Bias estimations and corrections of total column measurements are applied and evaluated with ozone data from satellite instruments providing near-real-time products during summer 2014 and 2015 and winter 2015. The developed standalone bias-correction system can be applied in near-real-time chemical data assimilation and long-term reanalysis. The instruments to which these bias corrections were applied include the Global Ozone Monitoring Experiment-2 instruments on the MetOp-A and MetOp-B satellites (GOME-2A and GOME-2B), the total column ozone mapping instrument of the Ozone Mapping Profiler Suite (OMPS-NM) on the Suomi National Polar-orbiting Partnership (S-NPP) satellite, and the Ozone Monitoring Instrument (OMI) instrument on the Aura research satellite. The OMI data set based on the TOMS version 8.5 retrieval algorithm was chosen as the reference used in the bias correction of the other satellite-based total column ozone data sets. OMI data were chosen for this purpose instead of ground-based observations due to OMI's significantly better spatial and temporal coverage, as well as interest in near-real-time assimilation. Ground-based Brewer and Dobson spectrophotometers, and filter ozonometers, as well as the Solar Backscatter Ultraviolet satellite instrument (SBUV/2), served as independent validation sources of total column ozone data. Regional and global mean differences of the OMI-TOMS data with measurements from the three ground-based instrument types for the three evaluated 2-month periods were found to be within 1 %, except for the polar regions, where the largest differences from the comparatively small data set in Antarctica exceeded 3 %. Values from SBUV/2 summed partial columns were typically larger than OMI-TOMS on average by 0.6 % to 1.2 %, with smaller differences than with ground-based observations over Antarctica. Bias corrections as a function of latitude and solar zenith angle were performed for GOME-2A/B and OMPS-NM using colocation with OMI-TOMS and three variants of differences with short-term model forecasts. These approaches were shown to yield residual biases of less than 1 %, with the rare exceptions associated with bins with less data. These results were compared to a time-independent bias-correction estimation that used colocations as a function of ozone effective temperature and solar zenith angle which, for the time period examined, resulted in larger residual biases for bins whose bias varies more in time.
The impact of assimilating total column ozone data from single and multiple satellite data sources with and without bias correction was examined with a version of the Environment and Climate Change Canada variational assimilation and forecasting system. Assimilation experiments for July–August 2014 show a reduction of global mean biases for short-term forecasts relative to ground-based Brewer and Dobson observations from a maximum of about 2.3 % in the absence of bias correction to less than 0.3 % in size when bias correction is included. Both temporally averaged and time-varying mean differences of forecasts with OMI-TOMS were reduced to within 1 % for nearly all cases when bias-corrected observations are assimilated for the latitudes where satellite data are present.
The works published in this journal are distributed under the Creative Commons Attribution 4.0 License. This license does not affect the Crown copyright work, which is re-usable under the Open Government Licence (OGL). The Creative Commons Attribution 4.0 License and the OGL are interoperable and do not conflict with, reduce or limit each other. © Crown copyright 2019
Total column ozone biases from satellite measurements are typically within a
few percent. Changes of a few percent over time or between instruments are
significant in affecting the correct identification of long-term trends.
Near-global reductions in column ozone were
The validation of satellite remote sounding products usually includes a
comparison to ground-based measurements, which provide a long-term reference
record. For satellite instruments measuring column ozone, this typically
consists of comparisons to Brewer and Dobson spectrophotometers, and
potentially filter ozonometers. The main advantage of ground-based versus
satellite total column ozone measurements is that they can view the Sun
directly as opposed to relying on the backscatter of solar radiation,
reducing the complexity and error sources of the retrievals. The final
resulting systematic errors of the calibrated ground-based total column
ozone daily averages for well-calibrated and maintained Brewer and Dobson
instruments are no larger than
Total column ozone bias estimation for observations can be performed in different ways and depend on different factors, such as the solar zenith angle (SZA), latitude, and season. Seasonal and related latitudinal changes in biases may result from limitations in retrieval algorithms. For example, the retrieval algorithm might not adequately account for the temperature dependence of the ozone absorption coefficients. Differences and limitations in accounting for clouds and surface albedos may also contribute to errors in total column ozone (e.g., Antón et al., 2009b). Bias parameterizations may range from being spatially and temporally global to more local.
The harmonization of different data sets through bias correction can be applied for standalone analyses, for reanalyses, and in near-real-time data assimilation. The assimilation process consists of introducing information from observations into model forecasts through the generation of analyses, the statistical blend of earlier forecast and observations, which serve as the initial conditions for subsequent forecasts. The assimilation of column and stratospheric ozone measurements for ozone-layer forecasting has been conducted mostly as of about 25 years ago, ultimately culminating in operational ozone-layer and UV-index forecasts (e.g., Lahoz and Errera, 2010; Inness et al., 2013). This typically involves the utilization of measurements from single to multiple satellite remote sounding instruments with the use of ground-based and other remote sounding data for independent verifications and, occasionally, bias correction.
Traditionally, the assimilation process assumes that both the model forecasts and observations are statistically unbiased following an initial spin-up time (unless biases are estimated within the analysis step). Unremoved biases or systematic errors in the observations or forecasting model can potentially impact the quality of the analyses and forecasts (e.g., Dee, 2005; Dragani and Dee, 2008). This is important for total column ozone when it comes to monitoring for multi-decadal trends, as referenced in van der A et al. (2010), for both trends inferred from just the observations alone or from their use within a data assimilation system. Generally, while the effectiveness of bias-correction schemes in removing biases is constrained by limited knowledge of the truth, their impact in reducing relative biases between different assimilated observations and/or correlated fields can potentially be just as significant for improved forecasting. An example of the latter is in multivariate assimilation, where ozone and meteorological assimilation can be coupled (e.g., Dee, 2008; Dee et al., 2011).
Ideally, the anchor used within a bias-correction scheme should be accurate,
have a wide range of coverage in both space and time, and for near-real-time
applications be available within a few hours or less after measurements are
taken. The summed partial columns from SBUV/2 satellite instruments have
been recommended as an anchor for long-term studies (Labow et al., 2013).
This is due to the long-term coverage provided by the series of SBUV/2
instruments, combined with the low variations in time of the differences
between these instruments and ground-based data (usually within
The focus of this study is bias estimation and correction of column ozone for multiple satellite sensors, towards eventual use in near-real-time data assimilation. The bias-estimation and bias-correction methods developed in this study may be integrated into an assimilation scheme that could be applied in near real time and can be utilized for other constituents. In this paper, we evaluate several different bias-estimation schemes used to correct observations of column ozone from satellite-borne instruments. Most of these methods utilize colocated observation sets for bias estimation. From this consideration, OMI-TOMS was chosen as the anchor for bias estimation and correction, as its dense spatial coverage allows for more colocations with measurements from other instruments. As part of this work, the OMI-TOMS column ozone data were evaluated using ground-based Brewer, Dobson, and filter ozonometer observations, as well as compared to SBUV/2 column ozone, for the limited time periods in this study. For these data sets, a target maximum residual bias of 1 % following bias corrections was selected. This satisfies the column ozone 2 % accuracy requirement from the European Space Agencies' Climate Change Initiate programme (van Weele, 2016) for random error levels of up to 1.7 %.
In this paper, we examine several bias-correction methods that use a discrete binning in latitude and solar zenith angle that, unlike a functional parameterization, allows for arbitrary nonlinear dependencies. In addition, an alternative estimation method involving the dependency on the ozone effective temperature (the mean temperature weighted by the ozone profile), as employed in van der A et al. (2010, 2015), was explored. As discussed later in the paper, dependencies on factors such as changes in cloud cover and viewing zenith angle were not examined.
Following bias estimation, data assimilations of column ozone observations from individual and multiple satellite instruments were conducted with and without bias correction. The impacts on the resulting 6 h forecasts were then assessed. The assimilations were conducted with the Environment and Climate Change Canada (ECCC) meteorological assimilation system adapted for constituent assimilation. These assimilations were univariate ozone assimilations and utilized operational ECCC meteorological analyses. The data sources assimilated in this study and correspondingly involved in the bias-estimation analysis are the GOME-2 instruments on the European MetOp-A and MetOp-B satellites (Munro et al., 2016; Hassinen et al., 2016), the total column measuring instruments of OMPS (Dittman et al., 2002a, b; Flynn et al., 2006) on the Suomi National Polar-orbiting Partnership (S-NPP) satellite, and OMI aboard the Aura research satellite (Levelt et al., 2018).
This paper is organized as follows: Sect. 2 describes the utilized ozone observations covering July–August 2014 and 2015 and January–February 2015. Following a general quality assessment of the OMI data based on the available literature, Sect. 3 evaluates the OMI-TOMS column ozone data for these periods against ground-based measurements. Having assessed the quality of the OMI-TOMS data for these specific periods, Sect. 4 describes and applies three different bias-estimation approaches with the column ozone measurements of different satellite instruments relative to OMI. Following a description of the assimilation system, the impact of column ozone assimilation on 6 h forecasts for individual and multiple sensors with and without bias corrections is examined in Sect. 5 for July–August 2014 using comparisons to both OMI-TOMS and ground-based data. Conclusions are provided in Sect. 6. The Supplement for this paper provides additional figures and tables supporting and complementing the discussed and presented results.
In this section, we give a brief description of the column ozone observations involved in the implementation and evaluation of bias correction, in data assimilation, and in the validation of short-term forecasts. Observational data sets were obtained for the periods of July–August of 2014 and 2015, and January–February 2015. The main data sources of interest are those specifically intended to provide satellite-based column ozone allowing near-real-time (NRT) assimilation. These consist of OMI, GOME-2, and OMPS-NM (total column Nadir Mapper) instruments that rely on optical solar backscatter of ultraviolet radiation in the nadir or near-nadir and provide data only during daytime. Ground-based Brewer, Dobson, and ozonometer filter instruments and additional satellite-based data from OMPS-NP (partial column Nadir Profiler) and SBUV/2 are included for evaluation and validation purposes.
The Ozone Monitoring Instrument (OMI) aboard the Aura research satellite has
been in operation since August 2004. The instrument stems from a
collaboration between the Netherlands Agency for Aerospace Programmes
(NIVR), now called the Netherlands Space Office (NSO), and the Finnish
Meteorological Institute (FMI). The OMI instrument provides a cross-track
width of about 2600 km on the ground and total column ozone mapping at a
spatial resolution of 13 km along, and 24 km across, the orbit ground track
at nadir (e.g., Bhartia and Wellemeyer, 2002; OMI Data User's Guide, 2012).
Some strips of the OMI measurement tracks were removed due to the row
anomaly of the OMI instrument, which for the time period under
consideration affects 23 of the 60 rows (see
This study employs the OMS-TOMS V8.5 standard science column ozone data produced by NASA based on the Total Ozone Mapping Spectrometer (TOMS) total column retrieval algorithm. The OMI-TOMS algorithm (Bhartia and Wellemeyer, 2002) principally utilizes only two different wavelengths, one with strong and one with weak ozone absorption, to estimate the total column ozone and surface reflectivity. The OMI-TOMS column ozone has estimated root-mean-squared errors of 1 %–2 % (OMI Data User's Guide, 2012). The OMS-TOMS V8.5 standard science product is close to, but can differ slightly from, the OMI-TOMS NRT data (OMI NRT Data User's Guide, 2010; Durbin et al., 2010). The OMI NRT Data User's Guide (2010) and Durbin et al. (2010) indicate a daily maximum percentage difference of 2.6 % between the standard science and NRT products, with a weekly average maximum difference of 1.4 %. Further comparisons by the authors show mean differences generally between 0.02 % and 0.04% in July–August 2016 and January–February 2017. While not used here, the other common OMI total column ozone retrieval products are based on the Differential Optical Absorption Spectroscopy (DOAS) algorithm (Veefkind and de Haan, 2002; Veefkind et al., 2006; Kroon et al., 2008) from the Royal Netherlands Meteorological Institute (KNMI).
Global Ozone Monitoring Experiment-2 (GOME-2) instruments are on the MetOp-A
(GOME-2A) and MetOp-B (GOME-2B) polar-orbiting satellites, launched in
October 2006 and September 2012, respectively, and are operated by the
European Organization for the Exploitation of Meteorological Satellites
(EUMETSAT). As of 15 July 2013, GOME-2A has been operating with a swath
width of 960 km and a 40 km
The GOME-2 NRT products used here, as well as those for OMPS and SBUV/2, were acquired from the National Environmental Satellite, Data, and Information Service (NESDIS/NOAA) and stem from the TOMS approach. More specifically, the GOME-2 retrieved data products are from the TOMS V8 algorithm (Zhang and Kasheta, 2012). Alternatively, EUMETSAT provides GOME-2 total column ozone data based on the DOAS retrieval approach (Loyola et al., 2011; GOME User Manual, 2012; GOME-2 ATBD, 2015).
The Ozone Mapping Profiler Suite (OMPS) on the Suomi National Polar-orbiting Partnership (S-NPP) satellite, launched in October 2011, consists of a combined nadir mapper (OMPS-NM) and nadir profiler (OMPS-NP) and a separate limb profiler (OMPS-LP), which provide total column, partial column profile, and limb profile products, respectively. A second suite was placed onboard the Joint Polar Satellite System JPSS-1 satellite (Zhou et al., 2016), renamed NOAA-20 and launched in November 2017. The retrieved data used in this study are from the OMPS S-NPP nadir measurements and are considered to be at a provisional product maturity level. They do not include improvements from the various corrections, calibration adjustments, and retrieval algorithm updates performed since the original near-real-time acquisition for the July–August 2014 period (Lawrence E. Flynn, NOAA, personal communication, 2016). The OMPS-NM and OMPS-NP ozone retrievals from the SBUV V8.6 retrieval algorithms (Bhartia et al., 2013; as referred to by Bai et al., 2016) became available after the completion of the assimilation experiments conducted for this work.
The OMPS-NM retrievals, summarized by Flynn et al. (2014), were made at the
NOAA Interface Data Processing Segment using the ratio of the measured Earth
radiances to solar irradiances at multiple triplets of wavelengths. The
nadir mapper has a cross-track width of about 2800 km and a 50 km
OMPS-NP profiles, each with a 250 km
Both ground-based and satellite-based column ozone data serve as independent verifications of the OMI-TOMS measurements, with the former also used for validation of the forecasts resulting from data assimilation. These data are described below.
The ground-based data consist of Brewer, Dobson, and filter ozonometer total
column ozone measurements (Fioletov et al., 1999, 2008; Staehelin et al.,
2003) from the World Ozone and Ultraviolet Radiation Data Center (WOUDC)
and of Brewer and Dobson measurements from the Global Monitoring Division
of the NOAA Earth System Research Laboratory (see Coldewey-Egbers et al.,
2015, for various references on the validation of column ozone satellite
data with ground-based Brewers and Dobsons). Only direct-Sun, clear-sky
daily daytime averages from these instruments were used. The error standard
deviations for Brewer and Dobson direct-Sun data are no larger than
Data from the Solar Backscatter Ultraviolet instrument (SBUV/2) were used
for verification purposes. The ozone data from SBUV/2 for the
period of interest are from the NOAA 19 satellite (Flynn, 2007; Bhartia et
al., 2013; McPeters et al., 2013). Two versions of the total column ozone
data are used here. The first is from the SBUV V8.6 profile retrieval using
wavelengths in the range of 250 to 310 nm (Bhartia et al., 2013; summarized
by McPeters et al., 2013; see also Flynn, 2007) for which the total column
ozone is the sum of the partial column layers, and the second is from the SBUV
V8 total column retrieval using two wavelengths between 310 and 331 nm
(Flynn, 2007; Flynn et al., 2009). The ozone measurements cover 170 km
Differences between OMI-TOMS and ground-based Brewer and Dobson data have
shown long-term stability and relatively little solar zenith angle and
latitude dependence (Balis et al., 2007a; Koukouli et al., 2012; Labow et
al., 2013; McPeters et al., 2008, 2015). Comparisons of OMI-TOMS V8.5
total column ozone with Northern Hemisphere ground-based data by Labow et al. (2013) and McPeters et al. (2015) based on multiple years indicate an
average underestimation of OMI-TOMS of about 1.5 %. Figure 2 of McPeters
et al. (2015) shows variations of weekly mean differences about the
long-term average underestimation mostly within about
To further examine the acceptability of using OMI-TOMS as a reference for
bias correction, a mean differences comparison of OMI-TOMS V8.5 with
near-colocated ground-based data at available sites over the periods of
study was conducted. The colocation requirements are the same as those
specified in Sect. 4.1 for the inter-comparison of satellite sensors.
Summary results are shown in Table 1 and Fig. 1 (see also Tables S1 to S3).
Bimonthly mean differences over regions, globally, and for the individual
stations, were produced for the three periods of Table 1 based on totals of
53 Brewer, 40 Dobson, and 20 filter ozonometer stations. Figure 1 shows the
station locations and mean differences for the July–August 2014 period. The
sizes of the global mean differences over the different periods are in the
approximate ranges of 0.0 % to
Regional and global relative mean differences (%) of total
column ozone between OMI-TOMS and the specified ground-based instrument types
over July–August 2014/2015 and January–February 2015. The averaging excludes
stations having outlier station mean differences for each period (see
Supplement tables S1 to S3 and the text of Sect. 3), except for the two rows
for the latitude region 60–90
Mean total column ozone differences (%) between OMI-TOMS and Brewer, Dobson, and filter ozonometer measurements over July–August 2014. The colours blue to purple denote negative differences and the colours yellow to red refer to positive differences.
The regional mean differences are within 1 %, with the exceptions being Antarctica for both the Brewer and Dobson instruments and the region of the North Pole for Dobson and filter ozonometer instruments. The mean differences for both polar regions are all negative, indicating an underestimation of OMI-TOMS column ozone in these regions for these periods relative to ground-based data, which is likely related to high SZAs. The mean differences for Dobsons and filters are similar to each other but slightly larger than for Brewers, despite error levels for the filter instruments being about 1.5 to 2 times larger (Sect. 2.4.1) and the small data sets.
More severe underestimations of OMI-TOMS relative to ground-based
observations of 3 %–6 % occur during July–August in Antarctica, which is
associated with SZAs close to or greater than 80
While not done here, a correction specifically for high SZAs could be envisaged, as done by van der A et al. (2015). While the OMI-TOMS data could be underestimating the total column ozone in the polar regions for these periods, there may be some uncertainty as to the actual OMI-TOMS bias. Factors that could affect the reliability of the comparison with the ground-based data at high solar zenith angles for Antarctica, beyond the low number of ground-based observations, include retrieval assumptions about the ozone layer (e.g., Bernhard et al., 2005), stray light sensitivity (especially for Dobsons; e.g., Moeini et al., 2019; Evans et al., 2009), and spatial gradients in the vicinity of the polar vortex.
Excluding the uncertainty in quantifying corrections in the region of the South Pole, the low mean differences of the OMI-TOMS V8.5 data with the ground-based data for most regions support not having to adjust the OMI-TOMS data before serving as an anchor in the bias estimation for the limited period covered in this study.
Observation biases can be examined as a function of various factors. In this study, the bias correction applied in the assimilation experiments used bias estimates for discrete SZA/latitude bins as a function of time. Different bias-estimation methods based on observation colocations and observation differences with forecasts will be examined. Solar zenith angle dependence is specifically included considering the varying sensitivities between the different instruments as shown in Koukouli et al. (2012). Latitude and time dependencies were introduced to capture other data processing biases as well as instrumental changes over time. The alternative method of using the dependence on the ozone effective temperature instead of latitude and time (e.g., van der A et al., 2010) was also explored. Any bias impact due to differences in spatial resolutions of the instruments or model forecasts would be part of the residual biases and associated representativeness errors. The effect of different resolutions between instruments in bias estimation would in part be mitigated by use of local averages of differences in space. While the dependency on other factors such as cloud cover and viewing zenith angle can vary with the instrument and retrieval algorithm, they are not included here as predictors. Their impact would then be reflected in the estimated standard deviations derived for observations. The bias-correction target is to reduce residual biases as a function of SZA and latitude relative to OMI-TOMS to within 1 %.
Both July–August and January–February periods are considered for a comparison of bias estimates between seasons within a yearly cycle. The two sets of SBUV/2 total column ozone values obtained from the two-wavelength retrieval (SBUV/2-TC) and the sum of the retrieved partial column profiles (SBUV/2-NP) are included in the comparisons to OMI-TOMS. These have been added to extend the evaluation of the OMI-TOMS data conducted in Sect. 3.
This method estimates the bias as the mean differences of colocated
observations with OMI-TOMS. Separate bias estimations are conducted for each
distinct instrument platform. Here, the criteria for observations to be
considered colocated are for the points to be within 200 km and
Mean differences for each latitude/SZA bin are generated for individual 6 h intervals with, as a precaution, the removal of outliers beyond 2 standard deviations about the initial mean when there are at least 100 points per bin. Instead of monthly mean bias estimation, a moving window using the previous 2 weeks of data was applied to better capture variations in time. The 6 h mean differences over the 2-week moving window were weighted in time with a Gaussian weighting function with a half width at half maximum of 4.7 d. The 6 h mean differences were generated starting 2 weeks prior to the start of assimilations to provide data over the full window at the start of the assimilation. Another 2 standard deviation outlier removal was applied, this time according to the variability of the 6 h mean differences over the 2-week period. A minimum of 25 total contributing differences originating from at least four 6 h intervals is imposed for valid bias estimates for each bin.
Global diagnostics of differences in total column ozone between satellite instruments and OMI-TOMS for July–August 2014 and January–February 2015. The diagnostics consists of global mean differences and percentages of non-empty SZA/latitude bins with mean differences exceeding 2 % in magnitude.
The time mean differences with OMI-TOMS for July–August 2014 and
January–February 2015 are shown in Figs. 2 and 3, respectively. The figures
indicate global averaged biases in the range of
Mean total column ozone differences (%) between GOME-2A/B, OMPS-NM/NP, SBUV/2-TC/NP and colocated OMI-TOMS data for the period of July–August 2014. The SBUV/2-TC total column ozone values stem from the two-wavelength retrieval, while those for SBUV/2-NP are the sums of the retrieved 21-layer partial columns. The colours blue to purple denote negative differences and the colours yellow to red refer to positive differences.
Same as Fig. 2 for January–February 2015.
The discontinuity appearing at 70
The pattern about the Equator in Fig. 3 (January–February) appears inverted as compared to Fig. 2 (July–August) for SBUV/2 and OMPS-NP (which can also be seen in Table 3). This suggests the possibility of some seasonally dependent differences with OMI-TOMS for these instruments that may be related to changes or differences in retrieval conditions as a function of season or spectral channels as a function of solar zenith angle.
The results for the differences of the provisional OMPS-NM data with
OMI-TOMS in Tables 2 and 3, considering the differences of OMI-TOMS with the
ground-based data, are roughly in the same magnitude range as the
differences provided in Bai et al. (2015, 2016) for the more recent OMPS-NM
total column ozone products based on the SBUV V8 and V8.6 retrieval
algorithms. Bai et al. (2016) provide a distribution of OMPS-NM minus
OMI-TOMS values with a mean of 7.6 DU (
Mean differences of the total column ozone (%) between
satellite instruments and OMI-TOMS for July–August 2014 and January–February
2015 for the Northern Hemisphere and Southern Hemisphere, for solar zenith angles below
and above 70
The SBUV/2-NP data set could have been an alternative candidate, as the anchor considering the temporal stability of the data quality and its level of agreement with ground-based data indicated in earlier studies. The comparisons of the SBUV/2 products with OMI-TOMS in Figs. 2 and 3 and Tables 2 and 3 suggest that OMI-TOMS may be generally closer to the ground-based data for these two periods (Table 1). OMI-TOMS also appears to be in better agreement with SBUV/2 in the Antarctic region than with the ground-based data. The agreement between OMI-TOMS and SBUV/2-NP was usually found to be slightly better than the agreement between OMI-TOMS and SBUV-TC.
The variations in time of the bias corrections for a selected single bin are
shown in Fig. 4 for the July–August 2014 period. The time variations for
many bins are most often within
Time series of total column ozone bias corrections (DU)
for July and August 2014 for GOME-2A/B, OMPS-NM/NP, and SBUV/2-TC/NP as
derived from the colocation method described in Sect. 4.1. Dashed vertical
lines show individual 6 h mean differences with OMI-TOMS, while the
solid curves of the same colour show the 2-week moving average bias
corrections. The particular (latitude, solar zenith angle) bins plotted are
5
An alternative bias-estimation approach utilizes the differences of the
original retrieved observation data with short-term model forecasts, with
the same binning in latitude and solar zenith angle averaged over a 2-week
moving window. This would be applicable for near-real-time or reanalysis
data assimilations. These bias estimates can be constructed by considering
observation (
where the angular brackets denote averages and the subscript “ref” denotes
differences for observations of the anchor set (OMI-TOMS for our case).
Option (a) provides the potential benefit of accounting for spatial
differences between paired colocation points, while options (b) and (c) bring the potential advantage of bias correction in the absence of
sufficiently close colocation pairs. If previous observations of the
reference or other bias-corrected instruments were assimilated into the
system that produces the short-term forecasts
All three of the above options for total column ozone bias estimation were
performed and compared to the estimates from Sect. 4.1. Mean differences
with forecasts would normally be determined and applied for bias estimation
during the assimilation and forecasting cycle. For convenience, here we
instead used the differences with 6 h forecasts from a separate
assimilation and forecasting run (the “OMI” assimilation run summarized in
Table 5), which is described in more detail in Sect. 5. In practice, the
forecasts used for this approach, if applied in a near-real-time setting,
would come from runs that assimilate the bias-corrected observations using
the correction method considered in this section. In this section, all
observational data sets used for bias estimation are thinned to 1
Bias estimates using options (a) to (c) above for July–August 2014 are
shown in Fig. 5, which also shows the colocations-only method of Sect. 4.1
for comparison, and are summarized in Table 4. Differences between the
biases resulting from options (a) to (c) and colocation alone are within 1 % over the 2-month period except for a few bins, which are mostly at
high SZA, and for GOME-2A in the Southern Hemisphere also at high SZA. The standard errors of the mean differences for all cases are
mostly less than 0.1 %, but can be as high as 1 % for the option (a)
to (c) cases at very high SZA for bins with little data. The time evolution
of these bias estimates from the 2-week moving window for two different
bins is shown in Fig. 6. All bias estimates (both those that do and do not
use forecast differences) follow the same general evolution in time, varying
within 1 % of one another. Figure 6a, b show
examples of bins that have larger and smaller evolutions in time,
respectively, where for these bins the bias estimates change by
Mean differences in total column ozone (%) between
satellite instruments and OMI-TOMS for July–August 2014 using the options
(a), (b), and (c) from Sect. 4.2, for the Northern Hemisphere and Southern Hemisphere and
solar zenith angles below and above 70
Time mean total column ozone biases (%) between
GOME-2A and OMI-TOMS for July–August 2014 from colocation alone and for the
options
The bias estimates that use differences with forecasts are largely consistent with estimates that use colocation alone. The estimates that utilize differences with forecasts can provide additional benefits over using colocations alone if the forecasts well represent the spatial variation in total column ozone for options (a) and (b) or if the forecasts have been sufficiently de-biased for option (c).
Time series of total column ozone bias corrections (DU)
for two latitude/SZA bins covering July–August 2014 for GOME-2A using
different bias-correction methods. All cases that include colocation methods
use thinned observation sets. The “
An alternative parameterization for the bias estimation consists of using ozone effective temperature and solar zenith angle, as done in van der A et al. (2010). A motivation for a dependency on ozone effective temperature is to compensate for any unaccounted temperature sensitivity of the ozone absorption coefficients used in retrievals. In this case, bias estimation is implicitly dependent on time through temporal changes in the ozone effective temperature (and solar zenith angle). This captures at least the seasonal variations of biases associated with changes in temperature in addition to constant offsets. In this section, we briefly consider such a parameterization. For these estimates, we return to the method of Sect. 4.1, in which mean differences with OMI-TOMS are computed using only colocated observations (i.e., no use of forecasts).
Mean total column ozone differences (%) between GOME-2A, OMPS-NM and colocated OMI-TOMS data as a function of ozone effective temperature (Kelvin) and solar zenith angle (degrees) for the periods of July–August 2014 and July–August 2015. The colours blue to purple denote negative differences and the colours yellow to red refer to positive differences.
Ozone effective temperatures were calculated from ECCC's Global Environmental Multiscale (GEM) meteorological model, with short-term ozone forecasts driven by the linearized ozone model LINOZ. These forecasts were launched from ozone analyses that assimilated total column ozone data. Both the GEM model and ozone analyses are described in more detail in Sect. 5.
Bias estimates for GOME-2A and OMPS-NM for July–August 2014 and 2015 using
an effective temperature parameterization can be seen in Fig. 7. By
comparing the bias estimates for the same months from different years, we
see that these bias estimates can differ notably for different time periods.
With this parameterization, the bias estimate for GOME-2A differs by roughly
3 %–4 % between 2014 and 2015 for SZAs less than 70
List of assimilation experiments and their corresponding shorthand identifiers. In the second column, an asterisk (*) next to an instrument denotes that the bias-corrected observations (using the colocation method of Sect. 4.1) were assimilated.
An equivalent time evolution of a latitude/SZA bin can be made from the time-averaged effective temperature/SZA bias estimate shown in Fig. 7. First, the ozone effective temperature of each observation falling within a selected latitude/SZA bin is used to map that observation onto the ozone effective temperature/SZA bias estimate (Fig. 7). Then the bias estimate at each observed ozone effective temperature/SZA point is averaged for each 6 h time period. The resulting curves are shown in Fig. 6 for the selected latitude/SZA bins. The small temporal evolutions of these curves (typically well within 1 %) reflect the slight changes in the ozone effective temperature–latitude relationship in time. The greater the variation in time of the bias estimates based on the time-varying latitude/SZA parameterization, the larger the differences with the estimates based on the temperature and SZA parameterization alone (an example of which is illustrated by comparing panels (a) and (b) of Fig. 6). Overall, this supports the use of an ozone effective temperature parameterization as an alternative to latitude (and time) parameterization, with the stipulation that one accounts for any remaining notable temporal changes in some fashion when necessary.
In this section, we examine the effects of bias correction on global ozone assimilation and compare the 6 h forecasts launched from these analyses to ground-based observations and to OMI-TOMS. Corrections of observation biases were updated every 6 h using a 2-week moving window from colocations with OMI-TOMS. Assimilation experiments were conducted for July–August 2014, with a start date of 28 June 2014, 18:00 UTC, with and without bias correction. All bias-corrected observations applied in assimilation used the colocation approach without use of forecast differences (Sect. 4.1) to obtain bias estimates.
Global mean differences (%) between Brewer and Dobson total column ozone measurements and short-term forecasts for July–August 2014. For bias-corrected observations, the colocated observation bias-correction scheme (Sect. 4.1) was used. The Dobson measurements used were adjusted as a function of the ozone effective temperature (see Sect. 2.4). The uncertainties denote the standard error of the mean differences or standard deviations (SDs) of the differences. The data from the two Antarctic stations have been included here even though their mean differences with OMI are outliers relative to most mean differences (Tables S1 and S2 in the Supplement).
The forecasting model used was ECCC's GEM numerical weather prediction model
(Côté et al., 1998a, b; Charron et al., 2012; Zadra et al.,
2014a, b; Girard et al., 2014) coupled to the linearized ozone model LINOZ
(McLinden et al., 2000; de Grandpré et al., 2016). The LINOZ model uses
pre-computed coefficients generated as monthly mean climatologies to
calculate the ozone production and sink contributions throughout the
stratosphere and upper troposphere down to 400 hPa. A relaxation towards the
climatology of Fortuin and Kelder (1998) was imposed between the surface and
400 hPa to constrain deviations away from the climatology, with a relaxation
timescale of 2 days. The GEM model was executed with a 7.5 min time step
with a uniform
Assimilation was done using an incremental three-dimensional variational
(3D-Var) approach with first guess at appropriate time (FGAT; Fisher and
Andersson, 2001). This assimilation system uses components of the ECCC
ensemble-variational data assimilation system (Buehner et al., 2013,
2015) adapted by the authors and Ping Du (ECCC) for constituent assimilation
and was run without ensembles. The ozone background error covariances
applied with this system are described in the Supplement, which has a
minimum error standard deviation equivalent to
The initial ozone field was an analysis from an earlier assimilation. Successive 3 h to 9 h forecasts were generated from analyses provided for 00:00, 06:00, 12:00, and 18:00 UTC synoptic times. The analyses are a composite of already available ECCC operational meteorological analyses and the ozone analyses generated from this assimilation study. Assimilation runs were compared to runs without ozone assimilation that used the same meteorological analyses as employed by the ozone assimilation runs.
Both individual and combined observation data sets were assimilated. Assimilating column ozone data from two or more sources ensures that data are continually available in the event of occasional to permanent interruption of data availability from specific instruments. For near-real-time assimilation, the interruption of the availability of the anchor data set implies the need for contingency planning for transitions of bias-correction references. One might opt to assimilate data from some sensors and monitor the data from others through comparisons with the assimilation analyses. While not necessarily negating the need for bias correction, one could always select to assimilate data from sensors with retrieval products having the smallest initial biases as compared to other products. The effects of bias correction on assimilation when separately assimilating individual and multiple sensors will be examined.
The applied evaluation metrics consist of mean differences, standard
deviations, and anomaly correlation coefficients (ACCs), i.e.,
Zonal mean total column ozone statistics of mean differences (%), standard deviations (%), and anomaly correlation coefficients (ACC; unitless) as a function of latitude (degrees) for the comparison between OMI-TOMS measurements and short-term forecasts for July–August 2014. The legends in the top plots indicate the assimilation run (see Table 5 for description) and apply to all plots in the same column.
We first examine the global differences of Brewer and Dobson total column ozone measurements with 6 h forecasts following assimilation with and without bias correction. These differences are located mostly in the northern midlatitude and tropical regions (see Fig. 1). The mean and standard deviations of these differences are shown in Table 6. Assimilating GOME-2A observations alone without bias correction actually increases the absolute size of the global mean differences relative to the no-assimilation case to over 2 %. Runs assimilating GOME-2A and OMPS-NM alone, as well as GOME-2A/B and OMPS-NM, have the global mean biases from both Brewer and Dobson reduced from above to well below 1 % when bias correction is introduced. Bias correction reduced the global mean differences to less than 0.3 % in size for all cases. Introducing ozone assimilation with and without bias correction, as compared to the no-assimilation case, reduced the standard deviations in the range of 0.5 % to 1.5 %. Introducing bias corrections results in only a small reduction of the standard deviation of differences. The remaining contributors to the standard deviation of differences include the variation of inter-station ground-based instrument calibration errors, the effect of residual bias features of the assimilated data such as from cross-track variations, and/or representativeness errors associated with the model resolution, in addition to forecast errors and random errors from the ground-based instruments.
Zonal mean differences (%) and anomaly correlation coefficients (unitless) for total column ozone between OMI-TOMS observations and short-term forecasts as a function of time (date). Results are shown for the case without assimilation as well as with the assimilation of OMI, GOME-2A/B, and OMPS-NM (both with and without bias correction). The legend indicates the assimilation run (see Table 5 for description). Each value plotted was calculated using a 24 h time window.
Comparisons of OMI-TOMS measurements with forecasts for the various
experiments with and without bias correction and without any assimilation
are shown in Figs. 8 and 9 for the July–August 2014 period. The GOME-2A and
OMPS-NM data sets show the largest reductions in mean differences from bias
correction, as would be expected from Fig. 2. The upper-left mean difference panel of
Fig. 8 indicates that introducing bias correction to GOME-2A significantly
increases the benefit of the GOME-2A assimilation in the tropics and
northern extra-tropics. Also, the inclusion of bias correction in the
assimilation of the provisional OMPS-NM data reduced the mean differences
from as much as
Assimilation of total column observations improves the standard deviations
of differences between the 6 h forecasts and OMI-TOMS across all
latitudes, as seen in Fig. 8. Larger regional impacts in reducing standard
deviations are found in the tropics and Southern Hemisphere, while there is
relatively little impact from the GOME-2A/B assimilations in the southern
extra-tropics, where relatively few observations are available. The large
mean differences and standard deviations for GOME-2A/B assimilations below
60
The drift of the mean biases in time in the absence of assimilation, as seen in Fig. 9, is due to the tendency of the forecast to move toward the ozone model equilibrium state. For the GEM-LINOZ model, this results in a long spin-up period in which the ozone field moves from its initial state, based on an earlier assimilation, toward the ozone model equilibrium state. Beginning with an initial ozone field at the model equilibrium state would have increased the mean observation-minus-forecast differences and would likely not have improved the ACC of the control case, as implied by Fig. 9. Also from Fig. 9, we can see that the error of the total column ozone forecast increases by less than 5 % over the course of 15 days, reflecting the high predictability of ozone medium-range forecasts. This limited deterioration would not deter, for example, in properly forecasting the movement of low total column ozone regions during these periods.
For the ACC, forecasts from the assimilation of GOME-2B in the tropics appear better than from the assimilation of OMI-TOMS when compared to the OMI-TOMS observations. This occurs even though the OMI-TOMS data set is larger by factors of about 6 to 12 than the individual thinned data sets of the other sources. On the other hand, the GOME-2B data set, which has low biases in the tropics, provides a slightly extended longitudinal coverage over 6 h intervals, largely due to the missing strips of the OMI data set (Sect. 2.1). The ACC also demonstrates a more marked improvement in multiple sensor assimilation in the tropical region as compared to OMI-TOMS assimilation alone, which is not well seen in the mean differences. Multiple sensor assimilation with bias correction even further increases the ACC and thus the quality of the pattern and variation of the forecast fields.
Bias correction of total column ozone data from satellite instruments was
performed using three different approaches. Two of the methods parameterized
the bias estimation as a function of latitude, solar zenith angle, and time,
while the other method used the ozone effective temperature in place of
latitude and time. These approaches consisted of using observation
colocation between satellite-borne instruments and a reference, referred to
in this paper as the anchor. One approach also involved differences between
observations and short-term forecasts. The bias estimates from the methods
using the latitude/solar zenith angle parameterization were generally within
1 % of each other. The 2-month time-averaged bias estimates from the
ozone effective temperature parameterization were similar to those from the
other approaches. However, the lack of an explicit time dependence prevented
it from capturing changes in time which, depending on the observation set
and the location, could occasionally reach
The anchor used in the bias-estimation schemes was chosen as the OMI-TOMS data product, due to its wide coverage in both time and space and its good agreement with ground-based instruments. For the time periods examined in the study, OMI-TOMS was found to have global and regional mean differences with ground-based Brewer and Dobson spectrophotometers, and filter ozonometers within 1 %, except in the polar regions. Similar to larger mean differences were found between OMI-TOMS and SBUV/2 data, with OMI-TOMS generally being in better agreement with the ground-based data.
For the July–August 2014 and January–February 2015 periods, the observations
based on the TOMS algorithm for the GOME-2A instrument were found to have
the largest mean differences with OMI-TOMS, which could be as high as 8 %
in some regions of the parameter space for solar zenith angles below
70
It was demonstrated that the assimilation of total column ozone observations
that include bias corrections as derived in this study can improve the
agreement between short-term forecasts and ground-based measurements. Using
a three-dimensional variational assimilation system, the assimilation of
GOME-2A without bias correction gives global and time mean differences
between ground-based observations and short-term ozone forecasts of
The aforementioned results indicate that the reduction of biases to within the 1 % target was achieved for most regions and cases, an exception being for conditions with high solar zenith angles. For the assimilation of two or more satellite sensors, while it is possible that the cancellation of errors from different instruments could reduce forecast biases, harmonizing the different data sets through bias correction better ensures that target reductions in residual biases are achieved. The assimilation of bias-corrected observations from multiple sensors does not notably reduce the mean differences as compared to the assimilation of individual bias-corrected sensors. However, a notable improvement in multiple sensor assimilation was seen in the tropical region as compared to OMI-TOMS assimilation alone in the anomaly correlation coefficient metric. This improvement implies an increase in the quality of the pattern and variation of the forecast fields.
The bias-estimation and bias-correction software with related shell scripts can be provided with the understanding that users will need to adapt the code to their preferred input/output data file formats. The observations can be obtained from the different centres identified in the text and the acknowledgements section below. The assimilation and forecasting system relies on ECCC computing environment tools and file conventions. Also, the computing hardware used for these assimilation cycles has since been replaced at ECCC with accompanying changes to the cycling package. References of the system components are provided in this paper. The large sets of model analyses and forecasts, and the observation-minus-forecast data sets, are saved with an in-house binary file format. Subsets could potentially be made available by the authors upon request. In addition to also containing a few complementary figures, the Supplement provides tables of station-by-station mean differences of OMI-TOMS with ground-based data related to Table 1 and Fig. 1.
The supplement related to this article is available online at:
YJR directed and supervised this study, conducted some of the assimilation experiments and most of the final analyses, and is responsible for most of the manuscript text. MS significantly contributed to the manuscript text, wrote the colocation and bias-estimation and bias-correction software, conducted many of the assimilation experiments and produced the figures and the data for the tables other than for Sect. 3. Both MS and YJR contributed to the observation acquisition and preparation, and the setup of the assimilation experiments. YMC contributed to the comparison of OMI-TOMS with the ground-based data by generating the content of Tables S1–S3 and Table 1 accompanying Fig. 1. All the authors contributed to reviewing and revising the manuscript.
The authors declare that they have no conflict of interest.
The authors wish to thank Lawrence E. Flynn (NOAA) and Vitali Fioletov (ECCC) for information and advice regarding use of the different satellite data sets and the Brewer ground-based data, respectively, Jean de Grandpré and Irena Ivanova (ECCC) for the availability and assistance in use of the version of GEM with LINOZ, Ping Du and Mark Buehner (ECCC) for contributions in extending the variational assimilation code for use with constituents, Jose Garcia (ECCC) and Vaishali Kapoor (NOAA), among others associated with NOAA, for contributions in facilitating data acquisition and the various instrument teams having generated the different observation sets. Also, we are very appreciative of the contributions by the two referees toward improving the quality of this paper. We also gratefully acknowledge the following centres for access to the observations used in this paper: the National Environmental Satellite, Data, and Information Service of the National Oceanic and Atmospheric Administration (NESDIS/NOAA), the Earth Observing System Data and Information System of the National Aeronautics and Space Administration (EOSDIS/NASA), the World Ozone and Ultraviolet Radiation Data Center (WOUDC), and the Global Monitoring Division of the NOAA Earth System Research Laboratory.
This paper was edited by Ken Carslaw and reviewed by Maria-Elissavet Koukouli and one anonymous referee.