Previously we developed a method of applying online measurement data from the continuous emission monitoring systems (CEMS, http://218.94.78.61:8080/newPub/web/home.htm, last access: October 2025) for emission estimation at the unit/plant level (Zhang et al., 2019b). With this basis, we have improved the emission estimation method to enable the stable and continuous acquisition of near-real-time emission data lagged by one month. For the small number of power-generating units without CEMS data, we assumed that their pollutant concentrations in the flue gas were at the average level of units with similar installed capacity (Tang et al., 2019). The emissions were calculated based on the mean hourly flue gas concentration of air pollutant obtained from CEMS and the theoretical flue gas volume of each unit/plant: 1Ei,j,day=Ci,j,month×ALj,month×Vj,m0×Pi,j,m,day2ALj=FmRm where E is the emission of air pollutant; i, j and m indicate the specific pollutant species, individual power plant or unit, and fuel type, respectively; C is the monthly average concentration in the flue gas; AL is the activity level (here monthly coal consumption); F is the monthly electricity generation for various fuels, as reported by NBS (2023); R is the fuel consumption rate for power generation, taken from Tong et al. (2021), V0 is the theoretical volume of flue gas produced per unit of fuel consumption (Zhao et al., 2010); P is the temporal profile of emissions (the daily to monthly emission ratio), based on the hourly pollutant concentrations and volume of flue gas for the month and specific day.

With its gradually expanding penetration, CEMS has become able to support near-real-time emission estimation for industrial plants (Tang et al., 2022; Bo et al., 2021). Given its varying coverage across sectors, we have developed a method that can stably estimate the near-real-time emissions at the plant level with a lag of one month. This method classifies industrial plants into three categories based on their CEMS coverage, as described below. 1.

Industrial plants with CEMS information. The method is similar to power plants:3Ei,j,day=Ci,j,month×ALj,month×Vi,j,k0×Pi,j,m,daywhere k denotes the industrial sector; AL is the activity level (here represents monthly product output) as reported by NBS (2023), and V0 is the theoretical volume of flue gas produced per unit of product output, which can be found in the technical specifications for the application of emission permits (MEE, 2021).

Daily vehicular emissions were estimated utilizing the International Vehicle Emissions model (IVE) combined with the Gaode live congestion index (Zhou et al., 2019; Kholod et al., 2016). The level of traffic congestion was indicated by the additional time incurred during a trip under congested conditions, expressed as a percentage relative to uncongested conditions (Huo et al., 2022). The Gaode congestion index is available for over 350 cities in China, with a temporal resolution of 5 min (https://report.amap.com/index.do, last access: October 2025). By integrating the congestion index with a Greenshield's traffic density model (Yang et al., 2019), we estimated the traffic volume which serves as a temporal allocation factor to calculate the daily emissions. This approach assumes that vehicular activity data (e.g., mileage and fuel consumption) are accessible, albeit typically with a lag in reporting, as such information is usually provided on an annual basis. Consequently, the near-real-time emissions can be estimated based on the daily variations of congested index and EFs compared to the previous year (Eq. 7): 7Ei,m,day=(Iday,year-1)×Iday,(year-1)2×EFi,m,day,year(Iday,(year-1)-1)×Iday,year2×EFi,m,day,(year-1) In Eq. (), EF represents the emission factor calculated by the IVE model. The input parameters of IVE, such as vehicle population by type, registration dates, fuel types, and emission standards, can be obtained from the transportation management departments of individual cities. These historical data can be extrapolated to the present date utilizing the vehicle survival curve, thereby bridging any gaps in the current information (Sun et al., 2020). Because official high-frequency traffic activity data are unavailable in near-real-time, we introduced I, the Gaode traffic congestion index, as a dynamic activity scaling factor. This index reflects the comprehensive traffic volume and operational status of the overall road network, allowing us to dynamically scale the baseline emissions into daily-scale trajectories. The index serves as a generalized proxy for total road network activity, and the same scaling factor was applied uniformly for all vehicle types. Although different temporal operational patterns might exist for various vehicle types (e.g., larger volume for trucks during nighttime or on specific freight corridors), obtaining the near-real-time activity information by vehicle type remains a challenge at the provincial level in China. The baseline EF for vehicles in Eq. (7) were derived using the IVE model, which comprehensively accounts for the influences of complex driving conditions, including vehicle speed and engine load. However, continuous recalculation of real-time and speed-dependent EFs on a daily, province-wide scale is computationally intensive and remains as a challenge. For the near-real-time estimation of traffic emissions, therefore, the EFs were treated as baseline constants for 2022, and we predominantly focused on the dynamic adjustment of activity levels and treated them as the primary driving factor for the daily emission fluctuations.

Off-road transportation was divided into five categories: construction machinery, agricultural machinery, marine, railway, and aviation. Emissions from construction machinery were estimated based on assumed daily utilization rates derived from the operating rates of construction sites (Shen et al., 2023; Huang et al., 2021). The daily usage of agricultural machinery was assumed to correlate with the application of nitrogen fertilizers from agricultural sources (see the description of agriculture as below). Emissions from railway, marine and aviation sources were estimated using data from passenger/cargo turnover, individual ports and commercial flights, respectively. These data were obtained from the China Entrepreneur Investment Club (CEIC) (https://www.ceicdata.com.cn, last access: October 2025), Marine Traffic (http://www.marinetraffic.com, last access: October 2025) and Flightradar24 databases (https://www.flightradar24.com/data, last access: October 2025) (Huo et al., 2022; Liu et al., 2020a).

We followed Shao et al. (2023) and developed a Bayesian hierarchical model to estimate daily heating energy consumption by fuel type, based on two primary factors influencing residential energy consumption: temperature and GDP. The daily temperature data were taken from ERA5 products provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) (https://cds.climate.copernicus.eu, last access: October 2025), while GDP from the national statistics published quarterly by the National Bureau of Statistics (http://www.stats.gov.cn/, last access: October 2025). For the months without GDP data, we assumed a linear relationship between GDP and the nighttime light index (Xu et al., 2024), and applied the National Polar-orbiting Partnership Visible Infrared Imaging Radiometer Suite (NPP-VIIRS, https://www.earthdata.nasa.gov/, last access: October 2025) provided by National Aeronautics and Space Administration (NASA) to extrapolate the GDP for those months. We applied the gridded population dataset (1km×1km) released by a database of the Chinese Academy of Sciences (https://www.resdc.cn/Default.aspx, last access: October 2025) for 2020. To account for the effect of large-scale population migration, we integrated the Population Migration Index (PMI) developed by Baidu (https://qianxi.baidu.com/#/, last access: October 2025). This index calculates the proportion of incoming migrants relative to the local population.

NH3 emissions from fertilizer use can be largely influenced by meteorological conditions, soil environment, and farming practices. In our previous study, we quantified NH3 emissions using dynamic EFs associated with those factors (Zhao et al., 2020b). In this study, we expanded the methodology and estimated NH3 emissions by using daily EFs. Regarding the baseline activity data for NH3 estimations, the information was systematically derived from official statistics. For livestock and poultry breeding, we utilized the year-end stock. For synthetic fertilizers, the application amount was calculated as the product of the city-level sown area of cropland and the provincial application rate per unit area obtained from national investigations. To convert these annual totals into dynamic near-real-time estimations, we integrated the temporal allocation of activity data with real-time meteorological conditions. Based on the regional farming database from the Ministry of Agriculture, we tracked the specific growing seasons of major crop types to determine the exact timing of basal dressing and top dressing. By combining the farming cycles with meteorological conditions and high-resolution soil pH databases, we generated the spatiotemporal pattern of NH3 emissions.

XGBoost model is an advanced and scalable machine learning framework based on gradient-boosted decision trees, widely recognized for its efficiency in handling structured data and modeling complex nonlinear relationships (Requia et al., 2020; Wang et al., 2023). XGBoost excels at processing high-dimensional spatiotemporal datasets, such as gridded emission inventories, by effectively capturing interactions among heterogeneous emission sources and temporal dependencies. Moreover, the inherent interpretability features facilitate seamless integration with explainable AI tools (e.g., SHapley Additive exPlanations (SHAP) to quantify the marginal contribution of each input feature to individual model predictions), enabling rigorous attribution analysis of air pollutant concentration variability (Zhao et al., 2025). To overcome the traditional “black box” limitation of tree-based machine learning models, we utilized the SHAP algorithm to interpret the XGBoost outputs. Based on the algorithm, we were able to explicitly attribute the day-to-day variations in ambient pollutant concentrations to precursor emission changes from specific sectors, thereby drawing physically meaningful and conclusions.

The SHAP value is calculated with following equation: 8yi=ybase+f(Xi,1)+f(Xi,2)+⋯f(Xi,n) where yi is the predicted value of the model for the ith sample; f(Xi,n) is the contribution of the nth eigenvalue in the ith sample to the final predicted value, with positive or negative representing that the eigenvalue makes the predicted value increase or decrease; and ybase is the baseline value of the predicted outputs for all types of predictions, representing the average prediction results for each category without the influence of any eigenvalue.

The XGBoost-SHAP modeling framework was implemented at the horizontal resolution of 3km×3km to capture the emission-concentration relationship. XGBoost regression models were independently trained for each grid cell. January and July were selected as typical months for ambient PM_2.5 and O3, respectively. Daily time series of 20 pollutant-sector combinations (4 pollutant (SO2, NOx, NMVOCs, PM_2.5) × 5 sectors (Power, Industry, On-road (Vehicles), Off-road, Residential) except for tiny On-road SO2, and agricultural NH3,) were set as predictors, and anthropogenic-driven variability of PM_2.5 or O3 concentrations as target variables. Similarly, the emission inputs were treated as anomaly (the difference between the current day's emissions and the moving average of 30 d around). A 10-fold cross-validation was applied (80 % training and 20 % testing), and the bias and correlation coefficient (R) were calculated to evaluate the model performance (Xiao et al., 2018).

SHAP values were calculated for each emission feature using the tree explainer algorithm, quantifying contributions of pollutant-sector combinations to variability of daily anthropogenic-driven concentrations. Note that SHAP values represented the deviation of individual predictions from the baseline expectation. Positive values indicated emission features that elevated pollutant concentrations above the baseline, while negative values indicated features that reduced concentrations below the baseline. Aggregation of daily SHAP values for various pollutant-sector combinations produced the daily-level contribution of total anthropogenic emissions to the changing ambient concentration, and the daily-level contributions could then be aggregated to the monthly level.

The total anthropogenic emissions of SO2, NOx, PM_2.5, NMVOCs, and NH3 in Jiangsu for 2022 were estimated at 246, 727, 298, 1186, and 377 Gg (Fig. S2), which were respectively reduced by 17 %, 33 %, 18 %, 7 %, and 11 % compared with those in 2019 (Gu et al., 2023). Note the emissions for multiple years (2015, 2019, and 2022) were estimated with a consistent methodological framework to make reasonbal interannual comparison. Our estimates indicated that the reduction rate of SO2 emissions was much lower between 2019 and 2022 than that at 53 % between 2015 and 2019. In particular, the emissions from the power sector were estimated to decline only 7 % during 2019–2022. The result confirmed that the abatement of SO2 emissions have been clearly decelerated following the full implementation of ultra-low emission retrofits, suggesting that the potential of further reduction of SO2 emissions for power sectors has become more limited. More energy structure adjustment instead of end-of-pipe controls is needed for the sector.

In contrast to SO2, the emissions of NOx and PM_2.5 were estimated to decline faster during 2019–2022 than 2015–2019. Industrial sectors contributed largely to these reductions, with the emission declining 27 % and 22 % for NOx and PM_2.5, respectively (Fig. S2). These reductions reflected expansion of intensified pollution control policies from power to other sectors, particularly the ultra-low emission standards implemented for steel (2019) and cement industries (2020) (https://sthjt.jiangsu.gov.cn/, last access: October 2025). By 2022, Jiangsu province had implemented ultra-low emission retrofits in over 80 % of iron and steel enterprises and approximately 60 % of cement clinker production lines (DEE, 2023). However, slower progress of emission controls in coking, glass, and chemical industries highlighted substantial emission reduction potential in these non-electrical industrial sectors. Meanwhile, the NOx emissions of transportation were estimated to decline by 41 % from 2019 levels (53 % for light-duty gasoline vehicles), driven mainly by the nationwide implementation of China VI vehicle emission standard and increasing penetration of renewable energy vehicles.

NMVOCs, as critical precursors of both secondary PM_2.5 and O3 formation, exhibited a slower decline in emissions and have emerged as the priority of emission controls in Jiangsu (Fig. S2). Industrial activities dominate NMVOCs emissions in Jiangsu, contributing 68 % of the provincial total emissions. It resulted from the heavy dependence of the province on chemical industries. For example, the province accounted for over 40 % of national pesticide active ingredient and dye production. Notably, more than 60 % of small-scale chemical enterprises persisted in utilizing solvent-based coatings, inks, and adhesives with high-VOCs content (Simayi et al., 2022; Hu et al., 2024). Furthermore, recent expansions in solvent consumption and chemical output within large-scale enterprises along the Yangtze River have largely offset the emission reductions through improvement of manufacturing and pollution control technologies (Li et al., 2019). Consequently, intensified emission controls should be urgently required for targeting key industrial sectors and critical regions for NMVOCs reduction. Agricultural NH3 emissions in Jiangsu have experienced a decline of 14 % during 2019–2022, primarily attributed to reduced nitrogen fertilizer usage. However, the absence of effective NH3 control measures prevented further substantial reduction of emissions for the sector (Zhou et al., 2023; Zhao et al., 2022).

Figure 2 show the daily variability of total and sectoral emissions of various pollutants (SO2, NOx, PM_2.5, NMVOCs, and NH3) in 2022, respectively (the time series of emissions (NOx as an example) for all the involved source categories are provided in Fig. S3). The results revealed distinct seasonal emission patterns of air pollutants driven by anthropogenic activities and/or meteorological conditions.

The emissions of SO2 and primary PM_2.5 followed the seasonal patterns of fossil energy consumption (Yun et al., 2021), with clear peaks in winter (from December to February) associated with the substantial coal combustion for residential heating and elevated industrial energy demand (Geng et al., 2021; Zhan et al., 2023). Regarding NOx, transportation has become the primary contributor to the emissions along with improved emission controls from the power and industrial sectors. Following the lifting of COVID-19 lockdown since June 2022, moreover, residents exhibited a strong desire to travel, which enhanced the emissions from transportation. Compared to the spring (from March–May), NOx emissions from transportation increased 12 % during the summer (from June–August), consistent with the elevated population mobility (Fig. S4). Additionally, the NOx emission peak in March reflected the resumption of industrial production and construction activities after the Chinese New Year. The area of construction for residential and commercial buildings increased 56 % from February–March, with these activities heavily dependent on diesel-powered machinery (Yang et al., 2015; Cliff et al., 2023). The NMVOCs emissions were the largest in summer. Enhanced volatilization of solvents and industrial chemicals by the warmer temperatures resulted in a 22 % growth of summer emissions compared to spring. Similar to NOx, the NMVOCs emissions in March rebounded with a 17 % growth compared to February, reflecting the resumption of coating, printing, and petrochemical industries. For NH3, the highest emissions in March and September were predominantly driven by the intensive spring sowing and autumn farming seasons. In contrast, although the total fertilizer amount decreased in summer (mainly limited to top dressing for specific crops like paddy rice), the high temperature in summer together with top dressing greatly elevated the NH3 volatilization rates, resulting in peak emission factors that kept the emissions at a high level.

Notably, the province has made great efforts on reducing emissions during the period with heavy pollution weather (DEE, 2022). The restriction measures included alternating operations of energy-intensive industrial plants, such as cement, steel, and glass production, in order to reduce the total production level and energy consumption during the period. Furthermore, industrial parks were required to temporarily shut down or to reduce the load of coal-fired boilers to mitigate regional precursor emissions under the unfavorable meteorological conditions. Compared to August 2022, mandatory restrictions on coal-fired boilers and industrial plants for September resulted in an 11 % reduction of coal consumption for major industrial sectors, leading to a decline of 7 %, 10 %, 15 %, and 12 % for anthropogenic emissions of SO2, NOx, PM_2.5, and NMVOCs, respectively. This demonstrated the effectiveness of pollution control measures conducted by the government on counteracting pollution episodes around August and September, despite persistent meteorological challenges (Wang et al., 2023). However, subsequent emission rebounds in winter for SO2 (+24% compared with those in Autumn) and PM_2.5 emissions (+19%) underscored the limitation of seasonal control strategies for combustion-derived pollutants, emphasizing the imperative for clean energy promotion to achieve sustainable emission abatement.

In April 2022, a great reduction in air pollutant emissions was estimated. Compared with March, the emissions of SO2, NOx, PM_2.5, and NMVOCs decreased by 11 %, 8 %, 6 %, and 12 % respectively. This abrupt decline was temporally associated with the COVID-19 induced lockdown implemented in Shanghai (28 March–1 June 2022). The lockdown substantially disrupted industrial production, transportation activities, and daily routines in neighboring Jiangsu Province. The results showed that short-term public health incidents exerted profound impact on air pollutant emissions (Zhang et al., 2024; Ma et al., 2023).

Based on the real-time geospatial information from the POI system (e.g., quarterly updated road networks, land use types, and monthly revised construction sites), we achieved the evolving spatial pattern of daily air pollutant emissions with a horizontal resolution of 3km×3km. Figure 3 presents the spatial distribution of daily average emissions of major sectors in Jiangsu Province for 2022. We selected NOx as an example to illustrate the sector heterogeneity. The NOx emissions from power, industrial, vehicle, off-road transportation and residential sources in Jiangsu were calculated at 144, 109, 247, 183 and 45 Gg respectively. Aviation emissions (less than 1 % of total NOx) were excluded due to their tiny contribution to the total emissions.

The spatial pattern of emissions was closely associated with corresponding anthropogenic activities. Agricultural machinery emissions were predominantly located in northern agricultural zones and coastal areas, correlating with the spatiotemporal distribution of farming activities. In contrast, emissions from other sources were more concentrated in the southern cities, especially along the Yangtze River with the most abundant power and industrial plants. The NOx emissions from five cities in southern Jiangsu (Nanjing, Suzhou, Wuxi, Changzhou, Zhenjiang) accounted for 59 % and 63 % of provincial power and industrial emissions, respectively. On-road transportation emissions demonstrated a strong dependence on the road network. Nanjing and Xuzhou, as critical national railway transportation hubs, contributed 24 % and 13 % of provincial NOx emissions from railways (Wang et al., 2016). In addition, Suzhou contributed 29 % of provincial marine emissions, attributed to its pivotal role in Yangtze River Delta inland waterway logistics (Shen et al., 2021). Unsurprisingly, the residential NOx emissions were closely correlated with the population density.

Figure 4 compares the monthly distributions of SO2, NOx, and PM_2.5 emissions estimated in this study with those in MEIC, as well as those of provincial averages of ambient concentrations of corresponding species obtained from the state-operating observation sites in Jiangsu. Due to the unavailability of MEIC for the year 2022, we used the result for 2020 instead.

For SO2 (Fig. 4a) and PM_2.5 (Fig. 4e), similar monthly variation patterns were found between emissions and observed concentrations in Jiangsu. The near-real-time emission estimates effectively captured the short-term fluctuations in anthropogenic activities, including the abrupt reduction in April associated with the COVID-19 lockdown and the seasonal change from the temporary pollution control measures in autumn. While ambient concentrations were influenced by meteorology and secondary formation as well, the relatively long atmospheric lifetimes of SO2 and PM_2.5 (typically several days) allow them to reflect the impact of primary emission variations. These results partly justified the capability of the approach to track the effect of changing anthropogenic activities on air pollutant emissions. In contrast, the highly reactive nature and shorter atmospheric lifetime of NOx resulted in a decoupling between its emissions and ambient concentrations. We found contrary monthly distributions between NOx emissions and the observed NO2 concentrations (Fig. 4c). The largest emissions were estimated in summer months but the lowest concentrations were observed for the same months across the year. This inconsistency likely resulted from the following factors. Increased transportation activity in summer, particularly mobility rebound after lockdown, elevated NOx emissions. Meanwhile, NO₂ was substantially consumed for O3 formation through rapid photochemical reactions under the intense solar radiation and high temperatures, and its atmospheric lifetime was reduced to merely a few hours. In winter, there was more NOx accumulation in the atmosphere with weaker photochemical reactions and reduced boundary layer heights (Ding et al., 2015; Wang et al., 2012).

Similar monthly distribution of emissions were found for the national (MEIC) and provincial emission estimates (this work), implying regular patterns of monthly anthropogenic activities could be captured by both inventories. Nevertheless, disparities existed in the overall emission totals and sector distributions between the two inventories. For instance, the contributions of industry to provincial emissions of SO2 and NOx were estimated at 45 % and 15 % in this work, greatly different from the MEIC estimation at 72 % and 41 %, respectively. These discrepancies might be attributed to that the national inventory (MEIC) for 2020 has not yet fully included the information of emission control technology upgrades (e.g., ultra-low emission retrofits) in the industrial sector. Taking the sintering process in the steel industry as an example, our facility-level estimations indicated that the average emission factors for SO2, NOx, and PM_2.5 were 0.143, 0.228, and 0.037 kgt-1, respectively, much lower than the recommended values of 1.34, 0.55, and 2.52 kgt-1 from the guidelines for development of national emission inventory (He et al., 2018).

Substantial discrepancies were revealed for off-road transportation of SO2 emissions. The provincial SO2 emission estimate from marine (12 877 Mgyr-1) were almost three times of that by MEIC (4690 Mgyr-1). As a major freight hub in the eastern coastal region of the country, Jiangsu Province played a pivotal role in marine transportation, and approximately 60 % of vessels utilized heavy oil with high-sulfur content as fuel (Dong et al., 2025). Application of national average EFs for the sector might lead to underestimation in emissions. Furthermore, the national inventory ignored the emissions from passing vessels at ports. Inclusion of such vessels would increase the SO2 emissions in the Yangtze River Delta region by a factor of 2.3 (Zhang et al., 2017). As power and industrial sectors have gradually completed ultra-low emission retrofits, marine emissions with less stringent controls may become more important in the future, requiring greater efforts on fuel quality improvement and stricter emission controls.

Figure 8 presents the contributions of the changing daily emissions to the monthly variability of PM_2.5 and MDA8 O3 concentrations based on the MLR model. The model performance was assessed with observed PM_2.5 and O3 concentrations (Fig. S11). The simulated concentrations were strongly correlated with observational data, with the correlation coefficient (R) of 0.79 for PM_2.5 and 0.88 for MDA8 O3. The validation indicated satisfying performance of MLR in capturing provincial air quality variability.

The anthropogenic-driven variability of PM_2.5 concentration was basically consistent with the temporal variation of estimated emissions. As shown in Fig. 8a, the abundant emissions in January resulted in a prominent enhancement of 12.7 µgm-3 for PM_2.5 concentration, followed by December (1.8 µgm-3) and June (1.6 µgm-3). In particular, the enhancement of June was driven largely by the post-pandemic economic recovery, as discussed in in Sect. 3.2. For most warm months (April–October, except June), negative impacts of anthropogenic activities on PM_2.5 level were found, ranging 1.1–4.2 µgm-3. Clear decline of PM_2.5 due to emission change was also found in February (5.5 µgm-3), resulting probably from the greatly reduced human activities (industry and transportation) during the Chinese New Year holiday. The PM_2.5 growth occurred during winter heating period highlighted the necessity of accelerating transition of clean household energy and improving management of industrial production after the short-term lockdowns.

The variation of anthropogenic emissions was found to elevate O3 concentrations in most months of the year, particularly for warm seasons (Fig. 8b). The enhancements during March-August ranged 0.8–3.8 µgm-3, suggesting the important role of human activities in aggravating O3 pollution. High temperature in summer promoted the emissions of temperature-dependent O3 precursors, particularly NMVOCs from various sources (Fig. 2d). In addition, the NOx emissions from certain sources were elevated in warm seasons, e.g., those from off-road machinery in the summer harvest season (Fig. 2a). The growing abundance of precursors, together with high temperature, enhanced the photochemical production rate of O3.

However, the anthropogenic emissions during winter demonstrated a net negative contribution to surface O3 concentrations (e.g., -6.2 and -2.4 µgm-3 for November and December, respectively), indicating a shift in the chemical regime of O3 formation. Although the NOx emissions were not enhanced in winter (Fig. 4), the weak photochemical production under low temperature and solar radiation made the NOx titration more dominating in O3 chemistry, primarily resulting in this net negative contribution. Simultaneously, reduced NMVOCs emissions and diminished photochemical activity restricted the efficiency of radical-driven O3 production. The resulting O3-depleting reactions overwhelmed potential formation mechanisms, leading to the estimated negative contribution from anthropogenic emissions. This pattern contrasted sharply with the net positive effect of anthropogenic activities in summer months, and underscored the complex season-dependent response of O3 level to the changing precursor emissions.

The impacts of anthropogenic emission fluctuations on variability of PM_2.5 and O3 concentrations were quantified by precursor and sector, with a machine learning framework integrating XGBoost and SHAP analysis. Derived from the 10-fold cross validation, the correlation coefficient (R) between machine learning prediction and observation reached 0.78 and 0.81 for daily PM_2.5 and MDA8 O3, respectively, suggested satisfying capability of the machine learning framework in predicting the anthropogenic-driven variability of PM_2.5 and O3 concentrations (Fig. S12).

Figure 9a and b illustrates the contributions of changing emissions from different pollutant-sector combinations to the variability of PM_2.5 concentration in January and that of MDA8 O3 in July, respectively. The temporal variability of PM_2.5 level attributable to anthropogenic emission changes was in general consistent with that of observed surface PM_2.5 concentration (Fig. 9a). For O3, there existed some discrepancy between the temporal distribution of anthropogenic-driven variability and observed concentration in summer. This discrepancy may be attributed to the substantial impacts of meteorological conditions and biogenic VOCs emissions on O3 formation (Gu et al., 2023).

Among all the pollutant-sector combinations, fluctuations in agricultural NH3 emissions accounted for 67.3 % of the variability of PM_2.5 concentrations in January, followed by off-road NOx (12.9 %) and residential PM_2.5 emissions (4.9 %). The contribution of NH3 emission variation significantly exceeded those of NOx (17.7 %), PM_2.5 (10.8 %), and SO2 (4.2 %), suggesting that Jiangsu may be transitioning to an NH3-rich regime following substantial reductions in SO2 and NOx emissions (Zhao et al., 2020b). Therefore, agricultural NH3 control has become the priority of the strategy design for PM_2.5 pollution alleviations, compared to traditional NOx abatement. The fluctuations in VOC-Industry contributed to 48.5 % of the variability of MDA8 O3 concentrations in July, followed by off-road VOCs (9.7 %) and NOx emissions (8.9 %). In total, the NMVOCs accounted for 69.7 % of the anthropogenic-driven variability of O3 concentration, exceeding the contributions from NOx (14.5 %), PM_2.5 (11.0 %), and SO2 (4.9 %). The positive contribution of NOx to MDA8 O3 indicated that the O3 formation mechanism in Jiangsu may be shifting from a VOCs-limited regime towards a transitional or NOx-limited regime. Regarding the sector contributions with various species aggregated, the agricultural emission fluctuations contributed most to anthropogenic-driven variability of PM_2.5 concentration (67.3 %, Fig. 9c), while industrial activities contributed most to that of O3 concentration (54.8 %, Fig. 9d). Notably, off-road transportation emerged as an important contributor to both pollutants (15.6 % for PM_2.5 and 24.4 % for O3), providing clear evidence for policy making of coordinating control of PM_2.5 and O3 pollution.