A detailed description of the GODFITv3 algorithm is given in or , here it is shortly summarised. The total ozone column data were generated in the framework of the ESA Ozone Climate Change Initiative (Ozone CCI) and are available on the CCI web page: http://www.esa-ozone-cci.org (March 2016).

The GODFIT algorithm minimises the difference between a sun-normalised calculated earthshine spectrum and the observation between 325 and 335 nm using standard minimisation tools. Therefore, a linearised forward model is used with the state vector including the TOZ, a temperature shift, the effective surface albedo (polynomial 3rd order), the Ring-effect correction term and an earthshine Doppler shift. Among other variables, the fit varies the TOZ which is then used to derive an ozone profile from a column classified ozone climatology based on TOMS data. For the lowest altitude layers, a better representation was found by replacing the TOMS data with the OMI/MLS tropospheric ozone climatology . Based on the ozone profile as well as the other atmospheric parameters (e.g. temperature profile), the radiative transfer model (LIDORT) calculates the intensity at the top of atmosphere as well as the Jacobian. During each minimisation step the intensity has to be calculated.

The cloud fraction and cloud height are taken from cloud products calculated before from the same instruments. In the effective scene approach as proposed by , the effective altitude results from the cloud fraction weighted mean of the cloud top height and the ground altitude. The effective surface albedo is included in the GODFIT minimisation retrieval. Even though the retrieval only includes the column above the effective surface, the total column still represents the complete column including the troposphere. The final profile is integrated between the surface and the cloud altitude to calculate the ozone column below and inside the cloud (ghost column).

For GOME, SCIAMACHY, GOME-2A and GOME-2B the cloud data (altitude and cloud fraction) are taken from the FRESCO v6 algorithm , which is based on the O2 A-band. Due to the shorter spectral range the OMI cloud data are derived form O2-O2 absorptions at 477 nm. The cloud albedo is fixed at 0.8 in both cloud algorithms. For S5P the cloud data will be calculated using the OCRA/ROCINN algorithm, which is also based on the O2 A-band . The TROPOMI offline total ozone data product will be retrieved with the GODFIT algorithm.

Although the same algorithm was applied to GOME, SCIAMACHY and GOME-2A including the cloud algorithm, the total ozone columns of the three sensors deviated from each other and possibly showed temporal drifts Fig. 7 in. Instrumental degradation e.g. causes errors in the absolute radiation of the level 1 data, thereby causing errors in the TOZ and the effective albedo retrieval. As a solution, a soft calibration of the data was introduced. The measured spectra are compared to simulated spectra in the wavelength range between 325 and 335 nm. The simulated spectrum depends on the ozone column; therefore for eight European stations Brewer ozone columns were included in the simulation of the respective spectra. Look-up tables of reflectance corrections factors for the different sensors were built, which depend on the wavelength, the time, the solar zenith angle and the instrument viewing angle. The measured reflectance is multiplied with the correction factor, properly interpolated through the look-up table.

A direct consequence of this soft calibration is to somehow align the retrieved ozone columns to the Brewer observations. In future versions a new calibration method based only on satellite observations is planned for the harmonisation of the total ozone columns.

To filter out outliers, the total column data are rejected if the rms fit residuals exceed an instrument-dependent threshold. In total roughly 2 % of the data were rejected.

The tropospheric column ozone will be operationally calculated from both the near real time and the offline total ozone columns for TROPOMI. The convective cloud differential algorithm summarised in the following originates from . Compared to the original algorithm some improvements and adaptions to the GODFIT data have been made. The tropospheric columns can only be measured by satellites during cloud-free observations. On the other hand this means that observations above high-reaching clouds hardly contain any tropospheric signal. Therefore, the measurements above deep convective clouds with a large cloud cover can be used to estimate the stratospheric column. For the retrieval of the total column (Sect. ), this effect is considered by adding a ghost column to the stratospheric signal. The ghost column is based on a climatology and includes the lowest part of the column, below as well as inside the cloud up to the effective cloud top height. Hence subtracting the ghost column from the total column results in the above-cloud columnar ozone (ACCO): ACCO=TOZ-ghost. Compared to (Eq. 3 therein) this is one adaption to the GODFIT data set and will also be used for offline tropospheric column ozone from TROPOMI.

To determine the stratospheric ozone only, clouds with a top height higher than 8.5 km are taken into account. Nevertheless the cloud top varies in a range from 8.5 km up to ≈15 km. The average cloud top height above 8.5 km in the current GODFIT data set is close to 10 km; therefore the ACCO are normalised to 10 km (≈ 280 hPa). For low cloud altitudes, the ACCO includes the partial ozone column between cloud top height and the 10 km level. This partial column is rated by a climatology-based column and subtracted prior to averaging all ACCO observations in a grid cell (Fig. ). For clouds with a cloud top height above this altitude level, a respective climatology-based correction column is added. For the ozone climatology, we used the sonde-based data set by . The pressure altitude grid was interpolated to an altitude grid using the climatology from . The correcting column is typically less than 1 DU. For GOME-2A (Sect. ) and January 2012 we checked the correction column in detail and found an average of -0.006±0.196 DU, the extreme values were -2.626 and 7.314 DU. The cloud top altitudes are taken from the cloud data retrieved in a separate step before the ozone retrieval. The cloud top determination is based on either O2 or O4 absorptions (Sect. ) and results in an altitude that is typically lower than the physical cloud top level. This causes an uncertainty in the ACCO, partly because the ozone inside the cloud is included in the ACCO .

The top height of 10 km is low for a tropospheric product, but shifting this altitude to higher levels only adds an offset which is given by the climatology profile between the higher altitude and the average cloud top height. Even though the tropical stratosphere begins at roughly 17 km (≤100 hPa) we call the corrected ACCO stratospheric column ozone (SCO). The optimal top height for TROPOMI is currently under investigation but can hardly be fixed prior to the launch.

The SCO is determined over a clean reference area with a sufficient frequency of high convective clouds 70∘ E and 170∘ W and averaged over a certain period. In this reference area, the error introduced due to in-cloud ozone is low. For TROPOMI a temporal resolution of several days (≈ 6) might be achieved. Thereby we assume that the stratospheric column ozone is constant in time and longitude. These assumptions are fulfilled in the tropics , which limits the algorithm to a range of 20∘ S to 20∘ N. Due to the seasonal migration of the ITCZ the data are binned to latitude bands of 1.25∘ each.

A measurement pixel is called cloud free if the cloud cover is less than 10 %. The CCD method is sketched in Fig. . Here the cloud-free observations as well as the convective cloud measurements are shown simultaneously. On the left, the ACCO is shown as the ozone column above the effective cloud top height given by the cloud retrieval. The correction term between the cloud top height and the fixed level (10 km) is subtracted from the ACCO. The other part of the figure shows the cloud-free case with less than 10 % cloud fraction, the TCO is the difference between the TOZ and the SCO: TCO=TOZcloudfree-SCO. The TCO and TOZs for cloud-free pixels were averaged over the time period used for the SCO and regridded to 2.5∘×1.25∘ resolution. Negative values in the averaged TCO were skipped. Due to the high resolution of the TROPOMI instrument and the expected large number of data points per grid cell, we assume that the complete ground pixel is in the same grid box as the centre point.

One of the basic assumptions of the CCD method is that the stratospheric column ozone is constant along the latitude bands and for the respective time period. In the winter months this assumption is not always fulfilled on both hemispheres. Stratospheric intrusions cause local changes in the stratospheric columns, which sometimes result in misleading tropospheric column ozone. Usually an automated quality control prevents this consequence for the TCO. The stratospheric reference data must meet four conditions of the quality control to be accepted.

The stratospheric column must not be lower than 200 DU.

The thresholds depend on the instruments and will be adapted to the real measurements as soon as they are available. Currently, default values are used based on the experiences gained in the data retrievals for the instruments mentioned below (Sect. ). In the final algorithm, the time resolution also has to be considered. The standard deviation of the TCO from the individual observations within a grid cell represents both the atmospheric inhomogeneity and the statistical error of the TCO. Therefore, it is an appropriate estimate of the error. The uncertainty in the TCO from TROPOMI can be estimated based on the current instruments (Sect. ) to a range of ≈ 3 to 5 DU. The uncertainty does not depend on the pixel size of the individual observers.

The algorithm as described in the previous section (Sect. ) was developed for the TROPOMI instrument on S5P. However, the launch is scheduled for early 2017; therefore the algorithm was applied to the data of the current European satellites: GOME, SCIAMACHY, OMI, GOME-2A and GOME-2B. All these instrument have a coarser resolution and less coverage. To some extent the algorithm had to be adapted to the different instruments. In a first step we reduced the temporal resolution from ≈ 6 days to 1 month, which increases the number of data points per grid cell. On the other hand the assumption of a temporal stable stratospheric column ozone might not be valid if the sampling period is longer than 1 month.

In the tropics, a grid cell of 2.5∘ longitude measures less than 280 km, hence it is smaller than a GOME pixel (≈ 320 km). In the latitudinal direction, however, the pixels are by far smaller than the grid cells (≈ 40 km vs. 139 km). Because of that, during the gridding process, the GOME data are weighted with the longitudinal fraction inside a grid box. The OMI footprint is 13 km × 24 km and hence small enough to apply the original S5P algorithm without any weighting. The TCO differ by ≈±0.3DU between the weighted and the non-weighted averaging even for GOME-2A and SCIAMACHY. The small difference supports the application of the faster and easier operational averaging, also for SCIAMACHY, GOME-2A and GOME-2B. Because for the stratosphere the column between 70∘ E and 170∘ W (120∘ or more than 13 000 km) is averaged, the weighting is not useful at this point for any of the instruments.

It is obvious that the threshold for the number of observations per latitude band differs between OMI and GOME. Also the number of tropospheric columns per grid cell varies between the individual sensors, for OMI the maximum numbers are above 1000 and for GOME the maximum sum of weights is usually between 40 and 60.

The standard deviation of the TCO is log-normal distributed with mean values between 3.3 and 4.39 DU, depending on the instrument. The width varies between 1.33 and 1.77 DU. In the merged product (Sect. ) the propagated standard deviations of the individual sensors determine the final error to 3.8±1.6 DU.

The algorithm was verified in the framework of the TROPOMI/S5P product development by a similar product from the University of Bremen . When applied to the same total ozone and cloud data, the difference between the two algorithms was typically less than 1 DU. A detailed error discussion based on the error propagation from the fits can be found in . While they apply the CCD retrieval to WFDOAS ozone columns, our data are derived from the GODFITv3 total columns (Sect. ,). Moreover the spatial resolution is 4 times coarser 5∘×2.5∘ compared to 2.5∘×1.25∘. Due to these differences the total error is underestimated compared to our variability based uncertainty: 1 to 2 DU compared to ≈3.8±1.6 DU.

The first results from TROPOMI are awaited at the beginning of 2017; until then we focus on the existing instruments. The TCO above the tropical Atlantic is strongly influenced by the sources of ozone precursors which originate from the forest fires in central Africa and are transported westwards with the trade winds. The migration of the ITCZ over the African continent causes similar seasonality of the rain season and the burning season, when harvested fields or parts of the rainforests are burned. A respective change in the location of the ozone maximum is visible in Fig. . The figure shows the 20-year average tropospheric ozone distribution for December, January, February (DJF) and June, July, August (JJA) as well as the propagated standard deviations from the monthly data. In JJA a clear ozone maximum on the central African coast is observed, in DJF when the burning season is further north, only moderated enhancement is observed there. A stronger maximum is found close to the South American coast and further south. This is probably caused by the biomass burning emissions from South America. Over the central Pacific (150∘ W to 150∘ E) the TCO reaches its minimum of less than 10 DU. In the same region, the standard deviation is also low.

The observed standard deviations are enhanced towards the northern edge in the boreal winter and southern edge in the austral winter. This effect is visible in both the merged data and in the data of the individual sensors. It might be related to some dynamical effects e.g. migration of the subtropical jet and the related stratosphere troposphere exchange e.g.. An enhanced downward transport of stratospheric ozone into the troposphere also causes a higher variability in the monthly means. Some of the TCO data at the winter edge of the tropics are dismissed and less data are considered in the averages and the deviations in Fig. . The latter indirectly contributes to the higher standard deviation in the winter compared to the summer.

After the harmonisation we compared our results with integrated soundings from SHADOZ project and the WOUDC project (http://www.woudc.org/, March 2016).

The sonde data from the stations listed in Table  were integrated up to 280 hPa (≈ 10 km) according to TCO=k⋅∑(ci(O3)⋅δPi), where k=0.789 DUppm-1hPa-1, P is the pressure in hPa and c is the average ozone mixing ratio at the respective pressure level (http://www.temis.nl/data/fortuin.html, February 2016).

The TCO data retrieved in the standard product have a monthly resolution. For most tropical stations four or fewer soundings per month are available. , however, showed that at least 12 soundings per month are required to adequately represent the monthly mean tropospheric column. Hence we extracted a 3-day product for the respective sounding days. Especially for GOME, the restriction on the exact soundings days was too strong and hardly any collocated TCO were found. Because of that we calculated a 3-day mean centred around the sounding day and area of 5∘×5∘ around the sonde station for the validation. The CCD method postulates the stratospheric column to be constant for roughly a month and zonally invariant, hence the monthly stratospheric columns were subtracted from the cloud-free total columns for the specific days and region. In total we have 4688 collocated observations. We added the same correction functions as for the harmonisation (Sect. ) to the 3-day means for the individual sensors before averaging to the merged data for each sounding station. The difference between the integrated sondes data and the merged satellite-based TCO shows a normal distribution with a mean of -1.7 DU and a width of 5.8 DU, indicating that the satellite data overestimate the TCO compared to the sondes. The top height uncertainty and the in-cloud correction (Sect. ) might be reasons for the observed bias.

The spatial and temporal averaged deviation was further investigated. For the four seasons the mean and the 10th and 90th percentiles are plotted in Fig.  for running 5-year periods. At the stations with 4 soundings per month this sampling period of 3 months and 5 years encompasses up to 60 soundings. For Nairobi, it was on average 46 soundings for the period 1995 to 2015. Within the 10th to 90th percentiles margin the data agree quite well. For the stations Nairobi and Natal the merged multi-sensor TCO tends to be higher than the integrated sounding data, as was already mentioned for the histograms.

Despite the algorithm for S5P, the focus of the manuscript is on the trend in tropospheric ozone data based on existing instruments. For example, for Natal a slight trend can be seen in Fig. , in JJA and SON. Furthermore, the sonde data seem to follow this trend, especially within the 10th to 90th percentile range (green band). A drift in the satellites' TCO might be misinterpreted as trend. So for each station in the tropics we subtracted the satellite TCO from the collocated sounding and fitted the combination of linear function and sine and cosine function as in Eq. (). We used the monthly mean differences as input to the fitting algorithm, rather than the complete data set of the differences. This increased the weight of the beginning and the end of the period where less collocated data are available. For the complete period the merged data product shows only a very small trend compared to the sondes with -0.014±0.226 DUdecade-1. Compared to the fit error it is negligible and will, therefore, no longer be considered in the trend analysis (Fig. ). The intercept of the fitted line equals -1.77 DU, hence the bias is confirmed. Also for the individual data points the slope was smaller than the fit error -0.151±0.169 DUdecade-1.

Besides the overall trend, the time series offers the possibility to study local or seasonal trends. The strong seasonal cycle for the TCO over the African coast is visible in Fig. . In this region a strong increase is found, but whether this increase is caused by increasing fire emissions or by an increase in the background TCO can hardly be explained with the figures above. In the overall trends the sine and cosine terms (Eq. ) reflected the seasonal cycle. When focusing on the individual seasons these terms must not be considered. TCOseason=a+b⋅t+e⋅ENSO3.4+f⋅solar+g50⋅QBO50+g30⋅QBO30

Again the fit parameters depend on the longitude and latitude, but also the trends for the tropical averages were fitted. Maps of the seasonal trends can contribute to clarifying this question (Fig. ). According to this figure the main increase in southern central Africa is found in the burning season, indicating that over the years more fields and forests might have been burned. showed that in the period 2000 to 2011 the burned area in Southern Hemispheric Africa increased by 1.8 %yr-1.

For other regions like the Indian Ocean the trends have opposite signs depending on the season. The TCO decreased between September and February and increased strongly between March and May. Overall the seasonal trend maps are quite noisy compared to the general trend map. This effect might be caused by the still limited number of data (21 years × 3 months = 63 data points in the maximum). The seasonal tropical average trends vary between 0.39 ± 0.91 DUdecade-1 in JJA and 0.58 ± 0.85 DUdecade-1 in SON. For all seasons the general fits are not robust and the fitted trends are less than the respective error.