Hourly gaseous pollutants and PM2.5 chemical compositions including water-soluble ions and trace elements were measured using online instruments during 1 January–31 March 2020 at a supersite in Tangshan. The supersite is located in a commercial region without short-distance industrial emissions (Fig. 1). SO2, NO2, and CO concentrations were determined by the ultraviolet fluorescence analyzer (TEI model 43i from Thermo Fisher Scientific Inc., USA), chemiluminescence trace gas analyzer (TEI model 42i from Thermo Fisher Scientific Inc., USA), and the correlation infrared absorption analyzer (TAPI model 300E, USA) (Li et al., 2017, 2019). The mass concentration of PM2.5 was determined using an oscillating balance analyzer (TH-2000Z, China) (Wang et al., 2014). The NH3 concentration and water-soluble ions including sulfate (SO42-), nitrate (NO3-), ammonium (NH4+), sodium ion (Na+), and chloridion (Cl-) were monitored with a gas and aerosol collector combined with ion chromatography (GAC–IC; TH-PKU-303, China) (Wang et al., 2014; Zheng et al., 2019). Organic carbon (OC) and elemental carbon (EC) were measured using an OC–EC analyzer (model RT-4, Sunset Laboratory Inc., Tigard, Oregon, USA). Nine trace elements including Hg, Pb, K, Ca, Cr, Cu, Fe, Ni, and Zn were determined by an online multi-element analyzer (model Xact 625, Cooper Environment Service, USA). The quality assurance of SO2, NO2, CO, and PM2.5 was conducted based on HJ 630-2011 specifications. For the quality assurance of NH3 and water-soluble ions, the concentration gradients of anion and cation standard solutions were set based on the pollution levels of target species, and correlation coefficients of the calibration curve must be higher than 0.99. Additionally, a standard sample was collected each day, and the relative standard deviation for the reproducibility test must be less than 5 %. The online device agreed well with the result determined by filter sampling coupled with inductively coupled plasma mass spectrometry (ICP-MS) and inductively coupled plasma atomic emission spectroscopy (ICP-AES).

The air pollutants were influenced by the combined effects of meteorological conditions and emissions. In order to quantify the contributions of anthropogenic emissions, the impacts of meteorological conditions should be removed. In our study, a random forest (RF) approach was employed to serve as the site-specific modeling platform (Chen et al., 2018). All gaseous pollutants and chemical compositions in PM2.5 were regarded as the dependent variables. The meteorological parameters including wind speed (WS), wind direction (WD), air temperature (T), relative humidity (RH), precipitation (Prec), air pressure (P), and time predictors (year, day of year (DOY), day of week (DOW), hour) served as the independent variables. The original dataset was randomly classified into a training dataset (90 % of input dataset) for developing the RF model, and the remaining one was treated as the test dataset. After the building of the RF model, the deweathered technique was applied to predict the air pollutant level at a specific time point (e.g., 1 January 2020 at 12:00 local time). The differences in original pollutant concentrations and deweathered pollutant concentrations were regarded as the concentrations contributed by meteorology. Some statistical indicators including R2 value, RMSE, and MAE were regarded as the major criteria to evaluate the modeling performance. In our study, the RF model with the R2 value lower than 0.50 was treated as the unreliable result and cannot reflect the impacts of emission and meteorology on air pollutants accurately because more than 50 % of the variability in the training model cannot be appropriately explained. After the model evaluation, only the species with cross-validation R2 values larger than 0.50 were selected to assess the respective contributions of emission and meteorology to their concentrations.

The positive matrix factorization (PMF 5.0) model version was used to perform the PM2.5 source apportionment. The deweathered gaseous pollutants and chemical compositions in PM2.5 were incorporated into the model. The objective of PMF is to resolve the issues of chemical mass balance between measured concentration of each species and its source contributions by decomposing the input matrix into factor contribution and factor profile. The detailed equation is shown in Eqs. () and (). Briefly, the basic principle of PMF is to calculate the least object function Q when the gik must be a positive definite matrix based on Eq. () (Chen et al., 2014; Sharma et al., 2016). 1xij=∑k=1pgikfkj+eij2Q=∑i=1n∑j=1mxij-∑k=1pgikfkjuij2, where xij and eij represent the concentration and uncertainty in the jth species, respectively; gik represents the contribution ratio of the kth source to the ith sample; fkj represents the ratio of the jth species in the kth source; and eij indicates the residual of the jth species in the ith sample. The uncertainties associated with factor profiles were evaluated using three error calculation methods including the bootstraps (BS) method, displacement (DISP) analysis, and the combination method of DISP and BS (BS–DISP). For the BS method, 100 runs were performed, and the result has been believed to be valid since all of the factors showed a mapping of above 90 %. DISP analysis also confirmed that the solution was considered to be stable because the observed drop in the Q value was less than 0.1 %, and no factor swap occurred. For the BS–DISP analysis, the solution has been verified to be useful because the observed drop in the Q value was less than 0.5 %. Furthermore, both of the results from BS and BS–DISP did not suggest any asymmetry or rotational ambiguity for all of the factors (Manousakas et al., 2017; Brown et al., 2015).

Figures 2–5 show the temporal variations in gaseous pollutants and chemical compositions in PM2.5 from 1 January–31 March, which could be divided into two periods including before and after the COVID-19 outbreak. In this study, 23 January was regarded as the breakpoint because China's government imposed a lockdown in Wuhan and surrounding cities. Before the COVID-19 outbreak, the average observed concentrations of SO2, NO2, CO, 8 h O3, and NH3 during 1–22 January were 34 µgm-3, 64 µgm-3, 2.0 mgm-3, 19 µgm-3, and 14 ppb, respectively. After the COVID-19 lockdown, the mean concentrations of these gaseous pollutants changed to 25 µgm-3, 39 µgm-3, 1.6 mgm-3, 49 µgm-3, and 18 ppb, respectively. Overall, CO, SO2, and NO2 concentrations decreased by 18 %, 27 %, and 39 %, respectively (p<0.05). However, the NH3 and O3 concentration increased by 35 % (p<0.05) and 160 % (p<0.01).

As shown in Fig. 2, the chemical compositions in PM2.5 also showed dramatic changes during 1 January–31 March due to the impact of the COVID-19 lockdown. The observed SO42-, PM2.5, Na+, and Cl- concentrations decreased by 6 % (p>0.05), 13 % (p>0.05), 29 % (p<0.05), and 48 % (p<0.01), respectively, while observed NO3- (2 %) and NH4+ (7 %) levels showed slight increases (p>0.05). In Shanghai, Chen et al. (2020) revealed that SO42- and NH4+ concentrations displayed significant decreases after the COVID-19 outbreak due to the obvious decreases in precursor concentrations (e.g., SO2, NOx). However, both observed NO3- and NH4+ concentrations in Tangshan even showed slight increases, though the NO2 concentration suffered from a remarkable decrease. It was assumed that the adverse meteorological conditions might be beneficial to the pollutant accumulation (Zheng et al., 2019; Y. Zhang et al., 2019). Additionally, the concentrations of nine trace elements were also determined. The observed values of Fe (25 %), Ca (39 %), Pb (41 %), Cr (41 %), and Zn (48 %) suffered from dramatic decreases (p<0.05), while the K (0 %), Ni (1 %), and Hg (8 %) concentrations still displayed slight increases (p>0.05). As a whole, the temporal variability in these elements in Tangshan before and after the COVID-19 lockdown was in agreement with the result in Beijing (He et al., 2017). However, the K concentration in Beijing showed rapid decrease after the COVID-19 outbreak, which was not coincident with our study (He et al., 2017). It suggested that the slight increase in K in Tangshan might be linked with the unfavorable meteorological conditions (He et al., 2017). The observed concentrations of OC (-19 %) and EC (-39 %) also suffered from rapid decreases after the COVID-19 lockdown (Fig. 4) (p<0.05), which was in good agreement with the sea-salt ions (e.g., Na+, Cl-) and most trace elements (e.g., Zn, Pb).

Although the observed concentrations of air pollutants can be applied to analyze the impact of the COVID-19 lockdown, the role of emission reduction in air quality might not be clearly revealed because the meteorological factors were also important variables influencing the air pollutant concentrations. In order to accurately reflect the response of air quality to emission reduction during the COVID-19 lockdown period, the meteorological conditions were isolated by the machine-learning model. In our study, we developed a random forest model to remove the impacts of meteorological conditions on air pollutants. Based on the results in Figs. S1–S5 in the Supplement, the RF models for most of the species showed better performance because their R2 values were higher than 0.50, and the slope of all of the fitting curves was also close to the R2 values. However, some other species such as Ag, Cd, and Mg2+ showed worse predictive performances, and thus these data cannot be utilized to distinguish the impacts of meteorology and emission on the concentrations of these species. Based on the cross-validation R2 value, the species with an R2 value higher than 0.50 were applied to assess the contributions of meteorology and emission to the concentrations. The deweathered concentrations of gaseous pollutants and chemical compositions in PM2.5 are depicted in Figs. 2–5. Compared with the period before COVID-19, the deweathered NH3, SO2, CO, and NO2 concentrations decreased by 27 %, 31 %, 32 %, and 42 % after the COVID-19 lockdown period outbreak, respectively (p<0.05), while the deweathered 8 h O3 concentration increased by 80 % (p<0.01). Meanwhile, the normalized-meteorology NH4+, NO3-, SO42-, Cl-, PM2.5, and Na+ and concentrations decreased by 14 %, 27 %, 35 %, 35 %, 38 %, and 47 %, respectively. For trace elements, deweathered Cu, K, Ni, Ca, Pb, Fe, Cr, and Zn levels reduced by 15 %, 23 %, 27 %, 54 %, 59 %, 61 %, 67 %, and 69 %, respectively (p<0.05). Nevertheless, the deweathered Hg concentration still kept a stable increase at a rate of 6 % compared with the period before the COVID-19 outbreak (p>0.05).

The deweathered concentrations for most of the pollutants showed significant decreases after the COVID-19 outbreak compared with the period before COVID-19 (Figs. 2–5). It was assumed that many cities proposed the lockdown measures, which significantly minimized industrial, transportation, and commercial activities. Among all of the pollutants, the deweathered Zn, Cr, Fe, Pb, and Ca experienced decrease rates of more than 50 % due to the lockdown measures. It was well known that Zn, Cr, and Fe originated mainly from the metallurgical industry (Sun et al., 2018; Zhu et al., 2018), while Pb might be derived from coal-fired power plants (Cui et al., 2019; Meng et al., 2020). During the COVID-19 outbreak, most of the industries have been shut down, and energy production by coal-fired power plants was reduced by one-third (Chang et al., 2020). Based on the adjustment factor estimated by Doumbia et al. (2021), the contributions of industrial activity and the power sector have decreased by 40 % after the COVID-19 outbreak, which was close to the decrease ratios of Zn, Cr, Fe, and Pb concentrations. It should be noted that the deweathered Ca concentration also decreased by more than 50 %. It was well documented that the Ca was often associated with the dust resuspension (Chang et al., 2018). In fact, the Ca was known as one of the most abundant elements in the upper continental crust, which likely originated from the fugitive dust (Chang et al., 2018; Shen et al., 2016). A reduction of more than 70 % in vehicle transportation and domestic flights facilitated the rapid decrease in Ca concentration (Chang et al., 2020). Although the observed K concentration did not show a marked decrease after the COVID-19 lockdown, the deweathered K level suffered from rapid decrease (-22 %) (p<0.05). It was widely acknowledged that K was considered to be a key fingerprint of biomass burning (H. Zheng et al., 2020), and thus the result suggested that the open biomass burning was also restricted during the period. Both of the deweathered concentrations of OC (-22 %) and EC (-45 %) also experienced remarkable decreases. In our study, both OC and EC concentrations showed significant correlation with K level (p<0.05), indicating that the restriction of biomass burning also led to the decreases in OC and EC. Additionally, PM2.5 and some water-soluble ions including deweathered SO42- and NO3- concentrations experienced marked decreases after the COVID-19 lockdown, which was in good agreement with their gaseous precursors. It might be attributable to the rapid decreases in precursor emissions. B. Zheng et al. (2020) verified that the SO2 emission in the industrial sector and NOx emission in the transportation sector in Hebei Province have decreased by 19 % and 13 %, respectively. The deweathered Na+ concentration showed the rapid decrease after the COVID-19 lockdown, which suggested that the Na+ in the PM2.5 of Tangshan was probably derived from waste incineration rather than sea-salt aerosol (Deshmukh et al., 2016).

Although most pollutant concentrations suffered from remarkable decreases, the decrease ratios of deweathered NH3 and NH4+ concentrations after the COVID-19 outbreak were far lower than those of many other gaseous pollutants and water-soluble ions. It was attributable to the fact that ambient NH3 was mainly sourced from the fertilizer application and livestock, which did not show significant decrease during the COVID-19 period (Kang et al., 2016; B. Zheng et al., 2020; Doumbia et al., 2021). Although the transportation volume suffered from dramatic decrease, the contribution of transportation to NH3 was generally less than 5 % (Kang et al., 2016). Furthermore, the contribution of urban waste sources slightly increased after the COVID-19 outbreak, offsetting the effect of traffic outage (Y. Zhang et al., 2020). Additionally, it should be noted that the normalized-meteorology 8 h O3 and Hg concentrations still maintained a stable increase. T. Liu et al. (2020) have confirmed that uncoordinated decreases in NOx and volatile organic compound (VOC) emissions (decrease ratio: NOx > VOCs) dominated the 8 h O3 increase in urban areas because most urban areas belonged to a VOC-limited region. Moreover, the excessive decrease in PM2.5 from primary emission significantly increased the HO2 radical concentration on the surface of aerosol, thereby promoting the O3 formation (Shi and Brasseur, 2020). The minor increase in deweathered Hg level was attributable to the fact that the coal combustion for domestic heating was not restricted during the COVID-19 lockdown period (Zhou et al., 2018). Based on the updated global anthropogenic emission adjustment factor during COVID-19, the contribution of the residential sector to air pollutants did not decrease after the COVID-19 lockdown (Doumbia et al., 2021).

Compared with the observed values, the deweathered concentrations of most pollutants were significantly reduced. Meanwhile, the deweathered decrease ratios of pollutants were significantly higher than those of observed values (Fig. 6). The result suggested that the meteorology conditions during the COVID-19 lockdown period were not favorable to the pollutant dispersion, as evidenced by some recent studies (Chang et al., 2020; Huang et al., 2020). In our study, six meteorological parameters including WS, WD, T, RH, Prec, and P have been integrated into the random forest model to assess the response of each species to different meteorological variables. The variable importance of each meteorological parameter to all of the species is shown in Figs. 7–10.

Among all of the gaseous pollutants, the meteorological conditions played significantly positive roles in NH3 (62 %) and 8 h O3 concentrations (80 %) (Fig. 6). As shown in Fig. 7, T was the most important factor for the rapid elevation of NH3 concentration after the COVID-19 lockdown. It was assumed that the higher T enhanced the emissions of NH3 from soil and urban wastes and promoted the volatilization of NH3 from aerosol NH4+ pools (Y. Zhang et al., 2020). In our study, the hourly mean air temperature increased from 0∘ before the COVID-19 outbreak to 5∘ after the COVID-19 lockdown, which strongly supported the inference. For 8 h O3 concentration (Fig. 7), T was also treated as the most important variable. On the one hand, the higher T generally enhanced biogenic isoprene emissions, which was the most abundant biogenic VOC and showed the highest ozone formation potential (Liu and Wang, 2020). On the other hand, high T often increased chemical reaction rates and accelerated the O3 formation (Shi et al., 2020). Additionally, WS also played an important role in the 8 h O3 concentration. Shi et al. (2020) have demonstrated that weaker winds often slowed down the advection and convection of NOx and VOCs, which was beneficial to O3 formation.

Moreover, the contributions of meteorological conditions to some secondary ions (e.g., SO42- (29 %), NO3- (29 %), and NH4+ (21 %)) were remarkably higher than those to other ions and some trace elements, suggesting that the chemical reactions and formation pathways of these species were more sensitive to meteorological variations. Deshmukh et al. (2016) confirmed that the high RH promoted the aqueous-phase oxidation of SO2 and the production of sulfate. Tian et al. (2019) also demonstrated that RH-dependent heterogeneous reactions significantly contributed to the sulfate generation, and the high RH enhanced gas- to aqueous-phase dissolution of NH3 and HNO3. These pioneering experiments suggested that secondary aerosols were often formed under the condition of high RH. Very recently, Chang et al. (2020) observed that the nitrate concentration in the YRD experienced unusual increase during the COVID-19 period, while Xu et al. (2020) obtained the opposite result in Lanzhou. It was assumed that the persistent increase in T and decrease in RH in Lanzhou during this period was not beneficial to the generation of secondary aerosol, while the high RH in the YRD significantly elevated local nitrate level. Although air temperature in Tangshan suffered from an increase after the COVID-19 lockdown, RH displayed a rapid increase from 47 % to 57 % during this period. Moreover, the increased O3 could promote the secondary aerosol formation and partially offset the decreased PM2.5 compositions triggered by the primary emission reduction (Liu et al., 2020). Similar to secondary ions, both OC and EC were also sensitive to RH. It was supposed that high RH could increase the secondary organic aerosol (SOA) levels, which accounted for the major fraction of OC (H. Zheng et al., 2020).

In addition, some trace elements such as Fe, Ni, and Cr were also significantly affected by the meteorological conditions. As shown in Fig. 9, these element concentrations were mainly sensitive to WD. It was assumed that the neighboring industrial points including cement plants and coal-fired power plants could influence the concentrations of trace elements via long- or short-range transport, which was strongly dependent on WD. Following WD, RH was also an important factor for the variation in these trace elements. Under the condition of high RH, Fe and Cr could catalyze the heterogeneous generation of sulfate and nitrate on the mineral or soot surface (Hu et al., 2015).

Unlike the trace elements, water-soluble ions and OC were less sensitive to WD. Major water-soluble ions in PM2.5 including SO42-, NO3-, and NH4+ were mainly derived from secondary formation rather than the direct emission (J. Feng et al., 2020; X. Zhang et al., 2020), and thus they were not very sensitive to WD.

The deweathered chemical compositions suggested that the sulfate and nitrate chemistry changed slightly after the COVID-19 outbreak. The oxidation ratio of sulfate (SOR; the ratio of sulfate concentration and the sum of sulfate and SO2 concentrations) decreased from 0.26 to 0.22, while the oxidation ratio of nitrate (NOR; the ratio of nitrate concentration and the sum of nitrate and NO2 concentrations) increased from 0.22 to 0.25 (Table 1). The decreased SOR after the COVID-19 outbreak indicated that the decrease rate of sulfate is higher than that of SO2. In contrast, the increased NOR during the COVID-19 lockdown period revealed that the decrease rate of nitrate is lower than that of NO2. The increased NOR after the COVID-19 outbreak suggested the consecutive nitrate production, though the NOx emission experienced tremendous reduction, which was in good agreement with the result observed by Chang et al. (2020). It was assumed that the persistently higher observed NH3 concentration during this period promoted the ammonium nitrate formation though the lower NOx emission (Y. Zhang et al., 2020), which also partially explained the abnormal increases in observed concentrations of secondary ions after the COVID-19 outbreak. In general, NH3 firstly tends to react with H2SO4 to form ammonium sulfate, and then the excess NH3 participated in the reaction with HNO3 (Chen et al., 2019; Q. Zhang et al., 2019). However, sulfate concentration suffered from a more dramatic decrease compared with SO2, which might be associated with the aerosol acidity during the COVID-19 lockdown period. The ratio of NH4+ and the sum of SO42-, NO3-, and Cl-, named C/A, was regarded as an indicator to reflect the aerosol acidity. In our study, the C/A value decreased from 0.33 to 0.28 after the COVID-19 outbreak, implicating that the aerosol acidity even showed slight increase during the COVID-19 lockdown period. It was well known that the higher aerosol acidity might prohibit the conversion from SO2 to sulfate (Liu et al., 2020; Shao et al., 2019), which yielded the lower SOR.

The emission control measures inevitably triggered the variation in source apportionment (Liu et al., 2017; Meng et al., 2020). In the present study, positive matrix factorization (PMF 5.0) was employed to identify the major sources of PM2.5 in Tangshan before and after the COVID-19 outbreak. About three- to nine-factor solutions were examined, and a five-factor solution obtained the lowest Q (robust) and Q (true) values. Additionally, the PMF analysis and error diagnostics also suggested that the result was robust (Tables S2–S4 in the Supplement).

The source apportionment profiles pre-COVID and post-COVID resolved by PMF are depicted in Fig. 11. For pre-COVID, the first factor contributed 36 % to the total species. The factor was characterized by high levels of NH4+ (41 %), SO42- (35 %), and NO3- (33 %). SO42- and NO3- were generally produced by oxidation of SO2 and NOx, respectively. The NH4+ was often formed through the heterogeneous reaction of NH3 and sulfate or HNO3. Thus, the factor was regarded as the secondary formation (SF). The second factor was characterized by high loadings of Zn (47 %), Cr (42 %), Fe (42 %), and Pb (31 %). Cr and Fe were mainly originated from fuel combustion and the metallurgical industry such as chrome plating and steel production (J. Liu et al., 2018), while Pb and Zn were derived from the roasting, sintering, and smelting process for the extraction of Pb and Zn ores (Wu et al., 2012). Therefore, the factor 2 was treated as the industrial process (IP) source. The predominant species in factor 3 included Na+ (42 %), K (40 %), OC (35 %), and EC (33 %). K was often regarded as the fingerprint of biomass burning (BB) (Chen et al., 2017; Zheng et al., 2019), whereas the Na+ was generally regarded as the tracer of waste incineration (Alam et al., 2019; Durlak et al., 1997). Hence, factor 3 was treated as the BB source. Tangshan suffered from remarkable increasing usage of biomass fuels for domestic heating in winter, which promoted the emissions of K and Na+ (Chen et al., 2017). The most abundant species in factor 4 were Hg (75 %), Pb (68 %), K (36 %), Cu (35 %), Cl- (33 %), and SO42- (27 %). Pb, Hg, and Cu were typical marker elements for coal combustion, and around 56 % of Pb and 47 % of Hg were released from coal combustion (Cheng et al., 2015; Zhu et al., 2020). In northern China, coal-based domestic heating was one of the most important sectors of coal consumption (K. Liu et al., 2018). Dai et al. (2019) also verified that the residential coal combustion was a major source of primary sulfate. Thus, factor 4 was regarded as the coal combustion (CC) source. The last factor was distinguished by high loadings of Fe (46 %), Ni (45 %), and Ca (38 %). Fe and Ca were main elements enriched in upper crust, and Ni was enriched in brake wear and tire wear dusts (Dehghani et al., 2017; Urrutia-Goyes et al., 2018). Thus, these elements in this factor were mainly sourced from traffic-related road dust (RD).

After the COVID-19 outbreak, the chemical compositions in PM2.5 were also classified into five sources including SF, IP, BB, CC, and RD. However, the contribution ratios of these sources varied greatly after the implementation of serious lockdown measures. The contribution ratio of IP experienced the largest decrease from 27 % to 20 %, whereas the apportionment of SF showed a marked increase from 36 % to 44 %. The contributions of the other three sources only suffered from slight variations. The rapid decrease in IP contribution might be associated with the shutdown of many industries during the COVID-19 period (Zheng et al., 2020), while the obvious increase in SF contribution was attributable to more heterogeneous or aqueous reactions of precursors (Chang et al., 2020). For nearly all of the species, the contribution ratios of IP suffered from remarkable decreases after the COVID-19 outbreak. Since the COVID-19 lockdown, the contribution ratios of SF to SO42-, NO3-, and NH4+ increased from 35 %, 33 %, and 41 % to 48 %, 44 %, and 52 %, respectively. However, the contribution ratios of SF for other species remained relatively stable. It was assumed that SO42-, NO3-, and NH4+ were mainly produced from secondary formation of precursors (Jiang et al., 2019; Yao et al., 2020), while other species, especially the trace elements, were mainly derived from the primary emission (Y. Wu et al., 2020). Although the COVID-19 pandemic led to the shutdown of many coal-fired power plants and industries and decreased the CC emissions from these sectors (Kraemer et al., 2020), the government-enforced home order might increase the electricity consumption (Venter et al., 2020), which offset the decreases in CC contributions to industrial activities. Therefore, the contribution ratios of CC did not experience dramatic variation after the COVID-19 outbreak.