The description of the mean state of the global ocean has a long history. Since the late 19th century, the continuously growing net of hydrographic observations has resulted in the production of increasingly detailed maps of temperature, salinity, and other parameters. All of these maps were hand drawn, often having the imprint of strong subjective data interpretation. The introduction of computers permitted the accumulation and analysis of large amounts of data and led to the construction of the objectively analyzed maps. The first climatology of the world ocean by Levitus (1982) has become a standard for the oceanographic community. Since then the NOAA (National Oceanic and Atmospheric Administration) NCEI (National Centers for Environmental Information, the former NODC) Ocean Climate Laboratory has regularly produced improved versions of the global climatology (Levitus et al., 1994, 1998; Locarnini et al., 2006, 2010). The last update (Locarnini et al., 2013) was based on the hydrographic data over the entire time period from the beginning of the hydrographic observations to 2013.

All NCEI climatologies possess a high degree of consistency and use similar quality control procedures in addition to the objective mapping method (Barnes, 1964). The interpolation is performed on a set of standard depth levels, with the response function defining the smoothing inherent in the objective analysis method. However, as noted by Lozier et al. (1994), averaging (smoothing) of oceanographic properties on isobaric surfaces results in the production of water masses with temperature–salinity (T–S) characteristics different from those of the observed data due to the nonlinearity of the equation of state for seawater. In order to avoid this artifact, it has been suggested that the data be averaged on isopycnal surfaces. The objective analysis on density surfaces mimics the process of isopycnal mixing and does not produce artificial water masses. Gouretski and Koltermann (2004) prepared the isopycnally averaged Global Hydrographic Climatology (WGHC) based on the high-quality data obtained during the World Ocean Circulation Experiment (WOCE). To achieve reasonable data coverage between the WOCE section lines, selected pre-WOCE hydrographic data were added to the WOCE dataset, which served as a reference dataset for the calculation of the systematic inter-cruise property offsets (Gouretski and Jancke, 2001). The WGHC was used in a number of applications (e.g., the WOCE Hydrographic Atlas of the Atlantic Ocean – Koltermann et al., 2011; and the calculation of the absolute salinity – IOC, SCOR and IAPSO, 2010). One of the faults of the WGHC climatology is the absence of seasonality: only data mean parameter distributions are available at all levels. More recently a global monthly isopycnal upper-ocean climatology with an emphasis on preserving a surface mixed layer was created (Schmidtko et al., 2013).

The purpose of the current study is to produce an update of the WGHC climatology. We use the advantage of the significantly improved data basis due to the implementation of the Argo programme to achieve monthly temporal resolution and increase the nominal spatial resolution to 0.25×0.25∘ latitude/longitude. We refer to this new climatology as the WOCE-Argo Global Hydrographic Climatology (WAGHC). However, the addition of the Argo data is not the novel feature of the new climatology, and the title simply highlights the importance of the Argo data.

Constructing the climatology consists of several steps which are briefly outlined here. First, the climatology time frame, spatial and temporal resolution, and observation types are selected. The automated quality control procedure is applied to the original temperature and salinity profiles, which are subsequently interpolated on a predefined set of depth levels. The interpolated profiles are then averaged in 1/4∘ bins on a monthly basis, with the binned data providing the input for the spatial optimal interpolation. Highly smooth gridded fields of water density, temperature, and salinity obtained by distance weighted averaging are generated and used as the first guess fields required by the objective mapping method. At each grid point, the covariance matrices for optimal parameter estimation take account of both the distance between the data points and the difference in the bottom depth, so that along isobath observations become greater weights than across isobath observations. It is assumed that the fields to be analyzed and the noise in the data are uncorrelated. Two versions of the climatology are subsequently constructed in two steps: (1) the isobaric climatology with the optimal interpolation (mapping) performed on depth levels and (2) the isopycnal climatology where mapping is carried out on local density surfaces. The new climatology is compared with the last version of the NOAA WOA13 atlas, which has the same temporal and spatial resolution and also includes Argo profiles. Finally, the new climatology is used as a reference for the calculation of the ocean heat content anomaly time series. It should be noted that even with the Argo data included, the 0.25∘ resolution should be considered as a nominal resolution for the greater part of the world ocean. Nevertheless, the increased resolution permits a better description of the ocean, both in the regions of complicated topography like the Indonesian seas and in the data abundant areas (Boyer et al., 2005).

The WOD13 database (Boyer et al., 2013) (including the January 2017 update) served as the main data source for the WAGHC climatology. The profiles of the four instrumentation types were used: ocean station data (OSD), conductivity–temperature–depth (CTD), Argo profiling floats (PFL), and the autonomous pinniped bathythermograph data (APB). The latter were only used in the Southern Hemisphere where data coverage is generally poorer compared to the Northern Hemisphere. All four data types normally report both temperature and salinity. As both of these parameters are required for the spatial interpolation on isopycnal surfaces, the expendable (XBT) and mechanical (MBT) bathythermograph data were not used. We then added 50 848 profiles obtained from the Alfred Wegener Institute, Bremerhaven, and 5340 profiles received from different institutions in Canada to the existing 4 665 330 temperature/salinity profiles from the WOD13. These two additional datasets helped to significantly improve the data basis for northern polar regions. Table 1 gives details regarding the data types and data sources that contributed to the WAGHC. The dismissal of some instrumentation types along with the stringent quality control criteria explain why the total number of retained profiles is less than the approximate 5.4 million salinity profiles available in the WOD13 archive.

Figure 1 shows the yearly number of profiles of each data type retained after quality control. Before 1990 OSD profiles prevail, whilst CTDs are the main data type between 1990 and 2003. In later years, observations were mostly delivered by Argo floats, with the implementation of Argo floats being marked by a step-like increase in the number of available data.

In general, for most of the global ocean, we used data from 1985 (the beginning of the pre-WOCE hydrographic programme) to the present; thus, we have incorporated data from the last 32 years, which is close to the 30-year period for calculating climate norms that is recommended by the World Meteorological Organization. In some regions (mostly at high latitudes and in several marginal seas) there are still no or insufficient data, meaning that older data (since 1925) were utilized. However, the time frame for the data selection was narrower, especially within the upper 2 km Argo float depth range.

Quality control is important for the construction of the climatology. Due to the large volume of data, an automated quality control (AQC) procedure was developed. It consists of a suite of the following quality checks:

crude parameter range check,

The overall performance of the AQC procedure is illustrated by two-dimensional histograms (Fig. 2). For both temperature and salinity, the time–depth histograms indicate a decrease of data rejection rates with time. The highest rejection rate is observed around the Second World War and may be attributed to the conditions being generally unfavorable for conducting high quality observations. A significant improvement in the quality of temperature and especially salinity observations took place due to the introduction of CTDs and electronic salinometers at the beginning of 1970s; the next data quality improvement was due to the introduction of profiling floats in the mid-2000s. The AQC procedure identified 3.745 % and 5.255 % of observed levels for temperature and salinity, respectively, as outliers, whereas for 14.201 % and 18.167 % of temperature and salinity profiles, respectively, at least one observed level outlier was identified. Further details regarding the quality control procedure are given in the Appendix. The implementation of manual quality control was restricted to several areas in the Arctic Ocean that had very poor data coverage.

The quality-controlled observed temperature and salinity profiles were finally interpolated on 65 unevenly spaced “standard” levels between the surface and 6750 m. The depth interval between the levels increased linearly with depth, so that a better vertical resolution was achieved in the upper layers, where higher vertical property gradients typically occur. Only levels with both temperature and salinity that passed all quality checks were retained for vertical interpolation using the weighted-parabola method by Reiniger and Ross (1968). The interpolation was not performed where the spacing between two levels exceeded the depth-dependent threshold value h: h=20 m within the upper 50m layer, h=20+0.24⋅z between 50 and 2000 m (z is the mean distance between the two levels in meters), and h=500 m for z>2000 m. The limitation on the spacing between the observed levels is necessary to minimize the creation of artificial water masses due to the interpolation procedure.

After the mid-2000s, the majority of the temperature/salinity profiles come from Argo floats (Fig. 1). Since the floats only measure within the upper 2000 m layer, a step-like decrease in the data coverage occurs around the 2000 m level, which would create a strong bias towards the observations above 2000 m when spatial interpolation is performed on isopycnal surfaces. To avoid this artifact, the profile extension method was developed. The method is based on the observational fact that the local temperature and salinity values are fairly constant below the main thermocline. First, in the vicinity of each profile potentially suitable for extrapolation, up to 10 deep CTD and OSD profiles are selected and the average deep profile is calculated using distance weighted mean values at each standard level (the influence radius for the deep profile selection does not exceed 333 km). For the last observed level Zm (the merging depth) of the profiles subject to extrapolation the temperature and salinity offsets (DTo and DSo) relative to the mean profile are calculated. If the parameter offset for the merging depth does not exceed a predefined threshold value (0.05 ∘C for temperature and 0.01 for salinity), and if Zm>1898 m (the deepest WAGHC standard depth level within the Argo depth range), the profile is considered to be suitable for extrapolation. The average profile is then modified as follows: at each depth level Z>Zm the offset value DP = DPo[1-(Z-Zm)/(Zmax-Zm)] is subtracted from the average parameter value (temperature or salinity). The modified mean profile is used to extrapolate the original profile below the level Zm.

Figure 3 illustrates the extrapolation procedure for three arbitrarily selected full-depth CTD profiles. In order to estimate the average extrapolation error, we selected 52 672 full-depth CTD and OSD profiles deeper than 2200 m obtained after 1984 and interpolated on standard levels. The extrapolation procedure was then applied to all of these profiles truncated at levels equal to or deeper than 1898 m. The respective mean absolute difference between the extrapolated and the original full-depth profile decrease with depth and are in the range between 0.03 ∘C at 3000 m and 0.002 ∘C at 6000 m for temperature and between 0.003 psu and 0.001 psu for salinity (Fig. 4).

Finally, the extrapolation procedure was applied to 720 839 quality controlled OSD, CTD, and PFL profiles that were obtained after 1984 and had their last level at or deeper then 1898 m. For most of the ocean area, the percentage of extrapolated levels exceeds 20 % from the total number of levels, with Argo extrapolated profiles comprising the largest group. The spatial distribution of full-depth and extrapolated profiles is shown in Fig. 5 along with the percentage of the interpolated levels and extrapolated profile frequency distributions vs. the number of extrapolated levels and the year of observation.

The significantly increased data basis since the introduction of the Argo floats permits a better temporal and spatial resolution compared to the earlier WOCE Global Hydrographic Climatology (WGHC) (Gouretski and Koltermann, 2004). For each standard depth surface, the quality controlled vertically interpolated/extrapolated data were gridded by bin-averaging the data separately for each calender month in each 1/4∘ grid cell. The binning procedure serves two purposes. Firstly, the binning reduces the overall number of observations and, secondly, it reduces noise in the data. The thinning of the input profiles permits the application of the classical optimal interpolation method without the use of a fast multiscale optimal interpolation algorithm proposed by Menemelis et al. (1997).

With the aim of producing a climatology for the most data abundant recent years, the data selection for each spatial bin was performed iteratively, with the data being selected first for the time period from 1985 to 2016; thus, the WOCE hydrographic survey was embedded in the analysis. If no data were available for a particular grid node, the time period was extended to 1957–2016. For a small fraction of the grid nodes, all data since 1925 were used to produce the bin-averaged values. The percentage of spatial monthly bins populated with two or more observations decreased from 30 % in the upper several hundred meters to about 20 % at levels deeper than 2000 m.

The mean climatological year changes with depth, in the range from 2007 to 2011. For the upper 2 km layer where the Argo float data prevail, the climatological year is within the 3-year range between 2009 and 2011 (Fig. 6). Below the Argo depth range, the mean year is within the range from 2007 to 2009. Differences between the mean climatological year for different calendar months do not exceed 1 year. It is important to mention that the WOA13 climatology is created by averaging six decadal climatologies between 1955 and 2012, with 1984 being the median year. This method prevents biases toward more recent and data abundant years. The difference between the median years of both climatologies does contribute to the temperature and salinity differences of the two gridded products and is discussed later in the text.

The contemporary database does not provide enough data to obtain bin-averaged values for each 1/4∘ monthly bin, meaning that an interpolation procedure to fill gaps is needed. The bin-averaged temperature and salinity profiles serve as inputs for the spatial optimal interpolation method, which is used here in the form suggested by Gandin (1963). For the optimal interpolation on isobaric surfaces, the normalized spatial covariance Cxyh of temperature and salinity was represented through the negative squared exponential: Cxyh=exp⁡-rx/Lx2+ry/Ly2+(h/H)2, where rx and ry are zonal and meridional distances between the two points, Lx and Ly are the zonal and meridional decorrelation scales, h is the depth difference between the two points, and H is the decorrelation depth scale. Outside of the ±20∘ zonal belt around the Equator, Lx=Ly whereas, within this belt, Lx increases linearly from Ly at 20∘ N and 20oS to 4 Ly at the Equator, in order to account for the zonal elongation of correlation scale within the equatorial belt. The introduction of the (h/H)2 term in Eq. (1) represents the added distance penalty for crossing isobaths. The value of H was set to 2 km. Based on the evaluation of the test calculations the signal-to-noise variance ratio was chosen to be 0.5 as a trade-off between the smoothness and desirable feature resolution.

For the optimal interpolation on isopycnal surfaces the normalized spatial covariance Cxyz was also represented through the negative squared exponential: Cxyz=exp⁡-rx/Lx2+ry/Ly2+(ℜ/Z)2, where ℜ is the depth difference between the vertical positions of the same isopycnal surface, and Z=250 m is the decorrelation depth scale. The introduction of this term is aimed at reducing the depth bias that appears near the boundaries of the domain, where observations are biased to one side (above or below) of the analyzed grid level. The objective analysis is performed on deviations between the observations and first guess values. To provide the distance-weighted means for the first guess temperature, salinity, and density fields, Eq. (1) was used with Lx and Ly set to 555 km.

The spatial covariances of the analyzed temperature and salinity fields should be derived from available observations. However, the correlation length scale must at least be larger than the data spacing (Nuss and Tutley, 1994; Sokolov and Rintoul, 1999). We use the mean average distance to the four nearest bin-averaged neighbor profiles as the measure of data sparseness (Fig. 7). The distance between the observation points increases with depth from about 70–100 km within the upper 2000 m to about 200–300 km in the lower layer. These mean values were used as a guide for the choice of the decorrelation length scale for the optimum interpolation. After some experimenting, we decided on a decorrelation scale value of 333 km, which was used at all levels for the current version of the climatology (the decorrelation scale is the distance at which the autocorrelation function decreases to 1/e times the value of the zero lag). As noted by Sokolov and Rintoul (1999), the optimal interpolation produces a spatial average of the analyzed parameters, acting as a low-pass filter.

During the first step, the isobaric climatology is constructed with the binned data being spatially interpolated at preselected standard levels for each calendar month. We do not perform spatial interpolation for temperature and salinity separately. Instead, to avoid the undesirable effect of artificial water mass production, we first perform spatial interpolation of sea water density. Subsequently the optimal estimate of temperature on isobaric surfaces is obtained. The interpolation of salinity is not performed: the salinity is inferred from the isobarically interpolated density and temperature values. We note, that the approach described above differs from the method used for the construction of earlier versions of the World Ocean Atlas (Levitus et al., 1994, 1998), where isobaric interpolation (averaging) was performed separately for temperature and salinity. The calculated density profiles are checked for hydrostatic stability and the stabilization is performed if necessary by introducing small adjustments to temperature and salinity.

At each grid location, the stabilized isobarically averaged density profile defines the set of local density surfaces on which the interpolation of temperature is subsequently performed. As in the isobaric case, the salinity is inferred from density and temperature values.

The advantage of the isobaric method is that it can be applied in exactly the same way throughout the water column. However, the averaging (smoothing) of data along levels of constant depth does not correspond to the process of the water mass mixing in the real ocean which takes place along isopycnal, or more correctly along the neutral density surfaces. In contrast, the isopycnal averaging does not produce artificial water masses; however, in the regions where isopycnals outcrop at the surface or bottom, the isopycnally averaged parameters are biased toward the ocean interior (Schmidtko et al., 2013).

Differences between parameter distributions on selected levels between the isopycnally and isobarically averaged WAGHC climatologies are shown in Fig. 8a–e. As expected the largest differences occur in regions of strong spatial temperature and salinity gradients, like the Gulf Stream, the Kuroshio, Antarctic Circumpolar Current, and the equatorial and tropical Pacific Ocean. In such regions the absolute difference in temperature and salinity can exceed 1 ∘C and 0.2 psu, respectively. The differences diminish with increasing depth. Thus, at the level of 1050 m, only the North Atlantic and the belt of the Antarctic Circumpolar Current show systematic differences exceeding 0.05 ∘C and 0.01 psu for temperature and salinity, respectively. Integrated over the whole ocean area (Fig. 8g), the climatological isobaric temperature values are higher than the isopycnally averaged values. The same is true for the salinity except for the upper 100 m layer. Below 2000 m, typical absolute differences between the isobarically and isopycnally averaged temperature and salinity values remain below 0.25 ∘C and 0.005 psu, respectively.

We compared the WAGHC monthly temperature and salinity fields with respective fields from the NOAA WOA13 atlas (Boyer et al., 2013). This atlas represents the last version of the NOAA temperature and salinity climatologies. For the upper 1500 m, the 1/4∘ resolution monthly WOA13 climatology was used, below this level we used the annual WOA13 temperature and salinity fields.

This paper introduces the new WOCE-Argo Global Hydrographic Climatology (WAGHC) and describes it in detail. The climatology was conceived as an update of the former WOCE Global Hydrographic Climatology, WGHC (Gouretski and Koltermann, 2004). Unlike its predecessor, the new climatology has a finer 1/4∘ spatial resolution and resolves the annual cycle of temperature and salinity on a monthly basis. Two versions of the climatology are available, with the spatial interpolation being performed on isobaric and isopycnal surfaces, respectively.

The WAGHC climatology is further compared to the widely used 1/4∘ resolution isobarically averaged climatology WOA13, produced by the NOAA Ocean Climate Laboratory (Locarnini et al., 2013). We note generally good agreement between these two gridded products. The differences between the two climatologies are basically attributed to interpolation method (isopycnal vs. isobaric averaging) and to the considerably improved data basis (the WAGHC includes an additional 4 years of the Argo float and other data). The inclusion of additional data into the WAGHC climatology significantly improved the representation of the thermohaline structure in polar regions. However, a significant improvement was also achieved for several data abundant regions like the Baltic Sea, the Caspian sea, the Gulf of California, the Caribbean Sea, and the Weddell Sea. Further investigations are needed to identify the causes of differences between the two climatologies in these regions.

We also tested the dependence of the ocean heat content anomaly (OHCA) time series on the baseline climatology. Since the 1950s, both WAGHC and WOA13 used as baseline means produce almost identical OHCA time series. Even for the earlier data-poor decades, the largest differences do not exceed 10 % of the full OHCA range.