Rome is the capital and the most populous city of Italy with almost 3 million inhabitants and one of the most densely populated cities in the European Union (ISTAT, 2021). It is located about 24 km east of the Tyrrhenian Sea, surrounded by an extensive undulating plain, and crossed by the Tiber and Aniene rivers. The city is part of the Lazio administrative region in the central part of the Italian Peninsula. The economic activities in the metropolitan area are characterized by the absence of heavy industrial facilities and are related mainly to the services and high-technology sectors, as well as commercial activities and tourism. The city air quality is strongly affected by local emission sources, such as transportation and domestic heating, but it is also markedly affected by local circulation and mid- to long-range transport events of sea salt, wildfires, and Saharan dust (e.g., Ciardini et al., 2012; Gobbi et al., 2013; Barnaba et al., 2017; Valentini et al., 2020; Di Bernardino et al., 2021).

Rome's air quality is monitored on a regular basis by standard in situ instrumentation. These measurements are complemented by multiplatform, long-term observations of aerosol and trace gases performed by a variety of ground-based remote sensing instruments such as sun–sky radiometers, Raman and elastic lidars, automated lidar ceilometers, Pandora, Brewer, and differential optical absorption spectroscopy (DOAS) spectrophotometers (e.g., Di Ianni et al., 2018; Iannarelli et al., 2021; Diémoz et al., 2021). In this study, we used remote sensing measurements of columnar NO2 and aerosol properties performed in two stations located in the greater area of Rome. More specifically, observations were obtained from an urban station (APL-SAP hereafter) located at the Atmospheric Physics Laboratory of the Physics Department of the Sapienza University of Rome in the city center (41.90∘ N, 12.52∘ E; altitude 75 m a.s.l. – above sea level) and a suburban site at the southern east edge of the city in the National Research Council (CNR) Institute of Atmospheric Sciences and Climate (ISAC) Rome Atmospheric Supersite (CIRAS) in Tor Vergata, Rome (41.84∘ N, 12.65∘ E; altitude 117 m a.s.l.). These two observational sites, along with the rural station of the CNR Institute of Atmospheric pollution Research (IIA) in Montelibretti, contribute to the Boundary-layer Air Quality-analysis Using Network of Instruments (BAQUNIN) supersite (Iannarelli et al., 2021) and to several national and international observing networks.

In this study, we also evaluate the impact of modified AOD and AE retrievals, as described in Sect. 2.4.1 and 2.4.2, on aerosol temporal trends. This is only a first attempt to investigate the possible effect of NO2 absorption on the AOD and AE trends, since the data sets used here are quite short for statistically meaningful calculations.

The annual trends in AOD and AE were estimated by applying the weighted least squares fitting technique introduced by Weatherhead et al. (1998) and previously adopted in several aerosol trend analysis studies from space and the ground (e.g., Zhang and Reid, 2010; Yoon et al., 2012; Logothetis et al., 2021). The applied linear trend model is based on the following formula: 13Yt=μ+ωXt+εt,t=1,…,T, where Yt is the monthly average aerosol property of interest, μ is a constant term representing the linear fit offset at the start of the time series, ω stands for the magnitude of the trend per year, and εt is the monthly average noise not represented by the linear fit. Xt=t/12 is the decimal number of years since the first month of the time series, t is the month index, T denotes the total number of months, and T/12 is the total number of years in the time series.

In order to account for data variability due to severe aerosol events and cloud disturbance, we introduced a monthly weighting factor wt into the linear fitting procedure (Eq. 14; Yoon et al., 2012). This weighting factor is defined as the square root of the number of observations available each month nt divided by the monthly standard deviation σt (Eq. 15). 14χ2μ,ω=∑t=1Twt⋅Yt-μ-ωXt2,15wt=ntσt In order to derive statistically significant monthly mean values, a minimum number of 10 observations in a daily basis was ensured. In addition, qualified monthly averages require the availability of measurements from at least 10 d per month. Data were filtered based on the above criteria, and days and/or months that did not fulfill them were excluded from the data sample used in the trend calculations. It should be noted that the data sets employed in this study are quite short for statistically meaningful aerosol trend analysis. However, this is a first attempt to investigate the impact of modified AOD and AE calculations on the derived temporal trends.

The Generalized Retrieval of Atmosphere and Surface Properties (GRASP; Dubovik et al., 2021) is a state-of-the-art inversion algorithm based on a statistically optimized multiterm least squares method (LSM) proposed by Dubovik (2004). GRASP has been applied to numerous applications covering a vast variety of instruments and, interestingly, to very different combinations between them. Among the different applications of GRASP, it is possible to find GRASP/POLDER-3 (Chen et al., 2020), GRASP/AOD (Torres et al., 2017), OLCI/GRASP (Chen et al., 2022), the combination of active lidar measurements and ground-based radiometry (Lopatin et al., 2013, 2021; Román et al., 2018; Herreras et al., 2019), the retrieval of all-sky cameras (Román et al., 2017, 2022), or, for example, applications to in situ measurements including polar nephelometers (Espinosa et al., 2017, 2019; Schuster et al., 2019).

The GRASP scientific core was borne from the heritage of the AERONET inversion algorithm (Dubovik and King, 2000; Dubovik et al., 2000; Dubovik, 2004; King and Dubovik, 2013). At the same time, as discussed above and by Dubovik et al. (2011, 2021), the possibilities of GRASP have been extended due to the totally generalized nature of the inversion module and the continuous developments of the forward model.

For this study, GRASP has been used to mimic AERONET standard retrieval in order to understand the effects of the NO2 concentration on the retrieved SSA at 440 nm. In this case, two different approaches were followed for the GRASP algorithm. First of all, GRASP has been used as close as possible to the standard AERONET retrieval, which means that the input measurements of the algorithm are the total optical depth (TOD) and the almucantar sky measurement routine at 440, 675, 870, and 1020 nm. In the first approach (GRASP/AERONET NO2 hereafter), the NO2 absorption is taken into account, using OMI climatology, exactly as done for AERONET. On the other hand, GRASP flexibility allows the use of different assumptions of the gaseous properties. Therefore, in addition to the standard approach, the aerosol retrieval has also been done using the total columnar NO2 concentrations provided by the Pandora spectrometers co-located with AERONET instruments at the two stations selected for this study. This methodology will hereafter be referred to as GRASP/Pandora NO2. Thus, in addition to the standard AERONET retrieval products, GRASP has provided aerosol retrieval using these more accurate NO2 concentrations. The NO2 absorption features were calculated more precisely from those concentrations by using a k-distribution approach or the “kbin” code (Doppler et al., 2014a, b) to speed up the calculations.

The differences in AOD (Δτ) at 440 nm and, thus, of its spectral variability through the AE (Δα at 440–870 nm) correcting for measured NO2 effects with respect to the standard AERONET retrievals are illustrated in Fig. 6 for both the Rome CNR-ISAC and APL-SAP stations. The frequency distributions of AOD, Δτ, and Δα are also included in Fig. 6. Δτ is defined as the standard minus the modified AOD (τAER–τAER_mod; see Eqs. 4–6). Similarly, Δα is defined as αAER–αAER_mod (Eq. 12). The derived values are presented versus the AOD at 440 nm and are color coded with respect to the Pandora NO2 retrievals. The dependency of Δτ on NO2 is quite clear. As expected, higher Δτ absolute values are obtained for higher NO2 concentrations, regardless of the initial measured AOD. Also, the absolute percentage of Δτ with respect to the AOD is higher for lower aerosol loadings, which means that the impact of the NO2 correction is more significant on lower AODs. This fact is also clear from Δα, which is higher not only for higher NO2 but also for lower AOD values as well. Interestingly, based on Fig. 6, the highest Pandora NO2 retrievals (reddish colors) are not associated with the highest AOD values, indicating that in Rome the high AOD loadings are not strictly associated with high NO2 pollution events. In fact, high AODs are frequently related to the long-range transport of elevated layers of desert dust, fire plumes, or a combination of both (e.g., Barnaba et al., 2011; Gobbi et al., 2019; Campanelli et al., 2021; Andrés Hernandez et al., 2022). Hence, it might be worth modifying the aerosol retrievals for high NO2 in those pollution-related events with low to medium AOD levels. More about AOD and aerosol type climatology for the Rome area can be found in Di Ianni et al. (2018) and in Campanelli et al. (2022).

In general, considering the climatological value chosen for Rome in AERONET retrievals, the use of actual, coincident NO2 measurements on the calculations of aerosol properties still seems to be useful for AOD < 0.3, while being quite low (less than 10 %) for AOD > 0.5 and almost negligible for AOD > 0.8. In most cases, AERONET retrievals seem to overestimate AOD and AE. However, there are cases of underestimation, especially in AE retrievals, which seems to be higher for lower AODs. Those underestimations correspond to overestimation of NO2 from satellite monthly climatological values used in AERONET retrievals. The estimated AOD and AE deviations are below 0.01 and 0.1, respectively, for the majority of observations, i.e., about 96 %–98 % of occurrences for both CNR-ISAC and APL-SAP (see also distributions in Fig. 6). The average AOD bias is between 0.002 ± 0.003 and 0.003 ± 0.003 (with the higher values observed at 380 nm), while the average AE bias is ∼ 0.02 ± 0.03. Overall, the mean AOD bias is low compared to the estimated uncertainties for the standard AERONET product, i.e., 0.01–0.02 (with the higher errors observed in the UV; Sinyuk et al., 2020). However, the mean AOD bias for the cases of high NO2 levels (> ∼ 0.7 DU) is ∼ 0.011 ± 0.003 at 440 nm and ∼ 0.012 ± 0.003 at 380 nm for APL-SAP and ∼ 0.009 ± 0.003 at 440 nm and ∼ 0.010 ± 0.003 at 380 nm for CNR-ISAC, which is comparable to the AERONET reported uncertainties. The estimated mean bias of AE retrievals for the cases with high NO2 (> ∼ 0.7 DU) is ∼ 0.08 ± 0.04 for both Rome sites. The threshold for NO2 has been selected as being the average Pandora NO2 (∼ 0.4) calculated from the whole data set plus 2 times the standard deviation.

The results for SKYNET observations are similar (Fig. 7), but only positive Δτ and Δα values are derived, indicating the overestimation of the aerosol properties, since the NO2 optical depth is not considered in the standard retrieval processes (see Eqs. 7–8). Δτ is defined as τSKYNET–τSKYNET_mod (see Eqs. 7–8), and Δα stands for αSKYNET–αSKYNET_mod (Eq. 12). In addition, the derived deviations in aerosol properties reach higher values compared to AERONET. Especially AE differences extend up to a value of about 0.7, which is more than double compared to AERONET results. Interestingly, these quite large Δα values (> 0.3) correspond to relatively low NO2 loadings (< 1.2 DU). The differences observed between the two networks can be partly attributed to the different wavelength channels used for AOD and AE retrievals. Similar to AERONET, the derived AOD and AE biases for SKYNET are below 0.01 and 0.1, respectively, for the majority of observations (i.e., about 85 % of occurrences for AOD and about 90 % for AE; see also distributions in Fig. 7). The overall average AOD bias is ∼ 0.007 ± 0.003, which can be assumed to be low, considering that Nakajima et al. (2020) have estimated a root mean square difference (RMSD) of about 0.03 for wavelengths < 500 nm in city areas in AOD comparisons with other networks. However, the mean AOD bias for the cases with high NO2 levels (> ∼ 0.7 DU) is found to be about 0.018 ± 0.003, which is comparable to the RMSD value reported by Nakajima et al. (2020). The overall average AE bias calculated in this study is ∼ 0.05 ± 0.04, whereas the AE bias averaged over the high NO2 cases is about 0.10 ± 0.05.

The World Meteorological Organization (WMO, 2005) states that, when comparing AOD retrieved from sun photometers, 95 % of the AOD differences should lie within ±(0.005+0.01/m) of AOD, where m is the optical air mass. The first term of the equation (0.005) represents the maximum tolerance for the uncertainty due to the atmospheric parameters used for the AOD calculation (additional atmospheric trace gas corrections, i.e., ozone and NO2 and Rayleigh scattering), while the second term (0.01/m) describes the calibration-related relative uncertainties, for which the WMO recommends an upper limit of 1 % (e.g., Cuevas et al., 2019; Kazadzis et al., 2018a). Based on the above, although the average deviations found in this study are low compared to the retrieval uncertainties, they cannot be considered negligible, especially the average systematic underestimation of AOD of about 0.007 from SKYNET, also bearing in mind that there are locations with much higher average NO2 compared to the city of Rome.

The statistics showing mean differences in AOD and AE AERONET and SKYNET retrievals using actual, coincident NO2 measurements are presented in Table 1. AERONET AOD retrievals at 380 nm are also included in the table. In addition, deviations of AOD and AE using daily or monthly averages of NO2 in AERONET and SKYNET observations are also investigated. Table 1 shows that the average deviations of AOD and AE values do not change significantly, regardless of whether the actual Pandora NO2 measurements or the daily or monthly mean values are used for the retrievals. The percentage differences for AOD lie within the range 1.2 %–1.9 % for AERONET, while they are more than doubled (5.3 %–5.7 %) for SKYNET. For the standard aerosol products of the latter, NO2 optical depth is not considered. The estimated percentage differences for AE are within 1.2 %–1.7 % and 2.6 %–2.9 % for AERONET CNR-ISAC and APL-SAP, respectively, and between 7 %–7.9 % for SKYNET APL-SAP. It should be noted that the spectral channels used in AERONET retrievals are 380 and 440 nm for AOD and 440–870 nm for AE, whereas SKYNET data refer to 400 and 400–1020 nm for AOD and AE, respectively.

Satellite sensors perform measurements globally and provide information on the air quality, even over regions that lack ground-based observations. However, as already mentioned for OMI in Sect. 2.3.1, the spatial resolution of the satellite retrievals is limited by the pixel size. Co-located S5P/TROPOMI observations, characterized by an improved spatial and temporal resolution compared to previous satellite missions (e.g., OMI), were also employed to investigate whether the ground-based retrievals of aerosol properties could be improved on a global scale. Again, the approach described in Sect. 2.4.1 and 2.4.2 was applied by replacing the Pandora total NO2 (cNO2PGN) with the corresponding columnar retrievals from TROPOMI. Based on the current satellite footprint (5.5 km × 3.5 km), a radius of 5 km around each ground-based station was selected for the spatial co-location. The TROPOMI NO2 data were time-interpolated to AERONET and SKYNET measurements. Despite the improved spatial resolution of TROPOMI, the NO2 corrections using TROPOMI data are expected to be less accurate than those performed with the Pandora product. For example, Lambert et al. (2021) showed a bias between TROPOMI and Pandora total NO2 column ranging from -23 % over polluted stations to +4.1 % over clean areas, with a median bias of -7.1 %, in the frame of the standard validation process of TROPOMI Level 2 NO2 products. Other studies have concluded similar results. For example, Zhao et al. (2020) showed a negative bias for the standard TROPOMI total NO2 product in the range 23 %–28 % over urban and suburban environments and a positive bias of 8 %–11 % at a rural site, while Park et al. (2022) showed 26 %–29 % negative bias and R2 within 0.73–0.76 over the Seoul metropolitan area in South Korea.

The statistical metrics of the averaged deviations of the modified AERONET and SKYNET AOD and AE retrievals using actual, co-located TROPOMI NO2 measurements from the network standard products are presented in Table 2. Similar to Sect. 3.1 and Table 1, the deviations of AOD and AE retrievals derived by employing daily or monthly mean TROPOMI total NO2 were also investigated. The average deviations of AOD and AE values do not change significantly, regardless of whether the actual TROPOMI NO2 measurements or the daily mean values are used for the retrievals. This behavior is expected, considering that TROPOMI overpasses occur once or twice per day, and hence, they do not capture daily variations in NO2. In the case of the monthly averaged TROPOMI NO2 data, the estimated differences between the standard and modified aerosol products drop notably for AERONET. However, there are still differences compared to OMI NO2 climatology due to the improved spatial resolution of the TROPOMI pixel. The average AOD bias is ∼ 0.001 ± 0.001 (with the higher values observed at 380 nm), while the average AE bias is ∼ 0.01 ± 0.01 for both AERONET stations. For the cases of high NO2 levels (> ∼ 0.7 DU), the mean AOD bias is ∼ 0.004 ± 0.001 at 440 nm and ∼ 0.005 ± 0.002 at 380 nm for APL-SAP and ∼ 0.003 ± 0.001 at both 440 and 380 nm for CNR-ISAC. The estimated mean bias of AE retrievals for the cases with high NO2 (> ∼ 0.7 DU) is ∼ 0.05 ± 0.04 and ∼ 0.02 ± 0.01 for APL-SAP and CNR-ISAC, respectively. In the case of SKYNET, the overall average AOD bias is ∼ 0.005 ± 0.002 for AOD and ∼ 0.04 ± 0.03 for AE. For the high NO2 cases, a mean AOD bias of about 0.011 ± 0.002 and an average AE bias of ∼ 0.07 ± 0.04 were calculated. Interestingly, the deviations of SKYNET retrievals using monthly TROPOMI data are very similar to those derived using the actual overpasses or daily averaged TROPOMI NO2, probably due to the fact that the NO2 optical depth is not included in the standard network AOD retrieval processes.

The percentage differences for AOD lie within the range 0.2 %–0.9 % for AERONET and are about 3.8 %–3.9 % for SKYNET, which are much lower compared to those derived using Pandora NO2 (see Table 1). The estimated percentage differences for AE are ∼ 0.8 %–0.9 % and ∼ 1.6 %–1.7 % for AERONET CNR-ISAC and APL-SAP, respectively, and about 4 % for SKYNET APL-SAP using actual or daily TROPOMI data. It should be noted again that the spectral channels used in AERONET retrievals are 380 and 440 nm for AOD and 440–870 nm for AE, whereas SKYNET data refer to 400 and 400–1020 nm for AOD and AE, respectively.

In order to investigate further the impact of high NO2 during pollution events on the retrieval of relatively low levels of AOD, we used measurements performed at APL-SAP on 25 June 2020, the morning of which there was a high NO2 event. In the upper panels of Fig. 8, the total NO2 measured from Pandora during that day is illustrated. For AERONET (left panels of Fig. 8), the satellite climatological values used in the retrieval of standard AOD product and their deviations from Pandora NO2 are also displayed. The standard and NO2-modified AOD and AE data from both AERONET and SKYNET (see also Sect. 2.4.1 and 2.4.2), in addition to the magnitude of the respective differences (Δτ and Δα), are presented in the middle and lower panels of Fig. 8.

The differences in the AOD and AE retrievals from both networks are significant only within a time span of about 3 h around the high NO2 event (∼ 07:00–10:00 UT) and can be assumed to be negligible for the rest of the day when the NO2 levels remain quite low. The median AOD bias for AERONET is about 0.003, with a maximum of about 0.02 at the peak of the event. The median and maximum AE biases are 0.014 and 0.11, respectively. It can be also noted that, in the case of SKYNET, both AOD (median value of ∼ 0.008 with a maximum of ∼ 0.03) and AE deviations (median and maximum values of ∼ 0.03 and 0.10, respectively) are a bit higher compared to the respective AERONET deviations of synchronous data. This can be mainly attributed to the fact that SKYNET standard AOD retrieval processes do not account for the NO2 absorption and can be partly explained by the different channels used in the detectors of the two networks.

In this section, a first attempt is made to investigate the effect of the modified AOD and AE retrievals based on the Pandora total NO2 observations on the annual trends of those aerosol properties. The annual trends of AERONET/SKYNET AOD and AE over both APL-SAP and CNR-ISAC sites, calculated by applying the approach described in Sect. 2.5, and their uncertainties (standard errors in the regression slope) are presented in Table 3.

It should be noted here that the aerosol data sets from the two networks correspond to slightly different time periods. In addition, there are significant gaps in the time series from CNR-ISAC due to instrument problems, and the COVID-19 lockdown period (February–May 2020) has been excluded from the data analysis. Therefore, the results in Table 3 are mainly intended to highlight how a different NO2 correction may affect the aerosol trends and should be interpreted separately for each individual site. Interpretation of the trend significance for the Rome area is not possible using only this short period of time (∼ 5.5 years), considering that the estimated trends are quite small and the uncertainties introduced by linear regression are relatively high.

One aspect shown here is that the difference in the AOD and AE trends for the two data sets (original and modified NO2) is comparable with the calculated trends. As expected, AE trends with and without NO2 correction show relatively higher differences, as AE is much more sensitive to spectral AOD changes. However, the linear fitting uncertainty of AE is also high. NO2 effects on AOD trends would be more obvious in the case of a significant NO2 trend during a certain period. A thorough long-term trend analysis is out of the scope of this work and could be the topic for a future study.

In this section, we have analyzed a potential effect of considered NO2 corrections on the agreement of AERONET and SKYNET AOD products with relevant satellite data. Indeed, it is well known that most satellite retrievals are validated against ground-based measurements of AOD that are considered to be a ground truth. Moreover, most satellite retrieval algorithms are substantially tuned to closely match AERONET observations. For example, all MODIS algorithms, including DB, rely, in one way or another, on AERONET dynamic aerosol models and climatologies of AERONET retrievals. Nonetheless, since MODIS retrievals fundamentally rely on MODIS radiances that are fully independent of AERONET data, some inaccuracies in the assumptions, such as those regarding the NO2 amount, can cause some additional biases between AERONET and MODIS AOD results.

To evaluate the effects of the proposed correction, we have compared AERONET and SKYNET AOD products against MODIS DB AOD products at 470 nm for the 2017–2022 period. In the intercomparison, we considered only MODIS DB AOD products for which the distance between the center of the pixel and the AERONET site location (APL-SAP or CNR-ISAC) does not exceed 5 km. Furthermore, we considered all the AERONET (or SKYNET) AOD data within ±30 min from the MODIS satellite overpasses. In order to guarantee the quality of the data, we used MODIS DB AOD with a QA index ≥ 2, which corresponds to good and very good products (Wei et al., 2019).

The intercomparison has been performed using MODIS DB AOD at 470 nm. Consequently, we computed the AERONET and SKYNET AOD at 470 nm, thus exploiting the AE. The AERONET AOD at 470 nm was calculated using the standard AERONET AOD at 440 nm and AE at 440–870 nm. Similarly, the SKYNET AOD at 470 nm was computed using the standard SKYNET AOD at 400 nm and AE at 400–1020 nm. The NO2-modified AERONET and SKYNET AOD values at 470 nm were also computed with the same approach, and the AOD and AE retrievals have been modified using the Pandora NO2 data.

We observe a generally satisfactory agreement between the ground-based (both AERONET and SKYNET) and MODIS DB AOD data, with a Pearson correlation (r) higher than 0.7. In general, MODIS DB AOD slightly overestimates the AOD observed by the sun photometers. The bias (calculated as satellite minus sun photometer AOD) between MODIS DB and the different ground-based data sets before the correction (upper panels of Fig. 9) varies from -0.009 for SKYNET APL-SAP data (-0.008, considering AERONET) to 0.027 for AERONET CNR-ISAC. AERONET data, available for both sites, highlight a lower agreement for the CNR-ISAC site, with a bias about 3 times larger with respect to the APL-SAP site. The correction introduces a slight change of about 0.003 in the agreement between MODIS DB and AERONET AOD products and of 0.006 between MODIS and SKYNET data (lower panels of Fig. 9). Figure 9 also shows an improvement in the percentage of MODIS AOD data falling within the expected error (EE) of ±(0.05 + 20 %; Hsu et al., 2013) for APL-SAP by adopting the correction for both AERONET and SKYNET.

In Fig. 10, we show the absolute correction (computed as the difference between original AERONET/SKYNET AOD data at 470 nm and modified ones) as a function of the MODIS DB AOD and the NO2 column retrieved by the Pandora instruments located at APL-SAP and CNR-ISAC sites (upper panels). As already highlighted, we observe that the correction only depends on the NO2 amount and not on the AOD. Figure 10 also highlights that, although the improvement is relatively low on average, the correction can be larger than 10 %/15 % in many cases.

This intercomparison exercise demonstrated that the proposed correction slightly improves the agreement between MODIS DB AOD data and AERONET and SKYNET AOD products, even if, on average, it is not statistically significant. Nevertheless, as shown in Fig. 10, the improvement becomes significant when the differences between the NO2 values observed by Pandora and the OMI NO2 climatology are also significant (lower panels in Fig. 10). Furthermore, since the proposed correction depends on the amount of NO2, the improvement is more evident in the correspondence of high values of NO2 (upper panels in Fig. 10), typical of highly polluted areas such as the urban area of Rome (APL-SAP). Also, a slight improvement is also achieved in the suburban area of Rome (CNR-ISAC). Finally, in the case of SKYNET AOD products, the systematic overestimation, due to neglected NO2 extinction in the official retrieval chain, is eliminated.

One of the main impacts of accurate characterization of the columnar NO2 concentration is certainly expected on the retrieved values of SSA in spectral ranges coinciding with NO2 absorption. In order to quantify this effect, the sensitivity of the AERONET retrieval of SSA at 440 nm has been tested. As previously explained (Sect. 2.6), two different GRASP approaches have been applied to this purpose, namely the GRASP/AERONET NO2 and the GRASP/Pandora NO2. Despite the close methodological basis between GRASP and AERONET retrievals, the divergence in the development of both algorithms has led to some differences in the retrieved products. Thus, in order to ensure that the difference in the retrieved SSA at 440 nm is produced exclusively by the changes in the description of NO2 absorption and to avoid the inclusion of any other sources of discrepancy, the GRASP code has been used in both approaches instead of the standard AERONET SSA product.

The comparisons of the SSA at 440 nm obtained with both methodologies for the two stations for the complete data set (not shown) do not show a clear influence of the change in the NO2 concentration. High correlations (R > 0.98) and a mean bias error (MBE < 0.002) very close to zero are obtained. The mean NO2 column concentration for the retrievals presented here is 0.4 DU. Thus, in general, the analyzed improvements are not expected to produce an important change in the retrieved parameters at 440 nm in conditions with relatively low NO2 absorption. However, in the cases where NO2 concentration is elevated compared to the climatologically expected range, significant changes in the SSA at 440 nm retrievals can be appreciated. Figure 11 shows the comparisons of the SSA at 440 nm obtained with GRASP, following an AERONET-like approach (x axis) and the approach with the new NO2 concentrations provided by Pandora (y axis) and filtered for NO2 concentrations higher than 0.7 DU, which corresponds to the average NO2 plus 2 times the standard deviation. The two stations are correspondingly represented in the left and right panels. As it can be noted, for both stations in conditions of high NO2 concentrations, there is a consistent positive bias of ∼ 0.02 (∼ 2 %). However, a high correlation (R > 0.96) and root mean square errors (RMSE < 0.03) are also observed. Previous studies found SSA retrieval uncertainties in the range of 0.02–0.03 (Eck et al., 2003; Corr et al., 2009; Jethva et al., 2014; Kazadzis et al., 2016), whereas the correction, when high NO2 is recorded, is usually higher. Thus, it is clear that in conditions of high NO2 concentrations an accurate characterization of this gas is necessary in order to avoid noticeable bias in the affected AERONET channel around 440 nm.