Sea surface temperature (SST) is the observational variable that can be presented for the longest period over the global ocean. Since the 18th century, SST in the global ocean has been observed mainly by ships and buoys and has been archived by several national agencies. Because the sea surface is the boundary between the atmosphere and the ocean, SST reflects the thermal conditions of both the lower troposphere and the upper ocean. Therefore, SST is widely used as a fundamental variable for investigation of climate change. SST is also important as a lower boundary condition for atmospheric general circulation models (AGCMs).
Several institutions and research groups have made and released gridded SST datasets from observational reports and/or satellite data. Significant differences are known to exist among the datasets resulting from the source data, gridding processes and other factors (Trenberth et al., 1992; Folland et al., 1993; Hurrell and Trenberth, 1999). Because the number of observations tends to decrease backward in time, the uncertainty and differences among the datasets are more serious in the past than in recent years (Trenberth et al., 1992). However, intercomparisons among SST datasets have mainly been done since the beginning of satellite measurements in the 1980s; intercomparisons in the past were done only in terms of global or regional means and secular trends, and rarely in terms of various climate signals.
The purpose of the present study is to compare SST datasets not only in the present but also in the past. In addition to the differences among climatologies, their standard deviations and correlation of anomalies, the differences in the representations of climate phenomena such as the El Niño/Southern Oscillation (ENSO) and the Pacific Decadal Oscillation (PDO), are also investigated. Datasets compared in the present study and their gridding procedures are described in Section 2. The results of the statistical comparisons are presented in Section 3, and the findings for climate phenomena are presented in Section 4. Section 5 contains concluding remarks.
2. Datasets and differences among gridding procedures
In the present study, seven historical gridded SST datasets are examined that have been widely used for climate research: the Hadley Center sea ice and SST dataset (HadISST), version 1 (Rayner et al., 2003); the centennial in situ observation-based estimate of SSTs (COBE; Ishii et al., 2005); the extended reconstruction of global SST (ERSST), version 3 (Smith and Reynolds, 2004; Smith et al., 2008); the optimal smoothing analysis by the Lamont-Doherty Earth Observatory (LDEO; Kaplan et al., 1998, 2003); the global monthly summaries of the International Comprehensive Ocean-Atmosphere Data Set (ICOADS) Release 2.1 (Worley et al., 2005); the second Hadley Center SST (HadSST) (Rayner et al., 2006); and SSTs prepared by the authors at Tohoku University (TOHOKU; Yasunaka and Hanawa, 2002). Characteristics of each of the seven SST datasets are summarised in Table I. Each contains monthly data but the datasets differ in their spatial resolutions, interpolation methods for missing data, instrumental bias correction methods, treatments of satellite-derived data and so on, which are described below.
Table I. Historical SST datasets investigated in the present study
Reduced-space optimal interpolation (Kaplan et al., 1998).
Reduced-space optimal smoother (Kaplan et al., 1998).
Although SSTs are observed and collected over the global ocean, the coverage is far from worldwide, especially before 1880 and during the two World Wars. To interpolate missing grid values, HadISST, ERSST and LDEO are reconstructed using empirical orthogonal functions (EOFs) or empirical orthogonal teleconnection modes, and COBE is interpolated with an optimal interpolation. ICOADS, HadSST and TOHOKU, being uninterpolated, have many missing grid values. Reconstruction is not applied in LDEO in the Southern Ocean and north of 70°N. In the periods and regions in which in situ observations are limited, the accuracy is low even if some grid value is given.
2.2. Elimination of extreme values and variance deflation
Extreme grid values are eliminated not only using quality control of individual observations but also by smoothing and/or interpolating processes, yet these processes also have the unexpected effect of variance deflation. In HadISST, COBE, ERSST and LDEO, extreme values are eliminated with interpolating processes. In HadISST, the spatial resolutions of reference EOF modes change before and after 1949. In the optimal interpolation used in COBE, interpolated anomalies tend to be underestimated in regions where sampling is inadequate. The number of anomaly modes used for the ERSST reconstruction is limited before the 1870s. No smoothing process is applied to ICOADS and HadSST, whereas in TOHOKU, extreme values are eliminated with a median filter and smoothing by a binominal filter.
2.3. Version of historical databases
ICOADS observations provide a representative historical database in which available global surface marine reports from the late 18th century have been assembled and quality controlled (Woodruff et al., 1998; Worley et al., 2005). After the first release in 1985, ICOADS observations were updated several times. COBE, ERSST and HadSST used the ICOADS observations, release 2.0. ICOADS and TOHOKU used the ICOADS observations, release 2.1. Ship reports included in HadISST and LDEO are taken from the Met Office Marine Data Bank (MDB), which are now included in ICOADS. For quasi-real-time updates, the Global Telecommunications System (GTS) reports have also been used in recent years in HadISST, COBE, ERSST, LDEO and HadSST.
2.4. Bias corrections for historical SSTs
The method of measuring SSTs changed over time, from uninsulated (or partly insulated) buckets to insulated buckets, engine intakes and hull-mounted sensors. Generally, the difference between an engine intake and a bucket temperature amounts to 0.3 °C (Folland et al., 1984; Jones et al., 1986). Folland and Parker (1995) estimated time- and area-varying corrections using model for evaporative cooling effect in bucket measurement, and their bias corrections have been widely used. Among the SST datasets compared in this study, HadISST, COBE, LDEO and TOHOKU adopt this correction. Rayner et al. (2006) adjusted the Folland and Parker (1995) correction method using ICOADS observations and applied it to HadSST. Smith and Reynolds (2002) proposed another correction method based on nighttime marine air temperatures and SST observations, and a modified version of this by Smith and Reynolds (2004) was applied to ERSST. These three corrections are generally similar. The main difference is the termination time of the correction; corrections by Folland and Parker (1995) abruptly end after 1941, while corrections by Rayner et al. (2006) and Smith and Reynolds (2004) are gradually reduced to zero over the period 1939–1941 to account for a different mix of observations. Other differences are found in the Northern Hemisphere winter and in the tropical Pacific; corrections by Smith and Reynolds (2004) are larger and smaller, respectively, than those of Folland and Parker (1995).
2.5. Satellite-derived data
To achieve almost complete observational coverage for recent years, some datasets use satellite-derived SSTs from Advanced Very High Resolution Radiometer (AVHRR). Although satellite-derived SSTs are adjusted to in situ measurements of bulk SST through a calibration procedure, some biases remain (Rayner et al., 2003). AVHRR SSTs are merged into in situ observations before the interpolating process in HadISST. SSTs blended from both in situ and satellite observations by Reynolds et al. (2002) are used in LDEO to make projections onto a reduced analysis space. Although the original ERSST also uses merged SSTs from satellite and in situ observations, satellite data has been removed from the ERSST that is disclosed to the public on the internet, which was used in the present study. COBE and HadSST use satellite-derived SSTs to obtain realistic climatologies in the polar region via assimilation by Japan Meteorological Agency and the global sea ice and SST (GISST) dataset by Rayner et al. (1996), respectively.
2.6. Sea ice information
To improve the data coverage at high latitudes, HadISST, COBE and ERSST add sea ice information to the analysis. SSTs near sea ice are estimated using statistical relationships between sea ice concentrations and SSTs. SSTs are set to the missing value in 100% sea ice covered grid boxes in HadISST, and they are set to the freezing point of seawater in grid boxes covered with more than 95% (90%) ice in COBE (ERSST). Discontinuities between different sources are removed in the sea ice fields used in HadISST and ERSST, whereas those in COBE are not adjusted.
2.7. Long-term changes
Long-term changes cannot be reconstructed adequately if only short-term statistical information is used for reconstructions. In HadISST, the leading EOF mode of low-pass filtered global SSTs is subtracted before the covariance EOFs are created for the reduced-space optimal interpolation. In ERSST, the low-frequency component is separated from the high-frequency component and it is interpolated using smoothing and filtering methods. On the other hand, LDEO does not pay special attention to long-term changes.
3. Comparison of statistic
All comparisons among the datasets are made over common time periods and sometimes on common spatial grids. The latter requires degrading the resolutions by simple averaging techniques. Thirty-year climatologies from 1971 to 2000 and anomalies from them are used in the present study. Intercomparisons were also performed using the anomalies from a common climatology, but the changes were not significant (not shown here). As HadISST is widely used for analytical studies and model forcing, HadISST is used for reference in the present study.
Differences between the 30-year climatologies of HadISST and the other datasets are shown in Figure 1, and the zonal means of differences are shown in Figure 2. Climatological means are not included in LDEO. Although the differences between HadISST and the other datasets are less than 0.5 °C over most of the oceans, the differences can exceed 1 °C in western boundary current regions and at high latitudes.
Systematic differences are seen in Northern Hemisphere mid-latitudes where in situ data are well sampled. HadISST is systematically colder in January and warmer in July by about − 0.2 and 0.4 °C, respectively, in the zonal means, and by greater than ± 1 °C along the east coast of the Eurasian continent and in the Gulf Stream region. This means that the seasonal cycle of the climatology in HadISST is larger than those of the other datasets in those places. This relatively large annual cycle in HadISST is consistent with the comparison by Reynolds et al. (2002) and Rayner et al. (2003). It might stem from the bias adjustment of AVHRR SSTs in HadISST. Figure 6 of Rayner et al. (2006) shows that the zonal mean adjustment has a prominent annual cycle component in Northern Hemisphere mid-latitudes, which is similar to the time-latitude difference plots between HadISST and the other datasets (not shown here).
Differences are large in high latitudes: greater than 1.5 °C with uninterpolated datasets and up to 1 °C with interpolated datasets. On the other hand, differences among the interpolated datasets are small in the polar region, as similar algorithms to convert sea ice concentration into SST are used. As in situ observations are extremely limited in these regions, the accuracy of every dataset is questionable. ICOADS shows many small patchy differences all over the ocean, which are related to noisy data in ICOADS.
3.2. Standard deviations
Here, the temporal variations of standard deviations are examined. Figure 3 shows a time series of the globally averaged root-mean-square standard deviations of SST anomalies in January for an 11-year running window. Standard deviations are computed in each grid, and then the root mean square is calculated over the globe with area weighting. Other months show similar variations. As the data coverage of uninterpolated datasets varies tremendously in time, the simple global means of standard deviations in all data-filled grids for uninterpolated datasets are contaminated by temporal variation of both data-filled grid distributions and standard deviations (Figure 3(a)). To avoid this problem, standard deviations are also calculated using common fixed grids that include no gaps from 1900 to 2002 (Figure 3(b)). The distributions of the grids are not global (not shown here). However, although the absolute values from the fixed grids differ from those of the unfixed grids, the temporal variation tendencies are consistent with each other.
Monthly standard deviations of HadISST increase a bit after the 1950s, which implies that the influence of different resolutions of reference EOF modes before and after 1949 remains. COBE shows a weak increasing trend, which stems from the number of observations. Monthly standard deviations of ICOADS and HadSST are larger than those of the other datasets through the whole period of analysis. The maximum in the 1940s is prominent in ICOADS, HadSST and TOHOKU, especially for the fixed grid, because of limited observations during the war.
Figure 4 shows the spatial distribution of SST standard deviations in two specific 30-year periods: 1871–1900 and 1971–2000. The standard deviations are commonly large in the North Pacific, the central and eastern tropical Pacific and the Gulf Stream region, associated with ENSO and/or the strong currents described in Hurrell and Trenberth (1999). In COBE, the standard deviations are small in the central North Pacific in the period of 1871–1900, which stems from the limited number of observations. The main ship routes can be seen in ICOADS through the small standard deviations in the Indian and Atlantic Oceans even in recent periods, because many observations are needed to suppress extreme values in ICOADS. ICOADS and HadSST show many areas with large standard deviations in the Southern Hemisphere extratropics because many extreme values are included. On the other hand, the standard deviations in the Southern Hemisphere extratropics in TOHOKU are small, whereas those in the equatorial Pacific in recent years are smaller than in the other datasets. This result means that both extreme values and localised climate signals are suppressed in TOHOKU by the smoothing process.
3.3. Correlation and root-mean-square difference
Figure 5 shows the correlations between SST anomalies of HadISST and those of the other datasets. High correlations (>0.8) are retained in the main shipping routes of the Indian and Atlantic Oceans from the early periods of the datasets and in the low to mid-latitudes in recent years. The central and eastern tropical Pacific shows the highest correlation coefficients (>0.8) throughout the periods from the 1880s to present even though the root-mean-square differences are large (>0.4 °C), except in recent years (the root-mean-square differences are not shown). In earlier periods, the correlations decreased and root-mean-square differences became larger. COBE, ERSST, LDEO and TOHOKU show better agreement with HadISST than ICOADS and HadSST do because of the extreme values of ICOADS and HadSST. Around 60°S, COBE and ERSST have low correlations with HadISST (<0.2 before the 1970s and < 0.4 even after the 1970s) because the observational data are very limited. On the other hand, correlations with COBE and ERSST are high (>0.8) near sea ice in the recent period, as similar algorithms are used to convert sea ice concentrations into SSTs. In ICOADS, high correlations (>0.6) are limited just to the main shipping routes and buoy-rich areas even in recent years, because of the presence of extreme values. High correlations are also limited in HadSST.
3.4. Autocorrelation of SST anomalies
Figure 6 shows 1-month-lagged autocorrelations of SST anomalies. Correlations are on the same level throughout time in the reconstructed datasets (HadISST, ERSST and LDEO), whereas correlations become lower in earlier periods in the other interpolated or uninterpolated datasets (COBE, ICOADS, HadSST and TOHOKU). Correlations for the reconstructed datasets are over 0.6 in most regions except in the very high latitudes. In the polar regions of HadISST, COBE and ERSST, anomalies are constant, as SSTs are fixed at the freezing point (−1.8 °C). Correlations are lower in ICOADS and HadSST than in the other datasets, for which high correlations (>0.6) are confined to the main shipping routes, such as in the Indian and Atlantic Oceans, and in the buoy-rich area of the tropical Pacific. On the other hand, correlations for TOHOKU are relatively high (>0.6) over most regions.
4. Comparison of climate signals
4.1. Global mean time series
Figure 7 shows globally averaged SSTs for each dataset. As LDEO does not include climatological means, the HadISST climatology is substituted for the LDEO climatology. As for the standard deviations, global means are calculated using common fixed grids that include no gaps since 1900. The differences from time series calculated using unfixed grids are mostly less than 0.05 °C (not shown here). This difference is consistent with experiments by Folland et al. (1990, 2001). Warming trends are captured in 1900–1940 and after the 1970s in all datasets. Year-to-year variations are also similar among datasets. All differences between the datasets, except before the 1940s for ICOADS, lie within ± 0.2 °C. However, several systematic differences can be found as described below.
Global mean of ICOADS is systematically cooler than the other datasets by 0.4 °C before the 1940s, which is caused by the lack of bias correction. Global mean of ERSST is about 0.1 °C warmer than the other before 1900 because of the different bias correction used. Global mean of TOHOKU is 0.2 °C warmer than the other in 1940, which stems from a too-large bias correction in TOHOKU during this period. Rayner et al. (2006) showed that the bias correction of Folland and Parker (1995) was 0.2 °C too large for the ICOADS observations. HadISST data are cooler before 1880 and in 1920–1940 and are warmer in 1945–1960 than COBE, ERSST, HadSST and TOHOKU, from improved data coverage in ICOADS observations as pointed out by Rayner et al. (2006). Global mean of LDEO is warmer than the others in 1890–1930, resulting from no extra treatment of long-term change, as Rayner et al. (2003) inferred, and a little cooler than in recent years, which might stem from a cold bias derived from the satellite-derived SSTs. Differences among datasets are larger in the 2000s than in the 1980s and 1990s, which stem from uncertainty of GTS reports.
4.2. Long-term trends
Long-term linear trends are widely used to capture overall changes in the globe, and often to inspect details of global warming, although the value of such trends changes depending on the periods used. Figure 8 shows the linear trends in °C per decade in each grid.
Warming trends are seen almost all over the ocean in every dataset. ICOADS shows a huge trend because no bucket corrections were performed. Unlike the other datasets, a warming trend appears in the high latitudes of the North Atlantic in COBE, which stems from having no adjustment between different sources of sea ice data. Almost no trends exist in ERSST in the North Pacific, whereas positive trends appear off the coast of California, in the Gulf of Alaska and around Japan in the other datasets. On the other hand, ERSST shows a positive trend in the tropical Pacific that is larger than the others. The difference between ERSST and the others stems mainly from having different bucket bias corrections.
4.3. El Niño/Southern Oscillation
The ENSO is the most dominant mode of interannual coupled atmosphere-ocean variability (Weare et al., 1976; Hsiung and Newell, 1983). The 5-month running mean of SST anomalies in the Niño 3.4 region (5°N–5°S, 120°W–170°W) has been widely used as an index of ENSO events (the Niño 3.4 index; Trenberth, 1997; Figure 9). The correlation coefficients in the running 30-year window of the Niño 3.4 indices from various datasets are above 0.98 from 1971 to 2000 and 0.85 from 1871 to 1900. As discussed before, the tropical Pacific is one of the regions that has the highest intercorrelations between the datasets. However, correlations before 1870 are low. The variabilities of COBE and ERSST are very small in the 1860s, which stems from the limited number of observations for interpolation or modes for reconstruction. Before the 1940s, the index of ICOADS is lower than the other datasets because it has no bias correction. The amplitude of TOHOKU is smaller than those of the other datasets, which is caused by smoothing (see Section 3.2).
Figure 10 shows time-longitude sections of the tropical Pacific averaged from 5°N to 5°S for the period from January 1925 to January 1928. The 1925/1926 El Niño event starts in the spring of 1925, achieves a mature phase at the end of 1925 and terminates in the spring to early summer of 1926. The amplitude of the positive SST anomaly is large in HadISST, LDEO and HadSST (>2.0 °C at maximum), and small in ERSST (<1.5 °C). At the mature phase of the 1925/1926 El Niño event, HadISST, COBE, LDEO and HadSST show a double maximum along the equator, whereas the other datasets show a single maximum. In the following La Niña event, the amplitude of the negative anomaly is small in HadISST and LDEO (<− 1.0 °C), and large in ICOADS and HadSST (>− 2.0 °C at the maximum). The relatively weak maximum and strong minimum of ERSST and ICOADS stem from having different or no bucket bias correction.
When we define ENSO events as having a Niño 3.4 index exceeding 0.4 °C for 6 months or longer (Trenberth, 1997), the durations and peak values of ENSO events are different among the datasets. For example, the 1925/1926 El Niño event appears to be an event of 12 months' duration and a peak value of 1.37 °C in HadISST, whereas it does not identify as an ENSO event in ICOADS.
4.4. Pacific Decadal Oscillation
The PDO, which has an ENSO-like spatial pattern with a longer time scale, is one of the most well-known long-term climate variations that manifest in the SST field (Tanimoto et al., 1993; Zhang et al., 1997). The PDO index is defined as the leading principal component of SST anomalies for the Pacific Ocean north of 20°N (Mantua et al., 1997). The period used here in the principal component analysis is the 30 years from 1971 to 2000. Although some grids were eliminated in the uninterpolated data, analysis domains covered almost the whole North Pacific.
Figure 11 shows the regression patterns of the PDO. All of them have positive anomalies over the central North Pacific and negative anomalies along the west coast of the North America and in the tropical Pacific. The central South Pacific shows weak positive anomalies, which suggests some symmetry about the equator. ICOADS shows many small patchy anomalies, especially in the Southern Ocean, and negative anomalies over the equatorial Pacific in ICOADS are weaker than in the other datasets because of extreme values remained. Small patchy anomalies also can be seen in HadSST.
The PDO indices are extended throughout all the data periods by projecting SST onto the corresponding EOFs. Figure 12 shows the winter PDO indices averaged from November to March. Indices are set to missing values if more than 10% of the grids for the EOF projection are empty. Indices correspond well after 1950 with each other [correlation coefficients in the running 30-year window are greater than 0.97 except for ICOADS (>0.88)], become gradually worse backward in time and are scattered before 1880. The variance of COBE is small before 1900 because of inadequate observations. ICOADS shows a consistently negative bias before the 1940s because it has no bias correction.
4.5. Indian Ocean Dipole
The Indian Ocean Dipole (IOD) is characterised by SST dipole anomalies in the southeastern equatorial Indian Ocean and in the western equatorial Indian Ocean, with accompanying anomalous easterlies along the equator (Saji et al., 1999; Webster et al., 1999). The IOD index is defined as the difference in SST anomalies between the tropical western Indian Ocean (50°E–70°E, 10°S–10°N) and the tropical southeastern Indian Ocean (90°E–110°E, 10°S–Equator) (Saji et al., 1999; Figure 13). Correlation coefficients between the IOD indices from various datasets in the 30-year period are above 0.74 from 1971 to 2000 and 0.56 from 1871 to 1900. Before 1870, the variations are different from each other, not only in their magnitudes but also in the signs of the IOD events. The amplitudes from HadSST and ICOADS are larger than those from the other datasets over the whole period, especially before 1880, in the last half of the 1910s and in the 1940s when the observational data are very limited. The amplitude from TOHOKU is also large before 1880.
4.6. Atlantic Multidecadal Oscillation
The multidecadal SST changes over the North Atlantic are referred to as the Atlantic Multidecadal Oscillation (AMO) (Kerr, 2000). The AMO index is defined as the SST anomalies averaged in the North Atlantic (Enfield et al., 2001). All datasets show a strong cold phase in the 1900s–1920s, a weak warm phase in the 1940s–1960s, a weak cold phase in the 1970s–1980s and a strong warm phase after the late 1990s, regardless of the data coverage (Figure 14). The index of ICOADS is systematically lower than the other datasets before 1940, which was caused by having no bias correction. The index of COBE is systematically lower than the other datasets before 1980, which stems from having no adjustment among different sea ice data. The index of ERSST is higher than the other datasets before 1900 because of having a different bias correction. HadISST and LDEO show a local minimum around 1940, which may stem from not including new data in ICOADS observations.
4.7. Antarctic Circumpolar Wave
The Antarctic Circumpolar Wave (ACW) is described as an approximately 4-year-period pattern of variability in the southern high latitude ocean-atmosphere system, characterised by the eastward propagation of anomalies in the Antarctic sea ice extent in a wave train that is coupled to anomalies in SST, sea surface height, sea level pressure and wind (White and Peterson, 1996). Figure 15 shows the monthly anomalies averaged over the latitudinal belt from 55°S to 57°S, which are 3-year low-pass filtered to highlight the ACW. HadISST shows the eastward propagation of SST anomalies in this period. On the other hand, such propagation is not prominent in ERSST, although the large anomalies are consistent with those in HadISST. The propagation in COBE is also vague. LDEO, ICOADS, HadSST and TOHOKU have insufficient (or no) data in this region. The ACW cannot be captured in almost any datasets that lack satellite data. In HadISST, a separate Reduced-space optimal interpolation (RSOI) analysis of the extratropical Southern Hemisphere was also carried out to preserve variability in the Southern Ocean (Rayner et al., 2003).
5. Concluding remarks
In the present study, various historical SST datasets that have been widely used for climate analyses are compared with each other, and differences among them in statistical features and climatic signals were described with reference to their gridding procedures.
Seven datasets are categorised into two groups: one includes fully interpolated datasets (HadISST, COBE, ERSST and LDEO), and the other includes simply averaged data (ICOADS, HadSST and TOHOKU). The latter group has many missing grid values, and it includes extreme values. The standard deviation and correlation of ICOADS and HadSST depend on the number of observations. In TOHOKU, instead of having a weak dependence of properties on the number of observations, the amplitudes of the localised variations are relatively small, especially in the tropical Pacific. On the other hand, it must be noted for fully interpolated datasets that the grid values include uncertainties, and correlations between the datasets are sometimes lower than 0.2, even if values are present.
Bucket bias corrections affect various climate signals. As ICOADS does not apply any bucket bias correction, systematic biases are prominent in the global mean, the long-term trend, ENSO, PDO and AMO. The bias correction applied on ERSST is different from the others; therefore, differences appear in the global means, the long-term trend, ENSO and AMO. The global mean and the AMO index of TOHOKU in 1940 are larger than in other datasets because of the too-large bias correction.
The amplitude of the HadISST climatology is larger in the mid-latitudes than in the other datasets, which might stem from the bias adjustment of AVHRR SSTs in HadISST. COBE include large gaps in the marginal sea ice zone of the Northern Hemisphere in 1987/1988, when the sea ice data source changed, and long-term trends and the AMO index show different features from those of the other datasets. LDEO does not pay special attention to long-term changes; therefore the warming trend of its global mean would be smaller than the others.
To avoid counterfeit signals that arise from gridding procedures, enough attention must be paid to the characteristics of the dataset. In some cases, comparison of the results from various datasets is advisable to confirm the results obtained. We especially need to be cautious in a quantitative treatment. The discrepancies among datasets can also cause physically important differences in atmospheric circulation when SST datasets are used in the lower boundary conditions of an AGCM (Hurrell and Trenberth, 1999; Rayner et al., 2003).
There is much room for refinement of SST datasets, in terms of having a suitable climatology for the generation of SST anomalies (Casey and Cornillon, 1999), applying a sophisticated bias correction between different measurements (Kent and Kaplan, 2006; Thompson et al., 2008), integrating historical observational data (Woodruff et al., 2005), synthesising satellite-derived data (Reynolds et al., 2005) and so on. Critical evaluations and comparisons of datasets need to be continued.
We thank Nick Rayner for scientific and editorial suggestions. We also wish to express our thanks to two anonymous reviewers for their many helpful comments. S. Yasunaka was financially supported by the Japan Society for the Promotion of Science (JSPS) as a research fellow. K. Hanawa was financially supported by Japan Fisheries Agency. This work is partially supported by the Japanese Ministry of Education, Culture, Sports, Science and Technology, through the Innovative Program of Climate Change Projection for the 21st Century.