The Global Ozone Monitoring Experiment 2 (GOME-2) are a series of nadir viewing instruments mounted on the Metop satellite series. These instruments fly in a polar orbit around the Earth and have an equator overpass time around 09:30 local solar time. The instrument uses back scattered light from the Sun in the UV-VIS spectral region (240–790 nm). The nominal swath width is 1920 km on the Earths surface with 24 pixels in the forward scanning direction of 80×40 km2 (cross track × along track). With the (nominal) wide swath width, global coverage can be achieved in 2 d.

In Table  the measurement modes and time periods included in this study are given. The GOME-2A instrument has gone through a number of modes. In December of 2008 the wavelength separation of band-1a/band-1b was shifted from 300 to 283 nm, resulting in stratospheric information with smaller footprints. In mid-July 2013, the swath width was changed from the nominal 1920 to 960 km, which remained until the End of Life de-orbit and shut down of the Metop-A platform in 2021. GOME-2B and GOME-2C have run in the same wide configuration since they were launched.

While the details of the ozone profile retrieval algorithm used to produce the ozone data are given in the Algorithm Theoretical Baseline Document (ATBD) of the Near Real Time (NRT), Offline and Data Record Vertical Ozone Profile and Tropospheric Ozone Column Products , we will provide a description of the relevant aspects of the algorithm below.

The Ozone Profile Retrieval Algorithm (Opera) retrieves vertical ozone profiles from GOME-2 data using the 265–330 nm (UV-VIS) spectral range from the Main Science Channels in band-1a, 1b and 2b. Because of the longer integration time, band-1a is used across multiple smaller band-1b/2b measurements.

The retrieved information consists of a vertical ozone profile of 40 ozone partial column values (in DU) on a fixed pressure grid from the surface to 0.001 hPa, a fitted albedo value (either surface albedo or cloud albedo) and an additional offset to correct for a systematic bias in band-1a (possibly a result of stray light or signal leakage) . The fixed pressure grid is adjusted to the surface height and, in case of clouds, to the cloud top height.

The LIDORT-A radiative transfer model is used in combination with a polarization correction. For the inversion we use Optimal Estimation , iteratively reaching a converged state of the atmosphere.

Based on the GOME-2 measurement alone, the ozone profile is under-defined, and therefore an a priori ozone profile climatology is used (resolved in latitude bands and months) .

The (initial) surface albedo comes from the GOME-2 version 3 Directional Lambertian-Equivalent Reflectivity (DLER) database (see ). The initial cloud albedo is set to 0.8.

Spectral spikes, such as those caused by highly energetic particles common over the South Atlantic Anomaly region, are removed by starting at 290 nm and filtering out all peaks below that wavelength above 4σ (where σ is the reflectance error), relative to the higher (valid) reflectance measurement. If there are more than 30 peaks removed, the remaining spectrum below the last valid reference measurement is removed completely because it is no longer representative.

An averaging kernel (AK) of an optimal estimation inversion is a matrix containing the weights of the linear effect of a change in the fitted parameters on the retrieved value at a specific position in the state vector . In other words, it represents how changes in the ozone partial column at various layers along the vertical model atmosphere affect the outcome of a particular retrieved layer. In an optimal retrieval, the AK would be close to the identity matrix (peaking near the nominal layer), but in real world situations vertical resolution of a retrieved profile is wider, and can span multiple layers or even the whole atmosphere.

The retrieved ozone profile depends on the a priori, the true ozone profile and the averaging kernel according to the following relationship (Eq. ): 1xretrieved=xa priori+AK(xtrue-xa priori) where x denotes a state vector (i.e.: the list of values: the ozone profile, albedo and offset), and the subscripts indicate the type of vector. Unfortunately, inverting the averaging kernel is not possible due to the matrix being singular (and also the pseudo inverse method does not produce sensible results), so the xtrue cannot be calculated on its own; it is always dependent on the other components and the relationship described above. Therefore a reference data set with a high vertical resolution can be easily compared to a satellite data set, not the other way around. In Sect.  we will revisit the topic of averaging kernels.

The gridded level-3 ozone profile product is created based on a grid definition, which sets the name of the variable, the name of the source parameter, and the number of latitudes, longitudes and vertical layers in which to divide the horizontal and the vertical dimension. The default horizontal resolution is 0.25°×0.25° with 40 vertical layers from the surface to 0.001 hPa.

The ground footprint of each retrieved vertical ozone profile is divided into equal footprint-sub-cells (4 along track ×8 across track, creating 10 ×10 km sub-cells) which centers are projected (binned) onto the fixed grid. At the end of the data binning, the averaged value, the standard deviation and other variables are calculated.

In the gridded ozone profile product, we provide the arithmetic mean, and the population standard deviation associated with the arithmetic mean.

The arithmetic mean value in a grid cell is calculated as follows: 2Arithmetic Mean=∑n=1NxiN where xi is the individual data point and N is the number of data points. In order to calculate the arithmetic mean, the number of data points in a grid cell and the sum of the data values in that grid cell are stored. This enables daily grids to be merged into monthly grids and monthly grids into yearly grids, without loss of information. This also applies to re-gridding onto coarser horizontal resolutions which can also be calculated without loss of information.

The population standard deviation σ in a grid cell is calculated as follows: 3σ=∑(x-x‾)2N4σ=1N∑n=1Nxi2-1N∑n=1Nxi25σ=N∑n=1Nxi2-∑n=1Nxi2N In order to calculate the standard deviation, the sum of the squared data values (∑n=1Nxi2) needs to be stored, in addition to the sum of the values (already stored for the Arithmetic Mean).

Table  lists the primary and derived values that are stored in the gridded ozone profile product.

As explained earlier in Sect. , the AK of the ozone profile describes the linear effect (weights) of ozone in all atmospheric layers on a particular (nominal) retrieved layer. In the level-2 ozone data, each individual retrieved profile has 40 layers and the associated AK matrix has 40×40 elements. While it is technically possible to store gridded/averaged AKs for all 0.25°×0.25° grid cells, this would make a regular level-3 gridded data product 40 times larger and unwieldy to use. As an alternative we have chosen to grid/average all AKs in 18 bands of 10° latitude.

An example of such a gridded/averaged AK is shown in Fig. , where for March 2019 the GOME-2B AK matrix for the latitude band at 45° N is shown on the left, and the smoothing weight curves for the nominal layers are shown in the middle. In the matrix plot, the x-axis (xi) can be seen as the nominal retrieved layer index, and the y-axis then represents the index of the smoothing weights of the contributing layers to the particular nominal layer at index xi. Also shown in Fig.  on the right is the ozone profile and a priori profile at 45° N/0° E for the same month, which gives a perspective in the shape of the profile and the position of the ozone maximum in the gridded data.

What stands out is that the bottom four layers have very little sensitivity (the matrix values are green across the vertical). In other words: the GOME-2 retrieved ozone profile is not very sensitive in the troposphere where (on average) relatively little ozone is present. For the next few layers, there is a positive sensitivity close to the nominal (indexed) layer, and a negative or positive weight higher up. Above layer 15 (31 hPa), the positive values start a little above the diagonal, which means that the profile is more sensitive to ozone above the nominal layer. For the top eight layers there is also little sensitivity, which corresponds with the upper stratosphere with diminishing ozone as one goes up.

In order to assess the the impact of these positive and negative effects indicated in the AK, we should also take into account the ozone content at those layers location. Even though the weights of the AK curves are fanning out at the top layers, their impact on the retrieved value is relatively small because there is very little ozone above layer 30 (1 hPa).

As mentioned above, we longitudinally average the AKs so they are representative for bands of 10° latitude in order to save space in the output product. While this is unconventional, a study done by shows that using a representative AK results in relatively small differences in a wide range of ozone profile cases (generally <5%–10 % across the whole profile). To explore this a little further, we present as an example the variability of the averaging kernels in the 40–50° N latitude band in Fig. . This figure shows the mean of the colocated averaging kernels (top) and the 10th, 50th and 90th percentiles (bottom row) of a full year of level-2 GOME-2A ozone profile retrievals in the 40–50° N latitude band. The percentile plots provide an indication of the spread of the AK values in the level-2 data used as a source for the level-3. While the change in the AK value along the percentiles is most prominent in the upper stratosphere (above layer 25 (3 hPa)), this is also the region with diminishing ozone and therefore has a smaller effect on the retrieved ozone. At the same time, the main structure along the diagonal remains intact across the whole range of percentiles in the altitude region where most of the ozone is present. The mean and median (i.e. the 50th percentile) of the averaging kernels are very similar, so in the level-3 gridded product we provide the mean because of ease of calculation. In the Appendix , the 10th, 50th and 90th percentile are shown for all 18 latitude bands for 2009 as a reference.

Figure  shows a time series of the global arithmetic mean of some of the parameters in the level-3 data set, which give a general impression of its behaviour over the mission period. It shows the atmosphere mole content of ozone (i.e.: the full atmosphere), the atmosphere mole content of the lower troposphere (surface to 500 hPa), the number of spectral detector pixels used in the (level-2) retrieval and the degrees of freedom of the resulting profile (the DFS is the trace of the averaging kernel). Due to specific validity of the spectral ranges of each of the GOME-2 instruments, the number of spectral pixels (NMeasurements) is also different. The degradation of the GOME-2 instrument contributes to a reduction of valid spectral information going into the level-2 retrieval over time. In the case of GOME-2A, the degradation (and quality control of the spectrum) causes an increasingly steep drop-off of the number of valid spectral information of GOME-2A after 2018. The effect of this can be seen in an increasing deviation of the total ozone column compared to the other instruments, and especially in the troposphere. The degradation in general also has an effect on the degrees of freedom of the retrieval. The DFS of GOME-2A has a discontinuity around September 2009, when an instrument de-contamination attempt caused a permanent degradation of the transmission and sensing of light through the instrument and detector.

Even though the retrieved level-2 and gridded level-3 data is available to users, we recommend to not use the ozone data from GOME-2A after 2018 and restrict our validation of this instrument therefore up to the end of that year.

Ground-based ozonesonde observations used in this study are available from the World Ozone and Ultraviolet Data Center (WOUDC, http://www.woudc.org, last access: 20 March 2026) and the NILU Atmospheric Database for Interactive Retrieval (NADIR) hosted by the Norwegian Institute for Air Research (NILU, https://ebas.nilu.no, last access: 25 March 2026). Ground-based lidar, FTIR and MWR ozone profiles were obtained from the Network for the Detection of Atmospheric Composition Change (NDACC, https://ndacc.larc.nasa.gov/, last access: 23 March 2026). All NDACC lidar, FTIR, and MWR instruments undergo a standardized evaluation process and extensive quality control .

Figure  provides an overview of the stations used for this validation. Ozonesonde stations are indicated by crosses, lidar stations by triangles, FTIR stations by circles, and MWR stations by squares. The latitudinal classification is as follows: northern polar stations (green, 67–90° N), northern mid-latitude stations (black, 30–67° N), tropical stations (red, 30° N–30° S), southern mid-latitude stations (gray, 30–70° S) and southern polar stations (blue, 70–90° S). For each station, we computed a monthly mean ground-based ozone profile, and compared it to the monthly mean level-3 ozone data obtained with a horizontal distance smaller that 100 km around the location of the ground-based stations.

The ground-based ozone lidar, FTIR, and MWR profiles were pre-processed and harmonized with the satellite retrieval grid in order to enable a consistent comparison. The lidar, FTIR, and MWR data files contain monthly mean ozone number densities (cm-3) between approximately 15 and 50 km. The lidar, FTIR, and MWR profiles were converted into partial ozone columns expressed in Dobson Units (DU) per layer according to: 6DUlayer=n(z)Δzcm2.687×1016, where n(z) is the ozone number density (in cm-3) and Δzcm is the layer thickness in centimeters.

For each month, the corresponding satellite data (retrieved ozone, a priori ozone, averaging kernel, and standard deviation) were extracted from the level-3 NetCDF ozone profile data record files. The satellite quantities, given as ozone column amounts in molm-2, were converted into Dobson Units by using 1DU=4.462×10-4molm-2. Ozonesonde data, already available in DU, were integrated over the GOME-2 pressure layers. The vertical coordinates of the satellite layers, originally defined in pressure, were converted to geometric altitude using the U.S. Standard Atmosphere (1976).

The averaging kernels (AKs) provided with the satellite product were applied to the ground-based ozone profiles from lidar, MRW, and ozonesonde following the standard smoothing approach described by Eq. (). Specifically, the smoothed ground-based profile x^gb was calculated following as x^gb=xa+A(xgb-xa), where xgb denotes the original ground-based ozone profile, xa the satellite a priori profile, and A the satellite averaging kernel. This operation ensures that the ground-based profiles are smoothed to the vertical sensitivity and smoothing characteristics of the satellite retrieval. All ozone profiles (satellite retrieved, a priori, ground-based, and smoothed ground-based) were subsequently interpolated and re-binned onto a uniform 1 km vertical grid up to 60 km altitude. The re-binning of partial columns from the irregular satellite layers to the uniform grid conserved the total amount of ozone in DU by weighting each layer according to the geometric overlap between the native and target bins. The described processing ensures that the ground-based profiles are expressed on the same vertical grid and smoothing characteristics as the satellite retrievals, thus allowing a direct and consistent validation of the satellite ozone profiles. FTIR ozone profiles are themselves retrieval products characterized by their own averaging kernels and a priori profiles. To ensure a consistent comparison with the level-3 satellite retrievals, a bidirectional averaging-kernel approach is applied following . In this framework, the satellite averaging kernels are applied in the Dobson Unit (DU) domain to the FTIR profiles while explicitly accounting for the FTIR a priori state. This procedure minimizes inconsistencies that arise from differences in vertical resolution and retrieval characteristics between the two data sets. Such treatment is consistent with established methodology for intercomparison of retrieval products with differing vertical sensitivities . A comparable strategy has been applied in previous satellite validation studies involving FTIR data, for example, in the validation of IASI ozone products using GOME-2 and ground-based observations , where the role of averaging kernels and a priori information in profile comparisons is explicitly addressed. For high-vertical-resolution reference measurements (lidar and ozonesondes), only the satellite averaging kernels are applied, as these measurements can be considered close approximations of the true atmospheric state.

For the validation the satellite data as well as the ground-based data were integrated over several altitude bands (e.g. 0–8, 8–14, 14–30, 30–40, and 40–50 km for mid-latitude levels). These band-integrated ozone columns were used to obtain time series for the ground-based, satellite, and smoothed ground-based datasets.

To assess the consistency of the GOME-2A/B/C level-3 ozone profile product in the troposphere, UTLS-zone and lower stratosphere, we compare the vertically integrated ozone partial profiles with the corresponding pressure levels derived from ozonesonde data, following Table . The averaging kernels are applied to high-resolution ozone profiles and those values are used in the evaluation. Table  summarizes the validation statistics as function of latitude and vertical regimes specified the three instruments respectively. In addition, Fig.  supplements the validation results from the mid-latitude stations in terms of time-series approach for the different latitude belts.

In the troposphere, all three GOME-2 instruments show predominantly positive biases in the Northern Hemisphere, with relative differences reaching 14 % to 15 % in the northern mid-latitudes for all sensors and up to 57.6 % in the tropics for Metop-C. Absolute differences in the northern mid-latitudes are consistently around 2 to 2.4 DU. Tropical biases are moderate for Metop-A and Metop-B (22 %–27 %) but substantially larger for Metop-C as earlier mentioned. In the southern mid-latitudes, positive biases are smaller (5 % to 12 %), while systematic negative biases are observed in the southern polar region (−7% to −10%). Overall, the tropospheric results indicate consistent Northern Hemisphere overestimation and enhanced variability compared to the UTLS, with notable inter-sensor differences in the tropics. For the UTLS region, all three GOME-2 instruments show consistent hemispheric differences. Positive biases are observed in the Northern Hemisphere, particularly in the northern mid-latitudes, where relative differences reach +26.0% (Metop-A), +23.1% (Metop-B), and +30.8% (Metop-C), corresponding to absolute differences of 4 to 5 DU. For northern polar stations, the biases are smaller but remain positive for Metop-A and Metop-B, while Metop-C shows near-neutral absolute bias with elevated variability. In the tropics, all sensors exhibit small positive differences (2 % to 5 %) and low variability, indicating stable performance. In contrast, systematic negative biases are found in the southern polar region (−9% to −15%). For the lower stratosphere, the intercomparison with ozonesondes shows that all three sensors produce comparable results, with relative differences within 10 %. A seasonal dependence is observed in the dataset, consistent with the Level 2 product. also reported this seasonal dependence in both the lower stratosphere and the troposphere. In the lower stratosphere, all three instruments exhibit systematic negative biases in the Northern Hemisphere and tropics, with relative differences typically between −2.5% and −6.8% and absolute differences of −3 to −7 DU. The strongest negative biases are observed in the northern mid-latitudes for all sensors. In contrast, the southern polar region shows small positive or near-neutral differences (0 % to 1 %), while southern mid-latitude biases remain weakly negative. Variability is comparatively low across all latitude bands and markedly smaller than in the troposphere and UTLS. Overall, the lower stratospheric results demonstrate consistent underestimation in most regions but relatively stable performance with limited dispersion. For the first two years of the GOME-2A mission, there was a particular instrumental integration time setting for the spectral bands 1a and 1b in the far- and near-UV regions, that made it difficult to distinguish the troposphere from the stratosphere on small ground pixels. This was changed by the end of 2008 (also listed in Table ). This instrument setting has affected the ozone profile across the full atmosphere for the first two years, but this effect should not cause the sustained bias for the first four years of the mission, as seen in Fig. . Our hypothesis is that the degradation correction, which spans multiple years in order to get the seasonal effect right, has a lingering effect.

To assess the consistency of the GOME-2 monthly mean level-3 ozone profile product in the upper stratosphere, monthly averaged satellite partial columns were compared with ground-based observations. In this altitude range, ground-based observations from lidar, MWR, and FTIR provide complementary information with different vertical sensitivities (see Table ). The analysis focuses on the 30–50 km altitude range, where nadir-viewing UV instruments provide good sensitivity to stratospheric ozone due to Rayleigh scattering. However, when comparing vertically resolved ozone profiles or partial columns in narrow altitude ranges, differences in vertical resolution and smoothing between satellite and ground-based observations need to be taken into account. Therefore, the application of the averaging kernel is required to ensure a consistent comparison between satellite and ground-based profile measurements. This is consistent with previous studies showing that averaging-kernel effects are less critical for total or broad stratospheric column ozone but become important for profile-based or layer-resolved comparisons e.g.,. The results of the comparison for mid-latitude and tropical stations are presented in Figs.  and , respectively. The mean smoothed ground-based profiles (left panel) closely follow the satellite retrievals, particularly between 40 and 50 km.

The time series of monthly mean ozone partial columns between 30–40 and 40–50 km (middle panels) show good agreement in both long-term variability and seasonal cycles, with maxima in summer and minima in winter. In the 40–50 km layer, the overall seasonal behavior is also reproduced, although a systematic phase shift between the satellite and ground-based time series is apparent. Although a bias in absolute values is apparent between the satellite and ground-based datasets, the satellite product reliably reproduces the temporal evolution of ozone in the upper stratosphere. The correlation coefficients for each altitude range are indicated in the middle panels. The higher correlation between the smoothed ground-based and level-3 time series, compared to that between the original ground-based and level-3 data, highlights the effectiveness of the apriori and averaging kernel-based smoothing in taking into account differences in vertical resolution. Here r denotes to the correlation coefficient calculated from the monthly mean time series. In the 40–50 km altitude range, the combination of narrow layer integration, occasional crossings of satellite and ground-based profiles, and the increased influence of the apriori profile can lead to reduced or even negative correlation coefficients, despite the overall consistency of the mean profiles. Such behavior is a known feature of nadir profile retrievals, where the vertical sensitivity and averaging kernels redistribute information across the profile and relative to the apriori, affecting the apparent temporal phasing and trend structure when compared to high-vertical-resolution references .

The relative bias (left panels in Figs.  and ) between the three satellite sensors and the smoothed ground-based data (right panels) remains within ± 10 % in the 30–40 km altitude range. In the 40–50 km region, the relative differences are larger, around 20 %, particularly during periods of ozone maxima and minima. Nevertheless, the mean relative difference, as reported in Table , is below 5 %. Overall, the ground-based validation confirms the robustness of the GOME-2 level-3 ozone profile product and its suitability for cross-validation with other satellite datasets and the construction of long-term ozone time series.

Figure   illustrates the inter-sensor differences between the three GOME-2 instruments respectively in the upper stratosphere for the altitude ranges 30–40 km (gray line) and 40–50 km (black line). Figure  is showing this for the troposphere within the altitude range 0–8 km (gray line) and for the lower stratosphere within the altitude range 14–30 km, averaged across all mid-latitude stations. The comparison reveals a high degree of consistency among the sensors. The smallest discrepancies are observed in the 40–50 km layer, indicating robust long-term stability in this altitude range. It is also shown in Fig.  that for GOME-2A, the data after 2018 is not consistent anymore and should not be used after that date.