The slant column densities (SCDs) of the oxygen dimer (O4), NO2, SO2 and HCHO are retrieved from scattered sunlight spectra measured by the MAX-DOAS instrument using the DOAS technique (Platt and Stutz, 2008) implemented by the WINDOAS software (Fayt and van Roozendael, 2001). SCD represents the TG concentrations integrated along the effective atmospheric light path. The TG cross sections, wavelength ranges and additional properties of the DOAS analyses are provided in Table 1. Figure 2 shows typical DOAS fit examples. We skip data for solar zenith angle (SZAs) larger than 75∘ because of stronger absorptions of stratospheric species and a low signal-to-noise ratio. We also skip the data with large root mean square (RMS) of the residuals and large relative intensity offset (RIO). All thresholds of the quantities used for filtering the results and the percentages of screened data of the total number of observations are listed in Table 2. Detailed discussions of the DOAS fit parameters for each species can be found in Sect. 2 of the Supplement.

Tropospheric vertical profiles of AE and trace gas volume mixing ratios (VMRs) from the ground up to 4 km are retrieved from the SCDs by using the PriAM algorithm, which was originally introduced by Wang et al. (2013a, b). Below we summarize the basic concept of the PriAM algorithm and its specific settings for this study, while details can be found in Sect. 3 of the Supplement. Like for other algorithms, a two-step inversion procedure is used in PriAM. In the first step, profiles of AE are retrieved from the O4 differential SCDs (dSCDs). Afterwards, profiles of NO2, SO2 and HCHO VMRs are retrieved from the respective dSCDs in each MAX-DOAS elevation sequence. The retrieval problem is solved by the Levenberg–Marquardt modified Gauss–Newton numerical iteration procedure (Rodgers, 2000). Considering the frequent variation of aerosols and TGs, very little is known about the expected profiles. Thus a set of fixed a priori profiles is used for each species. A smoothed box-shaped a priori AE profile (Boltzmann distribution) (Yilmaz, 2012), exponential a priori profiles of NO2 and SO2 (similar to Yilmaz, 2012 and Hendrick et al., 2014), and a Boltzmann distribution a priori HCHO profile (based on the MAX-DOAS and aircraft measurements in Milan during the summer of 2003 reported in Wagner et al., 2011) are used in this study and denoted by the black curves in Fig. S8. Besides these standard a priori profiles, we tested the effect of changing the profile shapes and absolute values on the fit results. The description of these sensitivity tests is given in Sect. 3.1 of the Supplement. We conclude that the standard a priori profiles are well suited to application to the long-term MAX-DOAS measurements in Wuxi. We also find that improper a priori profiles can strongly impact the aerosol profile retrievals, but only slightly impact the TG results.

PriAM uses the radiative transfer model (RTM) SCIATRAN version 2.2 (Rozanov et al., 2005). Based on the wavelength intervals of the DOAS fits, the RTM simulations are done at 370 nm for the retrieval of aerosols and NO2, at 339 nm for HCHO and at 313 nm for SO2. The surface height and surface albedo are set as 50 m a.s.l. and 0.05, respectively. The fixed single scattering albedo of 0.9 and asymmetry factor (Henyey and Greenstein, 1941) of 0.72 are chosen according to average inversion results from the Taihu AERONET station (see Sect. 2.2.4) from 2011 to 2013 (the data in 2014 is unavailable). The retrieved AE at 370 nm is converted to those around 313 nm for the SO2 and 339 nm for the HCHO retrieval using Ångström exponents also derived from the Taihu AERONET data sets.

In addition, here it should be noted that the Levenberg–Marquardt modified Gauss–Newton procedure is based on the assumption that the probability distribution function (PDF) of the atmospheric state (x) can be described by a Gaussian PDF (P) around the a priori state (xa) (Rodgers, 2000): -2ln⁡P(x)=x-xaTSa-1x-xa+c. Here c is a constant value and Sa is the covariance matrix of the a priori. Thus the solution can not reach the true state if the PDF of the atmospheric state (x) is skewed or asymmetric (Rodgers, 2000). In this study the retrieval of the AE for extremely high aerosol loads (e.g. fog and haze) belongs to cases, which probably do not fulfil this assumption. In such cases the AE is underestimated by the inversion (see Sect. 2.2.5).

The totally averaged averaging kernels (AKs) for retrievals of AE and the NO2 VMRs for favourable measurement conditions, namely cloud-free sky with relatively low aerosol loading (the sky condition is directly identified by MAX-DOAS observations as described in Sect. 2.2.5), are shown in Fig. 5. AKs for SO2 and HCHO are similar to those of NO2. They indicate that the inversions are sensitive to the layers from the surface up to 1.5 km. The degrees of freedom (DoF) are about 1.5 for aerosols (similar to Frieß et al., 2006), 2 for NO2 and 2.3 for SO2 and HCHO. The detailed discussion of the performance of the profile retrievals is given in Sect. 3.2 of the Supplement by comparing the measured and modelled dSCDs for different elevation angles, and comparing profiles, AKs and retrieval errors in different seasons. In general the retrievals performed consistently for different elevation angles and seasons.

In previous studies (Clémer et al., 2010; Hendrick et al., 2014; T. Wang et al., 2014), fixed temperature and pressure (TP) profiles are usually used (e.g. obtained from the US standard summer atmosphere for the measurements in China). However, for locations with a significant and systematic annual variation of TP, as in this study, this simplification can affect the retrieved AODs and AE profiles (and thus also the TG profiles) systematically, yielding virtual seasonal variations. The time series of TP near the surface from the weather station near the MAX-DOAS instrument are shown in Fig. 3 for the year 2012 (similar patterns are found for other years, see Fig. S16). A regular annual variation of surface TP is obvious, with amplitudes between winter and summer of about 20 K and 30 hPa, respectively. The O4 VCDs derived from the fitted curves of surface TP (the method is described in Sect. 3.3 of the Supplement) is also shown in Fig. 3. The O4 VCD in summer is systematically lower than in winter by about 15 % of the yearly mean O4 VCD. Ignoring this systematic seasonal variation can cause a 20–30 % bias of the AOD and near-surface AE (see Fig. S17). The error of the aerosol retrieval can further nonlinearly impact the TG profile retrievals. To account for this effect, the seasonal variation of TP and the O4 VCD is parameterized and explicitly considered in the forward model during the MAX-DOAS retrievals by the PriAM algorithm. Figure 4 shows the AODs retrieved by PriAM using either explicit TP information or the TP profiles from the US summer standard atmosphere. The consistency of the AODs retrieved based on the explicit TP data with the simultaneous Taihu AERONET level 1.5 AOD data sets (see Sect. 2.2.4) is better than for TP profiles from the US standard summer atmosphere.

The systematic variation of TP could also be considerable in many other locations of the world. A seasonal variation of temperature occurs in many locations, especially outside the tropics. However, the temporal variation of the pressure is usually more complex and strongly depends on the location. The variation of pressure in Wuxi (and also many other parts of eastern China) is related to the East Asian Monsoon and shows a systematic seasonal pattern. The monsoon is a general phenomenon in eastern China. The pressure over the continent is systematically lower and higher than that over the ocean in summer and winter, respectively. Thus a similar seasonal variation of the O4 VCD is expected in general in eastern China.

To validate the results from MAX-DOAS observations, the column densities and averaged concentrations in the lowest layer from 0 to 200 m are compared to independent measurements: a.

AODs at 380 nm (level 1.5) from the sun photometer at the AERONET (Holben et al., 1998, 2001) Taihu station. The data are downloaded from the website of http://aeronet.gsfc.nasa.gov/. The AERONET sun photometer is located 18 km south-west of the MAX-DOAS instrument. AERONET data in the period from May 2011 to October 2013 are included in the study. In the level 1.5 data, a cloud screening scheme is used to filter most of the cloud-contaminated data (Smirnov et al., 2000). Here it should be noted that AERONET Taihu station is located in a more remote area in Wuxi city than the MAX-DOAS instrument. This could contribute to a systematic bias between both data sets. However, the long residence time of up to several days (Ahmed et al., 2004) and the relatively homogeneous horizontal distribution of aerosols (illustrated by the weak dependence of AOD on wind direction; see Sect. 3.4.2) implies that the differences between both measurements should be small.

MAX-DOAS results are compared to the available independent measurements within 15 min time difference. In this section only the data recorded during clear-sky conditions with low aerosol load (the sky condition is identified by MAX-DOAS observations as described in Sect. 2.2.5) are compared (comparisons for different cloud conditions are shown in Sect. 2.2.5). For the comparisons of AODs, near-surface AE, NO2 and SO2 VMRs, the averaged absolute differences, standard deviations, correlation coefficients (R) (Pearson's product moment correlation coefficient is applied in this paper) and the slopes and intercepts derived from the linear regressions are shown in Fig. 6 for the different seasons. The histograms of the absolute differences in different seasons are shown in Fig. S18.

For the comparisons of AODs from MAX-DOAS and the AERONET sun photometer, almost symmetrical Gaussian-shape histograms of the absolute difference are found for different seasons, except for summer (see Fig. S18a). The mean differences of AODs are smaller than 0.16 (about 20 % of the average value). The correlation coefficients are within 0.56 to 0.91 (see Fig. 6). The highest R of 0.91 is found in summer, probably related to the wider range of AODs covered, but in that season the largest absolute difference of -0.16 is also found, probably due to the stronger aerosol load than in other seasons. Underestimation of high aerosol amounts by MAX-DOAS will be discussed in Sect. 2.2.5. Several previous studies applied a correction factor to measured O4 dSCDs to improve the consistency between the AODs derived from MAX-DOAS and those from AERONET (e.g. Wagner et al., 2009; Clémer et al., 2010; Frieß et al., 2016). So far there is no good explanation for this correction factor. In this study we do not apply any correction factor, because we achieve reasonable consistency between MAX-DOAS and AERONET results without the application of a correction factor.

The averaged AE in the lowest layer derived from the MAX-DOAS are compared with those from the visibility meter. Here it has to be noted that both instruments do not probe exactly the same air masses: the visibility meter is sensitive to air masses at the measurement location while the MAX-DOAS is sensitive to the air masses along the line of sight for up to several kilometres away from the instrument and up to a few hundred metres above the ground. Figure S18b shows almost symmetrical Gaussian-shape histograms of the absolute differences of the AE between the two techniques. The mean differences are < 0.18 km-1 (about 33 % of the average value), as shown in Fig. 6. The highest R of 0.74 is found in summer (see Fig. 6), probably related to the wider range of values and the stronger vertical convection, which causes a higher boundary layer and possibly a smoother vertical distribution of aerosols than in other seasons. The worst correlation is found in spring, which might be related to the occurrence of long-distance transport of dust with elevated aerosol layers (see Sect. 3.1).

The VMRs of NO2 and SO2 in the lowest layer derived from MAX-DOAS are compared with the values from LP-DOAS measurements for the individual seasons. Like for the AE, it has to be noted that both instruments do not probe exactly the same air masses, as the LP-DOAS yields the mean TG concentration for the light path defined by the set-up of instrument and reflector. In general the mean absolute differences are smaller than 5 ppb (about 50 % of the average value) for NO2 and 6 ppb (about 60 % of the average value) for SO2 (see Fig. 6). Almost symmetrical Gaussian-shape histograms of the absolute differences are also found for NO2 and SO2 in different seasons (Fig. S18c and d). The range of R is from 0.4 to 0.7 for NO2 and from 0.7 to 0.8 for SO2 in all seasons (see Fig. 6). The higher R value for SO2 than for NO2 is probably related to the longer lifetime and thus more homogeneous vertical and horizontal distribution of SO2 compared to NO2, especially in the layer from 0 to 200 m. The worst correlation of NO2, especially in the afternoon (see Fig. S19), is found in summer probably because of the low NO2 VMR near the surface, the small value range and the steep vertical gradient in the layer from 0 to 200 m (see below). The generally positive absolute differences of NO2 and SO2 shown in Fig. 6 could be attributed to strong gradients in the layer from 0 to 200 m as, e.g., found from tower measurements in Beijing (Meng et al., 2008): they concluded that the largest values of the NO2 and SO2 concentrations are not directly located at the surface, but at an altitude of about 100 m, especially in summer. However, it should be noted that the vertical gradients around Wuxi might be different from those in Beijing and thus other reasons might also contribute to the observed differences.

The retrieval of AE by PriAM from O4 absorptions is based on a RTM, which does not include the effects of clouds. In principle it should be possible to also include cloud effects in the RTM (at least for horizontally homogenous clouds), but in the current version of our retrieval this is not yet accomplished. In this section, we investigate how strongly different types of clouds affect the MAX-DOAS retrieval results of aerosols and TGs. For that purpose we compare the MAX-DOAS results with independent data sets for different cloud types. For the characterization of the cloud conditions we use the cloud classification scheme described in Wang et al. (2015) (based on the concept of Wagner et al., 2014) to classify the sky conditions from the MAX-DOAS observations, i.e. radiance, colour index and O4 absorption. The scheme differentiates between eight primary sky conditions (varying between clear skies with low aerosol load to continuous cloud cover) and two secondary sky conditions of fog and optically thick clouds. In this study we condense the eight primary sky conditions to five primary conditions by merging two types of cloud holes and two types of continuous clouds and ignoring the rare condition of “extremely high midday CI” (colour index)(Wang et al., 2015). The remaining five primary conditions are clear sky with low aerosol load (“low aerosol”), clear sky with high aerosol load (“high aerosol”), “cloud holes”, “broken clouds”, and “continuous clouds”. Please note that “low” and “high” in the marks of sky condition categories describe the aerosol load but not the aerosol height. Each MAX-DOAS measurement scan is assigned to one of the five primary sky conditions. In addition, they can be assigned to the two secondary sky conditions of “fog” and “optically thick clouds”. Here it should be clarified that the “fog” sky condition does not exactly belong to the meteorological definition, but represents a sky condition derived from MAX-DOAS observations with a low visibility. It should also be noted that the distinction between “low aerosol” and “high aerosol” conditions is based on the colour index measured by the MAX-DOAS instrument, but not on explicit AOD threshold values. The studies of Wang et al. (2015), however, demonstrated that the AODs observed by the Taihu AERONET sun photometer are mostly smaller and larger than 0.6 for the “low aerosol” and “high aerosol”, respectively. In addition to the cloud effect, the effect of high aerosol loads is also evaluated (due to the unrealistic assumption of the PDF of the atmospheric state in the OE algorithm for high aerosol loads (see Eq. 1)).

Firstly, measured and modelled dSCDs (results of the forward model) are compared under various sky conditions. In Fig. 7 (grey columns), the histograms of the differences between the measured and modelled dSCDs are shown for the four species (note that Fig. 7 represents the differences for all non-zenith elevation angles). The histograms are symmetric and the maximum probabilities occur around zero for all four species, i.e., overall, there is no indication for a significant systematic retrieval bias. In the same figure, the relative frequencies for the different sky conditions are shown in different colours. In general, for cloudy-sky conditions, especially for continuous clouds and optically thick clouds, larger discrepancies are found compared to cloud-free sky conditions. The effect of clouds on the inversion is stronger for aerosols than for TGs. For the aerosol inversion, more negative differences are found for “fog”, which indicates that the strong extinction in “fog” is not well represented by the forward model (the phenomenon is also found in Fig. 9 and discussed below). To skip those inverted profiles, which probably differ largely from the real profiles, we only keep the profiles, for which the differences between measured and modelled dSCDs are smaller than 2 × 1042 molecules2 cm-5 for the O4 dSCDs (90.6 % of the total observations) and 5 × 1015 molecules cm-2 for NO2 (89.8 %), SO2 (90.4 %), and HCHO dSCDs (97.9 %) for each elevation angle in one elevation sequence.

After this screening of potentially bad profiles, the mean profiles of AE and TG mixing ratios as well as the corresponding total AKs (which represent the sum of all AKs at the individual altitudes) are shown in Fig. 8 for different sky conditions. While the total AKs differ only slightly, the resulting profiles are quite different for different sky conditions. There are two interesting findings for the retrieved profiles: first, for all cloudy scenarios (including fog), the maximum AE is not found at the surface, but at higher altitudes, as observed also by Nasse et al. (2015). This can be explained by the fact that clouds act as a diffusing screen. The effect on MAX-DOAS observations is that the light paths, especially for low-elevation angles, become longer than for cloud-free conditions. Consequently, also increased O4 absorptions are measured for such conditions. A similar effect can also be caused by elevated aerosol layers. Since the forward model does not explicitly include clouds, elevated “cloud-induced” aerosol layers are usually derived in the profile inversion under cloudy conditions. The diffusing screen effect depends on the cloud optical thickness. The most pronounced cloud-induced elevated aerosol layers are retrieved for optically thick clouds.

Interestingly, for measurements under “fog” conditions, elevated aerosol layers are also obtained from the MAX-DOAS inversion. This is at first sight surprising, but can be explained by two aspects: first, for most measurements classified as “fog”, a weak systematic dependence of the O4 dSCDs on elevation angles is still found, indicating that during most “fog” events the visibility is still not close to zero. Second, for most of the measurements classified as “fog” the presence of clouds (including thick and broken clouds) was also detected (Wang et al., 2015). This finding indicates that, for most observations classified as “fog”, increased aerosol scattering close to the surface did indeed occur, but at higher altitudes, even larger extinction was present. We also found a generally larger value of the cost function under cloudy conditions (consistent with Fig. 7) and a systematic variation of the TG VCDs and near-surface VMRs for the different cloud scenarios. Besides measurement errors, these variations are probably also due to different photolysis rates and atmospheric dynamics (see Fig. S20).

In the following we compare the results from MAX-DOAS and other techniques under different sky conditions. Since the frequencies of different cloud conditions depend on season (Wang et al., 2015), and the agreement between MAX-DOAS and other techniques was also found to be different for different seasons (see Sect. 2.2.4), the comparisons are done for individual seasons. In Fig. 9 the comparison results in autumn are shown (similar conclusions are found for other seasons and the relevant figures are shown in Figs. S21–S24). Based on the comparisons of the retrieved profiles under different sky conditions (Fig. 8) and the comparison results with independent data sets (Fig. 9), we have developed recommendations: under which sky conditions a given data product might still be useful or should not be used. These recommendations, as summarized in Table 3, should not be seen as binding, but rather as a general indication of the usefulness of a given observation, and might change for improved inversion algorithms in the future.

In general we find that the aerosol results are more strongly affected by the presence of clouds than the trace gas results. This is especially true for the retrieved AOD for which large mean difference and worse correlation are found under “continuous clouds” (see Fig. 9). The impact of clouds on the retrieved AE profiles is also significant, as illustrated in Fig. 8a. Thus we recommend that retrieved AOD and AE profiles (except those close to the surface) should not be used for cloudy conditions. However, AE close to the surface can still be well retrieved under most cloudy conditions, except for “thick clouds” and especially for “fog”, because a significant increase of the mean difference and decrease of the linear correlation are found for “fog” (Fig. 9). The TG results are less affected by clouds. No significant effects of clouds on the profiles of TGs are found, as can be seen in Fig. 8c, e, and g. Only larger mean differences and worse linear correlations of the surface mixing ratios are obtained for “thick clouds” and “fog” (Fig. 9). Thus not only surface mixing ratios, but also TG profiles and tropospheric VCDs, can still be well retrieved for most cloudy situations (except for thick clouds and fog). The MAX-DOAS data used in Sect. 3 are filtered by the recommendations listed in Table 3.

In comparison with other sky conditions (except for “fog” and “thick clouds”), outlier values of the mean differences and slopes under “high aerosol” conditions are found for both aerosols and TGs, as shown in Fig. 9. This phenomenon could be attributed to two factors: a degraded performance of aerosol profile retrieval and a reality of vertical distributions of aerosols and TGs. On the one hand, aerosol profiles with high aerosol loadings could hardly be well reproduced by the retrieval algorithm due to the constraint of the a priori profile and the assumption of Eq. (1). This is indicated by the systematic overestimation of the modelled O4 dSCDs compared to the measured O4 dSCDs as shown in Fig. 7 for “high aerosol” conditions. On the other hand, a “high aerosol” condition usually indicates a polluted period; therefore it is reasonable to expect a very different inhomogeneity of the horizontal and vertical distributions of aerosols and TGs in the lowest layer (0–200 m) of MAX-DOAS profile retrievals compared to other relatively clean conditions. Thus different air mass observed by a MAX-DOAS instrument and other techniques could play a different role under “high aerosol” conditions. Further studies on the evaluation and improvement of profile retrievals of MAX-DOAS under heavy aerosol pollution conditions need to be carried out in the future. In this study, we decide to keep the MAX-DOAS retrievals under “high aerosol” condition because, despite their lower reliability, they can still provide useful information.

For the MAX-DOAS results, we derive the error estimates from different sources. First we estimate the error budgets for the near-surface values and column densities of the TGs and aerosols, which are summarized in Table 4. The following error sources are considered: a.

Smoothing and noise errors (fitting error of DOAS fits) on the near-surface values and column densities are derived from the averaged error of profiles from the retrievals (shown in Figs. S11b–S14b), and amount on average to 10 and 6 % for aerosols, 12 and 17 % for NO2, 19 and 25 % for SO2 and 50 and 50 % for HCHO, respectively.

Huang et al. (2014) showed that secondary aerosols including organic and inorganic aerosols (nitrates and sulfates) contribute to about 74 % of the PM2.5 mass collected during high-pollution events in January 2013 at the urban site of Shanghai. The aerosols in Wuxi (which is close to Shanghai) are expected to be similarly dominated by secondary aerosol. NOx and SO2 are the precursors of secondary inorganic aerosols through their conversion into nitrates and sulfates, respectively. HCHO can be used as a proxy for the local amount of VOCs, which are precursors of secondary organic aerosols (Claeys et al., 2004). To identify the dominant precursor, we investigated the relationship between aerosols and their precursors through a correlation study as in Lu et al. (2010), Veefkind et al. (2011) and T. Wang et al. (2014). Table 5 lists the correlation coefficients between the TG VCDs and AODs as well as the TG VMRs and AE near the surface. The correlations of near-surface values are always higher than those of the column densities. This finding could be probably explained by the effect of long-range transport, which typically occurs at elevated layers. For long-range transport, the effect of different atmospheric lifetimes is especially large, probably leading to weaker correlations between the aerosol and its precursors. In contrast, close to the surface, local emissions dominate the concentrations of TG and aerosols, and the effect of different lifetimes is negligible.

In general, correlations in spring are the lowest, probably due to the transport of dust and biomass burning aerosols. The correlations between aerosols and HCHO are higher in winter than in summer. This finding may be explained by the fact that anthropogenic emissions dominate the (primary and secondary) sources of HCHO and aerosols simultaneously in winter. Meanwhile the correlations between aerosols and HCHO are higher than those between aerosols and NO2 or SO2 in winter and autumn. This finding can be possibly explained by the fact that both HCHO and aerosols are dominated by secondary sources, while NO2 and SO2 are mostly from primary emissions in this region.

The MAX-DOAS station is located on the boundary of urban and suburban areas as shown in Fig. 1b. Several iron factories, cement factories and petroleum industries are operated in the south-west industrial area. The industrial activities and vehicle operations in the industrial area lead to significant emissions of NO2, SO2 and VOCs as well as aerosols (Huang et al., 2011). In the urban centre area, traffic, construction sites and other anthropogenic emissions emit significant amounts of NO2 and VOCs as well as particles. Some factories, such as an oil refinery, are located in the north-west of the urban centre, emitting pollutants including SO2 and VOCs. In addition, one power plant located at about 50 km in the north and the Suzhou city in the south-east direction of the MAX-DOAS station might contribute to the observed pollutants in Wuxi depending on the meteorological condition.

We analysed the distributions of column densities and near-surface values of the TGs and aerosols for different wind directions in Fig. 16. In principle, the near-surface pollutants are expected to be dominated by nearby emission sources, while the column densities can be additionally affected by transport of pollutants from remote sources. Long-range transport can weaken the dependence of the column densities on the wind direction because of the complex trajectories the air masses might have followed. For all four species, the highest values are observed for south-westerly winds, especially for the near-surface pollutants. This finding implies that the industrial area emits large amounts of NOx, SO2, VOCs and aerosols. Figure 16c shows that the HCHO south-west peak is only present in winter. This finding is probably caused by the fact that, in winter, anthropogenic sources (of precursors and direct HCHO emissions) dominate the HCHO amounts, while in other seasons natural sources dominate the HCHO amounts. Another peak of NO2 and SO2 is found in the north-west, obviously in winter, indicating considerable emissions in the urban centre. Figure 16d shows a weaker dependence of AODs on the wind direction than the VCDs of the TGs, which probably indicates the stronger contribution of long-range transport to the local aerosol levels compared to the TGs. In addition, for daily-averaged wind speed of smaller than 1 m s-1, the averaged TG VCDs and near-surface VMRs are higher than those for larger wind speeds (shown in Fig. S31a and b), indicating that dispersion of local emissions is more important than the transport from distant sources. For aerosols, a wind-speed dependency is only observed for near-surface AE, but not for AODs (see Fig. S31c), indicating the higher importance of transport for aerosols than for TGs.

Although local, this study shows several general important features: (1) the dependence of the measured TG VMRs on the wind direction indicates that the dominating sources of the pollution are local, and not from long-range transport. Also, a strong horizontal gradient appears. Because of the expected similar lifetime, meteorological conditions and emission sources, the conclusion could apply to the whole YRD region. (2) The study provides an example on how to use ground-based MAX-DOAS observations to find strong emission sources in an urban-sized area. (3) The seasonality of the wind dependence of the trace gases, especially for HCHO, indicates the different sources in different seasons.