This paper describes a blended sea-surface temperature
(SST) data set that is part of the National Oceanic and Atmospheric
Administration (NOAA) Climate Data Record (CDR) program product suite. Using
optimum interpolation (OI), in situ and satellite observations are combined
on a daily and 0.25
Sea-surface temperature (SST) is an essential climate variable (ECV). The Global Climate Observing System (GCOS) project developed a list of ECVs to focus worldwide observation efforts on a limited set of variables that are climate relevant, technically feasible, and cost effective (Bojinski et al., 2014). Collectively, ECVs can help develop adaptation and mitigation strategies, assess risks, allow attribution and prediction, and support climate services. SST is useful for monitoring El Niño events and multi-decadal ocean changes. It is also relevant to quantification and modeling of many other aspects of climate such as air–sea interaction, ocean acidification to determine solubility of carbon dioxide, biophysical processes, and marine organism distributions. However, models require not just observations but also complete data fields, also referred to as analyses. Today, satellites offer high spatial and temporal coverage and are, therefore, the main source of SST observations. Additional processing is applied to satellite data to form analyses to allow for bias corrections and gap-filling and thereby increase spatiotemporal consistency.
The objective of this work is to describe the National Oceanic and
Atmospheric Administration (NOAA) 1/4
Here, precursors to the dOISST.v2 that have evolved into the current CDR are
briefly reviewed to highlight the original motivation and subsequent
modifications. Historically, the widely-used name “Reynolds SST” has been
applied to all current and precursor products, and is therefore ambiguous
and not used here. Reynolds (1988) first introduced the concept of a blended
SST analysis that takes advantage of the sea truth offered by in situ data
and the high coverage of satellite data. Prior to 1980, ships were the only
source of observations, and the spatial–temporal coverage was sufficient
only for a coarse-scale analysis. Starting in late 1981, satellite-based SST
observations became available daily from an infrared instrument, the
Advanced Very High Resolution Radiometer (AVHRR), with Global Area Coverage
(GAC) resolution at
Reynolds and Smith (1994) adopted the optimum interpolation (OI) method to
increase the effective resolution of the blended analysis to 1
Reynolds et al. (2007) introduced major methodological changes to increase
the OISST resolution to the current daily, 1/4
Reynolds et al. (2007) referred to the above product as AVHRR-only, in
reference to the source of satellite SSTs. The same paper describes a
companion analysis that uses the same methodology but includes data from the
Advanced Microwave Scanning Radiometer on the Earth Observing System
(AMSR-E). This product, called AVHRR
As discussed in Appendix A, the 1/4
The daily OISST is available in netCDF and binary (FORTRAN IEEE big-endian) formats. In this paper, the archived netCDF files, publicly available at the National Centers for Environmental Information (NCEI) website, are described. However, the same data are repackaged and distributed elsewhere for specific projects or organizations such as the Group for High Resolution SST (or GHRSST) and Observations for Model Intercomparison Projects (Obs4MIPs), with accompanying metadata and documentation, but are not described here. The heritage binary format will be eventually phased out.
Examples of the four variables in a singles file:
A single netCDF file contains four global gridded fields (
Input data sets to the daily OISST version 2. The data sources are explained in detail in Reynolds et al. (2007). For version 1, ICOADS 2.1 was used.
Three other gridded fields at the same 1/4 Anomalies (i.e., the daily OISST minus the 1971–2000 climatological mean;
units in The standard error (with units in The 7-day median of daily sea-ice concentrations (expressed as a real
fraction from 0.0 to 1.0; Fig. 1d) is the basis of the proxy SST estimate in
the marginal ice zone. Aside from reducing noise, the temporal median
populates the time series in the early 1980s when satellite sea-ice
observations were available only every other day. There are no sea-ice data
from 4 December 1997 to 14 January 1998. This field is effectively also an
ice mask when the user opts to exclude areas with high ice concentrations.
The input data sets to dOISST.v2 are listed in Table 1 and have been evaluated in more detail in Reynolds et al. (2007). While reprocessed inputs are used whenever possible, only operational data sets meet the low latency needs of the daily updates. Users should be aware that sensor problems are typically cannot be addressed in near real time. The release date of the dOISST.v2 was November 2008. To minimize the impact of near-real-time sensor problems, data from two AVHRRs are used from 2007 onward (Table 2).
Platform time spans of AVHRR inputs to the daily OISST. Note that two satellites at a time are used beginning January 2007.
The analysis for the first day in the record used climatology as a first guess. For all other days the previous analysis is used as a first guess. For the daily update, a 1-day delayed analysis is produced. Two weeks later, after more data have become available, the analysis is repeated to produce higher-quality “final” product. The final and preliminary runs can be identified in the global attributes of the netCDF file, and the preliminary filename also contains the word “preliminary”. Only the “final” product is archived.
The daily OISST is available for the full period of record from September 1981 to the present. The data set is similar to other global daily SST analyses in that monthly, seasonal, and multi-year averages can be computed on global, regional, and local scales. For climate applications, the daily OISST is unique because it extends from late 1981 to the present and therefore spans over 30 years, often cited as the minimum period needed to distinguish interannual variations from long-term variations. The characteristic seasonal SST cycle, represented here by the 1982–2011 climatological mean, varies by location. In the tropics, it is exemplified by the NINO3.4 region (Fig. 1a), where the seasonal signal is weak. The start of the 1997 El Niño event is marked by SSTs that are more than 1 standard deviation greater than the climatological mean for over 3 months. A stronger seasonal cycle occurs in the temperate zone, as seen in the Gulf of Maine (Fig. 2b). The SSTs over the entire year 2012 exceed the climatological mean plus 1 standard deviation. The daily progression shows particularly elevated May–June temperatures, which initiated a season of anomalous lobster catch (Mills et al., 2013). Of course, these atypical events can also investigated by examining the anomalies.
Long time series are ideal for computing multi-decadal trends. On an annual
scale, the 1982 to 2014 global linear trend using dOISST is
Global OISST trends (1982–2014) using
Global validation of the dOISST using buoy and ship data is not an
independent assessment because in situ data are used to make the product,
although the amount of satellite data incorporated is much greater.
Comparisons with other SST analyses would have the same issue since most
analyses also use in situ data. With that caveat, Reynolds and Chelton
(2010) showed that, relative to buoys, the dOISST.v2 and other analyses all
exhibit regional variability in performance, reflecting their methodological
differences. For the dOISST.v2, the root-mean-square error relative to the
buoys is about 0.3
Argo data, which are not used as an input to dOISST, can also be used for
validation. However, Argo observations are available only after 2000 and
are located deeper (
Power spectra of three analyzed SST fields from the first 2 months of 2016. See text for explanation of data sets used. At smaller scales, the spectrum of the RSS product that ingests high-resolution inputs (1 km MODIS SSTs) continues to display the same spectral slope into the smaller-scale range. MODIS SSTs are available only from 1999.
In terms of feature resolution, i.e., the ability of an analysis to
reproduce mesoscale ocean features and capture SST gradients, Reynolds et
al. (2007) showed dOISST performs well. Reynolds and Chelton (2010) also
found that feature resolution of an analysis is not necessarily related to
the grid size. To illustrate this point here, the power spectral densities
of three SST data sets examined by Reynolds and Chelton (2010) are shown
(Fig. 4). The plot is similar to Reynolds and Chelton (2010) except that the
latest versions of the three data sets (from the first 2 months of 2016)
are used and the spectra are smoother because they are the average of
several areas rather than a single area, in order to provide a global
representation of each data set. The three products shown differ in grid
resolutions: dOISST is on a 1/4
In general, the higher-resolution SST analyses combine higher-resolution
infrared data (with low spatial coverage due to inability to penetrate
clouds) and lower-resolution but more spatially complete microwave data,
which have quasi-all-weather coverage (e.g., Vázquez-Cuervo et al., 2013).
The ability to provide good feature resolution in a particular area is
constrained by the availability of finer resolution (
A long-term sea-surface temperature climate data record consisting of in
situ and satellite data blended daily on a 1/4
Compared to the precursor weekly OISST at NCEP, the CDR has many updates
including higher spatial resolution, reprocessed inputs, and adjustment of
ship data to match buoys. The CDR is also used as an ancillary field in
reprocessed and operational satellite algorithms including the Pathfinder
AVHRR SST, Tropical Rainfall Measuring Mission (TRMM) rain rate, and
Aquarius salinity. The CDR version of the dOISST.v2 is available in netCDF
format
The dOISST.v2 data set described in this paper is available at the National
Centers for Environmental Information, under the name “NOAA Optimum
Interpolation 1/4 Degree Daily Sea Surface Temperature (OISST) Analysis,
Version 2” with
By Richard W. Reynolds, 30 January 2009
The purpose of this note is to discuss the upgrade of the version 1 (v.1) daily OI SST analysis (Reynolds et al., 2007) to version 2 (v.2). These changes are relatively small and mostly consist of additional temporal smoothing. In addition, preliminary Pathfinder data (following Kilpatrick et al., 2001) have been processed using NOAA-7. This allows the analysis to be extended backward in time. The daily OI AVHRR-only v2 analysis now begins on 1 September 1981; v1 began on 4 January 1985.
Other than the extension of v2 backward in time to September 1981, there are seven analysis changes in v.2.
Day-to-day analysis differences are discussed by Reynolds et al. (2007) on page 5491 and illustrated there in Fig. 13 by four partial snap shots of the Gulf Stream from the AMSR and the AVHRR instruments during 1 day. The day-to-day differences are due to a limited number of observations in regions of high variability. Observations are limited by the spatial width of the satellite swath as well as by cloud cover for AVHRR and by precipitation and the vicinity of land for AMSR.
In v.1 observations used in the daily OI were taken from the day analyzed. To temporally smooth the analysis, 3 days of data were used where the off days (the day before and after the analysis day) have doubled noise-to-signal ratios (standard deviation) compared to the center day. The doubled noise-to-signal ratio reduces the impact of the off days.
See Reynolds et al. (2007) page 5480 for a discussion of noise-to-signal ratios.
The linear least squares fit of the ship and buoy data shown in Fig. A2.
Comparison of different versions.
To verify the impact of this smoothing, 43 moored buoys were selected which had daily data for at least 99 % of the days for the period 2003–2005. These buoys were located off the coasts of North America and Europe and in the tropical Pacific and Atlantic. Auto spectra were computed for the 2003–2005 period at each of buoy locations from the daily-averaged buoy data and from four daily OI analyses: the OI using either only AVHRR or AMSR and AVHRR data with 1 day or 3 days of data. The spatial averaged spectra are shown in Fig. A1. The low frequencies (< 0.2 cycles per day) are nearly identical. The buoy data and the 3-day OI analyses have similar variances at higher frequencies although the buoy variance is being slightly higher. However, the 1-day OI analyses have considerably larger variance at higher frequencies than the others.
Globally averaged daily spectra for 2003–2005 computed at 43 moored
buoy locations and averaged. “AVHRR-only” and “AMSR
As discussed in Reynolds et al. (2007), the random and bias errors of ship SST data are larger than the random and bias errors of buoy SSTs. Furthermore, as shown in Fig. A2 from Reynolds et al. (2002), the coverage of buoys tends to increase with time while the coverage of ship tends to decrease. To determine the variability of a globally averaged bias, monthly averaged ship biases were computed with respect to buoys. However, even with temporal smoothing, differences occurred at irregular intervals and did not seem to be related to seasonal or El Niño–Southern Oscillation events.
Scatter plot of global collocated average monthly ship vs. buoy anomaly for January 1989–December 2006. The first 9 years are shown in the black and the second 9 years in red. Least squares linear fits for the two periods are also shown.
Monthly scatter plots of the collocated average global ship and buoy anomaly
SSTs are shown in Fig. A2 for two 9-year periods. The least squares linear
fit for the two periods is also shown with the slope and intercept given in
Table A1. These results strongly suggest that a spatial and temporal
constant bias correction is needed. However, finer space and time
corrections do not seem to be possible with the limited in situ data
available. The fit indicates that the average intercept is
As discussed on page 5482 of Reynolds et al. (2007), the daytime and
nighttime satellite observations are adjusted to the daily average of the in
situ (ship and buoy) data. This is done using EOTs which are similar to rotated empirical orthogonal
functions. The method produces an anomaly SST EOT for in situ data,
Average July 2006 difference between the daily AVHRR-only OI using Pathfinder NOAA-17 data and operational Navy NOAA-17 data. All versions use bias-corrected satellite data. In the top panel the Pathfinder daily OI uses no preliminary zonal bias correction; in the bottom panel the Pathfinder daily OI uses a preliminary zonal bias correction.
Spatially averaged nighttime AVHRR bias correction spectra for 2000–2005. Binomial three-point, five-point, and seven-point temporal smoothing are shown; an unsmoothed version is labeled “Nt 1 Fld”.
Daily OI Nino-3 anomalies using EOT bias correction with 15 and 7 days of data. “N-7” indicates that NOAA-7 satellite SST data are used.
AMSR extra quality-controlled SST data anomalies for 9 February 2003. The black regions show where data have been rejected by the extra quality control.
Figure 12 from Reynolds et al. (2002) shows that Pathfinder AVHRR SSTs have
cold biases with respect to operational Navy AVHRR. If the bias correction
has a residual, long-term differences will indicate it. This is shown in the
upper panel of Fig. A3 for July 2006. The tropical differences suggest
possible cloud Pathfinder contamination in the Intertropical Convergence
Zone. However, there are also high latitude differences where in situ data
are sparse. To correct these differences smoothed zonal in situ minus
satellite differences,
The biases,
A binomial filter using 3, 5, and 7 days was then used on the mode weights of the original 7-day bias corrections. To examine the impact, spectra were computed over a 6-year period. The spectral results were very similar for both day and night. The globally averaged nighttime bias spectrum is shown in Fig. A4 for each binomial filter along with the original unsmoothed spectrum. All spectra show some ringing which is roughly at frequency multiples of roughly one-seventh cycles per day and due to the use of 7 days of data.
All the binomial filters reduce the variance at higher frequencies. It is not clear which version of the binomial filter would be best. However, the 5-day binomial filter seemed to be a reasonable compromise and was selected.
Comments from John Stark, UK Met Office, and preliminary processing of NOAA-7 data indicated that the daily OI Niño-3 time series were noisy with periods of about a week due to the EOT bias correction. The time series was especially noisy in the earlier half of the record before 1990 when buoy data were sparse. Additional filtering of the weights (medians, nine-point box car, etc.) did not give much improvement. Thus, the EOT data period was increased from 7 to 15 days. Figure A5 shows that the Niño-3 anomalies using 7 and 15 days. In particular note the 7-day anomaly sign change centered near 15 January 1982. It is clear that this type of variability is reduced using 15 days.
Figure 12 from Reynolds et al. (2007) shows that the daily OI interpolates
the analysis across the region of missing AMSR data near 130
There were some errors in the quarter-degree land/sea mask. The major change
was to eliminate some inland fiords by setting these points to land. These
points occurred at the edge of the Arctic in Russia and Greenland, in the
Inside Passage area of Alaska south of Juneau, and in the Strait of
Magellan. In these regions winter sea ice was often the only data available
to the analysis and often lead to large anomalies in summer. In addition,
one badly represented small island in the Red Sea and one spurious island
off Antarctica near 75
The use of 3 days of data in the OI and smoothing of the modes in the bias
correction is not possible in near real time. Thus, two versions will be
run: a real-time interim version followed by a final version after a 2-week
delay. The interim version uses 1 day of in situ and satellite data in the
OI with a satellite bias correction using 7 days (one sided) of data and
without smoothing of the EOT modes. The final version uses 3 days (centered)
of in situ and satellite data in the OI with a satellite bias correction
using 15 days (centered) of data and smoothing of the EOT modes over 5 days
(centered). Both versions have a ship bias correction, a preliminary zonal
correction of satellite data, and improved quality control of the AMSR data.
The interim version is replaced by the final version when the final version
is computed. The daily OI using AVHRR only is available from September 1981
to present; the daily OI with AMSR
All authors contributed to the text. V. Banzon wrote the draft, with significant text added by T. Smith. M. T. Chin performed the spectral analysis for Fig. 4 and provided accompanying text. Processing details were provided by C. Liu and B. Hankins, who ran the operational production.
The authors would like to thank R. W. Reynolds (retired), H.-M. Zhang, and G. Peng for providing comments that greatly improved this paper. R. W. Reynolds also authored the material presented in Appendix A and gave permission to include it in this article. Members of the OISST Integrated Products team, including C. Hutchins, P. Jones, R. McFadden, V. Toner, and D. Wunder, helped meet CDR program requirements for transition and maintenance of dOISST.v2. Edited by: G. M. R. Manzella