Introduction
Obtaining accurate estimates of recent changes in the ocean carbon cycle,
including how these changes will influence climate, requires the
availability of high-quality data. The fully quality-controlled and
internally consistent Global Ocean Data Analysis Project (GLODAPv1.1; Key et
al., 2004) has for the past decade been the only global interior ocean
carbon data product available. GLODAPv1.1 has been and continues to be of
immense value to the ocean science community, which is reflected in the
almost 500 scientific studies that have used and cited it so far. GLODAPv1.1
has been used most prominently for calculation of the global ocean inventory
for anthropogenic CO2 (e.g., Sabine et al., 2004) and for validation of
global biogeochemical or earth system models (e.g., Bopp et al., 2013).
The GLODAPv1.1 data product is dominated by quality-controlled and de-biased
data from the World Ocean Circulation Experiment (WOCE; e.g., Thompson et al.,
2001) program in the 1990s, and it contains additional, generally not
de-biased, data from the entire period 1972–1999 but very few data north of
60∘ N in the Atlantic and no data in the Arctic Ocean or
Mediterranean Sea. Many more seawater CO2 chemistry data have been
collected on research cruises after 1999, particularly within the framework
of the global repeat hydrography program CLIVAR/GO-SHIP (Feely et al., 2014;
Talley et al., 2016), so that significantly more interior ocean carbon data
exist today than were available in 2004.
In response to the shortcomings of GLODAPv1.1 in terms of data coverage and
quality control of historical data, and to include more recent data, an
updated and expanded version has been developed: GLODAPv2.2016 (Key et al.,
2015; Olsen et al., 2016; see Appendix A for a note on naming of the data
products). This new data product combines GLODAPv1.1 with data from the two
recent regional synthesis products: CARbon dioxide IN the Atlantic Ocean
(CARINA; Key et al., 2010), and PACIFic ocean Interior CArbon (PACIFICA;
Suzuki et al., 2013). In addition, data from 168 cruises not previously
included in any of these data products – both new and historical – have been
included. Notably, 116 cruises in GLODAPv2.2016 cover the Arctic
mediterranean seas, i.e., the Arctic Ocean and the Nordic Seas (> 65∘ N). GLODAPv2 data are available in three forms: (1) as
original, unadjusted data from each cruise in WOCE exchange format files;
(2) as a merged and internally consistent data product, where adjustments
have been applied to minimize measurement biases (hereafter referred to as
G16D); and (3) as a mapped climatology. This paper presents the methods used
for creating the mapped climatology and its main features, while the
assembly of the data and construction of the product, including the broad
features and output of the secondary quality control, are described by Olsen
et al. (2016).
As opposed to a gridded data product, which, for example, the Surface Ocean CO2
Atlas (Pfeil et al., 2013; Bakker et al., 2014) provides (Sabine et al.,
2013), we have created mapped climatologies. The difference is that gridded
data are observations projected onto a grid, using some form of binning and
averaging, and where no interpolation or any other form of calculation is
used to fill grid cells that do not contain measurements. In mapped data
these gaps have been filled, in the case of GLODAPv2.2016 using an objective
mathematical method. The method used to create the mapped climatologies from
the merged and de-biased data product is presented in Sect. 2.2. Some of
the resulting data fields and their associated error estimates are shown in
Sect. 3 to highlight important features in the data product. Finally,
some notes regarding the mapped climatologies of oxygen and macronutrients
are given in Sect. 4. Please note that, unless otherwise stated, this
paper describes the mapped climatological fields which are called
GLODAPv2.2016b at CDIAC (hereafter referred to as G16M). Two versions are
available at CDIAC (http://cdiac.ornl.gov/oceans/GLODAPv2/),
named GLODAPv2_Mapped_Climatology and
GLODAPv2.2016b_Mapped_Climatology,
respectively. Most of the information regarding method and input data is,
however, identical and the main differences between the versions are
provided in Appendix A.
Methods
Observational data inputs
Whereas G16D contains many more variables (Olsen et al., 2016), we only
mapped its primary biogeochemical variables: total dissolved inorganic
carbon (DIC or, here, TCO2), total alkalinity (TAlk), pH, the
saturation state of calcite and aragonite (ΩC and ΩA), nitrate (NO3-), phosphate (PO43-), silicate
(Si(OH)4), dissolved oxygen (O2), salinity, and temperature. The
latter two variables are to be used as a reference for the biogeochemical
variables and are not suggested to be of sufficient quality to be useful for
physical oceanographic applications. In addition, we provide maps of
anthropogenic carbon (Cant) and preindustrial TCO2
(TCO2preind), which are not part of G16D. The G16D data product
includes vertically interpolated data for the nutrients, oxygen, and salinity
where any of those were missing from a bottle data point, as well as calculated
seawater CO2 chemistry data whenever pairs of measured CO2
chemistry parameters were available. All such interpolated and calculated
data were included in the mapping. Moreover, we opted to additionally
include all data that did not receive secondary quality control. Such data
are generally collected on cruises that did not sample deep waters or did
not cross over with other cruises (both conditions precluded crossover
analysis). The majority of these data are likely to nonetheless be of high
quality and their inclusion into the mapping procedure was found to improve
the result – often specifically because of their remoteness from other
cruises
Monthly data density of TCO2 in the years 1972–2013. The
figure includes all data shallower than 150 m.
![]()
Open-ocean biogeochemical measurements tend to have a seasonal bias at high
latitudes because of prohibitive wintertime weather. Figure 1 shows the data
distribution in G16D in each month of the year. The seasonal bias is clear:
the North Atlantic is almost exclusively sampled in May–August (boreal
summer); the Southern Ocean is almost exclusively sampled in December–March
(austral summer); the eastern North Pacific has significant amounts of data
only in June, August, and September; the tropical Atlantic Ocean has a
January–April bias; and the only region with a reasonably full annual data
coverage is the western North Pacific. No attempt has been made to correct
for a possible seasonal bias in the mapped climatologies. Due to the limited
data coverage, such corrections would have to rely on relationships with
ancillary variables and different temporal gap-filling methods. This is an
endeavor in itself and something that may be attempted for future versions
of GLODAP but would also increase the errors and uncertainties in the
mapped climatologies. We therefore concede that the G16M data product as
presented here is for many regions more representative of summertime mean
conditions than annual mean conditions, and users should take this into
account when using the data product.
Maximum distance criteria used when vertically interpolating the
input data.
Range (dbar)
Maximum distance allowed (m)
0–200
100
201–750
200
751–1500
250
1501–12 000
500
The following pre-mapping data treatments were carried out on the data of
G16D:
The original cruise data were quality-controlled using a crossover and
inversion method, which entails identifying biases between stations within a
fixed distance and then evaluating all biases for all cruises to determine
the corrections necessary to make all cruises consistent. The details of the
quality control process are given in Olsen et al. (2016).
All bias-minimized data were vertically interpolated onto 33 surfaces: 0,
10, 20, 30, 50, 75, 100, 125, 150, 200, 250, 300, 400, 500, 600, 700, 800,
900, 1000, 1100, 1200, 1300, 1400, 1500, 1750, 2000, 2500, 3000, 3500, 4000,
4500, 5000, and 5500 m. These depths are the same as those used in GLODAPv1.1
and originally chosen by Levitus and Boyer (1994). The interpolation was
done station by station, using a cubic hermite spline function. This
interpolation method is generally robust, but it can give unreliable results in
a few unusual circumstances, especially near the surface. Consequently, if
this interpolation gave values more than 1 % different from those produced
using a simple linear vertical interpolation, the linear results were used.
We used the maximum distance criteria specified in Table 1 to avoid
interpolation over excessive vertical distances between data points. These
maximum distance criteria are the same as those used by Key et al. (2004)
for GLODAPv1.1.
The vertically interpolated data for each depth surface were then gridded by
bin-averaging all data in each 1∘ × 1∘ grid
cell. The mapped climatologies are thus based on gridded (or “bin-averaged”)
data, not individual measurements. We do this because the repeat hydrography
program yielded several transects in the ocean that may have differing
observations at nearly exactly the same location, which in frontal regions
may otherwise have deleterious effects on the results of mapping.
G16D includes data for the period 1972–2013, during which the atmospheric
levels of CO2 increased strongly. Consequently, within this time frame,
oceanic TCO2 and related properties (e.g., pH and the saturation states of
CaCO3) have seen strong increases, particularly at shallower levels
(e.g., Orr et al., 2001; Lauvset et al., 2015; Sabine and Tanhua, 2010). To
correct for this, all TCO2 data in G16D were normalized to the year 2002
prior to mapping. The details of this procedure are given in Appendix B. The
employed methodology likely has substantial uncertainty, partially due to
(local) invalidity of assumptions and its practical implementation.
Nevertheless, for the purpose of creating a global mapped climatology of
TCO2 for G16M this uncertainty was deemed necessary in order to reduce
the risk of converting time trends into spatial variations. Being able to
use all 42 years of observations in the mapping significantly reduces the
mapping error.
ΩC and ΩA were calculated from the
TCO22002 and TAlk pair at in situ temperature and pressure, while pH
was calculated from the TCO22002 and TAlk pair at both in situ
temperature and pressure and at constant temperature (25 ∘C) and
pressure (0 dbar). All calculations were performed using the MATLAB version
(van Heuven et al., 2009) of CO2SYS (Lewis and Wallace, 1998). We used
pressure, temperature, salinity, phosphate, and silicate from G16D; the
dissociation constants of Lueker et al. (2000) for carbonate, and those of
Dickson (1990) for sulfate; and the total borate–salinity relationship of
Uppström (1974).
Derived variables
In G16M we provide mapped climatologies of both anthropogenic carbon
(Cant) and preindustrial carbon (TCO2preind). These are
based on using a classical application of the transit time distribution
(TTD) method (e.g., Hall et al., 2002; Waugh et al., 2006) on all available
CFC-12 data in G16D. Users of these fields should be aware that they were
obtained by a relatively crude application of the TTD methodology, and that
both rely on first-order approximations and assumptions (Appendix B).
However, the TTD methodology followed is not different from, for example, that used by
Waugh et al. (2006) based on GLODAPv1.1 and may, due to the higher data
coverage and quality in G16D, be considered an improvement of that earlier
work. Subtracting Cant from TCO2 provides TCO2preind,
which is of considerable relevance to ocean biogeochemical modeling studies
and thus also mapped and included in the G16M data product.
Mapping method
The Data-Interpolating Variational Analysis (DIVA) mapping method (Beckers
et al., 2014; Troupin et al., 2012) was used to create the mapped
climatologies. DIVA is the implementation of the variational inverse method
(VIM) of mapping discrete, spatially varying data. A major difference
between this and the optimal interpolation (OI) method used in GLODAPv1.1 is
how topography is handled. DIVA takes the presence of the seabed and land
into account during the mapping and gives better results in coastal areas
and around islands. In addition, the entire global ocean can be mapped at
once, for example, DIVA does not propagate information across narrow land barriers such
as the Panama isthmus. Consequently, there is no need to split the data into
ocean regions which then must be stitched together to form a global map
(like in, for example, GLODAPv1.1; Key et al., 2004), or to apply basin identifiers
which tell the software where to exclude nearby data because they are on the
other side of a topographical feature (like in, for example, World Ocean Atlas, Locarnini
et al., 2013). In DIVA the topography is used to generate a finite-element
mesh inside the topography contour on a given depth surface and the cost
function DIVA uses to create an analysis field (see below) is then solved
for each triangle of the mesh. The mesh size is directly dependent on the
correlation length scale (see below) in the open ocean, but near topography
the mesh is also dependent on the shape of the topography contour. The
analysis is then output on the specified regular grid. Each grid center point has to be
inside a mesh triangle to be deemed “ocean”. Whether a cell at a given depth
surface is masked as land or not is therefore a result of the choice of
topography and the detail of the finite-element mesh.
Apart from the data, the most important DIVA input parameters are the
spatial correlation length scale (CL) and the data signal-to-noise ratio
(SNR). The CL defines the characteristic distance over which a data point
influences its neighbors. For G16M.2016 this was defined a priori as
7∘ for all parameters. Since the oceans generally tend to
mix more easily zonally than meridionally, a scaling factor was used to make the
east–west correlation length twice the defined input (see the DIVA user
manual available at http://modb.oce.ulg.ac.be/mediawiki/index.php/Diva_documents for details). These settings give zonal and meridional
correlation length scales that closely match those used for GLODAPv1.1, but
this implies having to use Cartesian coordinates. Note that setting the CL is
partly a subjective effort, aiming to strike the optimal balance between
large values that tend to smooth the data fields and reduce mapping errors
and small values that lead to more correct rendering of fronts and other
features. We also want to stay within the physical constraints set by ocean
dynamics and natural spatial variability. It is possible to locally optimize
CL in DIVA, but this works well only when the data density is reasonably
high. The sparse global data distribution in G16D gives optimized CL in the
order of 25∘. Doing a cruise-by-cruise analysis following
Jones et al. (2012) to get spatially varying CL is possible, but this would leave
large gaps where we would have to guess CL. For G16M it was therefore
decided to use a globally uniform a priori value since this is the most
transparent and reproducible.
The SNR defines how representative the observations are for the
climatological state. For spatially varying data sets like G16D, it is the
assumed ratio of climatological spatial variability (“signal”) to the
short-term variability (“noise”) in the data. For G16M the SNR was defined
a priori to be 10 (i.e., the noise is 10 % of the signal), following Key et al. (2004). To understand the importance of SNR, and the reason for a subjective
a priori choice, a brief discussion of the differences between interpolation
and approximation/analysis is necessary. When interpolating between points in a data set,
gaps between data points are filled but the existing points are not
replaced. When approximating, a function (e.g., a regression line) is applied that describes
the original data points to some degree. The resulting approximated data set
has new values at every point and is smoother than the interpolated data
set. None of the approximated data points exactly match the original ones.
That assumes more uncertainty – or non-climatological variability – in the
data. In the case of very high SNR, the observed values are retained in the
mapped climatology and DIVA interpolates between them, while smaller values
allow for larger deviations from these and an increasingly smooth
climatology.
Working with real-world observations, we know that the observations are
indeed affected by shorter-term variations and in addition have analytical
uncertainties associated with them. They do not represent the “true”
climatological value. For this reason the SNR should always be kept quite
small when making mapped climatologies, but this needs to be balanced by the
need to keep the error estimates reasonable. The lower the SNR the further
the approximation is allowed to deviate from the original data and the
higher the error associated with the approximation becomes. The SNR can be
calculated from observations using generalized cross validation, but for
G16D such calculations give very high SNR (in the order of 100). This is
maybe not completely unreasonable, since G16D has been carefully quality-controlled and we have high confidence that the measurement uncertainties
are small. However, both increasing the SNR and increasing the CL will
decrease the error estimates, because this assumes small representativity
errors (i.e., that what is observed is the true climatology) and a large circle
of influence. We know that there is a seasonal bias in the observations
which therefore do not represent the true climatological values, and if the
assumptions above are wrong the errors will be significantly underestimated.
Therefore, the mapping errors are likely to be underestimated if we use the
SNR calculated from general cross validation.
A DIVA analysis is created by minimizing a cost function which is defined by
the difference between observations and analysis, the smoothness of the
analysis, and the physical laws of the ocean (Troupin et al., 2012). The
result is thus the analysis with the smallest global mean error, but
determining the spatial distribution of errors is important as well. In DIVA
this is non-trivial as, in contrast to OI, the real covariance function,
which is necessary to obtain spatial error fields, is not formulated
explicitly but is instead the result of a numerical determination (Troupin
et al., 2012). Determining the real covariance to get error estimates is the
most exact method, but it is computationally expensive. There are several error
estimation methods implemented in DIVA, from the very simple to the very
exact (Troupin et al., 2012; Beckers et al., 2014). Due to computational
cost, for G16M the error fields are based on the “clever poor man's error
calculation”. A detailed description of this method is given in
Beckers et al. (2014). Based on detailed studies of the Mediterranean Sea, Beckers et al. (2014) estimate that this method underestimates the real error by
∼ 25 %. The underestimation is not uniform, though, and
larger in areas with high data density, while in areas where observational
data are more than one CL distant, the clever poor man's error represents the
true error well (Beckers et al., 2014). The error fields are thus
appropriate to determine where the error is too large for the user's planned
use of the mapped climatologies.
Output and post-processing
Each of the G16M climatologies are global analyses for the range
-180 to 180∘ E with a 1∘ × 1∘ spatial resolution and 33 vertical layers, performed on
a Cartesian coordinate system from -340 to
20∘ E. To ensure that the analyses converges on the analysis
boundary at 20∘ E the input data (i.e., the observations) were
duplicated for 10∘ on either side of 20∘ E
before mapping. This duplication is a combination
of (i) mirroring the
observations for 2∘ on either side of 20∘ E, (ii) copying the observations in the area 10–20∘ E to 15–25∘ E and to
20–30∘ E, and (iii) copying the observations
in the area 20–30∘ E to
15–25∘ E and to 10–20∘ E. This is necessary mostly due to the paucity of data
along the boundary. The topography file has a longitudinal distance of
400∘ to ensure that the analysis neither begins nor ends at
the end of the finite-element triangular mesh. In addition, a weighted
average over 10∘ on either side of the boundary was used to
smooth the analysis in post-processing. This removes most discontinuities
along these boundaries. Where discontinuities are still visible, they are
not significant (i.e., not greater than the mapping uncertainty). The
topography is a priori known by DIVA, so no land masks need be applied in
post-processing. However, other masks have been applied:
A mask removing the results in all grid cells where the relative mapping
error exceeds 0.75 was applied. This mask effectively masks closed regions
with no observational data, in addition to masking regions where the DIVA
analysis is considered too uncertain to be useful.
Masks covering the Red Sea, Gulf of Mexico, Caribbean Sea, and parts of the
Canadian Archipelago were applied. In these regions the mapping error is not
overly large relative to the standard deviation in the global data, but a
general lack of data and our knowledge of oceanographic features lead to the
conclusion that the mapped climatologies do not appropriately reflect the
real world in these regions.
In the Arctic Ocean, cells where the DIVA analysis results in values that
differ from the observational mean in the different basins by more than
2.5 % were masked. Further details regarding the Arctic Ocean are given in
Sect. 2.5.
Mapping the Arctic Ocean
Mapping data in the Arctic Ocean is challenging for several reasons: (i) our
choice to increase the zonal correlation length (relative to the meridional
CL) required us to work with an equidistant world which is very
inappropriate for the high-latitudes; (ii) the Arctic Ocean contains the
northern boundary of DIVA's (Cartesian) finite-element mesh, which causes
peculiar artifacts; and (iii) there is a limited amount of observations. The
combination of these three factors can lead to unrealistic results for the
Arctic. We found that the DIVA analysis yields useful results in the top
1000 m of the Arctic Ocean. However, we have applied an additional mask
based on deviation from the observational averages to remove spurious
boundary effects (Sect. 2.4). Deeper than 1000 m, however, DIVA is not
able to provide useful climatological fields in the Arctic Ocean. It has
therefore been decided to replace the DIVA analysis with the basin averages
for each depth surface below 1000 m (i.e., surfaces 20–33). For these surfaces
the mapping error has been replaced with the standard deviation of this
average. The boundaries we have used for the four basins in the Arctic Ocean
are shown in Fig. 2.
Map of the Arctic Ocean with the four basins' definitions used to
calculate averages shown in different colors.
![]()
Results
Data fields
The mapped climatologies are available from CDIAC (http://cdiac.ornl.gov/oceans/GLODAPv2/) as one folder named
GLODAPv2.2016b_Mapped_Climatology. This folder
contains netCDF files for each parameter. Each of these netCDF files contain
the global 1∘ × 1∘ climatology of a
parameter, the associated error field, and the gridded input data for the
parameter in question (Table 2).
List of information available in the netCDF data files.
Variable name
Description
lon
Longitude in degrees east, range 20.5–19.5∘ E. This is the longitude in the center of the cell.
lat
Latitude in degrees north, range 89.5∘ S–89.5∘ N. This is the latitude in the center of the cell.
TCO2, TAlk, pH, ΩC, ΩA, NO3, PO4, silicate, oxygen, salinity, temperature, Cant, TCO2preind
Mapped climatology with all masks applied.
_error
Mapping error associated with the mapped climatology.
_relerr
Mapping error scaled to the standard deviation of the global input data at that depth level. Relerr = 1 means the error at that point equals 1 standard deviation of the global mean at that depth level.
Input_mean
Average of the observations located within each grid cell.
Input_std
Standard deviation of the observations located within each grid cell.
Input_N
Number of observational data points located within each grid cell.
Figures 3–5 show the mapped climatologies for TCO2, TAlk, and nitrate,
respectively, at two different depth surfaces. These all show the spatial
patterns expected from biological dynamics, evaporative processes, and
large-scale circulation: higher values of each in the deep Pacific Ocean
than in the deep Atlantic Ocean illustrate the greater age of water in the
deep Pacific Ocean; a higher concentration of remineralized carbon and
nutrients at depth than at the surface; and higher surface ocean TAlk in the
subtropics, where salinity is highest. The rather low nitrate at the surface
highlights the summer bias in the climatologies, except in the Southern
Ocean, where the biological production does not fully utilize available
nutrients and where there is both upwelling and strong vertical mixing.
Figures 6–8 show, for the same parameters and depth surfaces, the difference
between the gridded input and the mapped climatologies, which is relatively
large and variable near the surface and generally within the data
uncertainties in the deep ocean. Figures 9–11 show the error fields
associated with the climatologies shown in Figs. 3–5.
Mapped climatology of TCO2 at 10 m (a) and 3000 m (b).
![]()
Mapped climatology of TAlk at 10 m (a) and 3000 m (b).
![]()
Mapped climatology of nitrate at 10 m (a) and 3000 m (b).
![]()
Difference between the gridded TCO2 input data and the mapped
climatologies at 10 m (a) and 3000 m (b).
![]()
Difference between the gridded TAlk input data and the mapped
climatologies at 10 m (a) and 3000 m (b).
![]()
Difference between the gridded nitrate input data and the mapped
climatologies at 10 m (a) and 3000 m (b).
![]()
Mapping error for TCO2 at 10 m (a) and 3000 m (b). Notice how
the error increases with distance from transects, creating a spatial pattern
of square-like features in the Pacific.
![]()
Error fields
While the mapping error reflects the data distribution and the choice of
input variables (i.e., CL and SNR), it represents only the errors due to the
mathematical mapping of the input data, and does not take into account all
the uncertainty in the input data (although some of this uncertainty is
accounted for in the choice of SNR). The error fields moreover do not
include calculation errors for those variables that are calculated from
other measurements (e.g., pH and CaCO3 saturation states). For details
regarding the accuracy and precision of the G16D data product the reader is
referred to Olsen et al. (2016); briefly, the uncertainties in the input
data overall are smaller than the mapping errors. Overall, the spatial error
distribution is as expected: relatively small where there are observations
and larger elsewhere. All grid cells where the relative mapping error (i.e., the ratio
of the absolute mapping error at a grid point over the standard deviation of
all the data on that depth level) exceeds 0.75 have been masked. This step
nonetheless leaves several regions with high absolute mapping errors. The relative
error fields are provided in the netCDF files, making it possible for the
user to create alternative masks if desired. The range from minimum to
maximum error for the different parameters (Table 3) reflects the
variability in data values and data density on different surfaces. The
spatial variability in the mapping errors in large part depends on the
layout of the observational network, and further study of this variability
would be of great use in optimizing existing and future observational
networks.
Table of the range in mapping error averaged spatially over all
depth surfaces for the individual parameters.
Name
Unit
Range in mapping
error (min–max)
TCO2
µmol kg-1
4.5–12.5
TAlk
µmol kg-1
2.9–13.2
pH @ 25 ∘C and 0 dbar
0.006-0.032
pH @ in situ temperature and pressure
0.007–0.020
ΩC @ in situ temperature and pressure
0.01–0.26
ΩA @ in situ temperature and pressure
0.007–0.177
Nitrate
µmol kg-1
0.4–1.2
Phosphate
µmol kg-1
0.03–0.08
Silicate
µmol kg-1
1.2–4.3
Oxygen
µmol kg-1
2.9–8.0
Salinity
0.007–0.168
Temperature
∘C
0.04–1.15
Cant
µmol kg-1
0.30–1.7
TCO2preind
µmol kg-1
4.6–13.2
The mapping errors for TCO2 and TAlk in G16M have been compared with
those of GLODAPv1.1 (Figs. 12–13). Differences are expected from the
different methods used to calculate the errors and do not necessarily
indicate one product to be of superior quality. The most obvious result is
the large spatial variability in the differences, particularly at depth,
which seem to correlate with the data distribution. Very generally, there
are large differences (∼ 10 µmol kg-1) in error
estimates between G16M and GLODAPv1.1 for both TCO2 (Fig. 12) and
TAlk (Fig. 13) in the top 200 m of the Southern Ocean (exemplified by the
10 m surface in Figs. 12a and 13a). In this case the error estimate of
both TCO2 and TAlk in G16M is frequently 15 µmol kg-1 higher
than in GLODAPv1.1. For the rest of the world, however, the G16M errors are
smaller than the GLODAPv1.1 errors for both TCO2 and TAlk. Below 1000 m
(exemplified by the 3000 m surface in Figs. 12b and 13b) the mapping
errors are globally larger by 0–10 µmol kg-1 in G16M than in
GLODAPv1.1. Generally, for the deep ocean the error difference is greatest in
the Pacific and Southern oceans, while the Atlantic Ocean is relatively
comparable (Figs. 12b and 13b). A more comprehensive study of the
differences in mapping error between GLODAPv1.1 and G16M and the mechanisms
behind these would be worthwhile, and may improve future climatologies of
the marine CO2 chemistry, but is beyond the scope of this paper. The
exact reasons for the differences seen in Figs. 12–13 are currently not
evident, but several things are likely to contribute: (i) differences in the
methods used, and particularly that the method used by G16M is known to
underestimate the real errors by as much as 25 %, and (ii) there are
differences in data density and data distribution, with the improved
distribution in G16D revealing more small-scale natural variability and thus
larger, more realistic, mapping errors.
Mapping error for TAlk at 10 m (a) and 3000 m (b). Notice how the
error increases with distance from transects, creating a spatial pattern of
square-like features in the Pacific.
![]()
Mapping error for nitrate at 10 m (a) and 3000 m (b). Notice how
the error increases with distance from transects, creating a spatial pattern
of square-like features in the Pacific.
![]()
Mapping error for TCO2 in GLODAPv2.2016b minus mapping error
for TCO2 in GLODAPv1.1 at 10 m (a) and 3000 m (b). Note that the
differences are mainly attributable to differences in method and not real
reductions/increases in mapping error.
![]()
Mapping error for TAlk in GLODAPv2.2016b minus mapping error for
TAlk in GLODAPv1.1 at 10 m (a) and 3000 m (b). Note that the differences are
mainly attributable to differences in method and not real
reductions/increases in mapping error.
![]()
GLODAPv2.2016b nitrate gridded input data minus WOA09 annual
mapped nitrate climatology at 10 m (a) and 3000 m (b).
![]()
For the macronutrients (nitrate, phosphate, silicate) G16M can be compared
to the World Ocean Atlas 2009 (WOA09) nutrient climatologies (Garcia et al.,
2010). Before doing so several things need to be considered: (i) methods
used for mapping WOA09 (Garcia et al., 2010) are very different from those
used in G16M and (ii) WOA09 does not provide mapped error estimates for their
climatologies (only the difference in the 1∘ × 1∘ cells with observations), precluding a direct comparison
of errors. Instead, we compare G16M with WOA09 by plotting the difference
between (i) the annual NO3 climatology of WOA09 and (ii) the
bin-averaged data of G16M. For the purpose of this comparison we roughly
converted the WOA09 data from µmol L-1 to µmol kg-1 by
dividing by 1.024. Differences will be due to a combination of the
differences in input data and the differences in mapping methods and the
comparison reveals that at 10 m depth the G16M bin-averaged observations are
generally lower than the WOA09 climatology in high latitudes (Fig. 14a).
This is most likely a manifestation of the summertime bias in GLODAPv2
observations. In the tropics and subtropics the differences are within the
data and mapping uncertainties. In the deep ocean (Fig. 14b) the
differences between the G16M bin-averaged observations and WOA09 are
occasionally quite pronounced (for instance in the southeastern Atlantic
Ocean, and south and southeast of Australia, where G16D performed
significant adjustments to some cruises). However, in general, the match
between GLODAPv2 and WOA09 at depth is encouraging. This suggests that, below
the seasonally influenced surfaces, the differences between G16M and WOA09
stem mainly from differences in mapping method and input data but that the
climatologies otherwise are comparable. The biggest difference between G16M
and WOA09 is that the latter has considerably more input data and is thus
able to provide monthly climatologies. We have compared the G16M nitrate
climatology to the WOA09 nitrate climatology since the more recent WOA13 has
a very different vertical resolution of 137 surfaces (Locarnini et al.,
2013), compared to 33 in G16M.
We note that there is a significant difference in total ocean volume between
G16M and WOA09, with WOA09 having a 10 % larger volume than G16M. This
difference likely comes from differences in how topography is handled in the
two methods used and, related to that, the application of land masks, but
the exact details of how and where these differences occur are beyond the
scope of this paper.