A sea surface temperature (SST) and sea ice reanalysis has been produced at the Met Office based on the Operational SST and Sea Ice Analysis (OSTIA) system. The OSTIA reanalysis produces daily, high-resolution, global foundation SST and sea ice concentration fields from 1 January 1985 to 31 December 2007. The SST reanalysis uses reprocessed satellite and in situ observations that are assimilated using a multiscale optimal-interpolation-type scheme similar to that used in the near-real-time OSTIA system. Validation of the SST analysis using assimilated in situ observation-minus-background statistics shows that the accuracy of the analysis increases throughout the reanalysis period; the global root-mean-square difference is approximately 0.50 K by 2007. This approach to validation is supported in the recent period by results from comparisons with independent near-surface Argo data against which a global standard deviation error of 0.55 K was calculated. Assessment of the OSTIA reanalysis at high latitudes demonstrates that the SST and sea ice fields are more consistent with one another in the Southern Hemisphere than in the Northern Hemisphere. Comparison of the sea ice extents to those in a similar reanalysis shows OSTIA to have larger extents in the Northern Hemisphere, and the Southern Hemisphere extents are similar. The OSTIA reanalysis SSTs are shown to be regionally comparable with similar reanalyses, with the largest differences occurring at high latitudes in the summer hemisphere. Differences are observed around the ice edge and in regions with high SST gradients. The OSTIA reanalysis provides a valuable high-resolution addition to the satellite period SST data record that makes use of the (Advanced) Along-Track Scanning Radiometer [(A)ATSR] multimission data.
The need for information about global sea surface temperature (SST) comes from a wide range of users who require information about both the historical SST record and the current SST in near–real time (NRT). For instance, numerical weather prediction (NWP) centers require information about the SST to provide a lower boundary condition for their models in both forecast and reanalysis modes, seasonal forecasters require SST data to constrain the air–sea interface within their coupled ocean–atmosphere models, and climate groups monitor the current state of the SST within the context of historical SST records. Most of these users require the SST information as global gap-free analyses (otherwise known as level 4 data).
There are a number of long-term historical SST datasets, with intercomparison carried out through the Global Climate Observing System (GCOS) SST intercomparison project. The Met Office Hadley Centre Sea Ice and Sea Surface Temperature, version 1 (HadISST1), reanalysis is a global, monthly analysis produced on a 1° grid for 1871 to the present. It uses the International Comprehensive Ocean–Atmosphere Data Set (ICOADS) in situ observations and the Advanced Very High Resolution Radiometer (AVHRR) Pathfinder data (Rayner et al. 2003). The Reynolds optimal interpolation (OI) SST analysis covers the historical satellite period and is also available as a NRT product, produced at the National Oceanic and Atmospheric Administration (NOAA)’s National Climatic Data Center. The daily Reynolds OI, version 2 (v2), reanalysis is a global, daily analysis produced on a ¼° grid for 1981 to the present using ICOADS in situ data and the AVHRR Pathfinder and Advanced Microwave Scanning Radiometer for Earth Observing System (EOS) (AMSR-E) microwave data when it is available from 2002 onward (Reynolds et al. 2007).
The Operational Sea Surface Temperature and Sea Ice Analysis (OSTIA) system (Donlon et al. 2012), developed and run at the Met Office, was originally designed for NRT applications. For those purposes it is widely used, particularly by NWP centers, operational ocean forecasting systems, and climate monitoring groups. A homogeneous, daily, open ocean reanalysis using OSTIA has been produced as part of the MyOcean European project, covering the period 1 January 1985 to 31 December 2007 on a grid. Lake surface water temperatures (LSWTs) are not included in the OSTIA reanalysis despite now being included in the NRT OSTIA system (as of November 2011). The OSTIA reanalysis uses ICOADS in situ (Worley et al. 2005), AVHRR Pathfinder (NODC 2009), and the (Advanced) Along-Track Scanning Radiometer [(A)ATSR] multimission series (NEODC 2009) observations. The aim is to produce a reanalysis that is as homogeneous as possible, given the evolution of the observation network over time. The input data sources chosen have undergone temporally consistent reprocessing, and new sources of satellite data are generally not added to the reanalysis system as they become available.
As the OSTIA reanalysis provides SST information at a higher spatial grid resolution than the Reynolds OI SST and the HadISST1 reanalyses, differences in the analyses are expected in ocean regions with steep SST gradients, such as the Gulf Stream and Kuroshio regions. Differences in the analyzed SST will also exist due to different data assimilation and bias-correction techniques being applied by the different reanalysis systems. Differences also exist in the input SST data sources used, with only OSTIA using the (A)ATSR data. The input sea ice data also differ between analyses, as do the methods by which the ice data constrains the SSTs. OSTIA provides an estimate of the foundation SST, which is defined as the SST free of diurnal warming (Donlon et al. 2002). Since HadISST1 uses nighttime data only, it can also be considered a foundation SST product. However, Reynolds OI SST does not try to remove the effect of diurnal warming on the SST observations.
A number of OSTIA reanalysis products are freely available through the MyOcean project (http://www.myocean.eu.org). Daily high resolution () and reduced resolution (¼°) files are available, with the daily SST anomaly from Pathfinder climatology (NODC 2009) available at the reduced resolution. Secondary products have been calculated at both resolutions from these daily files, including daily and monthly climatologies and spatial maps of temporal trends in SST.
This paper provides a description of the OSTIA reanalysis system as well as an assessment of the outputs. The various data inputs provided by external projects used by the reanalysis are described in section 2. The method used to produce the reanalysis is described in section 3 with a focus on the differences to the NRT OSTIA system, together with a description of the additional quality control (QC) performed. Various aspects of the reanalysis output are investigated in section 4, including an assessment of the accuracy of the reanalysis, a comparison with the NRT OSTIA system, the impact of aerosols from the Mount Pinatubo eruption on the reanalysis, an assessment of the high-latitude regions where sea ice is important, and a comparison with other SST reanalyses. Section 5 summarizes the results and describes future work to improve the OSTIA reanalysis.
2. Input data sources
Figure 1 presents a timeline showing the observational datasets included in the OSTIA reanalysis. The AVHRR Pathfinder and ICOADS in situ datasets have been used throughout the entire period. The (A)ATSR dataset is comprised of observations from three different satellite missions measured using different instruments. In the overlap periods between these different missions, data from the newest mission are used. The European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) Satellite Application Facility on Ocean and Sea Ice (OSI SAF) concentration reprocessed dataset is also used in the reanalysis. These datasets are described in the following sections, followed by an overview of the observational data coverage.
a. AVHRR Pathfinder data
The NOAA–National Aeronautics and Space Administration (NASA) AVHRR Oceans Pathfinder SST, version 5, dataset has been used in the OSTIA reanalysis. It is available on a twice-daily period with a spatial resolution of 4 km on an equal-angle longitude–latitude grid. The dataset is a daily global composite of AVHRR data compiled from individual satellite overpasses (Kilpatrick et al. 2001; NODC 2009). The measurements over the time series are derived from different AVHRR instruments on board NOAA-9, NOAA-11, NOAA-14, and NOAA-16–NOAA-18 polar-orbiting satellites. Pathfinder was developed to reprocess this AVHRR database using a consensus SST algorithm to produce a consistent time series and ensure homogeneity across the sensors aboard different spacecraft (Kilpatrick et al. 2001; NODC 2009).
OSTIA is an analysis of the foundation SST and thus nighttime data only were used to minimize the diurnal warming signal, where nightime data are classified as observations from descending passes in the AVHRR Pathfinder dataset. It was found through investigations of a preliminary run of the OSTIA reanalysis that only using data flagged as the highest quality led to reduced data volumes, particularly at high latitudes. The introduction of lower-quality data did not greatly affect the OSTIA reanalysis error statistics, and thus the decision was taken to use a minimum quality flag of four for the AVHRR Pathfinder data. The standard deviation provided for each cell with the data was used as an estimate of the observation error in the OSTIA reanalysis. A minimum standard deviation of 0.33 K in each grid cell was assumed, following Kearns et al. (2000), to account for grid boxes with few observations. Information on sea ice is used in the AVHRR Pathfinder processing to ensure that ice-covered pixels should only be assigned quality flags of three or below. It was found that this processing was unsuccessful in the Arctic in the Northern Hemisphere summer, where a considerable number of pixels in the ice-covered region are erroneously flagged as high quality in the summertime. Section 3a provides a description of how the OSTIA system quality controlled these erroneous SST observations.
b. (A)ATSR multimission series data
The (A)ATSR multimission archive is a European Space Agency (ESA)–Natural Environment Research Council Earth Observation Data Centre (NEODC) project to produce an archive of SSTs using the (A)ATSR series of instruments (NEODC 2009). The dataset begins in 1991 with the launch of the ATSR-1 instrument, onboard European Remote Sensing Satellite-1 (ERS-1). This was followed in 1995 by ATSR-2 onboard ERS-2, and AATSR onboard Environmental Satellite (Envisat) in 2002.
Archive products have been reprocessed using the processing system that was operational in 2007 to produce a common format archive at a spatial resolution of 1 km. Further information on changes and improvements made to the datasets during the reprocessing can be found in NEODC (2009).
When available, ATSR-2 and AATSR data are used as a reference dataset in the analysis bias-correction scheme along with in situ data, as described in section 3b. The calibration procedures and the dual-view sensors make this dataset ideal for use as a reference dataset over other satellite instruments (Corlett et al. 2006). However, because of the loss of the 3.7-μm channel and problems associated with increased stratospheric aerosol resulting from the Mount Pinatubo eruption (Reynolds 1993), the ATSR-1 data are not used as reference data but are included in the reanalysis. When used as a reference dataset, only data flagged as the highest quality (flag 5) were used; however, for the ATSR-1 period, lower-quality data (flags 3, 4, and 5) were included in the reanalysis and bias corrected against in situ data, as for the AVHRR data. Results of investigations into the impact of bias correction for ATSR-1, detailed in Roberts-Jones et al. (2011), showed bias correcting the ATSR-1 data was necessary. The bias and standard deviation statistics provided with the (A)ATSR data as single sensor error statistics (SSES) are used in the OSTIA reanalysis: the SSES bias as a correction prior to the assimilation of the data and the SSES standard deviation as an estimate of the observation error.
c. ICOADS in situ data
The ICOADS, version 2.1, dataset contains in situ observations from ships and buoys from diverse data sources that span several centuries (Worley et al. 2005; Woodruff et al. 2010). ICOADS observations are resolved spatially to and undergo a series of QC procedures, including consistency checks, checks against climatologies, trimming of outliers, and removal of data with positional errors using a land–sea mask (Worley et al. 2005). Version 2.1 was updated from 1998 onward by the Met Office Hadley Centre, which also performed additional QCs comprising of checks for valid date, time and position, day/night flagging, a ship track check, comparisons to climatology, a freezing SST check, and a buddy check (J. Kennedy 2011, personal communication).
To estimate the foundation SST, nighttime data only were used for the OSTIA reanalysis. For ICOADS, this is set by the solar zenith angle: if the sun was above the horizon one hour previous to the measurement time, then the observation is classified as daytime and is not used. All in situ data types were used as a reference in the bias correction of satellite data in the OSTIA reanalysis (see section 3b). The use of ship data in addition to buoy data is consistent with the NRT OSTIA system and prior to the mid-1990s, few drifting and moored buoy observations are available.
d. OSI SAF sea ice concentration reprocessed data
The OSTIA ice field is derived from the OSI SAF sea ice concentration dataset. This is composed of data from the Scanning Multichannel Microwave Radiometer (SMMR) instrument between January 1985 and July 1987 and the various Special Sensor Microwave Imager (SSM/I) instruments between July 1987 and December 2007. Lists of the satellites carrying the instruments and overlap periods between instruments are given by Eastwood et al. (2010). The OSI SAF sea ice fields were bilinearly interpolated from a 10-km-resolution polar stereographic grid to the OSTIA grid. A QC procedure was performed by eye on the ice concentration fields before inclusion in the reanalysis to remove those files with regions of missing or poor data.
When the ice concentration field was not available or did not pass our QC checks, for gaps of less than 7 days, the previous day’s ice concentration field was persisted. For larger gaps the first field available after the outage was copied to the date in the middle of the outage. This aimed to reduce the impact on the SSTs of large jumps in the sea ice edge.
e. Observational data coverage
Time series of the number of observations for each of the SST datasets used in the reanalysis are shown in Fig. 2. The AVHRR Pathfinder has the greatest number of observations and, in comparison to the satellite datasets, in situ data volumes are very small. The in situ dataset is, however, very important, as it is used as a reference dataset for bias correction of the AVHRR and ATSR-1 datasets. A small proportion of days in the reanalysis period have limited or no data available for one or more data types. In addition to the total number of observations, analyses are also impacted by the number of grid boxes with data. When calculated on a ¼° grid, approximately 30% of ocean grid boxes contained (A)ATSR observations, approximately 50% contained AVHRR observations, and approximately 1% contained in situ observations over the 3-day assimilation window, detailed in section 3d. These calculations are averaged over the full reanalysis period. The temporal evolution of the number of grid boxes with data matched those of the number of observations shown in Fig. 2, which supports these numbers being representative of the data coverage.
Figure 2b shows that the number of ship-based observations has decreased over time, while the contributions from moored and drifting buoys have increased, particularly rapidly for drifting buoys from the mid-1990s onward. Automated platforms, such as moored or drifting buoys, are able to provide higher-frequency observations, at 3-hourly or hourly intervals, than ships that generally report at 6-hourly synoptic periods (Worley et al. 2005). Therefore, from 1990 onward the total number of observations increases as the dominant observing platform gradually changes from ships to drifting buoys. There is a distinct seasonal cycle in the number of nighttime in situ observations available, that is, there are fewer measurements available during the Northern Hemisphere summer than the winter due to the shorter nighttime period in the summer and the Northern Hemisphere bias of the observation network. Owing to an error in the date QC, the in situ data for 29 February in leap years has erroneously been flagged as bad and was not used in the reanalysis.
Areas that are predominantly cloud free, such as the equatorial regions, the Mediterranean, and the subtropical high pressure zones, have considerably more observations than those regions where cloud cover is more prevalent, such as the intertropical convergence zone (ITCZ) and the subpolar low pressure zones, as shown in Fig. 3a. Owing to the differing magnitudes of the number of satellite-based and in situ observations described previously, Fig. 3a represents in effect the spatial variation of the satellite data only.
Figure 3b shows that the coverage of the ICOADS in situ dataset, which includes observations from drifting buoys, moored buoys, and ships, has large spatial variations. The number of in situ observations in the Northern Hemisphere is larger than for the Southern Hemisphere, with the North Atlantic Ocean being particularly well observed. There is, relatively speaking, greater coverage at higher latitudes than for the satellite data, although coverage close to the ice edge is worse.
The Northern Hemisphere bias of the in situ observation network in 1985 can be seen in Fig. 3c. The majority of observations are at mid–high northern latitudes, and large areas of the Southern Hemisphere are poorly observed. The dominance of ship observations manifests itself in the discernible ship tracks. Figure 3d shows that by 2007, the in situ observation network has become truly global, mainly due to the expansion of the drifting buoy network.
3. Description of the OSTIA reanalysis system
This section details the steps performed within the OSTIA reanalysis system to produce the daily SST fields shown schematically in Fig. 4. Details of the NRT OSTIA analysis procedure are provided in Donlon et al. (2012), and the focus here is on the aspects specific to the OSTIA reanalysis.
a. Quality control
To minimize the risk of erroneous observations being assimilated, a series of QC checks are carried out. These are in addition to the QC procedures performed by the input data producers.
The OSTIA analysis is an estimate of foundation SST, so observations expected to be contaminated by diurnal warming are not used by the system. For the (A)ATSR multimission data, observations are flagged as being at risk of diurnal warming if the wind speed is less than 6 m s−1 and the sun is above the horizon. In such conditions, it is likely that the satellite is observing a thin stratified layer of warm surface water that is not representative of the foundation SST (Donlon et al. 2002). The AVHRR Pathfinder data and the updated ICOADS data processed by the Hadley Centre that were used in the OSTIA reanalysis did not contain the ancillary information required to perform this check, so nighttime-only observations were used. Sections 2a and 2c detail how the data producers define nighttime.
A Bayesian background check is carried out on the observations against the analysis closest to the observation times. Ingleby and Huddleston (2007) provide a full description of the background check algorithm. The scheme calculates the probability of gross error for each observation, and if this is greater than 50%, then the observation is rejected (Donlon et al. 2012).
Observations defined as being under ice or over land are rejected by the QC process. The sea ice field interpolated from the OSI SAF data are used to mask out SST observations in grid boxes with a sea ice concentration greater than 15%. This was introduced because of the erroneous AVHRR Pathfinder observations discussed in section 2a. A minimum SST check is also carried out that rejects SST observations of less than −2.0°C.
b. Satellite bias correction
Satellite observations suffer from both random measurement errors and systematic biases. These biases must be removed prior to assimilation, as the observations are assumed unbiased by the objective analysis scheme. The (A)ATSR data, which are observations of skin SST, are adjusted to compensate for the skin temperature bias by adding 0.17 K to the SST observation, which is valid for wind speeds greater than 6 m s−1 (Donlon et al. 2002). The (A)ATSR data also contain an SSES bias value, provided by the producers, that is subtracted from the measured SST observation.
A further bias correction is applied to the AVHRR and ATSR-1 data using the in situ data and the ATSR-2 and AATSR data, when they become available after July 1995, as a reference dataset. A full discussion of these choices is provided in Roberts-Jones et al. (2011). The use of ATSR-2 and AATSR data as a reference dataset provides accurate SST measurements in regions of the globe poorly sampled by in situ observations, such as the Southern Ocean (Stark et al. 2007). Observations from a 3-day window are used to generate biases using the bias-correction procedure detailed in Donlon et al. (2012).
c. Background creation
The SST analysis assimilates the satellite and in situ observations onto a background SST field generated by persisting the previous day’s SST analysis with a relaxation to climatology. Grid points with a sea ice concentration of less than 50% are relaxed toward a weekly Pathfinder climatology (NODC 2009) with a relaxation e-folding time scale of 30 days. To try to ensure consistency between the SST and sea ice fields, grid points with a sea ice concentration greater than 50% are relaxed toward −1.8°C. The relaxation e-folding time scale decreases linearly depending on sea ice concentration, from 17.5 days at concentrations of 50% to 5 days at concentrations of 100%.
d. Analysis procedure
The background field and bias-corrected observations are used to produce an SST analysis using a multiscale OI-type scheme. The OI equation is solved using an iterative analysis correction method (Lorenc et al. 1991; Martin et al. 2007).
The background error covariance matrix is split into two components, which represent errors due to mesoscale ocean features and larger-scale errors, such as those introduced by synoptic atmospheric features. The background error covariance matrix used in the OSTIA reanalysis is the same as that used in the NRT OSTIA system (Donlon et al. 2012) and uses error correlation length scales of 10 and 100 km with associated spatially varying error variances.
The observation errors comprise both a measurement error and a representivity error. Within the OSTIA system, an assumption is made that the high-resolution grid implies that errors of representivity are small and measurement errors dominate. For both the AVHRR Pathfinder and (A)ATSR satellite data, each observation comes with an estimate of the observation error provided by the data producers, detailed in sections 2a and 2b. For the in situ data, the observation errors vary spatially and are obtained from a static 2D field that is specified a priori (Donlon et al. 2012). It is assumed that there are no spatial correlations in the observations. Although this is known not to be true, the observation correlations are hard to estimate and model. Future work on estimating and using information on the spatial error correlations within the OSTIA system may be carried out within the ESA SST Climate Change Initiative (CCI) project (http://www.esa-sst-cci.org).
The assimilation scheme is run using a rolling observation window of 72 h centered on 1200 UTC on the analysis day. Observations closest to the center of the analysis day are given a higher weight in the assimilation scheme. This is achieved by increasing the observation error, thus giving the observation less weight in the assimilation as the time of observation increases away from 1200 UTC on the analysis day. The observation error is scaled by a factor that increases linearly from 1 at 1200 UTC on the analysis day to 1.5 at ±36 h. See Roberts-Jones et al. (2011) for details of the method used to estimate these parameters.
The estimated error standard deviation of the analyzed SST is provided as a daily uncertainty estimate. This is calculated using an analysis quality (AQ) optimal interpolation technique as described by Donlon et al. (2012).
4. Assessment of the OSTIA reanalysis
The following section provides results from a series of assessments carried out on the OSTIA reanalysis. The reanalysis SST fields are validated by comparing them to both assimilated and independent observations. SST and sea ice fields are compared to those obtained from other reanalysis products. The consistency between the sea ice and SST fields are also investigated. More detailed results of these assessments are described in Roberts-Jones et al. (2011).
a. Comparison to in situ observations
To assess an SST reanalysis product, comparisons of the analysis fields with an independent reference dataset are desirable. In the production of the OSTIA reanalysis, surface observations have not been withheld from the assimilation scheme for validation purposes. This was due to both the scarcity of reference observations for the full reanalysis period and the likely detrimental impact of withholding observations on the reanalysis product. Validation using an independent dataset has been carried out using near-surface Argo data when they are available for the more recent period. However, validation of the SST reanalysis over its full time period is required, and in the absence of an independent reference dataset, assimilated data are used.
Within the reanalysis, the 3-day assimilation window means that two-thirds of the SST observations are shared between successive days’ analyses. This use of observations implies that the background field is not independent of the observations for any particular day. Validation using independent Argo data produces similar observation-minus-analysis statistics to those obtained from assimilated observation-minus-background differences for the recent period. Bearing in mind the caveat that the background data are not independent, and in the absence of a viable alternative for validation over the full reanalysis period, observation-minus-background statistics are used.
The differences between each SST observation and the background field are calculated by bilinearly interpolating the background to observation location. The mean and root-mean-square (RMS) of these differences is calculated within each ¼° grid box. The mean value of all the differences provides an estimate of the bias, and the RMS of the differences can be used to assess accuracy. These statistics are calculated for all grid boxes and then combined over the global ocean and regionally to provide spatial statistical information on both the observations and analysis.
A reduction in the global RMS differences for all in situ observations from approximately 1.00 K in 1985 to approximately 0.50 K in 2007 is observed in Fig. 5a. This result indicates that by the end of the reanalysis, the RMS is in line with that of the NRT OSTIA system of 0.57 K, calculated for the period 1 January 2007–31 December 2010 (Donlon et al. 2012). A seasonal cycle can be seen in the RMS with an increase in the Northern Hemisphere summer. This may in part be due to the Northern Hemisphere bias of the observation network, which means that less data are available during the shorter Northern Hemisphere nights.
The seasonal cycle described above can be discerned in the RMS of the ship, drifting buoy and moored buoy observations that make up the in situ data, shown in Figs. 5b–d. In the drifting buoy statistics, this decreases as the drifter network matures to provide near-global coverage (after 2002). The largest seasonal variation is apparent in the moored buoy statistics that contain observations from both coastal and tropical moorings. This is due to a Northern Hemisphere bias in the network of coastal moorings. For the three observation types, the validation statistics remain relatively consistent throughout the reanalysis period with an observation-minus-background RMS of 1.10, 0.50, and 0.55 K for ship, drifting buoy, and moored buoy observations, respectively. The statistics from the individual observation types along with Fig. 2 show that one of the factors contributing to the decrease in the in situ observation-minus-background RMS during the reanalysis period is the change from a ship-dominated observation network in 1985 (with high RMS values) to a drifting-buoy-dominated observation network in 2007 (with more accurate observations).
The changes in the satellite observations during the reanalysis have also had an effect and can be seen in the bias calculated against in situ observations. There is an increase in the observational bias after July 1995 once ATSR-2 satellite data are used to bias correct the AVHRR Pathfinder data (Fig. 5). Prior to this, all satellite data were bias corrected to the in situ observations, so this is not surprising, but it does highlight a possible disagreement in the ATSR-2 and in situ observations. The initial increase in the in situ observation bias is most marked in the ship observations where the bias increases from approximately 0.1 to approximately 0.25 K during the ATSR-2 period. The subsequent drop in full in situ bias once AATSR data comes online in July 2002 highlights a possible disparity in the ATSR-2 and AATSR observations and can be seen in all three in situ observation types. This disparity is best quantified using the drifting buoy statistics, which are the least biased of the in situ–measured types, and had a drop in bias from 0.06 to 0.02 K when AATSR data are used.
Near-surface Argo data from the Enhanced Ocean Data Assimilation and Climate Prediction–Ensembles-Based Predictions of Climate Changes and Their Impacts (ENACT–ENSEMBLES) (EN3) database (Ingleby and Huddleston 2007), version 2a, have been used to independently validate the OSTIA reanalysis between January 2003 and December 2007. For each Argo temperature profile, the top value that has passed the quality control falling within 3–5-m depth is taken. These data have been shown to be a good estimate of foundation SST by C. Merchant and G. K. Corlett (2011, personal communication), who performed a three-way comparison between Argo, surface drifter, and AATSR data. The geographical coverage of the Argo data in January 2003 is far from global, so statistics from this early period are less robust than those from later in the time series.
For each day of the reanalysis, the OSTIA data have been bilinearly interpolated to the location of the Argo observations valid on that day and the differences calculated. Statistics of these differences have been calculated globally and for various ocean regions (not shown). The global standard deviation error of the reanalysis is approximately 0.55 K with a cold bias of 0.10 K, as shown in Fig. 6. The standard deviation errors are smaller in tropical regions (0.40 K) and increased in areas of higher SST variability, such as the North Atlantic (0.70 K). The possibility that the cold bias could come from a diurnal effect in the Argo data has been tested by performing a similar comparison to Argo data that are valid at nighttime only (not shown). The resulting biases are very similar to those shown in Fig. 6a.
A comparison of the errors in the OSTIA reanalysis outputs and those in the NRT OSTIA system to the Argo data for 2007 is presented in Fig. 6b. Both the standard deviation and mean errors are higher in the reanalysis, which implies that the extra data being used in the NRT system [including Meteorological Operation (MetOp) AVHRR data, data from the Spinning Enhanced Visible and Infrared Imager (SEVIRI) geostationary satellite, and microwave data from AMSR-E and Tropical Rainfall Measuring Mission Microwave Imager (TMI)] provide a significant benefit.
b. Comparison to satellite observations
Figure 7 shows comparisons of the background field to the (A)ATSR and AVHRR Pathfinder observations. The observation values used in these calculations are those before the OSTIA bias correction is carried out, and so any observation bias that we are attempting to correct for will show in these results. The AVHRR and ATSR-1 data are bias corrected; the ATSR-2 and AATSR data are used in the reference dataset in addition to the in situ data (section 3b).
Statistics are shown for the three missions, ATSR-1, ATSR-2, and AATSR, in Fig. 7a. The impact of bias correcting to the ATSR-2 data is evident by the step change in the RMS in June 1995 when the data become available, with the RMS decreasing from approximately 0.75 to approximately 0.35 K. A discernible seasonal cycle can be observed in the RMS with highs during Northern Hemisphere summer, although the magnitude of this cycle decreases once ATSR-2 data are used. The noisier daily RMS and bias values for the ATSR-1 observations relative to the ATSR-2 and AATSR observations may in part be due to the ATSR-1 data volumes being less stable than the later missions, with a greater proportion of days with reduced numbers or no observations (Fig. 2). The peaks observed in the RMS in April 1998 and in February–June 2002 correspond to periods during which there was a significant reduction in the number of observations. The gap in the time series during which there were no observations between 1 January 1996 and 1 July 1996 is due to a known scan mirror problem with the ATSR-2 satellite (M. J. Pritchard 2009, personal communication). The observation-minus-background bias decreases from approximately 0.24 to approximately 0.00 K in June 1995 once the ATSR-2 data are used as a reference dataset in the bias correction, and the analysis is pulled closer to the ATSR-2 observations. The (A)ATSR observations remain unbiased for the remainder of the analysis, with the bias remaining close to 0.00 K. Occasional increases in the bias are seen due to days with low data volumes (Fig. 2).
A seasonal cycle can be observed in the AVHRR Pathfinder observation-minus-background RMS with peaks occurring in the Northern Hemisphere summer, as shown in Fig. 7b. Unlike the (A)ATSR observations, a similar pattern is also seen in the bias. The positive effect of including the ATSR-2 data in the bias correction can be observed in the AVHRR statistics. The RMS decreases from approximately 0.65 to approximately 0.50 K when the ATSR-2 data are used in the bias correction. This impact is most evident in the period of ATSR-2 data outage described previously between 1 January 1996 and 1 July 1997 when the AVHRR RMS returns to its pre-ATSR-2 values. A further reduction from approximately 0.50 to approximately 0.45 K occurs when the AATSR data are used as a reference.
c. Impact of the Mount Pinatubo eruption
The volcanic eruption of Mount Pinatubo in the Philippines in June 1991 resulted in a substantial quantity of aerosol being ejected into the stratosphere. This changed the radiative properties of the atmosphere and thus impacted the retrieval of SSTs from satellite instruments. Atmospheric aerosol contamination leads to negatively biased satellite SST retrievals, as the infrared radiation emitted from the sea surface is absorbed by the aerosol and reemitted at the lower temperature of the aerosol (Reynolds 1993). Of relevance to the OSTIA reanalysis is the effect of the Mount Pinatubo eruption on the accuracy of the satellite observations used and the ability of the OSTIA system to correct any biases associated with the Mount Pinatubo aerosol. During this period, both the ATSR-1 and AVHRR Pathfinder data were bias corrected to the in situ observations before being used in the analysis.
The initial large negative bias in the global ATSR-1 observations-minus-background difference of approximately −0.40 K is because the satellite was launched into the aftermath of the Mount Pinatubo eruption, as shown in Fig. 7a. As the aerosol disperses, the bias decreases, and averaged globally the bias stabilizes to approximately −0.15 K by January 1994. Figure 7b shows that the impact of the eruption can also be discerned in the AVHRR Pathfinder RMS and mean bias soon after the eruption. The errors peak globally in 1992, when the levels of aerosol in the atmosphere have the greatest effect on the AVHRR retrievals.
The global distribution of the two satellite observations-minus-background fields can be used as a proxy for the biases in the satellite data, assuming the in situ data to be unbiased. The observation-minus-background statistics were calculated on a ¼° grid for each calendar month. The effect of the aerosol ejected into the atmosphere from the Mount Pinatubo eruption can be observed in the monthly observation-minus-background spatial plots for August 1991, as it leads to a cold bias in the ATSR-1 and the AVHRR Pathfinder SST retrievals, as shown in Figs. 8a and 8b. At this time the aerosol is latitudinally confined and can be discerned as a zonal band of cold bias around the tropics in both plots. The finescale spatial pattern of the ATSR-1 and the AVHRR Pathfinder bias is similar with regions of increased negative bias, such as those off the western coast of South America and off the western coast of North Africa. The magnitude of the bias in the AVHRR Pathfinder data is larger than that of the ATSR-1 data due to the dual-view ATSR-1 retrieval being more robust to increased levels of aerosol (sections 2a and 2b). Regions of increased bias can be discerned in the AVHRR Pathfinder observations very close to the equator. There are differences in the biases between the two data sources in the western tropical Atlantic, where the ATSR-1 observations are warm biased, compared to the background, and the AVHRR Pathfinder observations are cold biased. There is also a difference in the Southern Ocean, where the ATSR-1 data have a colder bias than the AVHRR Pathfinder data. Not all the biases observed in the figures will be due to the Mount Pinatubo eruption. This may obviously be the case in regions away from the tropically confined aerosol, but there may be additional biases that are compounded by the aerosol bias in the tropical region, such as those due to Saharan dust. The shortage of ATSR-1 data in August 1991 can be seen in Fig. 8a, as the available observations in a month do not give full global coverage.
The cold tropical bias due to the Mount Pinatubo aerosol persists for both the ATSR-1 and AVHRR Pathfinder observations, as shown in Figs. 9a and 9b. The magnitude of the AVHRR Pathfinder bias decreases with the large biases close to the equator no longer discernible by October 1991. Throughout the period shown, the tropical bias is greater in the AVHRR Pathfinder data than in the ATSR-1 data. The continued dispersal of the atmospheric aerosol can be observed as the magnitude of the bias decreases with time for both the ATSR-1 and AVHRR Pathfinder observations. By January 1993, the tropical band of bias can no longer be discerned in the ATSR-1 observations. A persisting tropical bias in the AVHRR observations can be observed after January 1993, which was present prior to the Mount Pinatubo eruption (not shown). A greater cold bias at high latitudes is evident in the ATSR-1 observations compared to the AVHRR Pathfinder observations. This may be due to cloud contamination in the ATSR-1 data (C. Merchant 2010, personal communication).
d. Temporal stability of the OSTIA reanalysis
A time series of the spatial standard deviations of the daily OSTIA reanalysis SST anomalies from the Pathfinder climatology (NODC 2009) has been used to study the temporal stability of the OSTIA reanalysis. Daily standard deviations are averaged over the month for clarity. Figure 10a shows that the global standard deviation is dominated by a seasonal cycle. The increase in variability in the Northern Hemisphere summer is due to large anomalies in high northern latitudes. These anomalies are around the ice edge (caused by different ice extents being used in the OSTIA reanalysis to constrain the SSTs) and over regions of the ocean covered by ice (caused by the erroneous Arctic summertime Pathfinder data discussed in section 2a). A signal from ENSO can be observed in the global anomaly standard deviation, with increased variability associated with the strong La Niña of 1988/89 and the strong El Niños in 1987 and 1997. The record ice extent minimum in 2007 (NSIDC 2012) can also be discerned in the variability shown in Fig. 10a and was again due to large anomalies around the ice edge. Figure 10a shows the global anomaly standard deviation to be relatively stable as new satellite data are assimilated by the OSTIA reanalysis. There is a slight decrease in the anomaly standard deviation when ATSR-2 data become available in July 1995 and are used as a reference in the OSTIA bias correction.
To remove the effects of anomalies at high latitudes on the results, the spatial standard deviations were calculated between 40°N and 40°S only (Fig. 10b). An annual bimode can still be discerned in the spatial standard deviation, with increases seen in the middle of winter and summer. The large variabilities in the Northern Hemipshere summer seen in Fig. 10a can no longer be observed. The changes in standard deviation associated with the ENSO climate variability described for the global deviations are more apparent in Fig. 10b compared to Fig. 10a. An increase in the standard deviation can be discerned in 1991, which persists until 1993 due to the effects of the eruption of Mount Pinatubo discussed in section 4c. Despite this temporal variability, the spatial anomaly standard deviation is relatively stable to the introduction of new satellite data to approximately 0.04 K. The slight decrease in the standard deviation associated with the use of the ATSR-2 data in July 1995 is seen in Fig. 10b. The anomaly standard deviation decreases from 0.93 K (mean daily value from 1 January 1985–30 June 1995) to 0.89 K (mean daily value from 1 July 1995 to 31 December 2007). Over the whole period, the anomaly standard deviation is approximately 0.91 K (±0.002 K).
e. High-latitude SST and sea ice assessment
As described in section 3, the reanalysis procedure attempts to maintain consistency between the sea ice and SST fields during the production of the background SST field. In general the match between the sea ice concentration and the analyzed SST at freezing point is good. A representative day is shown in Fig. 11 that illustrates that OSTIA is able to capture well the Southern Hemisphere breakup of winter sea ice from within the ice pack. The −1.7°C contour is shown in Fig. 11 as the freezing contour rather than the −1.8°C contour, as the relaxation to −1.8°C is asymptotic and hence the SST never reaches this value.
SST and sea ice show a reasonably good match for the Southern Hemisphere. Figure 12a shows the match for the period 1 January 1993–31 December 2000, which includes, for example, the melt season of summer 1995, shown in Fig. 11. For the Northern Hemisphere, shown in Fig. 12b, the consistency between sea ice and SST is poorer; the ice extent is consistently larger than the freezing SST extent. This could be due to a bias in the ice concentration field, a bias in the SSTs, or the lack of data at high latitudes, particularly in the summer months in the Arctic, when there is little nighttime in situ data. It could be due to SST information being spread too far under the ice by the analysis. However, SST information spreading under the ice can be advantageous, as the freezing SST extent is able to continue to grow/decline in the absence of variations in the sea ice field due to missing data.
Figure 13 provides a comparison of OSTIA and HadISST1 (Rayner et al. 2003) sea ice extents calculated for regions with greater than 15% sea ice concentration. The HadISST1 ice field is a monthly dataset on a 1° grid composed of data from a number of different sources, including ice chart information and several passive microwave datasets employing different retrieval algorithms. Figure 13a shows that in the Southern Hemisphere, HadISST1 and OSTIA ice extents are similar, with OSTIA having a slightly smaller extent. Averaging OSTIA ice extents to monthly values and subtracting HadISST1 extents gives a mean difference of −0.68 × 106 km2 (standard deviation of 0.50 × 106 km2) in the Southern Hemisphere. This difference exhibits a seasonal cycle, with the largest mean seasonal difference occurring in September–November (SON) of −0.88 × 106 km2 and the smallest difference occurring in March–May (MAM) of −0.42 × 106 km2. OSTIA has larger ice extents than HadISST1 in the Northern Hemisphere, as shown in Fig. 13b. The mean difference between the two datasets is 1.79 × 106 km2 (with a standard deviation of 0.47 × 106 km2), which is significantly larger than the differences found in the Southern Hemisphere. Northern Hemisphere differences also vary seasonally between 1.89 × 106 km2 in December–February (DJF) and 1.71 × 106 km2 in SON. The magnitude of the year-round difference in the Northern Hemisphere could indicate there is an issue with the Northern Hemisphere ice extent in OSTIA. Possible sources of error in the OSI SAF ice concentration data used for the OSTIA reanalysis are discussed in Eastwood et al. (2010).
f. Temporal comparison to other reanalysis products
The temporal evolution of the OSTIA reanalysis SST fields has been compared to those from the HadISST1 (Rayner et al. 2003) and the Reynolds OI v2 SST reanalyses (Reynolds et al. 2007) using daily, regionally averaged SST. As HadISST1 is a monthly product, the monthly SST values have been linearly interpolated to produce daily values.
In terms of the global average SST for the majority of the period, the OSTIA and Reynolds OI SST reanalyses are generally consistent with one another with regard to long-term temporal patterns and interannual variability, as shown in Fig. 14a. The two analyses show a marked divergence in global average SST when the ATSR-1 data are assimilated by OSTIA after August 1991 (these data are not used in the Reynolds OI reanalysis). Global average temperatures are 0.30 K colder in OSTIA compared to Reynolds OI reanalysis during this period. Prior to the introduction of the ATSR-1 data, the OSTIA reanalysis was warmer than Reynolds OI reanalysis by 0.05 K, particularly in the Northern Hemisphere summer, with good agreement observed in the Northern Hemisphere winter. Once the ATRS-2 data are assimilated in July 1995, the two analyses are in closer agreement, although OSTIA is 0.05 K colder. This small difference persists when the AATSR mission replaces the ATSR-2 mission in July 2002 and throughout the remainder of the reanalysis period. The differences described above due to the ATSR-1 data are not apparent between 40°N and 40°S (Fig. 14b), indicating an issue with ATSR-1 data at mid to high latitudes.
The global average SST in the OSTIA and the HadISST1 reanalyses are very different, with OSTIA being approximately 0.50 K colder throughout the period, as shown in Fig. 14a. The cooling effect of the ATSR-1 data on the OSTIA reanalysis is evident, as it increases the divergence between the two analyses to −0.90 K during this period. Apart from the cold bias, there is a degree of consistency between the two analyses in terms of the long-term temporal patterns and the interannual variability. The largest differences again exist at mid–high latitudes in excess of 40°N and 40°S. This is shown in Fig. 14b, which shows closer agreement between the two analyses between 40°N and 40°S. Here, OSTIA varies between being approximately 0.30 K colder during the ATSR-1 period and being 0.10 K colder once the AATSR data are used.
There is good agreement between the OSTIA and Reynolds OI reanalyses in their representation of SST during ENSO cycles, as shown in Fig. 14c. The figure shows the average SST in the Niño-3.4 region, which was chosen as it is widely used as a climatic index to categorize whether an ENSO event is occurring (Trenberth and Hoar 1997). The daily variability matches well, and the strong El Niño events in 1987 and 1997 and the strong La Niña in 1988/89 can be discerned. Although HadISST1 as a monthly product is somewhat smoother than OSTIA, the long time-scale variability matches reasonably well with large-scale ENSO events being captured in both analyses. OSTIA is 0.15 K colder than HadISST1 in this region, although this is smaller than the differences discussed previously for the global and 40°N and 40°S comparisons.
g. Spatial comparison to other reanalysis products
The spatial pattern of the differences between the OSTIA reanalysis and the HadISST1 and the Reynolds OI v2 SST reanalyses were investigated using the mean and standard deviation of the monthly differences between OSTIA and the other SST analyses. These differences have been calculated using the reduced-resolution OSTIA data (¼° grid) regridded to the Reynolds and HadISST1 grids, respectively; data from the 1985 to 2007 period only are considered.
Figure 15a shows good agreement between OSTIA and Reynolds in January. OSTIA is colder than Reynolds around the ice edge in the Northern Hemisphere. This is due to the sea ice data used to constrain the SSTs in OSTIA having a larger extent than that used in Reynolds. The two reanalyses also differ slightly in high-gradient SST regions, such as the Gulf Stream, the Kuroshio, and at the edge of the Antarctic Circumpolar Current (ACC) in the southern Indian Ocean. The two analyses resolve the SST gradients slightly differently in these regions. These differences in feature resolution manifest themselves as the SST anomaly dipoles that can be observed in Fig. 15a. These differences around the ice edge and in high SST gradient regions exhibit a large amount of variability, which is illustrated in the relatively high monthly standard deviations shown in Fig. 15b.
The OSTIA reanalysis is colder in the Southern Ocean, which is in part due to the (A)ATSR data used in the OSTIA reanalysis but not in Reynolds. This difference is most evident during those years of the ATSR-1 mission, as described in section 4f. This is apparent when looking at the individual mean monthly differences between the two analyses (not shown). OSTIA is colder than Reynolds in the western tropical Pacific and tropical Atlantic, which is contrary to the general pattern for OSTIA to be slightly warmer in open ocean regions, illustrated in Fig. 15a. Unlike those described previously, these differences are relatively consistent and exhibit very little variability, which can be seen in Fig. 15b. This is also true of the cold difference observed in the Sea of Okhotsk.
Figure 15c shows a much greater disparity between the OSTIA and HadISST1 reanalyses in January compared to the OSTIA and Reynolds differences. OSTIA is colder over all ocean regions. This difference is most evident at mid–high latitudes in both the Northern and Southern Hemispheres. Differences in the ice extents of the two analyses in the Arctic are discernible through the SSTs, with OSTIA having a larger ice extent than HadISST1, as discussed in section 4e. The difference in resolution between the two analyses has manifested itself in a greater disparity in the SSTs in regions with high SST gradients than those shown for Reynolds. This has led to the dipolar banding, which can be observed in the Gulf Stream and Kuroshio regions. Similar patterns can be observed at the edge of the ACC off the southern coast of Africa and South America and at the western edge of the South Atlantic gyre. Again, these differences exhibit a large amount of variability, as shown in Fig. 15d.
Figure 16 shows the mean and standard deviation of the monthly differences for OSTIA minus Reynolds and OSTIA minus HadISST1 for July, illustrating the differences in Northern Hemisphere summer. Figure 16a is very similar to Fig. 15a with one major exception. OSTIA is colder at high northern latitudes right up to the pole under regions covered by ice during the Northern Hemisphere summer. It is likely that these differences are due to the effect of erroneous AVHRR Pathfinder data that were not used in the OSTIA reanalysis due to the ice masking of the observations as described in sections 2a and 3a. If these suspect observations were used in the Reynolds reanalysis, they would have led to the observed Reynolds warm bias. Comparison of Figs. 15a and 16a illustrate that OSTIA is no longer cooler than Reynolds in the Southern Ocean in July, apart from differences around the Antarctic ice edge, but it is now colder in the Northern Pacific and North Atlantic. This change indicates that OSTIA is colder than Reynolds at high latitudes in the summer hemisphere. The differences around the Northern Hemisphere ice edge and in high-gradient SST regions described previously for the January comparison are in the main still observed; however, the dipolar differences in the Gulf Stream are reduced. The variability of the monthly differences shown in Fig. 16b has the same patterns as those observed for January in Fig. 15b.
Comparison of the July and January OSTIA-minus-HadISST1 differences in Figs. 16b and 15b shows a similar seasonal cycle to that observed in the OSTIA-minus-Reynolds differences. As for the Reynolds differences, OSTIA is colder than HadISST1 at high latitudes in the summer hemisphere. However, the HadISST1 difference is greater than the Reynolds difference and extends to midlatitudes in the summer hemisphere. OSTIA is warmer than HadISST1 in the Sea of Okhotsk and north of Hudson Bay; this could be due to different ice extents in the two analyses, as the ice retreats in the Northern Hemisphere summer. The differences in these regions show a high degree of variability, as illustrated in Fig. 16d. The features in the mean difference associated with the ice edge and the high-gradient SST regions described previously are also apparent in July, as illustrated in Fig. 16b. The variability of the OSTIA-minus-HadISST1 differences is increased at high latitudes in the Northern Hemisphere in July relative to January, as seen when comparing Figs. 16d and 15d. Figure 16d also shows that a large degree of variability is exhibited in regions of high SST gradients.
5. Summary and conclusions
The OSTIA reanalysis provides a daily, high-resolution, global SST and sea ice analysis for the satellite period from 1 January 1985 to 31 December 2007. The data sources used in the reanalysis have been described, as has the OSTIA reanalysis system.
Assessment of the OSTIA reanalysis has shown via comparisons to assimilated in situ observations that the accuracy of the analysis improves throughout the reanalysis period. The global in situ observation-minus-background RMS decreases from approximately 1.00 K in 1985 to approximately 0.50 K in 2007. Once the drifting buoy network is mature, the accuracy is within the operational target uncertainty of the NRT OSTIA system of 0.50 K. The validity of using nonindependent data to validate the reanalysis for the full period is supported by results from comparisons with independent near-surface Argo data during the recent period. These comparisons show the OSTIA reanalysis to be biased cold by 0.10 K globally with a standard deviation error of 0.55 K. Ship observations have been found to be biased, which raises questions as to their inclusion in the reference dataset used in the bias correction. Options are limited, owing to the scarcity of other in situ observations prior to the drifting buoy network becoming mature after 2002. Future work may involve investigating the possibility of propagating recent ship biases back through the reanalysis. An in situ observation dataset that includes an estimate of the bias for each observation would help alleviate the problem.
The way in which the OSTIA analysis system makes use of the different data sources is evident in the validation statistics for the satellite observations. Bias correcting the AVHRR Pathfinder data using the ATSR-2 and AATSR data has been shown to have a positive impact on the analysis statistics. The Mount Pinatubo eruption impacted the satellite observations used in the reanalysis at a global level. Biases in the satellite observations due to the volcanic aerosols are apparent for up to 2 yr after the eruption. These biases were greater in magnitude and in latitudinal extent in the AVHRR Pathfinder data than in the ATSR-1 data due to different SST retrieval methods. Despite changes in the satellite observation network, over time the OSTIA reanalysis remains relatively stable throughout its duration to approximately 0.04 K when high-latitude anomalies are not included. A slight decrease in the spatial anomaly standard deviation occurs when the ATSR-2 data are introduced.
Assessment of the OSTIA reanalysis at high latitudes has shown that the SST and sea ice fields are more consistent in the Southern Hemisphere than in the Northern Hemisphere. The analysis at high latitudes also suffers from issues in the Arctic summer due to inconsistencies in how the AVHRR and in situ data sources flag day and night observations. This resulted in satellite—but not in situ—data being used during the Arctic summer. The feasibility of using daytime measurements in the Arctic summer while minimizing any diurnal warming signal should be investigated. Work to improve the consistency between the sea ice and SST will be undertaken and may involve investigations into the shortening of correlation length scales close to the marginal ice zone. Comparison of the OSTIA reanalysis ice extents to those in the HadISST1 reanalysis shows that although the extents are similar in the Southern Hemisphere, OSTIA has consistently larger ice extents than HadISST1 in the Northern Hemisphere.
Comparison to the HadISST1 reanalysis has shown the OSTIA reanalysis SSTs to be globally colder by approximately 0.50 K. A seasonal pattern is evident in the differences, which are increased at mid to high latitudes in the summer hemisphere. Large differences are observed around the ice edge due to differing ice extents being used in the reanalyses. Dipolar differences exist in regions with high SST gradients, as the two analyses resolve the gradients differently. Investigations into the cause of the global difference through comparison of the bias corrections applied to the data in the two analysis systems will be carried out.
Although differences with respect to the Reynolds OI SST do exist, particularly at high latitudes during the ATSR-1 data period, the OSTIA reanalysis is comparable in most other regions. Differences do exist around the ice edge and in regions with large SST gradients. Large differences at high ice-covered northern latitudes exist in the Northern Hemisphere summer. It is likely that these differences are due to suspect Pathfinder observations used by Reynolds OI SST. In analyzing the differences between reanalysis products, the true SST remains unknown. The OSTIA reanalysis does, however, provide a valuable high-resolution addition to the satellite period SST data record, and unlike the comparable reanalyses, it makes use of the (A)ATSR multimission data.
Although the OSTIA reanalysis ends on 31 December 2007, there is a period of overlap with the NRT OSTIA system. Data from the NRT OSTIA system can be viewed as an ongoing version of the reanalysis, but this does not constitute a homogeneous time series. The next reanalysis using the OSTIA system will be carried out as part of the ESA SST Climate Change Initiative (CCI) using data sources generated during the project. It is hoped that the OSTIA CCI reanalysis will be continuously updated in delayed mode to form a homogeneous time series up to present day. This reanalysis will probably not assimilate any in situ data, and thus an independent reference dataset will be available to validate the reanalysis for the full period. Lake surface water temperatures (LSWTs) are now included in the NRT OSTIA system (as of November 2011) and may be included in future reanalyses.
The research leading to these results has received funding from the European Community’s Seventh Framework Program FP7/2007-2013 under Grant Agreement 218812 (MyOcean). The authors thank Nick Rayner, John Kennedy, Steinar Eastwood, Chris Merchant, Gary Corlett, and Mike Bell for their inputs during the production of the OSTIA reanalysis.