In this section, the most prominent quantities that affect the methodology are studied and the results are analyzed. The aim of the derivation of satellite NO
2 surface concentrations is mainly to provide information at locations where ground-based stations do not exist. Therefore, each quantity involved needs to be studied in detail and cross-validated with in situ measurements. Three instances and their imprint on the results are examined: the vertical levelling scheme used in the LOTOS-EUROS CTM and the CSO operator simulations (
Section 3.1.1), the versions of the S5P/TROPOMI satellite data (
Section 3.1.2), and the new updated air mass factors estimated through the CSO process (
Section 3.1.3).
3.1.1. LOTOS-EUROS Vertical Leveling Scheme
Initially, the effect of the LOTOS-EUROS CTM vertical leveling scheme on the results is examined. There are three methods to define vertical layers in the LOTOS-EUROS configuration: the mixed-layer definition, the hybrid-layer definition and the meteo-level definition [
39]. The
meteo-level definition, used in this study, adopts the level definition of the meteorological data. Layer interfaces can be defined as pressures or heights above the surface, depending on the meteorological data. This option can be more realistic when using the model at high resolution, depending on the application and resolution of the input meteorology.
In this work, the meteo-level definition is used in order to keep the model as consistent as possible with the ECMWF meteorological data. Two different meteo-leveling schemes are applied on the model runs. The first setup of model simulations, which is the base setup (hereafter mentioned as
meteo12 leveling scheme), uses 12 vertical layers and the second setup (hereafter mentioned as
meteo34 leveling scheme) uses the same configuration as the ECMWF model with 34 vertical layers. The meteo12 model simulations extend to approximately 9 km whereas the meteo34 simulations extend to 30 km. Both schemes include eight layers on top of the model, 20 and 42 total layers in total, respectively, filled by the boundary conditions in order to have full atmosphere to simulate total NO
2 columns. The first three layers of both schemes are identical. In meteo12, a coarsening of the layers takes place after the first three model levels, whereas the second setup is more detailed, providing information on each vertical layer corresponding to the ECMWF vertical layers. The meteo12 scheme is very efficient in terms of computation time, while the meteo34 scheme is computationally more expensive [
40].
Figure 5 shows the LOTOS-EUROS NO
2 surface concentration of the first model layer for both leveling schemes for a zoomed-in area covering the regions of Belgium, western Germany and the Netherlands. At first glance, no significant differences can be spotted between the meteo12 (
Figure 5a) and the meteo34 leveling schemes (
Figure 5b) NO
2 surface concentrations. By observing the absolute differences between the NO
2 surface concentrations of the meteo34 and the meteo12 schemes (
Figure 5c), however, it is evident that the meteo34 leveling scheme results in slightly higher concentrations over high-emitting land areas, by a mean of +1.2 μg/m
3, and sharper gradients over hotspots but generally lower concentrations over the sea. The same pattern is also observed in the winter months (
Figure A2).
A more detailed view of the differences between the two leveling schemes is shown in the vertical profiles of the modeled NO
2 concentrations (
Figure 6) for June 2019 and in
Figure A3 for January 2020. Layer interfaces are defined as heights above the surface according to the ECMWF data.
Figure 6a and
Figure A3a show the vertical profiles of a hotspot pixel while
Figure 6b and
Figure A3b depict the vertical profiles of a rural pixel.
Figure 6c and
Figure A3c show the differences between the two vertical schemes. Both hotspot and rural pixels are selected as the closest to a traffic and a rural station in the Netherlands, within the city of Amsterdam. Differences between the vertical profiles of the two leveling schemes are calculated for the first 12 common reference heights, namely the top of each layer from the meteo12 scheme, for both hotspot and rural pixels (
Figure 6c and
Figure A3c). In both summer and winter, the meteo34 scheme shows higher concentrations for the first three layers. On the contrary, meteo12 shows higher NO
2 concentrations between the fifth and the ninth layer (between 0.12 and 1.5 km), while for higher layers the differences become negligible. Differences are more pronounced for the hotspot pixel, where the meteo34 leveling scheme shows higher NO
2 concentrations for the first three layers by 0.2 μg/m
3 in June 2019 and by 0.9 μg/m
3 for January 2020 (
Figure A3c). The rural differences are an order of magnitude smaller than for hotspot pixels. Overall, for the first model layer, which is used to derive satellite NO
2 surface concentrations, the meteo34 leveling scheme shows 2–4% higher concentrations over the hotspot pixel and 6–10% over the rural pixel for both periods. For the whole central European domain and the first model layer, the meteo34 leveling scheme shows approximately 5% higher NO
2 concentrations for the summer months and 3% for the winter months.
The LOTOS-EUROS NO
2 surface concentrations of the first layer from both leveling schemes are applied to the third setup (
Table 1). Inferred TROPOMI v2.3 NO
2 surface concentrations are then estimated for each station type and studied period. The output is two surface products, updated with the new air mass factors and averaging kernels, derived from the two different leveling schemes. Those newly estimated datasets are intercompared for all station types and studied periods in order to assess the effect of the leveling scheme to the implemented methodology.
Figure 7 shows the scatter density plots of rural background (
Figure 7a) and rural industrial (
Figure 7c) stations for the two leveling schemes (
Figure 7b,d) for the winter. NO
2 TROPOMI inferred surface concentrations show an overall good agreement with the in situ measurements for both station types with correlation coefficients between 0.53 and 0.7. More specifically, the rural background NO
2 surface concentrations derived from the meteo12 leveling scheme show a correlation of 0.53 and slope of 0.67, whereas the concentrations derived from the meteo34 leveling scheme show a slightly higher correlation (0.55) and a slope of 0.75. Rural industrial correlations are 0.70 and 0.67, and the slopes are 0.79 and 0.94, respectively.
Table 2 and
Table A1 summarize the correlation coefficient, slope, and relative bias of the comparison between the inferred NO
2 TROPOMI v2.3 surface concentrations, derived from each leveling scheme, and the in situ measurements for winter and summer, respectively. Overall, the use of the meteo34 leveling scheme leads to improvements for almost all statistical parameters examined. The urban and suburban traffic stations bias is reduced, from −24.55% to −20.70% and from −26.90% to −23.18%, respectively. Suburban industrial stations bias is lower in the case of the meteo34 leveling scheme (−9.70%) compared to the meteo12 (−15.66%) and the rural industrial stations bias is significantly improved from −15.57% to −4.32%. For background stations, the mean relative bias is slightly higher in the case of the meteo34 leveling scheme for all station types by ~5–7%. Slopes are closer to the 1:1 line for the NO
2 surface concentrations derived from the meteo34 leveling scheme, except for the urban background stations. Correlation coefficients are very similar for both leveling schemes with the highest being calculated for the industrial stations (0.63, 0.62 for the suburban–industrial and 0.7, 0.67 for the rural–industrial stations) and the lowest for the traffic stations (0.47,0.48 for the urban–traffic and 0.43, 0.45 for the suburban–traffic).
In the summer (
Table A1), correlations and slopes are generally lower compared to the winter. For traffic stations, correlations range from 0.10 to 0.32 and slopes from 0.03 to 0.14. Relative biases are extremely high for both leveling schemes (~−75%), showing an overall poor agreement with the in situ data. This might be attributed to the higher underestimation of the in situ measurements by the model during the summer. Background stations show better statistical indicators, especially for the NO
2 surface concentrations derived from the meteo34 leveling scheme. Relative biases are lower for the meteo34 leveling scheme by approximately 8% when compared to meteo12. Finally, industrial stations show the highest correlations of all the station types (ranging 0.58–0.63) and the relative bias is lower by ~5% for the surface concentrations derived from the meteo34 leveling scheme.
Overall, the meteo34 leveling scheme yields a better agreement with the in situ measurements. Slopes are closer to 1 and biases are lower for most station types, except for the background stations in winter, where a modest overestimation of the ground-based measurements is found. The only significant drawback of applying the meteo34 leveling scheme to a larger dataset and longer period is that this option is computationally more expensive.
3.1.2. S5P/TROPOMI Versions Comparison
Another quantity that affects the results is the product version of the TROPOMI tropospheric NO
2 VCDs.
Figure 1c and
Figure A1c have already shown differences between the TROPOMI v1.3 and the TROPOMI v2.3 NO
2 tropospheric VCDs. TROPOMI v2.3 NO
2 concentrations are higher by approximately 3% in summer and by 11–18% in winter. Both v1.3 and v2.3 TROPOMI tropospheric VCDs are used as input in the implemented methodology for all the possible setups (
Table 1). Here, the imprint on the derived NO
2 surface concentrations of the third setup is shown for winter. The meteo12 leveling scheme was applied for the LOTOS-EUROS simulations due to computational reasons.
Figure 8 shows the scatter density plots of the urban traffic and background stations between the inferred TROPOMI v1.3 and v2.3 NO
2 surface concentrations of the third setup and the in situ measurements, for the winter. Both versions of TROPOMI inferred NO
2 surface concentrations show nearly identical moderate correlations for both urban traffic and background stations (
Figure 8). However, the relative bias is much improved, and the slope is closer to 1 in the case of the TROPOMI v2.3 inferred data, indicating that the concentrations of the latter dataset are closer to the ground-based truth.
The statistical indicators of
Table 3 show a significant improvement for the TROPOMI v2.3-derived NO
2 surface concentrations for all station types except from the rural background stations. More specifically, the mean absolute bias of the urban and suburban traffic stations decreases from 15.46 to 10.46 μg/m
3 and from 20.19 to 11.53 μg/m
3, while there is a significant improvement in the slopes. Urban and suburban background TROPOMI v2.3 NO
2 mean absolute bias improves from 3.86 to −2.21 μg/m
3 and from 2.27 to −0.89 μg/m
3. Rural background TROPOMI v1.3 NO
2 inferred surface concentrations lay close to the ground-based stations measurements with a mean absolute bias of 0.05 μg/m
3, whereas the TROPOMI v2.3 slightly overestimates the in situ measurements with a bias of −1.97 μg/m
3. Slopes do not show a major improvement for the TROPOMI v2.3 background-inferred NO
2 surface concentrations. This is possibly related to the known lingering background values issues of the TROPOMI tropospheric NO
2 data [
22]. Finally, suburban and rural industrial TROPOMI v1.3 inferred products show a higher bias compared to the TROPOMI v2.3 inferred data, by approximately 15%. Slopes are higher for the industrial TROPOMI v2.3 inferred products, improving from ~0.58 to ~0.78, reinforcing the fact that TROPOMI v2.3 inferred NO
2 surface concentrations correlate better with the in situ measurements. Correlation coefficients are not included in
Table 3, as they are shown extensively in
Table 2 and are nearly identical for the comparison of both TROPOMI versions. For the summer (
Table A2), inferred TROPOMI v2.3 NO
2 surface concentrations are slightly higher for all station types by approximately 2 μg/m
3. Relative biases are lower by approximately 5%, 14% and 13% for the traffic, background and industrial stations, respectively.
Overall, the comparisons between the TROPOMI versions and the in situ measurements clearly show that the TROPOMI v2.3-inferred NO2 surface concentrations, after the application of the updated air mass factors, correlate much better with the ground-based measurements.
3.1.3. Application of the Updated Air Mass Factors
Another important ingredient in the derivation of NO
2 satellite surface concentrations is the application of the updated air mass factors and averaging kernels to the satellite retrievals (Equation (4)) and the model simulations (Equation (6)) described in the CSO operator process. Therefore, the NO
2 surface products of all possible setups (
Table 1) and their comparisons with the in situ measurements are examined in order to determine if the application of the updated air mass factors and averaging kernels improve the results. As concluded in the previous section, TROPOMI v2.3 is optimal for the derivation of satellite NO
2 surface concentrations, and it is therefore used as input in all setups. For the LOTOS-EUROS simulations, the meteo12 leveling scheme was applied to the model configuration, since computational time is important, as the differences between the two leveling schemes are not so important for this analysis.
Figure 9 illustrates the relative bias among the inferred TROPOMI NO
2 surface concentrations of all setups and the in situ measurements, and among the LOTOS-EUROS a priori NO
2 surface concentrations and in situ measurements. The relative bias is extremely high for all setups for the urban and the suburban traffic stations, improving from ~−90% in the first setup to ~−78% in the third setup. For all different types of background stations, we find an improved relative bias. Urban background relative bias improves significantly from ~−77% to ~−45% and from ~−72% to ~−30% for the suburban background stations in the third setup. Rural background stations relative bias also decreases for the third setup from ~−78% to −53%, albeit this improvement is not as remarkable as for the urban and suburban background stations. Finally, for the suburban and rural industrial stations the bias notably decreases from ~−73% to ~−51% and from ~−60% to ~−35%, respectively. Worth mentioning is the fact that LOTOS-EUROS relative bias is quite low (~−8%) for the rural industrial stations compared to the other datasets. This can be attributed to the fact that TROPOMI v2.3 NO
2 data seem to underestimate NO
2 levels over the selected rural industrial pixels.
During the winter (
Figure 9b), there is an obvious improvement in the relative bias of all the involved parameters with the in situ measurements for all the station types compared to the summer results (
Figure 9a). Overall, inferred TROPOMI v2.3 NO
2 surface concentrations derived from the third setup seem to provide a more realistic product, closer to the ground-based truth compared to the baseline setup. More specifically, the bias of the urban and suburban traffic stations shows a remarkable improvement, ~−25% and ~−27%, respectively, by approximately 20–30% when compared to the first setup. Suburban and rural industrial stations bias for the third setup is approximately ~−16%, almost 15% lower for both station types compared to the first setup. For the background stations, a reversal of the sign for the relative bias is evident, considering the absolute levels of NO
2 over the background stations, but still is the lowest compared to the other station types. Inferred TROPOMI v2.3 NO
2 surface concentrations slightly overestimate urban and suburban background in situ measurements by 7.4% and 3.9%, whereas the overestimation is higher (10.37%) for the rural background stations. It is apparent that the second setup (in blue) shows a lower bias for the background stations (0.49%, 1.40% and 5.96%). This can be attributed to the enhancement of the TROPOMI NO
2 tropospheric VCDs by the application of the updated air mass factors and averaging kernels. Finally, inferred TROPOMI NO
2 surface concentrations, derived from the baseline setup, show the highest discrepancies with the in situ measurements for all the station types.
The effect of the air mass factors is better illustrated in the scatter plots of
Figure 10. Suburban background (
Figure 10a–c) and industrial (
Figure 10d–f) stations are depicted.
Figure 10a,d, shows the scatter plots between the in situ measurements and the inferred TROPOMI NO
2 surface concentrations of the first setup derived from Equation (7).
Figure 10b,e and
Figure 10c,f show the same comparisons, including the inferred TROPOMI NO
2 surface concentrations of the second and the third setup, respectively. The TROPOMI-inferred NO
2 surface concentrations of the second setup are calculated with the application of the TM5-MP averaging kernels to the LOTOS-EUROS simulations, whereas the inferred NO
2 surface concentrations of the third setup are estimated with the application of the updated air mass factors and averaging kernels on both the TROPOMI data and the model simulations. An improvement of the slope is found for both the second and third setup, from 0.71 to 0.77 and 0.78 and from 0.63 to 0.73 and 0.76, respectively. This statistical indicator shows that the third setup performs better compared to the other two setups, and the derived NO
2 surface concentrations are closer to the ground-based data. Correlation coefficients are moderate for the suburban background station (R~0.5) and good for the suburban industrial stations (R~0.65), varying insignificantly for each setup.
Concluding, it is obvious that the TROPOMI-inferred NO2 surface concentrations of the third setup perform better overall. Biases are significantly lower insummer. In winter, there is a remarkable improvement for the traffic and industrial stations; whereas, for the background stations, a slight overestimation is found, which, however, does not exceed the threshold of 10%. We should underline the fact that the second setup performs better for the background stations.