Optimizing ammonia emissions for PM_2.5 mitigation: environmental and health co-benefits in Eastern China

Ammonia (NH₃) is a key precursor of PM_2.5, contributing to the formation of secondary inorganic aerosols and playing a crucial role in haze events. However, current bottom-up emission inventories in China often underestimate NH₃ emissions, particularly with significant uncertainties in urban areas. This study developed a “top-down” iterative algorithm that integrates the IASI satellite observations with the WRF-Chem model to optimize bottom-up NH₃ emissions, and further quantified the impacts of source-specific emission reductions on PM_2.5 pollution. The result reveals that the updated NH₃ emissions in Eastern China for 2016 amounted to 4.2 Tg yr⁻¹, 27.3 % higher than prior estimations. The optimized NH₃ emissions peak in summer at 463.1 Gg month⁻¹, with agricultural sources accounting for 85 %, while winter emissions drop to 217 Gg month⁻¹ when the contribution from non-agricultural sources (e.g., industry, vehicle) significantly increases. The optimized NH₃ emission significantly improved the simulation of both total column and surface NH₃ concentrations, with improvements in magnitude (31 %–42 %) and variations (17 %–55 %). Sensitivity simulations show that a 30 %–60 % reduction in NH₃ emission led to decreases of 1.5–8.8 µg m⁻³ in city-level PM_2.5 concentrations and the potential effect of reducing non-agricultural emissions is comparable with that from agricultural sources. Furthermore, the NH₃ reduction positively impacts public health, resulting in a 6.5 %–10.3 % decrease in premature deaths attributed to PM_2.5 exposure. Our study evaluated NH₃ emissions from various sources in Eastern China, emphasizing the impact of reducing non-agricultural ammonia emissions on air quality and public health benefits.

1 Introduction

In recent years, China has continued to face significant challenges associated with PM_2.5 pollution (Geng et al., 2024; Lei et al., 2022). This issue adversely affects atmospheric environment via reducing visibility (Hu et al., 2021; Yang et al., 2022) and deteriorating air quality (Lei et al., 2024; Song et al., 2025), impacts climate change by altering radiation balance (Tang et al., 2025) and cloud formation (Gao et al., 2023; Yang et al., 2021), and poses substantial threats to human health (Du et al., 2024; Feng et al., 2016; Liu et al., 2025; Xiao et al., 2022; Zhu et al., 2025). Ammonia (NH₃), a key precursor of PM_2.5, neutralizes sulfuric acid (H₂SO₄) and nitric acid (HNO₃), leading to the formation of secondary inorganic aerosols (SIA), which contributes 19.4 %–55.0 % of the total PM_2.5 (Huang et al., 2014; Liu et al., 2022b; Wang et al., 2016; Wei et al., 2023; Zheng et al., 2015; Zhou et al., 2022). Reducing NH₃ emissions is a highly effective strategy for mitigation of PM_2.5 pollution (Bessagnet et al., 2014; Xu et al., 2022), particularly in light of the successful control of sulfur dioxide (SO₂) and nitrogen dioxide (NO₂) in China over the past decade (Li et al., 2023b; Wang et al., 2017; Zhang et al., 2019; Zheng et al., 2018).

The anthropogenic sources of NH₃ include agriculture, industry, power generation, transportation and residential activities. Numerous studies have estimated NH₃ emissions using a bottom-up approach, reporting emissions in China ranging from 9.7 to 13.2 Tg yr⁻¹ (Chen et al., 2021; Huang et al., 2012; Kang et al., 2016; Li et al., 2021; Ma, 2020). Among these sources, the agricultural (AGR) sector is identified as the dominant contributor nationwide, accounting for 75.0 %–94.5 % of total NH₃ emissions (Guo et al., 2020; Ma, 2020; Zhou et al., 2021). Additionally, some studies have highlighted that in densely populated regions, NH₃ from non-agricultural (non-AGR) activities, such as industrial production/slip, vehicles, and waste disposal, contributing up to 50 % of regional emissions and should not be overlooked (Chang et al., 2015, 2016; Chen et al., 2022; Feng et al., 2022; Pan et al., 2016, 2018b; Pu et al., 2020; Song et al., 2021; Sun et al., 2017; Van Damme et al., 2018; Wu et al., 2020). However, despite considerable progress, bottom-up estimates still exhibit considerable discrepancies and are often outdated, with a time lag of 1–2 years, mainly due to the lack of accurate and timely statistical data.

The uncertainty in the emission estimation further contributes to significant discrepancies, reflecting the range of results (1 %–50 %) reported in the literature, in assessing the impacts of NH₃ reduction on PM_2.5 level (Guo et al., 2018, 2024; Li et al., 2024; Liu et al., 2019, 2021, 2023; Pan et al., 2024; Zhang et al., 2022). Cheng et al. (2021) employed WRF-Chem simulations to demonstrate a 24.6 % reduction in PM_2.5 from the removal of AGR NH₃ emissions. Concurrently, Ti et al. (2022) determined that a 74 % decrease in AGR NH₃ resulted in a 34.9 % reduction in PM_2.5 in China.

To enhance the accuracy and reliability of bottom-up emission estimations, air quality monitoring satellites are increasingly regarded as valuable tools from a top-down perspective, offering advantages in both magnitude and timeliness (Chen et al., 2025, 2021; Guo et al., 2020; Jin et al., 2023; Qi et al., 2017; Xia et al., 2025; Zhou et al., 2021, 2017). Many studies have estimated optimized NH₃ emissions in China to be between 10.0 and 18.9 Tg yr⁻¹ by coupling chemical transport models, mass balance approaches, or machine learning techniques with various NH₃ measurements (satellite retrieval or ground monitoring). Some studies have also improved the description of the spatial and monthly variations of NH₃ emissions (Kong et al., 2019; Liu et al., 2022a; Paulot et al., 2014; Zhang et al., 2018, 2017). However, most top-down studies lack further investigation into the source-specific allocation of emissions based on the optimal total emission assessment (Fu et al., 2015; Sun et al., 2017; Zhang et al., 2024). Hence, a more comprehensive understanding of NH₃ emissions from diverse sources across varying seasons is needed to improve existing top-down inventories and enhance the scientific accuracy of NH₃ emission reduction assessments.

In this study, we used satellite and surface NH₃ measurements alongside the regional chemical model WRF-Chem to constrain bottom-up and source-specific NH₃ emission estimates over Eastern China, with the aim of more accurately assessing the impacts of NH₃ emission reductions from different sources on PM_2.5 concentrations. The paper is structured as follows: Sect. 2 describes the detailed methodology, Sect. 3 presents the simulated NH₃ with prior emission, Sect. 4 provides a top-down estimate of NH₃ emissions, and Sect. 5 demonstrates the direct correlation between NH₃ emission reductions and PM_2.5 concentration levels, as well as the associated health benefits. Our work differs from previous studies in that we constrain NH₃ emissions by sector, season, and region, and further assess the potential mitigation effects of NH₃ based on the optimized NH₃ inventory.

Figure 1 Simulation domains of the WRF-Chem model used in this study (left). Right panel illustrates the four research regions in Eastern China. Names and locations are labeled with different colors in this panel.

2 Methodology

2.1 Air Quality Model

In this study, the chemical transport model WRF-Chem v3.9.1 (Grell et al., 2005) was utilized to constrain the NH₃ emissions and to assess the impact of reduced NH₃ emission on PM_2.5 concentrations. Spatially, two nested domains were configured with horizontal resolutions of 54 km × 54 km and 18 km × 18 km. The outer domain covered entire China and the inner domain focused on Eastern China, characterized by intensive anthropogenic activities and elevated pollution levels (Pendergrass et al., 2025; Peng et al., 2025), including the Beijing-Tianjin-Hebei (BTH) region, Henan, Shandong, and the Yangtze River Delta (YRD) region (Fig. 1). The initial and boundary conditions of meteorological parameters were derived from FNL reanalysis datasets provided by the National Centers for Environmental Prediction (NCEP) of the United States (https://rda.ucar.edu/datasets/, last access: 28 October 2025). The initial and boundary conditions of chemical species were obtained from the global chemical transport model MOZART (Emmons et al., 2010). We conducted simulations for the entire year of 2016. The physical and chemical parameterizations describing sub-grid processes, such as radiation, microphysics, and gas-phase reaction schemes, are listed in Table S1 in the Supplement.

We adopted the anthropogenic emissions from the Multi-resolution Emission Inventory for China (MEIC, version 1.3) developed by Tsinghua University (Li et al., 2017; Zheng et al., 2018). Furthermore, biogenic emissions were calculated online using the Model of Emissions of Gases and Aerosols from Nature (MEGAN, version 2.0.4) (Guenther et al., 2006). Our numerical simulations also incorporated offline biomass burning emissions of various air pollutants, based on the wildfire model Fire Inventory from NCAR (FINN, version 1.5) (Wiedinmyer et al., 2011).

2.2 Satellite retrievals and surface measurements

We obtained the total column density of NH₃ from the passive satellite remote-sensing product of the Infrared Atmospheric Sounding Interferometer (IASI) (version 3.0, https://iasi.aeris-data.fr/nh3/, last access: 28 October 2025) as the observational constraint. The IASI is a Fourier transform spectrometer on board the Metop series of meteorological satellites, which circle the Earth in a polar Sun-synchronous orbit (Van Damme et al., 2014). Consequently, the satellite-based IASI instrument can cover the entire globe and provide measurements twice a day at 09:30 and 21:30 LST (local solar time). The IASI instrument detects infrared radiation in the spectral range from 645 to 2760 cm⁻¹ emitted by Earth's surface and atmosphere with a 12 km circular footprint at nadir. This radiation absorption range includes the NH₃ signal near 950 cm⁻¹.

The daily NH₃ column concentrations are categorized into level-2 satellite data and are developed based on the ANNI-NH₃ inversion algorithm without averaging kernels, as presented by Van Damme et al. (2017). Specifically, their retrieval algorithm derives hyperspectral radiation indexes (HRI) from the direct satellite spectrum detection, which is then converted into final NH₃ column concentrations using an artificial neural network technique (Whitburn et al., 2016). For better data quality, the present study removed NH₃ column concentrations associated with cloud cover of more than 10 %. Furthermore, we preprocessed the IASI NH₃ column concentration data through averaging all daily values to obtain a monthly mean value. Spatially, we mapped the original satellite product data to the grid cells of the WRF-Chem model for further comparison with those simulated NH₃ columns.

In addition, surface in-situ NH₃ measurements reported by Pan et al. (2018a) were collected for model evaluation. These ground-based measurements were summarized into the seasonal mean concentrations of NH₃ at 53 sites in China from September 2015 to August 2016.

Additionally, surface meteorological data, including air temperature, relative humidity and wind speed was obtained from China Meteorological Administration website (https://data.cma.cn/, last access: 28 October 2025) to assess the meteorological simulations over the study region. Air pollutant concentrations associated with NH₃ (such as PM_2.5, NO₂ and SO₂) from public website of the Ministry of Ecology and Environment (MEE) of China (https://air.cnemc.cn:18007/, last access: 28 October 2025) were also derived for evaluation. Furthermore, speciated inorganic aerosol data from a representative site in Beijing were collected to evaluate the model's capacity in characterizing the formation of secondary inorganic aerosols (Tan et al., 2018). The complete information of the in-situ measurements used in this study is available in Tables S2–S4.

Figure 2 Prior and posterior NH₃ emissions from agricultural and non-agricultural sectors in the study region. The red numbers show the total emissions.

3 NH₃ simulations with bottom-up emissions

We applied the bottom-up NH₃ emissions from MEIC (Li et al., 2017; Zheng et al., 2018) to drive the prior simulation. As shown in Fig. 2, the prior NH₃ emission amounted to 3.3 Tg yr⁻¹ in Eastern China, among which 93.0 % emission is from AGR sources and the other 7.0 % emission is from non-AGR sources. The largest emissions are recorded in July at 366.8 Gg month⁻¹, while the smallest emissions are recorded in January at 206.5 Gg month⁻¹ (Fig. S1 in the Supplement).

We compared the prior model results with IASI NH₃ column concentration and surface NH₃ volume concentration observations. The detailed method for calculating NH₃ total column concentrations and surface volume concentrations from WRF-Chem is provided in Sect. S1 in the Supplement.

To quantitatively describe model performance, we adopted three statistical metrics, including root mean squared error (RMSE, 0 to +∞), index of agreement (IOA, 0 to 1) and mean fractional bias (MFB, −2 to 2) (Huang et al., 2021). The IOA quantifies the overall model skill, where a value of 1 indicates a perfect match and 0 denotes complete disagreement. The MFB diagnoses systematic model bias, where positive values indicate overestimation, negative values indicate underestimation, and 0 signifies no average bias. The RMSE represents the average model error in the same units as the variable under evaluation, with lower values indicating better performance. They were calculated following Eqs. (1)–(3), where C represents the concentration of the target pollutant (e.g., NH₃ total column or surface concentrations), and subscripts s, o and N represent simulations, observations, and the number of samples, respectively.

$$\begin{array}{}\text{(1)}& {\displaystyle }& {\displaystyle }\text{RMSE}=\sqrt{{\displaystyle \frac{{\sum }_{i=\mathrm{1}}^N(C_m-C_{\mathrm{o}}{)}^{\mathrm{2}}}{N}}}\text{(2)}& {\displaystyle }& {\displaystyle }\text{IOA}=\mathrm{1}-{\displaystyle \frac{{\sum }_{i=\mathrm{1}}^N{\left(C_{\mathrm{s}}-C_{\mathrm{o}}\right)}^{\mathrm{2}}}{{\sum }_{i=\mathrm{1}}^N{\left(\left|C_{\mathrm{s}}-\stackrel{\mathrm{‾}}{C_{\mathrm{o}}}\right|+\left|C_{\mathrm{o}}-\stackrel{\mathrm{‾}}{C_{\mathrm{o}}}\right|\right)}^{\mathrm{2}}}}\text{(3)}& {\displaystyle }& {\displaystyle }\text{MFB}={\displaystyle \frac{\mathrm{1}}{N}}{\sum }_{i=\mathrm{1}}^N{\displaystyle \frac{(C_{\mathrm{o}}-C_m)}{\left(\frac{C_{\mathrm{o}}+C_m}{\mathrm{2}}\right)}}\end{array}$$

As shown in Table S5, the annual average of NH₃ total column concentrations is simulated to be 17.4 × 10¹⁵ $\mathrm{molec}.{\mathrm{cm}}^{-\mathrm{2}}$ for Eastern China, with a 61 % underestimation of MFB compared to the observations from IASI satellite retrievals (29.0 × 10¹⁵ $\mathrm{molec}.{\mathrm{cm}}^{-\mathrm{2}}$). The IOA between observations versus simulations is 0.72. The seasonal simulations of NH₃ concentrations also exhibit significant discrepancies with observations, especially in spring. Specifically, the simulated NH₃ total column concentration in Eastern China is only 13.2 × 10¹⁵ $\mathrm{molec}.{\mathrm{cm}}^{-\mathrm{2}}$ in spring, with concentration in 67.5 % of the study region being underestimated by more than 50 %. These discrepancies are evidently exhibited in Fig. 3. Most simulated NH₃ total column concentrations are underestimated by more than 30 % compared with the observed values by satellite with the associated RMSE exceeding 10 × 10¹⁵ $\mathrm{molec}.{\mathrm{cm}}^{-\mathrm{2}}$.

Figure 3 Scatter plots of the prior and posterior NH₃ total column data versus IASI retrievals. Each point represents prior (or posterior) data for a specific season and a specific region. Circles, triangles, rhombuses, and rectangles correspond to the BTH, Henan, Shandong, and YRD regions, respectively. Orange and blue markers represent a prior and a posterior data, respectively. The red box indicates the performance area, with a model error within ± 30 % and an RMSE below 10 (× 10¹⁵ $\mathrm{molec}.{\mathrm{cm}}^{-\mathrm{2}}$).

As illustrated in Fig. 6, satellite-based observations reveal that the spatial high-value areas of NH₃ column are located at the junction of Henan, Shandong, and Hebei provinces. In contrast, the prior modeling results show that NH₃ column densities are more concentrated in Henan. This indicates a clear discrepancy in the spatial distribution of NH₃ column densities between the prior simulations and the observations.

Additionally, the comparison between the simulated and observed surface NH₃ volume concentrations also indicates a notable underestimation (Fig. S2). The mean simulated surface NH₃ volume concentration over the study region is 6.3 µg m⁻³, which is only half of the observation value (12.7 µg m⁻³), with an IOA of 0.57 and an MFB of −61 %, respectively (Table S5).

4 Top-down estimates of NH₃ emissions

4.1 Iterative algorithm for NH₃ emission estimation

We utilized an iterative algorithm (Fig. 4) to update the prior NH₃ emissions from different sources constrained by IASI observations. This process was carried out in January, April, July, and October in 2016 to represent four seasons. The posterior emission inventory derived for each representative month was then applied to all three months within its corresponding season to generate the full 12 month posterior inventory. This representative-month approach was adopted to allow for a robust validation against the full 12 month period, with the remaining eight months serving as an independent dataset, and to manage the substantial computational cost of the iterative process. We compared the prior simulation results with satellite retrievals and discussed the performance of prior emissions in detail in Sect. 3. Furthermore, we conducted a series of sensitivity simulations to obtain prior simulated NH₃ from disparate sources and which were then fed into the iterative algorithm along with satellite data for calculation. In each iterative calculation, the monthly average satellite-derived NH₃ column concentration served as the target, and multiple linear regression (MLR) was applied to calculate the corresponding regression factors for AGR and non-AGR emissions (Fig. S3). This separation of sectors by MLR is effective because their respective spatial distributions are distinct and largely uncorrelated (r = 0.35). Here, we take the i iteration in k month, j region as an example to calculate the regression factors, and the formula is as follows:

$$\begin{array}{}\text{(4)}& \begin{array}{rl}{\text{TA}}_{\text{satellite}}^{j,k}-{\text{SA}}_{\text{transport}}^{j,k}=& {\mathit{\alpha }}_i^{j,k}\times {\text{SA}}_{{\text{agriculture}}_{i-\mathrm{1}}^{j,k}}+{\mathit{\beta }}_i^{j,k}\\ & \times {\text{SA}}_{{\text{non-agriculture}}_{i-\mathrm{1}}^{j,k}}\end{array}\end{array}$$

where, ${\text{TA}}_{\text{satellite}}^{j,k}$ denotes the monthly average of total NH₃ column density retrieved from the IASI satellite data, and ${\text{SA}}_{\text{transport}}^{j,k}$, ${\text{SA}}_{{\text{agriculture}}_{i-\mathrm{1}}^{j,k}}$ and ${\text{SA}}_{{\text{non-agriculture}}_{i-\mathrm{1}}^{j,k}}$ stand for the simulated total column concentration of NH₃ contributed by outside transport, AGR emissions, and non-AGR emissions, respectively. We clarified this NH₃ concentrations contributed by different pathways by conducting sensitivity experiments with the WRF-Chem model (Table 1).

Table 1 List of sensitivity tests for optimized iterative algorithm.

Figure 4 Visualization of the workflow in this study.

In each experiment, we zeroed out AGR emissions, non-AGR emissions and regional external emissions to obtain the corresponding NH₃ column concentrations. The ${\text{SA}}_{{\text{agriculture}}_{i-\mathrm{1}}^{j,k}}$, ${\text{SA}}_{{\text{non-agriculture}}_{i-\mathrm{1}}^{j,k}}$, and ${\text{SA}}_{\text{transport}}^{j,k}$ are calculated by subtracting A_blank from A_agr, A_non-agr, and A_transport, respectively. Here, symbols A represent the total simulated NH₃ column concentrations that result from each of the sensitivity simulations listed in Table 1. Specifically, the modeling case A_blank refers to a simulated NH₃ total column in which all anthropogenic emissions within the study domain were zeroed out. The purpose of this simulation was to establish background concentrations, which represents the influence of the chemical boundary conditions provided to our model domain.

Furthermore, the MLR approach provided regression coefficients ${\mathit{\alpha }}_i^{j,k}$ and ${\mathit{\beta }}_i^{j,k}$, which function as scaling factors, respectively correspond to AGR and non-AGR NH₃ emissions in month j from region k, within the i iteration. To ensure the statistical robustness of the regression equation, we need to correct for this regression coefficient. The biases between the model simulation and the satellite retrievals were calculated as $D_i^{j,k}$. Specifically, it is the difference between the mean simulated column and the mean satellite retrieval, divided by the mean satellite retrieval. We considered the residuals of the MLR approach, the goodness of fit and $D_i^{j,k}$, and obtained the judgment coefficient $K_i^{j,k}$. The regression coefficients with excessive residuals, defined as cases where the 95 % confidence interval of the residual does not contain zero, are removed to increase credibility. Concurrently, the goodness of fit of the regression is calculated as the coefficient of determination (R-square, R²). To maintain algorithm stability, regressions with an R² less than 0.3 are deemed invalid and excluded from the emission update, as they exhibit insufficient explanatory power (indicating > 70 % unexplained variance) and introduce destabilizing noise into the adjustments. We further use it to make a trade-off for the regression coefficient. If a regression is valid, the adjustment factors a and b are set to the new regression coefficients; if invalid, the factors are kept unchanged from the previous iteration. The updated emissions for the next iteration are then calculated by multiplying the emissions from the previous step by these adjustment factors. Finally, the entire process is iteratively repeated, a framework that captures the overall non-linear atmospheric response by combining the dynamic simulation of non-linear chemistry within each WRF-Chem step with the collective behavior of multiple iterations. The iteration concludes when the mean bias between the simulated values and observations is less than 30 %, a criterion chosen to represent a significant improvement over the large prior bias while falling within the range of widely accepted model performance benchmarks.

4.2 posterior NH₃ emission estimates

The top-down constrained results (posterior) indicate that the annual NH₃ emission in Eastern China has been updated to 4.2 Tg yr⁻¹, representing a 27.3 % increase compared to the prior value (Fig. 2). The posterior AGR emissions increased slightly, from 3.0 to 3.1 Tg yr⁻¹, but the high-emission regions shift from Henan to Shandong, Jiangsu and northern Anhui (Ren et al., 2023). The posterior non-AGR emissions show a significant increase, from 0.2 to 1.1 Tg yr⁻¹, particularly in urban regions along the Yangtze River, as well as in southern BTH, central Shandong and northern Henan (Fig. S4). Analysis of emission inventories (An et al., 2021; Hoesly et al., 2018; Li et al., 2021, 2017; Ma, 2020; Wu et al., 2024) reveals that residential activities and waste disposal are dominant sources of non-AGR NH₃ emissions, particularly in densely populated regions (Fig. S5). In multiple iterations, the framework optimizes the relative mix of the two sources to better match the observed spatial patterns. For instance, the spatial correlation between model and observation in Henan increased from 0.47–0.58 (prior simulations) to 0.64–0.90 (posterior simulations).

Figure 5 Posterior emission characteristics. (a) Contribution from regional emission sectors. (b) Comparison of the posteriori and prior emissions (unit: Mg) in study region.

In terms of seasonality, as shown in Fig. 5, the posterior NH₃ emissions are highest in summer, with a total of 463.1 Gg month⁻¹, followed by spring (442.4 Gg month⁻¹), largely due to fertilizer application (Li et al., 2021; Lu et al., 2025; Ren et al., 2025), and lowest in winter (217.4 Gg month⁻¹). The seasonal variations in the posterior emissions is the net result of complex adjustments in both the AGR and non-AGR sectors. At the specific-source scale (Fig. S6), AGR NH₃ emissions show similar seasonal patterns with the total NH₃ emissions, higher in summer and spring. In contrast, non-AGR NH₃ are highest in winter and fall because fossil fuel combustion-related emissions are higher in cold season, while the lowest emissions occur in summer. In addition, the ratio of AGR and non-AGR NH₃ emissions significantly varies across different regions. The contribution of non-AGR NH₃ emissions range from 18.8 % to 35.8 %, which is higher than the proportion in the prior inventory (Fig. 5a). This shift can be attributed to the increased relative importance of fossil fuel combustion-related emissions under high PM_2.5 loadings, which in turn promote higher NH₃ emissions from these sources (Pan et al., 2018b). Meanwhile, AGR NH₃ emissions are relatively inactive in winter due to unfavorable meteorological conditions. Similar high fractions of non-AGR emissions have also been reported in other studies (Feng et al., 2022; He et al., 2021).

Table 2 NH₃ emission estimates in recent studies.

Table 2 compares the results with related studies focused on NH₃ emission estimates. Overall, the estimated NH₃ emission in this study is comparable to the estimates of the other studies based on both “top-down” and “bottom-up” approaches. In similar years and regions, the discrepancy between the estimates of this study and other studies ranges from 1.0 % to 19.6 %. The slight discrepancy can be partially explained by our estimate being a conservative lower bound, a consequence of the residual gap remaining with satellite retrieval. Additionally, uncertainties from the model's chemical mechanisms and the influence of nearby grid transport also contribute to this gap, but the overall impact on the final estimate is limited. Furthermore, the seasonal distribution of NH₃ emissions in this study aligns with the findings of previous studies (Kong et al., 2019; Liu et al., 2024; Zhang et al., 2018; Zhao et al., 2020).

In terms of sectors, other studies have indicated that the contribution of NH₃ emissions from AGR sources is more than 80 %, using the bottom-up approach (Chen et al., 2021; Huang et al., 2012; Kang et al., 2016; Li et al., 2021). The relatively small proportion of non-AGR emissions is likely due to overlooked industrial (e.g., NH₃ slip and indirect emissions) (Chen and Wang, 2025; Chen et al., 2022; Wei et al., 2022) and residential sources (e.g., from waste) (Shao et al., 2020), combined with unrepresentative transportation emission factors (Sun et al., 2017; Zhang et al., 2021). This study, however, reveals a proportion of 74.4 % for AGR emissions, thereby emphasizing the contribution of non-AGR emissions. Concurrently, the eastern developed industry is expected to exhibit an increase in the proportion of NH₃ emissions from non-AGR sources when compared to the national average. Our work attempts to quantitatively disentangle the emissions from AGR and non-AGR sectors directly within our top-down framework and facilitates a more comprehensive capture of neglected non-AGR sources.

It is important to note that discrepancies in results between studies may be attributable to methodological differences (e.g. the sensitivity of the top-down approach to target data selection) and uncertainty in the underlying data. For instance, the NH₃ emission estimated by Paulot et al. (2014) using the mass balance method based on ammonium wet deposition fluxes is significantly lower than that in other studies, which may be attributed to its fewer observation sites in China. These discrepancies underscore the necessity to enhance the reliability of NH₃ observations in forthcoming studies, with the objective of enhancing the precision of the estimates.

Figure 6 Distributions of NH₃ total column from prior simulation, posterior simulation and satellite retrieval in different seasons.

4.3 Simulated NH₃ with top-down emissions

Figure 6 compares the spatial distributions of NH₃ total column density from satellite retrievals, prior simulations and posterior simulations. The annual mean simulated NH₃ total column density improved from the prior result of 17.4 × 10¹⁵ $\mathrm{molec}.{\mathrm{cm}}^{-\mathrm{2}}$ to a posterior value of 23.7 × 10¹⁵ $\mathrm{molec}.{\mathrm{cm}}^{-\mathrm{2}}$ , with an increase of 35.9 %, and is closer to the observed value of 29.0 × 10¹⁵ $\mathrm{molec}.{\mathrm{cm}}^{-\mathrm{2}}$. IOA and MFB between the posterior simulations versus measurements are 0.9 and −30.0 %, respectively. Figure 3 also shows the improvement in model performance. More than 80 % of the points fall in the range where the simulation-to-observation ratio is between 0.7 and 1.3 and the RMSE is less than 10 × 10¹⁵ $\mathrm{molec}.{\mathrm{cm}}^{-\mathrm{2}}$. A more consistent seasonal distribution can be obtained in a posterior simulation, with associated temporal MFB of NH₃ column density on the seasonal scale is reduced from −53 % (prior) to −24 % (posterior). Simultaneously, the spatial distribution pattern of posterior simulation is more identical to the characteristics revealed by satellite-based observations (Fig. 6). The spatial MFB is also decreased from −52 % (prior) to −20 % (posterior), with an increase in spatial correlation coefficient from 0.79 to 0.92. The improvement is especially notable in the BTH region, where the simulated NH₃ column densities are doubled. In summary, the posterior simulation improves the agreement between the simulated NH₃ column concentrations and satellite observations in both overall magnitude and spatial distribution, although some deviations remain, particularly in the colder seasons. These can likely be attributed to methodological limitations, such as the inherent tolerance of our 30 % iterative stopping criterion and potential inconsistencies from aggregating monthly optimizations to a seasonal scale.

A similar improvement is also witnessed in the modeling of surface NH₃ concentrations, which were evaluated against in-situ measurements from 13 sites reported by Pan et al. (2018a) for the 2015–2016 period (Table S2). The posterior simulation significantly improves the annual mean, increasing the surface concentration from 6.3 µg m⁻³ (prior) to 9.4 µg m⁻³ (posterior), much closer to the observed average of 12.7 µg m⁻³ . As shown in the scatter plot in Fig. S7, the posterior simulation alleviates the underestimation at most sites, which is quantified by a 42 % reduction in the overall underestimation bias and a clear improvement in the IOA. On a seasonal basis, the posterior emissions also alleviate the large underestimation of the prior simulation across all seasons, though the degree of improvement varies (Table S6). The prior simulation showed significant underestimation in all seasons, with the MFB ranging from −0.37 in winter to −0.79 in spring. The posterior simulation demonstrates a particularly evident improvement in spring, where the MFB reduced from −0.79 to −0.24. While some underestimation remains in summer, the posterior results still show improved performance metrics (e.g., lower RMSE and higher IOA) for all seasons, confirming a better capture of the seasonal characteristics overall. The remaining discrepancy between the posterior simulation and surface observations can be attributed to several factors, such as the spatial representativeness of the surface sites and the accuracy of the secondary inorganic aerosol simulation.

Furthermore, improving the NH₃ simulation results in the other simulated air pollutants being closer to observed levels (Table 3). Specifically, we compare the annual mean concentrations of PM_2.5, SO₂, and NO₂ from the prior and posterior simulations against surface observations averaged from 80 monitoring sites across 9 major cities (Table S4). It was found that posterior NH₃ emissions effectively bridge the gap between simulated and observed PM_2.5. The average PM_2.5 concentration increased from 65.7 to 67.3 µg m⁻³, which is closer to the observed value of 67.1 µg m⁻³. To further characterize the model's chemical performance beyond total PM_2.5, we also evaluated the simulation of secondary inorganic aerosol (SIA) components against in-situ measurements from a representative site in Beijing (Table S7). The evaluation shows that the posterior NH₃ emissions improved the simulation of ammonium and nitrate, reducing the bias between simulated and observed concentrations. Although the model underestimates sulfate, likely due to missing formation mechanisms (Cai et al., 2024; Wang et al., 2021, 2020), the total SIA concentration is well reproduced with an overall bias of only −11.0 %. A similar improvement is also observed for SO₂, where the posterior simulated concentration (6.8 ppbv) better matches the observed value (6.5 ppbv), reducing the model's previous overestimation by 27 %. This improvement is most significant in autumn. The successful capture of air pollutants highlights a significant improvement in the NH₃ emission inventory for Eastern China. The evaluation of routine air pollutants in each city is detailed in Figs. S8–S10. The statistics of evaluation metrics for each city's meteorological simulations can also be found in Table S8.

Table 3 Simulated and observed air pollutant concentrations.

5 PM_2.5 and its health burden response to NH₃ reduction

To investigate the response of PM_2.5 to various NH₃ emission reduction scenarios, we conducted sensitivity experiments as outlined in Table S9. We formulated emission reduction scenarios of 30 %–60 % for January and July of 2016, considering the severe particulate pollution in winter and the higher NH₃ concentrations in summer. Emission reductions from both the AGR and non-AGR sectors were considered separately.

Figure 7 Response of NH₃ emission reduction in 30 %–60 % in (a, b) concentration of PM_2.5 and (c) premature death caused by different diseases. The IHD, Stroke, LC and COPD represent the premature death caused by ischemic heart disease, stroke, lung cancer, chronic obstructive pulmonary disease.

Figure 7 illustrates that reducing NH₃ emissions by 30 %–60 % can decrease the seasonal PM_2.5 concentrations by 1.5–5.7 µg m⁻³ (2.0 %–7.2 %) averaged for Eastern China in winter, mainly due to the reduction in SIA. Specifically, nitrate, ammonium and sulfate are reduced by 0.9–3.3, 0.4–1.3 and 0.3–1.0 µg m⁻³, respectively. It is worth noting that the reduction in sulfate is smaller than that in nitrate because NH₃ preferentially reacts with sulfuric acid during aerosol formation (Fig. S11). When ambient NH₃ concentrations are limited, nitrate concentrations decrease more significantly than sulfate concentrations. In summer, although aerosol pollution is relatively lower, NH₃ emissions and atmospheric reactivity are higher. Consequently, reducing emissions by the same percent results in a decrease in PM_2.5 concentration by 5.5–8.8 µg m⁻³.

In terms of special sources, reducing non-AGR NH₃ emissions is just as crucial as reducing AGR NH₃ emissions in mitigating PM_2.5. A 30 % to 60 % reduction in non-AGR NH₃ emissions during winter can lead to a decrease in PM_2.5 by 0.9–1.5 µg m⁻³, which is comparable to the effect of reducing AGR NH₃ emissions (0.9–2.0 µg m⁻³). It should be noted that the reduction in PM_2.5 resulting from both AGR and non-AGR NH₃ emissions is not proportional to the emission reduction across all sectors. This is due to the non-linear relationship between NH₃ emissions and PM_2.5 concentrations.

This study utilized the integrated exposure–response (IER) model to estimate premature mortality resulting from PM_2.5 exposure. Detailed methods and data can be found in our previous work (Li et al., 2023a). In the base case, PM_2.5 exposure exhibits a significant impact on premature mortality, leading to 698.4 thousand deaths in the study region. Specifically, premature deaths attributable to ischemic heart disease (IHD), stroke, lung cancer (LC), and chronic obstructive pulmonary disease (COPD) are 202.3, 347.9, 61.5, and 86.7 thousand, respectively. In other scenarios, the overall premature mortality burden decreases by 45.6–72.0 thousand instances (6.5 %–10.3 %) in Eastern China. Notably, the decline in premature deaths, especially those related to stroke, plays a significant role in the overall reduction.

6 Conclusions

An accurate NH₃ emission inventory is essential for developing effective air quality improvement policies. Numerous studies have demonstrated that the current bottom-up NH₃ emission inventories in China often underestimate the total NH₃ emissions, with significant uncertainties in the estimation of emissions from various sources. In this study, we used IASI satellite products and an iterative algorithm with the WRF-Chem model to optimize the bottom-up NH₃ emission inventory for Eastern China and further assessed the impacts of NH₃ emission reductions from different sources on PM_2.5 concentrations.

The posterior results indicate that the NH₃ emission in Eastern China for 2016 amounted to 4.2 Tg. The highest emissions occurred in summer (463.1 Gg month⁻¹), with AGR sources contributing 86.5 % and non-AGR sources contributing 13.5 %. In contrast, emissions were lowest in winter (217.4 Gg month⁻¹), and the proportion of emissions from non-AGR sources were higher than that from AGR sources. Spatially, the region with the highest NH₃ emissions was located at the intersection of Henan, Hebei, and Shandong provinces. This is attributed to a combination of high emission intensity from dense agricultural and industrial activities and topographical effects that hinder the dispersal of pollutants. The optimization of the NH₃ inventory further improved the simulation underestimation of the NH₃ total column (MFB from −61 % to −30 %) and surface concentration (MFB from −61 % to −19 %). It also indirectly improved the simulation of other air pollutants, such as PM_2.5, NO₂ and SO₂.

Based on the posterior emission inventory, we conducted a series of sensitivity simulations to investigate the response of PM_2.5 concentrations to NH₃ emission reductions. A 30 %–60 % reduction in NH₃ emissions resulted in a 1.5–8.8 µg m⁻³ decrease in PM_2.5 concentrations. In terms of sectoral contributions, reductions in AGR emissions led to a decrease in PM_2.5 ranging from 0.9 to 7.4 µg m⁻³, while the response to reductions in non-AGR NH₃ emissions ranged from 0.9 to 5.3 µg m⁻³. Furthermore, the reduction in NH₃ emissions had a beneficial impact on public health, with a 6.5 %–10.3 % decrease in premature deaths attributed to PM_2.5 exposure.

This study obtained a high-resolution NH₃ emission inventory for Eastern China and highlights the significant role of non-AGR NH₃ emission reductions in further decreasing PM_2.5 levels. The findings provide robust data support for air quality research and offer scientific insights for exploring the potential air quality and public health benefits of NH₃ emission reduction.