind city, quantify transported mass, or establish that one city’s policy affects another city’s concentration. Those claims require meteorological data, emissions inventories and a transport or identification model. The correlation evidence is used here to indicate that annual local performance and common regional fluctuations should be assessed separately.
5. Discussion and Urban Management Implications
The results distinguish the reduction of a regional mean from the narrowing of city differences. This distinction complements international work on the distribution of pollution and on explicit reporting of policy-related disparities [3], [8]. Every city improved on the three principal indicators of particulate matter and nitrogen dioxide between 2019 and 2025. Nevertheless, the relative distribution of PM$_{10}$ became more dispersed, and city rankings remained highly persistent. A regional average therefore provides an incomplete basis for allocating monitoring resources and technical support. Source inspection priorities require additional local evidence. Municipal comparisons need both a concentration measure and a measure of how the distribution has changed.
The PM$_{10}$ result is particularly relevant to performance reporting. A city can reduce its concentration and still remain relatively disadvantaged if other locations improve more quickly. Conversely, a low-concentration city may contribute to regional emissions or experience substantial within-city differences that its sampled mean does not reveal. Monitoring-based rankings should therefore trigger local investigation rather than automatically determine sanctions or rewards. Where persistent particulate disadvantages coincide with identified industrial, freight or construction sources, the appropriate response is a targeted source assessment followed by sector-specific action. The present results identify locations requiring attention but do not assign those sources to particular cities without additional evidence.
This interpretation complements the institutional concerns raised by recent studies of regional cooperation [23], [24], [25], [26]. Cooperation arrangements need mechanisms for following the distribution of outcomes, not only commitments to a common regional target. A practical annual review could report the city median and range alongside the regional mean, retain station-level values, and distinguish observed concentrations from any meteorologically normalised or event-adjusted indicators. Publishing both absolute and relative dispersion would make it harder for a favourable aggregate trend to obscure persistent local disadvantages.
The weak ozone trend means that a particulate-centred account of improvement would miss a continuing management problem. The 2025 eight-hour percentile results show that a stable annual ozone mean can coexist with elevated episodes at particular stations. Urban governments should consequently retain separate indicators for long-term particulate concentrations and ozone episodes. Combining them in a single unvalidated score would obscure the different atmospheric processes and intervention requirements.
Previous studies provide a basis for considering coordinated management of volatile organic compounds (VOCs) and nitrogen oxides (NO$_\text{x}$), together with the timing and spatial coverage of controls [32], [33], [34], [35], [36], [37]. The monitoring results presented here do not determine an optimal precursor reduction ratio. That decision requires chemical sensitivity analysis and emissions information. They do, however, show why reducing one pollutant cannot be taken as evidence that the broader urban air quality problem has been resolved. Further economic upgrading, transport electrification or fuel substitution should be assessed against its pollutant-specific outcomes and the relevant regional electricity and supply chains, rather than assumed to improve every ambient indicator equally.
For metropolitan planning, these findings favour joint review of industrial permits, freight activity and regional episode forecasts where source assessments establish their relevance. Municipalities can retain responsibility for local inspections and implementation while sharing analytical capacity for regional ozone formation and episode response. This division of tasks is consistent with the institutional collective action problem of coordinating decisions across fragmented jurisdictions [15]. The local effectiveness of a particular instrument still requires evaluation, as the international evidence on low-emission zones illustrates [16], [17].
The 2025 dust episode illustrates why accountability systems need an explicit event protocol. Using the observed network PM$_{10}$ mean alone would suggest deterioration from 2024 to 2025. The official event-adjusted regional figure gives a different interpretation at the published precision. Both statistics have a role: observed concentrations describe the air experienced at monitoring sites, while a carefully documented adjustment can inform assessment of controllable performance. Publishing one without the other would remove information needed by the public or the responsible authority.
A suitable protocol would retain the original observations, document the event dates and scientific attribution, state the adjustment method, and show how city comparisons change. It should be common enough across jurisdictions to support comparison, while respecting the standards applied by each authority. A dust label should not automatically exclude all high concentrations or be used to infer that local sources made no contribution. The present data establish the importance of this reporting distinction. Higher-frequency measurements and source attribution are required to reproduce the adjustment locally.
Strong co-movement in monthly changes provides a further reason to coordinate episode review. Several cities can experience an unusual month together even when their annual rankings differ. Shared forecasts and consistent public communication can address this common component, while municipal source assessments investigate persistent concentration differences. The distinction avoids assigning a region-wide shock entirely to individual local administrations and avoids allowing a common shock to conceal a persistent local problem.
The negative association between GDP per capita and PM$_{2.5}$ is consistent with a contemporary pattern in which some higher-income cities have lower sampled particulate concentrations. Its persistence after excluding the two special administrative regions makes it less dependent on their inclusion. However, industry composition, geography, regulation and monitoring placement remain possible explanations. A seven-year pollution panel paired with a single-year economic comparison cannot identify the separate contribution of these mechanisms or establish a development turning point.
This distinction is consequential for urban strategy. Assuming that growth will automatically solve pollution could delay action on precursor emissions or exposure hotspots. The pollutant-specific rank associations show that development level does not describe all monitored air quality outcomes in the same way. The economic comparison is best used to frame more specific questions about urban production, infrastructure and environmental management. It should not be used to claim an estimated income elasticity or the effects of a change in coal use.
Comparable longitudinal energy and industrial data would allow a later analysis to test these mechanisms directly. Such an extension would need to distinguish measured fuel consumption from industrial output proxies, use consistent city boundaries and accounting units, and address endogeneity and spatial dependence. It would allow economic mechanisms to be distinguished from the spatial and institutional differences described here.
Table 7 translates the observed patterns into proposed management tasks and specifies what remains to be established before an intervention is selected. The distinction between monitoring diagnosis and intervention assessment is central: a higher sampled concentration can justify further investigation, but does not identify an emission source or establish a municipality’s responsibility. The priorities below are therefore recommendations for the use of evidence, not evaluated policy effects.
| Observed Pattern | Proposed Task and Scale | Further Evidence Required |
|---|---|---|
| Persistent city ranks and a higher PM$_{10}$ coefficient of variation | Municipal site review with regional technical support | Local source inventories; additional spatial coverage |
| Declining particulate trends but no clear ozone decline | Regional precursor diagnosis and coordinated episode forecasts | Ozone sensitivity to VOC and NO$_\text{x}$ emissions; meteorology; emissions information |
| April 2025 materially affected annual PM$_{10}$ reporting | A shared episode protocol retaining observed and adjusted indicators | High-frequency observations; event attribution; adjustment method |
| Strong co-movement of particulate and ozone anomalies | Shared episode analysis and public communication | Weather and transport analysis separating common drivers |
| Pollutant-specific GDP per capita associations | City-level economic and infrastructure diagnosis | Harmonised longitudinal industry, energy and economic data |
The first limitation is spatial representativeness. Several cities are represented by only one station in this network, while other cities have multiple sites of different settings. Equal city weighting improves the balance between jurisdictions but cannot recover unsampled neighbourhoods or population exposure. The fixed-station check addresses documented continuity concerns; it cannot remove gradual changes in surrounding land use or all measurement uncertainty. The findings should therefore be interpreted as an assessment of the regional monitoring network rather than an exhaustive measure of environmental inequality.
Second, monthly seasonal controls and HAC intervals do not replace meteorological normalisation. The estimates retain the influence of unusual weather and other omitted time-varying conditions. With 84 months, the uncertainty intervals remain approximate. Third, the economic analysis is exploratory and based on a small cross section. Fourth, correlations between cities do not establish atmospheric spillovers, and no health burden was calculated. Finally, rounded annual source values limit interpretation of small year-to-year differences, especially at low concentrations.
The reporting approach is applicable to other metropolitan monitoring networks, subject to their coverage and measurement conditions. Other metropolitan regions can apply the same separation of network averages, city differences, monitoring continuity and exceptional episodes using their own data. The numerical trends and income associations should not be extrapolated beyond this case. The accompanying source records and code allow the present assessment to be updated when a further complete reporting year becomes available.
6. Conclusions
The Greater Bay Area achieved substantial monitored reductions in particulate matter and NO$_{2}$ from 2019 to 2025. The decline was shared by all 11 cities and survived checks using screened station identities and alternative monthly trend specifications. Progress was less uniform when judged by the distribution of remaining pollution: PM$_{10}$ relative dispersion increased, and city rankings for particulate matter and NO$_{2}$ remained persistent. Ozone showed no clear sustained monthly decline, and four stations exceeded the network’s 160 μg/m$^3$ eight-hour 90th-percentile benchmark in 2025.
These findings support pollutant-specific targets, monitoring of both regional progress and city differences, and a transparent distinction between observed pollution and event-adjusted performance. The 2025 dust episode shows why this distinction matters for local accountability. Higher GDP per capita was associated with lower particulate concentrations in the 2024 cross section, but the associations varied by pollutant and did not establish causation. Metropolitan environmental management should therefore combine coordinated episode response with locally grounded source assessment, rather than rely on a single concentration ranking or an assumed income–pollution pathway.
Conceptualization, D.L.T. and K.Z.P.; methodology, D.L.T. and K.Z.P.; software, D.L.T.; validation, D.L.T. and K.Z.P.; formal analysis, D.L.T. and K.Z.P.; investigation, D.L.T. and K.Z.P.; resources, D.L.T. and K.Z.P.; data curation, D.L.T.; writing—original draft preparation, D.L.T. and K.Z.P.; writing—review and editing, D.L.T. and K.Z.P.; visualization, D.L.T.; supervision, D.L.T.; project administration, D.L.T.; funding acquisition, D.L.T. All authors have read and agreed to the published version of the manuscript.
The data analysed in this study were obtained from publicly available official monitoring reports and economic statistics, all of which are cited in the reference list. The original data sources can be accessed through the corresponding references.
The authors declare no conflicts of interest.
