The TOC products validated in this work and the respective algorithms are described in the following sections. The TROPOMI dataset used here spans the time period from its launch in October 2017, until 30 November 2018, hence a full year of operation is covered, including the commissioning phase “E1” that concluded at the end of April 2018. This phase started immediately after the initial switch-on and acquisition of nominal orbit characteristics, in order to perform functional checking of the end-to-end system on board the Sentinel-5P, as well as engineering calibration and geophysical validation of the first observations.

The validation of the NRTI and the OFFL products was performed using both direct-sun (DS) measurements from Dobson and Brewer UV spectrophotometers, as well as zenith-sky scattered-light measurements obtained with ZSL-DOAS (Zenith Scattered Light Differential Optical Absorption Spectroscopy) instruments. It should be noted that zenith-sky measurements are also obtained from Brewers and Dobsons, but an advanced processing is required to match the quality of DS observations (e.g., Fioletov et al., 2011), which is not available at a large set of stations. Moreover, even with such processing, these measurements still show shortcomings in very cloudy conditions (low light) and at high AMF. As such, they provide little additional value in the current context. Brewer and Dobson TOC direct-sun ground-based measurements have been used for many years now as a solid means of comparison, analysis and validation of satellite data. Past publications that have used these kinds of measurements include Balis et al. (2007a, b), Fioletov et al. (2008), Antón et al. (2009), Loyola et al. (2011), Koukouli et al. (2012, 2015a), Labow et al. (2013), Bak et al. (2015), Garane et al. (2018), etc. The instrumentation and the measurement principles are thoroughly described in Koukouli et al. (2015a), Verhoelst et al. (2015), Garane et al. (2018) and in references therein.

Daily means of TOC measured by Brewer (Kerr et al., 1981, 1988, 2010) and Dobson (Basher, 1982) spectrophotometers, deposited to the WOUDC (World Ozone Ultraviolet Radiation Data Center) archive (http://www.woudc.org, last access: 6 September 2019), were used. Additionally, individual Brewer TOC measurements are used, acquired from (a) the European Brewer Network (Eubrewnet, Rimmer et al., 2018, http://rbcce.aemet.es/eubrewnet/, last access: 6 September 2019) and (b) the Canadian Brewer Network (http://exp-studies.tor.ec.gc.ca/, last access: 6 September 2019). The advantage of the two latter networks is that the Brewer measurements are processed by the same algorithm, which creates a “common ground” among the stations. The Eubrewnet network consists of 46 stations, mainly in Europe and South America but also in North America, Greenland, North Africa, Singapore and Australia. After quality control (QC) of their measurements, some Brewers were excluded from the validation datasets, while others did not have available measurements for the time period of interest, leaving the network with 25 Brewers. The Canadian Brewer Network is comprised of eight sites, plus Mauna Loa, Hawaii (MLO), and South Pole (SPO) observatories where Brewers are operated jointly with NOAA. Every site (except SPO) has at least two Brewers, including one double spectrometer, while each Arctic site has three Brewers. Due to very low stray light, double Brewers produce reliable ozone measurements when the sun is low above the horizon (air mass values up to of 7 at SPO and 5 at all other sites). All Canadian Brewers are calibrated against the World Brewer Calibration Centre (the Brewer triad), located in Toronto (Fioletov et al., 2005).

As discussed by Garane et al. (2018), Dobson TOC measurements are affected by a well-known dependency on the stratospheric effective temperature, which has already been seen numerous times in satellite TOC validation studies (for, e.g., Kerr et al., 1988; Kerr, 2002; Bernhard et al., 2005; Scarnato et al., 2009; Koukouli et al., 2016). Hence, when the assumed stratospheric temperature deviates strongly from what is assumed by the algorithms, which is a phenomenon usually occurring during winter months, the differences between ground and satellite measurements increase (see the recent work of Koukouli et al. (2016), and discussion therein, on this topic). For the case of the validation of the ESA GODFIT v4 long-term satellite record, the expected global mean difference between the two types of instruments (Brewer and Dobson) was found to be about 0.6 % (Garane et al., 2018).

TROPOMI TOC measurements were also validated against ZSL-DOAS measurements from 13 instruments that constitute part of the SAOZ network (Système d'Analyse par Observation Zénitale; Pommereau and Goutail, 1988) of the Network for the Detection of Atmospheric Composition Change (NDACC, http://www.ndaccdemo.org/, last access: 6 September 2019). For applications where processed measurements are needed as soon as possible, such as this validation of the recently launched TROPOMI instrument, the Laboratoire ATmosphères Milieu Observations Spatiales real time facility provides a first processing of the SAOZ measurements within a week of the actual observation. This data are called LATMOS_RT and are used here. In the context of satellite validation, the SAOZ measurements are complementary to the Brewer and Dobson measurements for several reasons: (a) they use spectral features of the visible Chappuis band, where the ozone differential absorption cross sections are temperature insensitive, (b) the long horizontal stratospheric optical path allows measurements of the column above cloudy scenes, and (c) measurements are always performed in the same small SZA range (86–91∘). For further details on the measurement procedures we refer to Balis et al. (2007a), Verhoelst et al. (2015), Garane et al. (2018) and references therein. Additional information on the specific collocation approach, taking into account the actual area of measurement sensitivity, is given in Sect. 2.4.

The uncertainty of the Dobson ground-based instruments is estimated by Van Roozendael et al. (1998) to be approximately 1 % for direct-sun observations under cloudless skies and 2 %–3 % for zenith-sky or cloudy observations. The respective uncertainty budget for a Brewer spectrophotometer is about 1 % (e.g., Kerr et al., 1988, 2010). Note that instrument uncertainties vary from site to site depending on the instrument state, calibration history and other factors (Fioletov et al., 2004). According to Hendrick et al. (2011), the total uncertainty of the SAOZ measurements is of the order of 6 %, which contains the systematic uncertainty of the absorption cross sections (3 %). The random uncertainty of SAOZ spectral analysis is less than 2 %, going up to 3.3 % when the random uncertainty on the air mass factor, mainly impacted by clouds, is added (Hendrick et al., 2011).

Another, possibly important, source of bias between the different datasets discussed in this paper is the use of different ozone absorption cross section coefficients; while the Dobson and Brewer TOC algorithms are based on the traditional Bass and Paur (1985; BP) ozone absorption cross sections, the TROPOMI NRTI TOCs are extracted using the so-called “Brion–Daumont–Malicet” (BDM) cross sections (Daumont et al., 1992; Malicet et al., 1995; Brion et al., 1998), whereas the TROPOMI OFFL TOCs using the more recent Serdyuchenko et al. (2014), henceforth Serdyuchenko, coefficients. It has already been shown that, for the Brewer wavelengths, the replacement of the BP with the Serdyuchenko cross sections would cause a minimal reduction of the extracted Brewer TOCs of less than 1 %, whereas a replacement with the BDM would result in a reduction of the nominal TOC by about 3 % (see Fragkos et al., 2013; Redondas et al., 2014). For the Dobson wavelengths, the calculated TOC changes by +1 %, with little variation depending on which of the aforementioned cross sections is used (see Redondas et al., 2014; Orphal et al., 2016). These findings illustrate the current uncertainty associated with the use of different ozone cross section measurements between platforms and should be considered when examining biases between the different TROPOMI TOC algorithms validated against the Brewer and Dobson observations.

The lists of the stations used in this validation work for each instrument and database category are displayed in Tables S1–S5 in the Supplement. In Fig. S1 the respective maps show the very good geographical coverage of the Earth by the ground-based measurement sites used herein. Specifically, in Fig. S1a the WOUDC Network is shown, in Fig. S1b and c the two Brewer networks (Eubrewnet and Canadian) are shown, and in Fig. S1d the SAOZ stations are displayed. It should be noted that when Brewer ground-based (GB) measurements from WOUDC are used, only the Northern Hemisphere co-locations are considered because of the limited number and poor spatial distribution of stations with Brewer instruments in the Southern Hemisphere (SH).

After the generation of TROPOMI overpass files for each station including all relevant parameters for each measurement (date, time, spatial coordinates, solar zenith angle, error, cloud cover, cloud height, ghost column, etc.), a co-location methodology, similar to the one described in Garane et al. (2018), is applied using direct-sun GB measurements from Dobson and Brewers for the comparisons. One major difference compared to previous validation publications, such as Koukouli et al. (2015a) and Garane et al. (2018), is the maximum distance permitted between the direct-sun instruments' coordinates and the projection of the satellite's central pixel on the Earth's surface, which hereafter will be referred to as the “search radius of the co-location”. Due to the unique, high spatial resolution of the TROPOMI observations, it is apparent that the 150 km maximum distance co-location criterion should be significantly decreased.

Figure 1 investigates the effect of different co-location search radii on the percentage differences between GB and satellite measurements. OFFL TOC from TROPOMI and nine Brewer GB stations from the Canadian Brewer Network are shown to demonstrate the dependency of the mean percentage difference (Fig. 1a) and its standard deviation (Fig. 1b) on the spatial criterion chosen. It can be noted that the mean difference for each site (in different colors) remains almost stable when increasing the co-location radius. However, this is not the case for the respective standard deviation, which increases with distance between the satellite pixel and ground-based station location. This testifies that the radius of co-location used in TROPOMI TOC validation exercises should be as small as possible to ensure that the same air parcels are compared, while at the same time reserving a sufficient amount of co-location points, as was already demonstrated for GOME-2 by Verhoelst et al. (2015; their Fig. 11).

Investigating the optimal solution for the distance criterion, the closest distance between the projection of the TROPOMI's central pixel and the station's location for all the available co-locations of each GB station were studied. The dataset for this investigation consisted only of the closest co-locations found within 50 km for each satellite orbit and its statistical analysis showed that the median of the closest distance spans between 2 and 3 km, while its 75th percentile goes up to 4 km. However, we decided to keep the co-location criterion for the validation at 10 km, since no obvious increase in variability was found for the 10 km distance (Fig. 1) but mainly to ensure that the number of co-locations is high enough to have statistically significant results.

It should be noted that when investigating the closest co-location distance it was also seen that, for each S5P CCD pixel, only 3 % of the total co-locations had a closest distance of 10–50 km. Out of those, almost 90 % were assigned to CCD pixels number 3 and 450, due to geometry reasons, i.e., the periodical capturing of some stations by the edges of the orbit's swath. As it is thoroughly explained in the OFFL and NRTI S5P Mission Performance Centre (MPC) product readme files (Heue et al., 2018; Lerot et al., 2018), no data from CCD pixels 1 and 2 are available, due to the lack of cloud information. As it is reported, this is caused by a misalignment of Band 3, used for the total ozone retrievals (450 pixels per scan line), and Band 6, used for deriving the cloud altitude information (448 pixels per scan line), which led to the application of a shift of two detector pixels between the two bands. Therefore, due to the lack of cloud information for the first two pixels, the respective data could not be analyzed.

Daily values of TOC retrieved from the WOUDC and the NDACC databases were widely used in previous studies for GOME2/Metop (Koukouli et al., 2015a), IASI/Metop (Boynard et al., 2018), OMI/Aura (Garane et al., 2018), and SBUV/NOAA (Labow et al., 2013) data validation. In addition to daily values, individual GB measurements from Eubrewnet and the Canadian Brewer Network are also used in this study. Thus, the effect of the time difference of the sensing between satellite and ground-based measurements had to be investigated. For this purpose, the mean percentage differences were computed for all co-located measurements with maximum temporal differences (Δtmax) varying between 5 and 60 min, keeping the search radius limit to 10 km. An example is presented in Fig. 2 for a middle latitude Eubrewnet station (Hobart, Australia, 42.9∘ S, 147.3∘ E), showing the mean and the standard error of the comparisons versus the Δtmax (blue data points with error bars). In this figure it was chosen to show the standard error instead of the standard deviation to take into account the effect of the number of co-locations for each case. The standard error of the mean decreases for temporal differences up to 40 min and after that the decrease is almost indistinguishable, even though the number of co-locations (displayed with the red squares) increases dramatically with Δt. The same conclusion was reached for all GB stations that were studied. Hence, it was decided that the temporal criterion applied to the individual measurements is to keep all co-locations within 40 min to ensure the reduction of the GB measurements' uncertainties and at the same time to have enough co-location points for statistically significant validation efforts.

The use of the quite strict spatial criterion of 10 km might seem contradictory compared to the rather relaxed criterion of 40 min temporal difference. However, we found this was the best option, especially for the high-altitude stations, where we need a strict spatial constraint to avoid biases due to the missing column, and the only way to have enough co-locations is to keep the temporal constraint moderate. The comparison between TROPOMI OFFL TOC and the Brewer GB measurements is presented in Fig. 3 for the example of the station in Manchester, UK, utilizing these coincident criteria. The blue open circles represent the comparisons of the satellite data to the individual measurements of the particular site (downloaded from Eubrewnet) with a maximum temporal difference of 40 min, while the red dots stand for the respective GB daily data acquired through the WOUDC repository. All co-locations included in the plot have a maximum search radius of 10 km and refer to the same time period of operation. In both cases, the mean bias is negative, even though it is different by 0.7 %, but the standard deviation of the mean is only slightly different between the two data series, which proves that even when daily means are used for the TROPOMI validation, the statistical results of the comparison are equally reliable.

Comparing TROPOMI to twilight SAOZ measurements is complicated not only by the different measurement times (TROPOMI overpass time versus the time of sunrise or sunset) but also by the large difference in horizontal resolution. It is well known that the air mass to which a twilight SAOZ measurement is sensitive spans many hundreds of kilometers towards the rising or setting sun (e.g., Solomon et al., 1987). Our co-location scheme takes this into account by averaging all TROPOMI pixels of a temporally co-located orbit (maximum allowed time difference of 12 h) within a so-called observation operator.

This 2-D polygon is a parametrization of the actual extent of the air mass to which the SAOZ measurement is sensitive. Its horizontal dimensions were derived using a ray-tracing code, mapping the 90 % inter-percentile of the total vertical column to a projection on the ground (Fig. 4), and then parameterizing it as a function of the solar zenith angle and azimuth angle during the twilight measurement, where the SZA during a nominal single measurement sequence is assumed to range from 87 to 91∘ (at the location of the station). Note that the station location is not part of the area of actual measurement sensitivity.

The average TROPOMI measurement over this observation operator can then be compared to the ozone column measured by the SAOZ instrument. An illustration of one such co-location is presented in Fig. 5. Note that at polar sites, the above-mentioned SZA range may not be covered entirely, in which case the observation operator is limited to noon or midnight depending on the circumstances (sunrise or sunset, close to polar day or polar night). For more details, we refer to Lambert and Vandenbussche (2011) and Verhoelst et al. (2015).

In the two following figures (Figs. 12 and 13) the TROPOMI OFFL TOC is compared to temporally common OMPS/NPP TOC measurements using the Brewer and Dobson spectrophotometer co-locations as reference. The blue and red lines represent the TROPOMI OFFL and OMPS GODFIT v4 TOC comparisons to GB measurements, respectively. Figure 12 shows the monthly mean time series of the percentage differences between the two sensors and the co-located GB measurements for the same temporal range. Figure 12a and b show the Northern Hemisphere and Southern Hemisphere comparisons to WOUDC Dobson GB measurements, whereas in Fig. 12c the Northern Hemisphere WOUDC Brewer comparisons are shown. The inter-sensor consistency is highly satisfying in terms of pattern. The enhanced annual variability for the Dobson comparisons is obvious here as well as in Fig. 6. The difference in the overall mean bias between TROPOMI and OMPS is less than 0.7 % for the NH, while in the SH the two sensors are almost identical. As for the mean standard deviation, TROPOMI has, in all cases, a lower variability in comparison to OMPS, which is within the product requirements, especially in the NH. One more interesting feature seen in Fig. 12a and c, is that for the NH comparisons the deviation between TROPOMI and OMPS seems to have a seasonality depending on the GB instrument type: for the Dobson comparisons the deviation is smaller in the summer months (June–August) and for the Brewer the same is true in winter months (November–February). Nevertheless, since we have only 1 year of available data, no solid conclusions about seasonality in the differences can be drawn.

Figure 13 shows the same temporally common co-locations for the two sensors but as a function of latitude. The comparisons to Dobson GB measurements and to Brewer GB measurements are shown in Fig. 13a and b, respectively. The latitudinal dependency is nearly the same for both sensors, which proves the good quality of the TROPOMI OFFL TOC measurements at all measurement sites, since the TOC from the OMPS instrument was repeatedly validated during its operational period. The inter-sensor consistency is very good in the midlatitudes of both hemispheres and in the NH high latitudes. This is likely because of (i) the higher number of stations (therefore co-locations) in these areas and (ii) the less variable atmospheric conditions in this part of the globe. Finally, in the NH, especially above 30∘ N, the TROPOMI OFFL TOC measurements are lower than those of the OMPS by 0.5 %–1 %, depending on the GB instrument type, which is a minor difference.

In line with the previous section, the inter-sensor consistency between the TROPOMI NRTI TOC and the GOME2/Metop-A and Metop-B (hereafter referred to as GOME2A and GOME2B) TOCs processed with the GDP 4.8 algorithm, is examined. The latter sensors were previously successfully validated and their validation report is published in Koukouli et al. (2015b). In the following figures the comparisons of the sensors to GB data are symbolized with a blue line for TROPOMI, green line for the GOME2A and orange line for the GOME2B percentage differences. Figures 14 and 15 show the time series and the latitudinal dependency of the comparisons, for the same temporal range and for common co-locations only, in accordance with the previous section.

In Fig. 14a, a quite different behavior is seen between TROPOMI and the other two sensors when compared to Dobson measurements in the NH. This can be attributed to the high overestimation of the NRTI TOC coming from the 70–80∘ N latitude bin that was previously discussed in Sect. 3. In the latitudinal dependency of the comparisons, seen in Fig. 15, a very good agreement between the three sensors is obvious in the NH, with deviations of up to ±1 %. The only exception is the highest latitude bin of the Dobson comparisons, as also seen in Fig. 7a. One would expect that since the NRTI product calculation is based on the GDP 4.x algorithm, the differences between the three sensors should be minor. However, the two algorithms (GDP 4.8 and NRTI) are different in some aspects, such as the surface albedo climatology used for the TOC retrievals, which is the main reason for the deviations discussed above. The other important updates are briefly discussed in Sect. 2.1.1 and are summarized in Table 3. Furthermore, it was found that the deviation between the two algorithms in this particular latitude bin is almost eliminated when TROPOMI data acquired during the commissioning phase of its operation are excluded from the dataset (not shown here). This is in line with the work of Inness et al. (2019) that detected enhanced discrepancies between TROPOMI NRTI TOC and other sensors in the high northern latitudes for this particular time period, when a lot of in-flight calibration and testing took place. Unfortunately, the 6 % difference between the NRTI and OFFL products in this area (Fig. 7a) is only reduced to 5 % when the same temporal restriction is applied.

The inter-sensor consistency is very good for the time series of the Brewer and the SH Dobson comparisons (Fig. 14c, b). The difference in the three sensors' mean bias is about ±0.7 % in both hemispheres and for both types of GB instruments. For the TROPOMI NRTI TOC product, the mean standard deviation of the comparisons is in all cases lower than that of the other two sensors used in this validation exercise, proving its good quality and its stability during this first year of operation. The seasonality pattern, already thoroughly discussed above, is evident here as well, mainly for the Dobson comparisons.

To summarize the results of Sect. 4.1 and 4.2, the statistical analysis of the comparisons between the four sensors (TROPOMI, OMPS, GOME2A and GOME2B) are shown in Table 4, where the differences of the mean bias between TROPOMI and GOME2A, GOME2B, or OMPS are shown along with the differences in mean standard deviation for each pair of sensors.

In this section we briefly present direct global TOC comparisons between TROPOMI and other UV–VIS sensors to directly exploit the global extent of the satellite-to-satellite comparisons, something not possible using only the GB measurements, due to their limited geographical coverage, especially in regions like the poles. The comparisons shown below are against the following sensors, already presented in the previous sections: (i) the NRTI TOC product will be compared to GOME2A and GOME2B processed with the GDP 4.8 algorithm and (ii) the OFFL TOC will be compared to OMPS processed with the GODFIT v4, as before. Additionally, since the GOME2A and GOME2B sensors are the European predecessors of TROPOMI, the OFFL TOC will be also compared to their measurements processed with the GODFIT v4 algorithm, as part of the C3S climate total ozone record production. The TOC datasets from the other sensors are restricted to the time period of the TROPOMI/S5P, namely from November 2017 to November 2018.

Daily NRTI observations, as well as the corresponding GOME2A/2B data records, were averaged on 2.5∘×2.5∘ latitude–longitude grid, while the OFFL data, and corresponding GOME2A/2B and OMPS data records, were placed on a 0.5∘×1.0∘ grid. For each pair of instruments, daily gridded relative differences were then computed for every grid cell containing measurements and all those daily difference grids were then either averaged in time to have a global representation of the spatial patterns of the differences (as shown in Fig. 16) or also averaged in space for certain latitudes bands. As such, Fig. 17 shows the gridded differences as a monthly mean time series for selected zonal belts.

In more detail, Fig. 16 shows the global distribution of the relative percentage differences between TROPOMI OFFL TOC and GOME2A (Fig. 16a), GOME2B (Fig. 16c) and OMPS (Fig. 16e) GODFIT v4 TOCs and between the TROPOMI NRTI TOC product and the GOME2A and GOME2B GDP4.8 TOCs in Fig. 16b and d, respectively. In general, total ozone columns from different satellite instruments agree quite well, especially at low and midlatitudes. The magnitude of those differences appear to be slightly smaller for the OFFL product than for the NRTI data, highlighting a better inter-sensor consistency. Differences tend to increase at higher latitudes where the more extreme geophysical conditions (large ozone optical depth, high variability in surface reflectivity, large observation angles) make the retrievals less accurate.

The OFFL product (Fig. 16, left column) appears to have a variable correlation to the other three sensors: i.

Compared to GOME2A (Fig. 16a), differences are generally very small (<±0.5 %). They are slightly larger only in high southern latitudes where they reach around -2 %.

The percentage differences of the OFFL TOC compared to the other three sensors demonstrate great temporal stability for every latitude belt, except for the belt south of 50∘ S (cyan line), where the variability is stronger. Those plots confirm the conclusions drawn previously, with differences generally lower than ±1 % at low latitudes and midlatitudes and slightly larger in polar regions. Recall also that the GODFIT v4 GOME2A and GOME2B datasets are produced with a Level 1b soft-calibration procedure, which introduces its own inaccuracies (Lerot et al., 2014). This might explain the slightly larger variability of the TROPOMI–GOME2A differences in the 50–90∘ S bin and the larger TROPOMI–GOME2B differences in August 2018. As shown in Fig. 16e and Fig. 17e, the OMPS TOCs are lower than the TROPOMI OFFL TOCs in the SH, where the cyan line shows differences up to +2 % during the polar winter and spring.

The TROPOMI NRTI TOC percentage differences exhibit a quite different behavior compared to the OFFL TOC product. The variability of the monthly mean time series seen in Fig. 17b and d, is more pronounced for all latitude belts except for the tropics. Each latitude belt has a different temporal dependency, which does not change when a different sensor is used for the comparison to TROPOMI.

Despite the differences between the two algorithms that emerged from this direct satellite-to-satellite comparison, it should be stressed that the mean bias of the percentage differences between TROPOMI and the other sensors is always within the product requirements, reproduced as the yellow and gray shaded areas in Fig. 17.