• Balmaseda, M. A., M. K. Davey, and D. L. Anderson, 1995: Decadal and seasonal dependence of ENSO prediction skill. J. Climate, 8, 27052715.

    • Search Google Scholar
    • Export Citation
  • Casey, K. S., and P. Cornillon, 2001: Global and regional sea surface temperature trends. J. Climate, 14, 38013818.

  • Dee, D. P., and Coauthors, 2011: The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553597, doi:10.1002/qj.828.

    • Search Google Scholar
    • Export Citation
  • Donlon, C. J., M. Martin, J. D. Stark, J. Roberts-Jones, E. Fiedler, and W. Wimmer, 2012: The Operational Sea Surface Temperature and Sea Ice Analysis (OSTIA). Remote Sens. Environ., 116, 140158.

    • Search Google Scholar
    • Export Citation
  • Elsner, J. B., J. P. Kossin, and T. H. Jagger, 2008: The increasing intensity of the strongest tropical cyclones. Nature, 455, 9295, doi:10.1038/nature07234.

    • Search Google Scholar
    • Export Citation
  • Embury, O., and C. J. Merchant, 2012: A reprocessing for climate of sea surface temperature from the along-track scanning radiometers: A new retrieval scheme. Remote Sens. Environ., 116, 4761.

    • Search Google Scholar
    • Export Citation
  • Embury, O., C. J. Merchant, D. Berry, and E. Kent, 2011: A reprocessing for climate of sea surface temperature from the along-track scanning radiometers. Proc. 2011 EUMETSAT Meteorological Satellite Conf., Oslo, Norway, EUMETSAT, P.59.

  • Embury, O., C. J. Merchant, and G. K. Corlett, 2012: A reprocessing for climate of sea surface temperature from the along-track scanning radiometers: Initial validation, accounting for skin and diurnal variability. Remote Sens. Environ., 116, 6278, doi:10.1016/j.rse.2011.02.028.

    • Search Google Scholar
    • Export Citation
  • Horrocks, L. A., A. R. Harris, and R. W. Saunders, 2003: Modelling the diurnal thermocline for daytime bulk SST from AATSR. Met Office Forecasting Research Tech. Rep. 418, 27 pp. [Available online at http://www.metoffice.gov.uk/media/pdf/i/k/FRTR418.pdf.]

  • Kantha, L. H., and C. A. Clayson, 1994: An improved mixed layer model for geophysical applications. J. Geophys. Res., 99 (C12), 25 23525 266.

    • Search Google Scholar
    • Export Citation
  • Kennedy, J. J., N. A. Rayner, R. O. Smith, D. E. Parker, and M. Saunby, 2011a: Reassessing biases and other uncertainties in sea surface temperature observations measured in situ since 1850: 1. Measurement and sampling uncertainties. J. Geophys. Res., 116, D14103, doi:10.1029/2010JD015218.

    • Search Google Scholar
    • Export Citation
  • Kennedy, J. J., N. A. Rayner, R. O. Smith, D. E. Parker, and M. Saunby, 2011b: Reassessing biases and other uncertainties in sea-surface temperature observations measured in situ since 1850: 2. Biases and homogenisation. J. Geophys. Res., 116, D14104, doi:10.1029/2010JD015220.

    • Search Google Scholar
    • Export Citation
  • Kilpatrick, K. A., G. P. Podestá, and R. Evans, 2001: Overview of the NOAA/NASA advanced very high resolution radiometer Pathfinder algorithm for sea surface temperature and associated matchup database. J. Geophys. Res., 106 (C5), 91799197.

    • Search Google Scholar
    • Export Citation
  • Lea, D. W., 2004: The 100 000-yr cycle in tropical SST, greenhouse forcing, and climate sensitivity. J. Climate, 17, 21702179.

  • Llewellyn-Jones, D., and J. Remedios, 2012: The Advanced Along Track Scanning Radiometer (AATSR) and its predecessors ATSR-1 and ATSR-2: An introduction to the special issue. Remote Sens. Environ., 116, 13.

    • Search Google Scholar
    • Export Citation
  • Llewellyn-Jones, D., M. C. Edwards, C. T. Mutlow, A. R. Birks, I. J. Barton, and H. Tait, 2001: AATSR: Global-change and surface-temperature measurements from Envisat. ESA Bulletin, No. 105, Ecological Society of America, Washington, DC, 10–21. [Available online at http://www.esa.int/esapub/bulletin/bullet105/bul105_1.pdf.]

  • Mackie, S., C. J. Merchant, O. Embury, and P. Francis, 2010: Generalized Bayesian cloud detection for satellite imagery. Part 2: Technique and validation for daytime imagery. Int. J. Remote Sens., 31, 25952621, doi:10.1080/01431160903051711.

    • Search Google Scholar
    • Export Citation
  • Merchant, C. J., A. R. Harris, E. Maturi, and S. Maccallum, 2005: Probabilistic physically based cloud screening of satellite infrared imagery for operational sea surface temperature retrieval. Quart. J. Roy. Meteor. Soc., 131, 27352755.

    • Search Google Scholar
    • Export Citation
  • Merchant, C. J., and Coauthors, 2008: Deriving a sea surface temperature record suitable for climate research from the along-track scanning radiometers. J. Adv. Space Res., 41, 111, doi:10.1016/j.asr.2007.07.041.

    • Search Google Scholar
    • Export Citation
  • Merchant, C. J., and Coauthors, 2012: A twenty-year independent record of sea surface temperature for climate from Along-Track Scanning Radiometers. J. Geophys. Res., 117, C12013, doi:10.1029/2012JC008400.

    • Search Google Scholar
    • Export Citation
  • O'Carroll, A. G., R. W. Saunders, and J. G. Watts, 2006a: The measurement of the sea surface temperature by satellites from 1991 to 2005. J. Atmos. Oceanic Technol., 23, 15731582.

    • Search Google Scholar
    • Export Citation
  • O'Carroll, A. G., J. G. Watts, L. A. Horrocks, R. W. Saunders, and N. A. Rayner, 2006b: Validation of the AATSR Meteo Product sea surface temperature. J. Atmos. Oceanic Technol., 23, 711726.

    • Search Google Scholar
    • Export Citation
  • O'Carroll, A. G., J. R. Eyre, and R. W. Saunders, 2008: Three-way error analysis between AATSR, AMSR-E, and in situ sea surface temperature observations. J. Atmos. Oceanic Technol., 25, 11971207.

    • Search Google Scholar
    • Export Citation
  • Rayner, N. A., D. E. Parker, E. B. Horton, C. K. Folland, L. V. Alexander, D. P. Rowell, E. C. Kent, and A. Kaplan, 2003: Global analyses of sea surface temperature, sea ice, and night marine air temperature since the late nineteenth century. J. Geophys. Res., 108, 4407, doi:10.1029/2002JD002670.

    • Search Google Scholar
    • Export Citation
  • Smith, D. L., J. Delderfield, D. Drummond, T. Edwards, C. T. Mutlow, P. D. Read, and G. M. Toplis, 2001: Calibration of the AATSR instrument. Adv. Space Res., 28, 3139.

    • Search Google Scholar
    • Export Citation
  • Stoffelen, A., 1998: Toward the true near-surface wind speed: Error modeling and calibration using triple collocation. J. Geophys. Res., 103 (C4), 77557766.

    • Search Google Scholar
    • Export Citation
  • Trenberth, K. E., and Coauthors, 2007: Observations: Surface and atmospheric climate change. Climate Change 2007: The Physical Science Basis, S. Solomon et al., Eds., Cambridge University Press, 235–336.

  • Veal, K. L., G. K. Corlett, D. Ghent, D. T. Llewellyn-Jones, and J. J. Remedios, 2013: A time series of mean global skin-SST anomaly using data from ATSR-2 and AATSR. Remote Sens. Environ., 135, 6476.

    • Search Google Scholar
    • Export Citation
  • Wentz, F., and T. Meissner, 2000: Algorithm Theoretical Basis Document (ATBD), version 2. Remote Sensing Systems, 66 pp.

  • Wentz, F., C. Gentemann, and P. Ashcroft, 2003: On-orbit calibration of AMSR-E and the retrieval of ocean products. Preprints, 12th Conf. on Satellite Meteorology and Oceanography, Long Beach, CA, Amer. Meteor. Soc., P59. [Available online at http://ams.confex.com/ams/pdfpapers/56760.pdf.]

  • Woodruff, S. D., and Coauthors, 2011: ICOADS release 2.5: Extensions and enhancements to the surface marine meteorological archive. Int. J. Climatol., 31, 951967, doi:10.1002/joc.2103.

    • Search Google Scholar
    • Export Citation
  • Závody, A. M., C. T. Mutlow, and D. T. Llewellyn-Jones, 1995: A radiative transfer model for sea surface temperature retrieval from the along-track scanning radiometer. J. Geophys. Res., 100 (C1), 937952.

    • Search Google Scholar
    • Export Citation
  • View in gallery
    Fig. 1.

    (a) Hovmöller diagram showing the monthly mean of the matchups over 3° latitude bands. (b) Hovmöller diagram showing the corresponding error estimate calculated using Eq. (1). Two-channel nighttime retrievals were used and the full Bayesian cloud mask was used for all ATSR-2 and AATSR data.

  • View in gallery
    Fig. 2.

    Year global means of the ARC (1-m depth)–buoy SST differences. The full Bayesian cloud mask was used for all ATSR-2 and AATSR data while for ATSR-1 the cloud mask used is detailed in the plot legend. Error bars are calculated using Eq. (2).

  • View in gallery
    Fig. 3.

    Year standard deviations of ARC (1-m depth)–buoy SST differences. The full Bayesian cloud mask was used for all ATSR-2 and AATSR data while for ATSR-1 the cloud mask used is detailed in the plot legend.

  • View in gallery
    Fig. 4.

    Histograms showing the distribution of differences globally for 1° grid boxes using two-channel nighttime SST for 1993 and three-channel SSTs for 1998, 2003, and 2009 (bin size = 0.1 K; BM = minimum Bayesian cloud mask).

  • View in gallery
    Fig. 5.

    Graphs showing the global weekly mean biases averaged over 1997–2009 for the Northern Hemisphere, Southern Hemisphere, and tropics using three-channel nighttime retrievals. Error bars are calculated using Eq. (2).

  • View in gallery
    Fig. 6.

    Zonal means using two-channel nighttime [minimum Bayesian (BM) cloud mask] SST for 1993 and three-channel SSTs for 1998, 2003, and 2009 (differences are averaged over 3° rings). Crosses indicate the mean value and the associated error bars are calculated using Eq. (1). The corresponding standard deviations of the means are marked by asterisks.

  • View in gallery
    Fig. 7.

    (a) Plot comparing the year mean biases of different ocean regions. Error bars are calculated using Eq. (1). (b) Plot of the standard deviation of the year biases for different ocean regions. Two-channel nighttime ARC SSTs and the minimum Bayesian cloud flag for ATSR-1 period were used.

  • View in gallery
    Fig. 8.

    (a) Relationship between uncertainty of ARC SST retrieval and the two-channel nighttime retrieval overplotted with the standard deviation of the differences used for each mean. (b) Relationship between uncertainty in ARC SST retrieval and the three-channel retrieval overplotted with the standard deviation of the differences used for each mean. Crosses mark the mean for that bin and the error bars are calculated using Eq. (1). Diamond symbols mark the corresponding standard deviation of the values in the bin. No ATSR-1 data were included in these plots.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 202 99 22
PDF Downloads 62 25 2

Validation of the ATSR Reprocessing for Climate (ARC) Dataset Using Data from Drifting Buoys and a Three-Way Error Analysis

Katie LeanMet Office, Exeter, United Kingdom

Search for other papers by Katie Lean in
Current site
Google Scholar
PubMed
Close
and
Roger W. SaundersMet Office, Exeter, United Kingdom

Search for other papers by Roger W. Saunders in
Current site
Google Scholar
PubMed
Close
Full access

Abstract

The Along-Track Scanning Radiometer (ATSR) Reprocessing for Climate (ARC) project aims to create an independent climate data record of sea surface temperatures (SSTs) covering recent decades that can be used for climate change analysis. Here, the ARC SSTs are assessed using comparisons with collocated drifting buoy observations and a three-way error analysis that also includes Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) data. The SSTs using the three-channel nighttime retrievals in the ARC data at 1-m depth are found to have a warm bias of 0.054 K (standard deviation 0.151 K) with respect to the drifting buoy data for the 1995–2009 time period using ATSR-2 and Advanced Along-Track Scanning Radiometer (AATSR) instrument data. However, when studying the two-channel retrievals, the ATSR-1 data are found to be less stable and with more extreme values than in later years. Some dependence on latitude, season, and fields such as total column water vapor is found in the ATSR-2 and AATSR period. An assessment of the ARC SST uncertainty shows a stable bias for low uncertainty values but more deviation above 0.6 and 0.35 K for the two- and three-channel nighttime retrievals, respectively. The three-way error analysis reveals a standard deviation of error of 0.14 K for the ARC 1-m depth SSTs using the three-channel nighttime retrieval. Estimates of the standard deviation of error for the drifting buoys are also produced and show evidence of improvement in the buoy network in the years 2003–09 from 0.19 to 0.15 K.

Corresponding author address: K. Lean, Met Office, FitzRoy Road, Exeter, Devon EX1 3PB, United Kingdom. E-mail: katie.lean@metoffice.gov.uk

Abstract

The Along-Track Scanning Radiometer (ATSR) Reprocessing for Climate (ARC) project aims to create an independent climate data record of sea surface temperatures (SSTs) covering recent decades that can be used for climate change analysis. Here, the ARC SSTs are assessed using comparisons with collocated drifting buoy observations and a three-way error analysis that also includes Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) data. The SSTs using the three-channel nighttime retrievals in the ARC data at 1-m depth are found to have a warm bias of 0.054 K (standard deviation 0.151 K) with respect to the drifting buoy data for the 1995–2009 time period using ATSR-2 and Advanced Along-Track Scanning Radiometer (AATSR) instrument data. However, when studying the two-channel retrievals, the ATSR-1 data are found to be less stable and with more extreme values than in later years. Some dependence on latitude, season, and fields such as total column water vapor is found in the ATSR-2 and AATSR period. An assessment of the ARC SST uncertainty shows a stable bias for low uncertainty values but more deviation above 0.6 and 0.35 K for the two- and three-channel nighttime retrievals, respectively. The three-way error analysis reveals a standard deviation of error of 0.14 K for the ARC 1-m depth SSTs using the three-channel nighttime retrieval. Estimates of the standard deviation of error for the drifting buoys are also produced and show evidence of improvement in the buoy network in the years 2003–09 from 0.19 to 0.15 K.

Corresponding author address: K. Lean, Met Office, FitzRoy Road, Exeter, Devon EX1 3PB, United Kingdom. E-mail: katie.lean@metoffice.gov.uk

1. Introduction

Accurately assessing the trend in sea surface temperature (SST) is very important for climate change studies (Trenberth et al. 2007), as changes in the processes by which the oceans store and transport heat may have significant impacts on the climate. SST is critical for weather forecasting, as in warm-water regions (T > 26°C) it appears to be a strong and sensitive factor for the formation of tropical cyclones (Elsner et al. 2008). SST is also important for operational oceanography [such as the Operational Sea Surface Temperature and Sea Ice Analysis (OSTIA) product described in Donlon et al. (2012)], for the estimation of the net air–sea flux of carbon (Lea 2004), for use in seasonal and decadal forecasts (Balmaseda et al. 1995), and also for atmospheric and ocean reanalyses (Rayner et al. 2003).

To understand better any variation in the SST, monitoring is needed on a global scale and over a reasonable time period covering recent years when climate change has been most rapid. Traditionally buoy and ship data have been used for looking at SST trends, such as in Casey and Cornillon (2001) (where the results are compared with several other studies using in situ data). However, there are several disadvantages to using these data, such as variation in the depth of the measurement—the drifting buoys represent the upper 1 m of water rather than one specific depth—and drifts in calibration, which lead to some inaccurate readings (Kennedy et al. 2011b). Retrieval of SSTs from instruments onboard satellites provides much better global coverage and has the potential to give more accurate results. However, as stated in the Intergovernmental Panel on Climate Change Fourth Assessment Report (IPCC AR4; Trenberth et al. 2007), satellite SST data alone have not been used as a major resource for estimating climate change because of their strong time-varying biases, which are hard to completely remove. The Along-Track Scanning Radiometer (ATSR) Reprocessing for Climate (ARC) project aims to develop an improved, highly accurate SST dataset suitable for use in climate change analyses and in the study of oceanographic processes. Further information about the ARC project can be found in Merchant et al. (2008). The use of independent datasets can be valuable in verifying and establishing confidence in results such as evidence of climate change. Often long-term satellite SST records are reliant on in situ data—for example, the Advanced Very High Resolution Radiometer (AVHRR) SSTs (Kilpatrick et al. 2001). However, the ARC SSTs are only linked indirectly to in situ data through their use in an SST analysis used as one of the NWP parameters in the cloud detection scheme (Merchant et al. 2012). This project therefore has the potential to provide a virtually independent check on estimates of the rate of warming of the ocean surface covering recent decades.

To verify the accuracy (bias, random error, and stability) of the ARC data, validation was carried out using drifting buoy data. While there are some drifting buoys that are not performing accurately (Kennedy et al. 2011b), quality control procedures aim to eliminate these outliers, leaving the majority of buoy observations available to produce bulk statistics of value. SSTs from the two data sources were compared using a collocation technique in which a buoy observation is matched to a satellite observation if certain criteria are met. During the analysis, regional, global, and latitudinal biases were investigated; we also considered biases as a function of variables such as wind speed and insolation for which data from the European Centre for Medium-Range Weather Forecasts (ECMWF) Interim Re-Analysis (ERA-Interim; Dee et al. 2011) were also used. In addition to this, a three-way error analysis was carried out that, with the inclusion of a third SST data source, allows the estimation of the standard deviation of the error in the ARC data. For this part of the study, SSTs from the Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E; Wentz et al. 2003) were also used.

The paper is organized as follows. The different datasets used are described in section 2, while section 3 contains details of the collocation technique and the method of assessing the data using a statistical analysis that includes the three-way error analysis. The results are presented in section 4 and finally section 5 summarizes the findings of the assessment.

2. Datasets

a. ARC dataset

The ATSR instruments were built with a specific aim to provide an accurate SST record (Llewellyn-Jones and Remedios 2012). The first instrument, ATSR-1, was launched on the first European Remote Sensing satellite (ERS-1) in July 1991 in a sun-synchronous orbit. Subsequently, ATSR-2 and the Advanced ATSR (AATSR) were launched on ERS-2 in April 1995 and on Envisat in March 2002, respectively, also in sun-synchronous orbits with overpass times corresponding to about 1000 or 2200 local equator crossing time (LECT). All the instruments were very well calibrated (Smith et al. 2001) and this has led to a continuous and accurate record of SST spanning more than two decades (Veal et al. 2013). The ARC data contain new developments to previous (A)ATSR products such as better cloud detection and an improved retrieval scheme (Embury and Merchant 2012). Data provided for the validation of the ARC SSTs extended from August 1991 to December 2009. ARC data using the AATSR instrument were accessible from late July in 2002 although files using ATSR-2 data were available until mid-2003. To maximize the use of data derived from AATSR, in the analysis in this report observations processed from ATSR-2 have only been used until 23 July 2002. In 1995 and 1996, ATSR-2 data are used where possible in preference to ATSR-1. Rare problems with the instruments meant that some days were unavailable such as for a period of almost three weeks in early 2001 when ATSR-2 shutdown due to a gyroscope failure on the spacecraft.

The ATSR instruments are dual view radiometers with one aperture directed in the nadir view and the second, forward view at a viewing angle of 55° to the zenith (Llewellyn-Jones et al. 2001). Previous SST retrieval algorithms have been based on radiative transfer models, for example as described in Závody et al. (1995). Details of the retrieval scheme specifically used in generating the ARC data can be found in Embury and Merchant (2012). When a SST retrieval is made, there is a choice to use only the nadir view or both views and either two channels (10.8 and 12 μm) or three channels (10.8, 12, and 3.7 μm). The three-channel retrieval is only valid at night as a result of sunlight contamination of the 3.7-μm channel during the day. The instruments were intended to give a dual view retrieval of the SST to an accuracy of 0.3 K while seeking a long-term stability of 0.1 K decade−1 (O'Carroll et al. 2006a). Throughout this report, the dual view was used for both two- and three-channel retrievals and all SSTs from three-channel retrievals were only from nighttime observations. Initial validation presented in Embury et al. (2012) compared two SST datasets produced through the ARC processing, but with one set using the new ARC retrieval scheme and the second using the operational scheme. The results showed a reduction of the biases in the majority of cases for various different properties including against latitude, wind speed, and viewing geometry.

While the (A)ATSR data are of a good quality, enhancements to the processing of the data have been needed (e.g., O'Carroll et al. 2006b) to reach a standard required for climate applications. The ARC project has made further improvements, such as better methods of cloud detection and further analysis of the instrument overlap periods in order to correct instrumental drift. It is the aim in this paper to assess the accuracy of this new dataset. The data were acquired from the Natural Environment Research Council (NERC) Earth Observation Data Centre (NEODC; http://neodc.nerc.ac.uk/browse/neodc/arc). Testing was carried out on version 0.9 of AATSR and ATSR-2 data while version 0.91 was used for the ATSR-1 period (differences between the two versions are minor and related to fields that were not used in the analysis presented here). This was the most up-to-date product at the time of testing but has now been surpassed with a later version. At the time of writing, data from 1991 to 2006 are now provided as version 1.0 while 2007–11 are available as version 1.1, which can be obtained online from NERC (http://badc.nerc.ac.uk/view/neodc.nerc.ac.uk__ATOM__DE_3abf8c96-a7d6-11e0-9cb8-00e081470265). Differences in the dataset versions mainly concern improvements to cloud masks. There is no expected difference from the currently released version for the combination of cloud mask and retrieval type (two or three channels) used for the analysis presented here using the beta version.

The new cloud mask is based on a probabilistic Bayesian method discussed in Merchant et al. (2005) and further developed using methods including those described in Mackie et al. (2010). The “full” Bayesian mask employs all three of the ATSR infrared channels during the night and uses 1.6, 11, and 12 μm during the day while the “minimum” Bayesian mask only uses 11 and 12 μm. Throughout the analysis presented here, ARC data using the full Bayesian cloud mask—with the most reliable detection—were used, apart from two-channel nighttime retrievals during the ATSR-1 period. The 3.7-μm channel failed on ATSR-1 in May 1992, which meant that three-channel retrievals were only available until this time. From May 1992 the full Bayesian cloud mask was not available for two-channel retrievals at night but the minimum Bayesian mask could be used for this period instead. However, for the daytime data either the full Bayesian, which could continue to use the 1.6-μm channel, or the minimum Bayesian cloud mask could be used throughout the ATSR-1 period. Unless otherwise specified, during the analysis data using the full Bayesian cloud mask were used in preference when both cloud masks were available. Before the failure of the 3.7-μm channel, there was a problem in switching between day and night modes so prior to this there were very few daytime retrievals.

The data contained the two- and three-channel retrieved skin SSTs and the adjustments needed to modify the values to produce the subskin (a depth of about 1 mm) temperature and the temperature at depths of 0.2, 1, and 1.5 m. The differences, due to thermal stratification, are based on the Kantha and Clayson model (Kantha and Clayson 1994) and are calculated using the method described in Horrocks et al. (2003). The 1-m measurement is taken here to compare with the drifting buoys as the depth of the drifting buoy is estimated to be representative of the top 1-m layer of water (Kennedy et al. 2011b). A depth of 1 m has also been used in other ATSR-drifting buoy comparisons such as in O'Carroll et al. (2006b, 2008), where it is stated that the heat capacity of the upper layers of ocean is best represented by the SST at this depth. Several other fields for each grid point were included such as the total column water vapor and the uncertainty of the SST retrieval, which is a combination of the theoretical performance of the retrieval, number of pixels present, and the variability in each cell. Further details regarding the contents of the ARC files can be found in Embury et al. (2011). All SSTs and accompanying fields were presented on a 0.1° × 0.1° grid. Values of solar flux and wind speed for each cell were acquired from the ERA-Interim data (Dee et al. 2011).

The ARC dataset was supplied in daily files and the time of the observations was given in seconds after midnight. The LECT for AATSR was 30 minutes earlier than for ATSR-2 and ATSR-1. To avoid any discontinuities in the long-term record, the AATSR SSTs at depth were adjusted before inclusion in the data files provided to be an estimate of the temperature 30 min after the observation time. However, the affected records in the ARC data were still provided with the original observation time, which corresponds to the skin SST retrieved. To carry out the validation with the buoys, 30 min were then added to the observation time before the collocation process was carried out.

b. Buoy data

Two different sources of buoy data were used for the collocation process and only drifting buoys were selected. From 1991 to 1996, the International Comprehensive Ocean–Atmosphere Dataset (ICOADS), version 2.5, was used (Woodruff et al. 2011). The ICOADS contains data acquired over a time scale of centuries and includes drifting and moored buoy, ship, and platform observations. The locations of buoy observations in the ICOADS files used here were only recorded to a resolution of 0.1°.

From 1997 onward, drifting buoy data acquired via the Global Telecommunications System (GTS) were used. More observations can be found in the GTS files in comparison to the ICOADS files and locations of measurements are given to three decimal places.

c. AMSR-E data

AMSR-E data are available from June 2002 and were used in this validation project as a third source of data in order to carry out the three-way error analysis. AMSR-E is a microwave instrument that was launched in May 2002 on the National Aeronautics and Space Administration's (NASA's) Aqua satellite and its data are used to retrieve information about SSTs, atmospheric water vapor, cloud water, and rain rate. Details of the algorithm and corrections used to produce the AMSR-E SSTs can be found in Wentz and Meissner (2000) and Wentz et al. (2003). Version 5 data produced by Remote Sensing Systems (www.remss.com) were obtained in daily averaged files and fields are given at a resolution of 0.25°. The SSTs correspond to an ocean depth of a few millimeters.

3. Method

a. ARC data and buoy collocation process

The Met Office monitored operational AATSR SSTs through comparison with collocated drifting and moored buoys with the analysis being updated daily with the most recently received data. This process was adapted to use the ARC data and to add relevant model fields from the ERA-Interim.

1) Initial quality control

Before the matchup process was carried out, the buoy and ARC data were initially subject to quality control. For all years, the unique identification numbers of the buoys were checked against a list of known unreliable buoys and areas of ocean prone to inaccurate readings and rejected if on the list. The buoy data also contained a quality control flag from previous tests carried out in processing at the Met Office; these considered, for example, if the buoy SST was more than 8 K from the climatology value of the area or if parameters such as the longitude and latitude were valid. For the years 2002–09, the buoy records also included corresponding NWP model predictions of SST from the Met Office for the same location and time. For these years, the matchup process was run on weekly batches of data and considered the differences between the measured SST and model SST for records from the same buoy throughout the week period. The mean and standard deviation of these differences were calculated and compared to thresholds so that if a buoy displayed too much variation within the week or showed a considerable bias, then all the records from that buoy would be rejected on the basis that it was unreliable.

The ARC data contained a quality control flag that was used to remove any records where no correction was applied to adjust the skin temperature to the temperature at various depths. A potential thermocline was flagged and the record removed where the ARC subskin (a depth of about 1 mm) temperature was predicted to be more than 0.2 K warmer than the 0.2-m-depth temperature.

2) Collocation criteria

The ARC data were given on a grid of 0.1° square cells so that the latitude and longitude of the cell boundaries were given to one decimal place. This meant that as the locations of the buoys in the ICOADS were reported to a precision of 0.1°, their given position corresponded to the boundary of four cells. In the matchup process, these four cells were considered for each buoy and if more than one provided a successful collocation, the one with the shortest time difference between the measurements was chosen. For the buoys in the GTS dataset, the three-decimal-place precision meant that a single corresponding cell could be found. In the case that a buoy might fall on the cell boundary the cell at the southwest latitude and longitude position was selected to check for collocation.

If a buoy SST was found in the same cell as a satellite measurement, then, to be considered for inclusion in the list of collocations, the time difference between the two observations was required to be less than 180 min. This time window is chosen to limit the diurnal effects while still providing a good volume of data.

Only one successful collocation was allowed for each unique buoy observation so in the event that it matched with more than one satellite measurement (e.g., from multiple overpasses of the satellite), then the one with the shortest time difference was selected. Duplicate matchups, where the same buoy produces more than one measurement that matches to the same satellite observation, were also screened by taking the records with the smallest time difference.

Collocations where there was any ice recorded in the field of view of the satellite were also discarded. The information on which grid cells contained ice was provided as part of the ARC dataset and was based on using sea ice concentration data from ECMWF where a threshold of 25% was used to define the cell as contaminated with sea ice. The ARC data were supplied with the SST taken from NWP fields provided by the 40-yr ECMWF Re-Analysis (ERA-40) and ECMWF operational data so matchups were successful if the buoy measurement was within ±5 K of the NWP estimate (this extra check could be more beneficial for buoys from before 2002 when less initial quality control could be carried out).

When carrying out the collocation process, the sensitivity to the matchup criteria was investigated. It was found that when duplicates were included and buoy data flagged as unreliable in the initial quality control were allowed to pass, there was still very little change in overall mean bias for the whole time period despite the rise in the number of matchups. The standard deviation was observed to increase, which corresponded to the larger amount of poorer data. The use of different time windows and spatial criteria was also explored, and it was found that small changes such as reducing the time window to 60 min or relaxing the precision of the buoy location to 0.1° did not have a significant impact. This suggests that the lower-resolution location of the ICOADS buoy data should not have much influence. More extreme changes—for example, reducing the precision of the buoy location to just 1°—did cause the statistics to become inconsistent with the results produced from all the experiments containing small changes.

b. Analysis of the collocations

1) Statistical analysis

The quality of the ARC data was investigated in a number of ways. Global and regional statistics concerning the ARC–buoy bias were considered. This included looking at weekly average biases as well as longer, yearly statistics. The zonal trend and seasonal variations were also studied. Histograms were constructed to gain a better understanding of the spread of the biases and comparisons between matchups from the different satellites were made. Further to this, an investigation was also carried out to look at the biases as a function of the ARC SST uncertainty, wind speed, insolation, total column water vapor, the time difference between the collocated measurements, and various other fields.

To account for the spatial variation when considering the global statistics, ARC–buoy differences were collected into 1° × 1° cells and averaged within each box. This helped to ensure that cells toward the high latitudes containing fewer matchups were not overwhelmed by the higher densities in lower latitudes. Using a simple cosine, area weights were calculated for each cell which were related to the area of the box on the surface of the earth. This prevents cells of different sizes from having an equal contribution and gives each cell a more appropriate weight. Before collecting the data onto the 1° grid, a mean and standard deviation were calculated from the differences. Those matchups exceeding ±3 times the standard deviation from the mean were rejected and successful collocations were then averaged in the 1° grid boxes.

The areas of ocean selected for the regional statistics are listed in Table 1. Since a smaller latitude range of the globe is selected, a simpler method of calculating the mean and standard deviation was used that involved the exclusion of biases more than ±3 times the standard deviation from the mean before recalculating the statistics. This method of two passes to find the mean and standard deviation should help to remove outliers and was employed wherever possible before further processing in other aspects of the analysis such as the zonal statistics. Zonal means were calculated using 3° bands of latitude.

Table 1.

Latitude and longitude limits for ocean basins used for regional bias statistics.

Table 1.

Uncertainties in the mean values are indicated by error bars calculated using an error model outlined in Kennedy et al. (2011a). The following formula is applied:
e1
where is the total measurement error variance, represents the variance in the random error, represents the variance in the systematic error, is the number of differences used in calculating the mean, and is the number of unique drifting buoys used in calculating the mean. In this way, the random and systematic errors are both taken into consideration; we also try to account for the further uncertainty that may be introduced by the collocations not all originating from unique buoys. It is possible to add a term accounting for sampling error as was carried out in Kennedy et al. (2011b). However, this contribution has been ignored here as the collocation criteria used here are stricter so that the terms in Eq. (1) will dominate. By assuming that the drifting buoys are of consistent quality, constant values were taken throughout for and of 0.26 and 0.29 K, respectively, as used for drifting buoys in Kennedy et al. (2011a).
The model is applied in a more complex way to situations where the mean is calculated globally, meaning that the data are first collected onto the 1° × 1° grid and area weights are applied to the means of these grid boxes before the global average is taken. For this case, Eq. (2) is used, which is based again on the work presented in Kennedy et al. (2011a):
e2
where is the variance of the total global error, is the area weight for grid box , and is the error variance for the differences in the grid box as calculated in Eq. (1). The multiplicative factor of 2.4 is introduced to account for correlation between grid boxes and was derived for use in the calculation of the global average in Kennedy et al. (2011a).

2) Three-way error analysis

The method of the three-way error analysis enables an empirical estimate to be made of the random uncertainty—the standard deviation of error—of each observation type. Work carried out by O'Carroll et al. (2008), using AATSR three-channel nighttime retrievals from 2003 (and similarly comparing with AMSR-E and drifting and moored buoy SSTs) found that the observation error variance can be calculated by the following equation:
e3
where is the variance of the error in observation type and is the variance of the difference between two observation types, x and y. The three-way error analysis has been developed in earlier work such as Stoffelen (1998) for use in studying wind speed, although derivation and discussion of this result is repeated in O'Carroll et al. (2008). It reported that when using drifting buoys with a 3-h time window, the standard deviation of error for the AATSR three-channel retrieval SST at 1-m depth was 0.14 K, for buoy SST was 0.24 K, and for AMSR-E SST was 0.42 K. The three data sources must be collocated in time and space although the calculation permits some small flexibility in the time window and allowance for the observations being on different spatial scales. The AMSR-E data have 0.25° resolution so the nearest grid point to the location of the ARC data in each of the matchups was chosen. The standard deviation of the differences between each combination of two data sources was calculated and substituted in Eq. (3).

The method relies on using scales for which the covariances of the errors of representativeness [errors concerning the “difference between the value of the variable on the space/time scale on which it is actually measured and its value on the space/time scale on which we wish to analyse it”; O'Carroll et al. (2008)] are negligible compared to the error covariances. This assumption allows use of the simplified equation above. However, the difference in temporal or spatial scales between the observation types causes some true geophysical differences that will add some variability of unknown extent to estimate of the standard deviation of error. Investigation showed that the statistics are reasonably insensitive to small changes in the matchup criteria such as the increased spatial resolution of ICOADS.

4. Results

a. Statistical analysis of the matchups

1) Global statistics

Figure 1a shows the mean bias for each monthly 3° latitude band of data. The two-channel night retrieval is used here in order to show more collocations from the ATSR-1 period. However, it is quite representative of the values and patterns observed in the other retrieval types. The full Bayesian cloud mask has been used for the ATSR-2 and AATSR data. The transition between the ATSR-2 and AATSR periods during 2002 is smooth but from ATSR-2 to ATSR-1 there is still a noticeable difference in the stability with more extreme values in the earlier ATSR-1 years. The polar regions tend to have a cooler bias for most of the time period while the midlatitudes (around ±30° latitude) generally show a warm bias.

Fig. 1.
Fig. 1.

(a) Hovmöller diagram showing the monthly mean of the matchups over 3° latitude bands. (b) Hovmöller diagram showing the corresponding error estimate calculated using Eq. (1). Two-channel nighttime retrievals were used and the full Bayesian cloud mask was used for all ATSR-2 and AATSR data.

Citation: Journal of Climate 26, 13; 10.1175/JCLI-D-12-00206.1

The number of drifting buoy observations per day increases quite dramatically throughout the study period leading to a rise in the number of collocations. When using the full Bayesian cloud mask, the ARC data generally contained more daytime than nighttime data, leading to more successful matchups during the daytime. This is due to the mask being stricter during nighttime conditions. The total time period for ATSR-2 contains only about a quarter of the number of collocations for AATSR and the global coverage is still not as complete. Figure 1b illustrates the corresponding error for the means in Fig. 1a calculated using Eq. (1) applied to the monthly collocations over 3° latitude bands for the two-channel nighttime retrieval. Note that the overall patterns observed for the two-channel nighttime retrieval are again typical for the different retrieval types. There is higher error for the collocations in the polar regions where there is also a decline in the number of drifting buoys. The earlier years also show higher error with many values around 0.2 K or higher, indicating that the extreme values seen in the mean are not reliable estimates and may be partly fluctuations caused by small sample sizes. This highlights the issue that statistics derived from data in the ATSR-1 period should be treated with more caution. However, in particular for the AATSR period, the errors have small values (less than about 0.1 K), especially in the midlatitude regions.

A summary of the mean global biases and standard deviations for the different instrument periods is presented in Table 2. The standard deviations decrease for more recent instruments. The low mean bias for ATSR-1 is accompanied by a lot of variability around this value as seen in Fig. 1a. The statistics of the whole time series also demonstrate how the later years dominate and mask some of the extreme values observed in the early years. All the overall mean biases for the different instruments are found to have a magnitude of less than 0.1 K although there is an element, particularly in the earlier years, of having a balance between cells containing warm and cold differences rather than collocations with a consistently small bias. The two-channel nighttime retrievals produce the lowest overall bias compared to the drifting buoys followed by the three-channel retrieval. However, the three-channel retrieval consistently gives the lowest standard deviation over the different instruments.

Table 2.

Summary of mean global biases and standard deviations using 1° grid boxes for the different instrument periods and the complete time series. For the complete time series, statistics are given for use of the full and minimum Bayesian cloud mask during the ATSR-1 period separately (bm = Bayesian minimum mask; bf = Bayesian full mask).

Table 2.

In calculating the statistics for the instrument periods, a three-sigma filter was employed to eliminate outliers. For the different instrument periods, the percentage of values discarded was relatively small (around 2%); for example, for the AATSR three-channel retrieval 1.4% of data was removed. Looking at histograms of the rejected data, there is a slight uneven distribution about the mean with a larger cold tail. It is also more prominent in the two-channel retrieval than the three-channel retrieval. This could imply some problems with residual cloud. Including the data generally has the effect of reducing the mean and also increases the standard deviation such as from 0.040 to 0.028 K and from 0.296 to 0.459 K, respectively, for the ATSR-2 three-channel retrieval.

The yearly global means are shown in Fig. 2 where, for the ATSR-1 period, the data using the full Bayesian cloud mask and the minimum Bayesian cloud mask are shown separately. The AATSR period is the most stable for all the types of the retrievals with the two-channel nighttime retrieval consistently producing yearly mean values smaller by around 0.01 K than the three-channel retrieval.

Fig. 2.
Fig. 2.

Year global means of the ARC (1-m depth)–buoy SST differences. The full Bayesian cloud mask was used for all ATSR-2 and AATSR data while for ATSR-1 the cloud mask used is detailed in the plot legend. Error bars are calculated using Eq. (2).

Citation: Journal of Climate 26, 13; 10.1175/JCLI-D-12-00206.1

The time period for the ATSR-2 instrument (mid-1995 to mid-2002) shows slightly larger variation in the global mean bias although the transition between years is relatively smooth. The fall in bias during 1997–98 may be a consequence of differential geographical sampling where a high proportion of the matchups are located in the higher-latitude regions where the satellite can sometimes produce a lower bias as shown in Fig. 1a. The ocean currents around the equator tend to act to sweep the buoys northward and southward leading to lower concentrations of collocations in the low-latitude regions. Later years tend to have more complete global coverage and higher densities of collocations reflecting the larger number of drifting buoys.

The yearly means for ATSR-1 vary more between years and show some reasonably large biases in the two-channel nighttime data compared to AATSR. There is also a larger difference between the two-channel nighttime and daytime data. The difference between using the full or minimum Bayesian cloud masks in the daytime data is not very significant.

The reduction in standard deviation of the bias seen after the earliest three years (Fig. 3) is partly correlated with the dramatic increase in the number of matchups, which also leads to lower uncertainty as seen in Fig. 1b. The greater variability seen in 1991–96 suggests that there is generally a tendency toward more extreme values (this is discussed further below when considering Fig. 4). The use of the full Bayesian cloud mask appears to produce slightly lower standard deviations, apart from during 1991, for the daytime collocations. This supports the concept that the full cloud mask should be more effective in cloud detection and therefore reduce the number of cloud-contaminated observations.

Fig. 3.
Fig. 3.

Year standard deviations of ARC (1-m depth)–buoy SST differences. The full Bayesian cloud mask was used for all ATSR-2 and AATSR data while for ATSR-1 the cloud mask used is detailed in the plot legend.

Citation: Journal of Climate 26, 13; 10.1175/JCLI-D-12-00206.1

Fig. 4.
Fig. 4.

Histograms showing the distribution of differences globally for 1° grid boxes using two-channel nighttime SST for 1993 and three-channel SSTs for 1998, 2003, and 2009 (bin size = 0.1 K; BM = minimum Bayesian cloud mask).

Citation: Journal of Climate 26, 13; 10.1175/JCLI-D-12-00206.1

The histograms in Fig. 4 give a clearer idea of the variation of the differences and support the trend of standard deviations discussed earlier. The graph for 1993 reveals the large spread around the mean value and the higher proportion of extreme values. As the years progress, a much higher percentage of the ARC–buoy differences can be found close to the mean and the percentage of differences further from the mean falls off more quickly. Similar magnitudes and spreads are also observed in the daytime differences.

It is worth noting that throughout the analysis, although the 1-m depth SST was used, statistics calculated using the ARC SSTs from 0.2 and 1.5 m were very similar. The nighttime retrievals using two or three channels showed very little change while the statistics for the daytime retrieval between 0.2 and 1.5 m displayed the largest difference, which is consistent with the expected influence of diurnal warming—some stratification still remains after removing the larger diurnal thermoclines. Table 3 shows an example of the statistics using ARC SST adjusted to different depths for the AATSR period data.

Table 3.

Mean and standard deviations of the global ARC–buoy SST differences using 1° grid boxes and using data from the AATSR period adjusted to different depths.

Table 3.

2) Seasonal variation

The seasonal variation in the three-channel ARC 1-m depth–buoy SST bias is shown in Fig. 5 where the SST differences for the same week of each year were averaged over the years 1997–2009 for different latitude bands of the globe. Data from before 1997 were not included and instead treated separately in order to investigate the more stable part of the time series independently. The Northern Hemisphere (20°–90°N) and Southern Hemisphere (90°–20°S) both show the bias decreasing during the summer months. In the tropics there is no obvious seasonal trend with the mean bias appearing slightly higher than the global average for the time period for most weeks. These higher biases near the equator may be as a result of higher total column water vapor values that prevent the longer wavelength channels at 11 and 12 μm from so effectively sensing the surface and increase the retrieval uncertainty. It is also worth noting that in removing data from 1997 to 1998 there is very little impact on the seasonal statistics.

Fig. 5.
Fig. 5.

Graphs showing the global weekly mean biases averaged over 1997–2009 for the Northern Hemisphere, Southern Hemisphere, and tropics using three-channel nighttime retrievals. Error bars are calculated using Eq. (2).

Citation: Journal of Climate 26, 13; 10.1175/JCLI-D-12-00206.1

The seasonal variation for the ATSR-1 period (not shown) has means covering a slightly larger range of values with the patterns still seeming to support the idea of higher biases during the winter months. However, because of the large errors associated with the mean values, the trend does not appear significant so there is no clear agreement or contradiction with the results from the later years.

3) Zonal variation of biases

Plots of the zonal mean biases using latitude bands of 3° (Fig. 6) show slightly higher biases are observed in the region around ±30° latitude and larger error ranges are typical in the high latitudes where there are smaller numbers of matchups present. For the earlier years the uncertainties are quite large and it is difficult to discern much trend in latitude. The selection of plots from different years also highlights how the increase in the density and global coverage of the drifting buoy network allows firmer conclusions to be made about the bias in SST. It is not clear whether the trend in the zonal mean bias is reducing toward later years. Each year's values were offset to have the same global mean as in 2009, allowing the overlap of the error bars to be used as an indicator for significance in the zonal trend compared to 2009. However, it was found that there was no pattern for which latitude bands were changing compared to the 2009 value. This suggests that there may not be a trend in the magnitude of the variation of the zonal mean bias through the years.

Fig. 6.
Fig. 6.

Zonal means using two-channel nighttime [minimum Bayesian (BM) cloud mask] SST for 1993 and three-channel SSTs for 1998, 2003, and 2009 (differences are averaged over 3° rings). Crosses indicate the mean value and the associated error bars are calculated using Eq. (1). The corresponding standard deviations of the means are marked by asterisks.

Citation: Journal of Climate 26, 13; 10.1175/JCLI-D-12-00206.1

4) Regional variation

From patterns shown in Fig. 1b and from the results in zonal variation there is a suggestion that there are varying trends in the biases in the different ocean regions. The statistics of the different regions (listed in Table 1) were investigated and the year means are presented in Fig. 7a. This plot shows the results from the two-channel nighttime retrieval in order to include results from ATSR-1. The three-channel retrieval (not shown) behaves in an almost identical way in the later years. The extreme value observed in the tropical Atlantic in 1998 was caused by an unusually small number of collocations containing some very low values. A corresponding large peak in the standard deviation is also seen for this point indicating that this deviation in mean value is not necessarily significant. The uncertainty for this point is not similarly large, which highlights a limitation of the error model used here. The model assumes that the behavior of the drifting buoys is consistent and does not account for situations where an anomalous set of differences arises. Results for the daytime retrieval are quite similar throughout the time period although this extreme value is not present.

Fig. 7.
Fig. 7.

(a) Plot comparing the year mean biases of different ocean regions. Error bars are calculated using Eq. (1). (b) Plot of the standard deviation of the year biases for different ocean regions. Two-channel nighttime ARC SSTs and the minimum Bayesian cloud flag for ATSR-1 period were used.

Citation: Journal of Climate 26, 13; 10.1175/JCLI-D-12-00206.1

Generally, the spread in the mean biases and the standard deviations from the different regions decrease as the years increase. The transition from the ATSR-2 to ATSR-1 instrument is quite noticeable with a larger divergence of the means from the different regions. However, for ATSR-1 there may be fluctuations as a result of small sample sizes as observed when discussing the uncertainty levels and geographical coverage in Fig. 1b. Throughout most of the time period, the North Atlantic and Southern Ocean regions frequently show lower mean biases than for the other ocean regions, in particular after 2000—further evidence of a cooler bias in the SSTs around the higher latitudes.

5) Assessment of estimated uncertainties

The relationship between the bias and the uncertainty of the ARC SST retrieval was considered. The SST uncertainty was calculated from a combination of the theoretical performance of the retrieval, the number of pixels present, and the variability in each cell (Merchant et al. 2012) and was provided as part of the ARC dataset. For a dataset intended for use in climate applications, it is important that good uncertainty information is provided. The uncertainty can be used to indicate the confidence users can place in each SST measurement. If the uncertainty is modeled well, at low values the standard deviation of the ARC–buoy difference should be dominated by the drifting buoy error. As the uncertainty increases, the standard deviation should then similarly rise. ATSR-1 data have not been included in this part of the analysis so that the more stable data, as demonstrated in earlier analysis, can be used and results may be produced with greater confidence.

The two-channel nighttime (Fig. 8a) and daytime biases remain quite stable to an uncertainty value of around 0.6 K and fluctuate for larger values. The standard deviation steadily rises in a linear trend as the uncertainty increases although there are some deviations at very high values (above an uncertainty of 0.8 K). In the case of the three-channel retrieval (Fig. 8b), the average bias only remains stable up to uncertainties of around 0.35 K before decreasing for values up to around 0.6 K and fluctuating beyond that. The locations of the matchups with high values are distributed over the globe and in virtually all cases of uncertainties greater than around 0.32 K the cells were adjacent to cloud edges (O. Embury 2011, personal communication). The standard deviation also increases as the uncertainty increases but this linear trend seems to break down earlier than in the two-channel case at uncertainty values above 0.6 K. Removing the ARC SSTs with large uncertainty in the retrieval may improve the accuracy. However, the large majority of observations have retrieval uncertainties less than 0.4 K so there is little impact on the overall statistics.

Fig. 8.
Fig. 8.

(a) Relationship between uncertainty of ARC SST retrieval and the two-channel nighttime retrieval overplotted with the standard deviation of the differences used for each mean. (b) Relationship between uncertainty in ARC SST retrieval and the three-channel retrieval overplotted with the standard deviation of the differences used for each mean. Crosses mark the mean for that bin and the error bars are calculated using Eq. (1). Diamond symbols mark the corresponding standard deviation of the values in the bin. No ATSR-1 data were included in these plots.

Citation: Journal of Climate 26, 13; 10.1175/JCLI-D-12-00206.1

6) Relationships between SST bias and other variables (wind speed, insolation, etc.)

As the ATSR-1 data are observed to be comparatively unstable it was decided to treat this part of the time series separately. The results discussed below only use ATSR-2 and AATSR data, but in each case the trend in the ATSR-1 data is the same or the uncertainties are too large to discern any dependence.

Considering the bias as a function of wind speed (not shown), there is a slight trend for lower biases at higher wind speeds, which is more defined for the daytime data. Potential thermoclines should have been removed but this result could be associated with their occurrence at low wind speeds. The wind data used were taken from the ERA-Interim rather than measurements taken by the drifting buoys. The in situ observations provided few measurements and an uneven spread of wind speeds, making it difficult to draw any conclusions. However, the results did not contradict those seen with the model data.

Virtually no trend is observed for the dependence of the two-channel daytime retrieval on solar flux (data provided by the ERA-Interim) although a small decrease in bias can be seen for lower solar fluxes (not shown). The collocations in these low flux regions are mostly found in the higher latitudes. A stricter threshold to screen potential thermoclines of 0.05 K gave no discernible difference in the results.

In the two-channel daytime retrieval a small increase in bias is observed for increasing total column water vapor (TCWV) (not shown) and small error ranges suggest that this trend is significant. Mean values rise from about 0.03 K at TCWV of 10 kg m−2 to 0.1 K at 50 kg m−2. Dependences are not present in the nighttime two- or three-channel retrievals although, for both day and night, at very high TCWV the value of the average bias is very unstable. The higher values of TCWV occur mainly in the tropics where the increased cloud cover can make retrievals more challenging. The larger amount of water vapor absorption may cause difficulty in calculating an accurate estimate of radiation from the sea surface. However, no relationship was found in two- or three-channel retrievals between the bias and the number of cloudy pixels in the field of view within each ARC grid cell.

The dependence of the bias on the time difference between the buoy and satellite measurement was also investigated. There was no overall trend in the bias as the time increased.

b. Three-way error analysis

The three-way error analysis was computed for each year from 2003, which is the earliest complete year of AMSR-E data, to 2009. This allows comparison with the results obtained by O'Carroll et al. (2008), who used the same method to investigate the standard deviation of error in AATSR using data from 2003. The three-way error analysis using the 180-min window and same cell collocation was also carried out for subsequent years in order to assess whether there are any trends in the standard deviation of error for any of the instruments.

Table 4 shows the standard deviation of errors calculated for the different instruments for each year between 2003 and 2009. The same error is found for the ARC data in 2003 as was found for the AATSR data in the study of O'Carroll et al. (2008). Similar errors for this year are also found for the AMSR-E SST in both studies while the error for the buoy SST in this report is slightly lower in comparison. The values for the ARC SSTs have the smallest range over the seven years and do not have any obvious trend. However, for the AMSR-E SSTs there seems to be a slight rise in error as the years progress, which could be as a result of the instrument degrading with age. The buoy error falls slightly in the early years then becomes more stable. The Data Buoy Cooperation Panel (DBCP) had a campaign to deploy over 1250 drifting buoys to improve the network and this work was completed in 2005. The gradual introduction of these new buoys may be the cause of the decreasing error.

Table 4.

Standard deviation of error for 2003–09 for the ARC (1-m depth), AMSR-E, and buoy SSTs.

Table 4.

5. Conclusions

The ARC data have been assessed using various statistical methods, which included a three-way error analysis. The results from the statistical analysis have shown a contrast between data using ATSR-1 and the other instruments. It was observed that the earlier years tend to contain larger ARC–buoy differences and show more variation (e.g., the yearly standard deviations were above 0.35 K while for AATSR values from 2003 were all below 0.25 K). At shorter time scales and when considering smaller areas this variation may be as a result of the sparser coverage of the buoys, as suggested by the plots of zonal means for a selection of years. The ATSR-2 period showed a smoother transition between years than ATSR-1 and the AATSR era showed the greatest stability. The large increase in the number of collocations with time contributed greatly to reducing the error associated with the mean values on time scales from a week to a year. Overall, the SSTs using two-channel nighttime retrieval produced the lowest mean bias followed closely by the three-channel retrieval and last the two-channel daytime SSTs; for example, the values for the AATSR period were 0.053, 0.059, and 0.071 K, respectively. However, the standard deviation of the ARC (1-m depth)–buoy SST difference was lowest for the three-channel retrieval such as 0.143 K for the AATSR period compared to 0.159 and 0.157 K for the two-channel night and day retrievals, respectively.

Regional variation was observed for the different retrievals: regions around ±30° latitude tended to show warmer biases. Higher mean biases around these latitudes were also generally observed in the zonal means with more variation in the polar regions caused by small numbers of collocations. Seasonal variation was also found when considering the weekly means in the Northern and Southern Hemisphere with lower biases found in summer months.

The relationship of the bias with other fields such as wind speed was also considered. A small dependence between the wind speed and SST from daytime and nighttime retrievals was found. The two-channel SST bias was also found to increase with TCWV, and the fall in bias at low levels of insolation corresponds to higher-latitude locations where colder biases are more often observed. The estimated uncertainty of the satellite SST retrieval was also considered. At low uncertainty values, it was encouraging to see stable biases with narrow confidence ranges. However, it was shown that for uncertainties greater than around 0.35 K, the biases using the three-channel retrieval could become less reliable while a higher threshold of 0.6 K could be set for the two-channel retrieval during both day and night. Through plotting the standard deviation of the differences used in each 0.02-K bin of uncertainty values an assessment of the uncertainty was made. The desired pattern of a linear increase in standard deviation with uncertainty was seen in both retrievals but at high values of uncertainty (above 0.8 and 0.6 K for the two- and three-channel retrievals, respectively) this correlation did break down.

The impact on the statistics because of the variation in buoy coverage throughout the time period was also important to consider. Until around 1997, the low density and poor global coverage meant that conclusions regarding the quality of the data were more difficult to make. Larger calculated errors, particularly for the ATSR-1 period and for the polar regions in some of the later years, indicated that the mean values were less reliable as a good representation of the true value. The impact on the statistics because of fewer buoys was also evident when considering smaller spatial or temporal resolutions such as weekly mean biases or the zonal averages.

In the three-way error analysis, the ARC data for 2003 were shown to agree with the standard deviation of error of 0.14 K that was found in the previous study of O'Carroll et al. (2008) when using the same criteria. Similar ARC errors were also found when considering subsequent years. The AMSR-E error appears to increase as the years progress; this may be as a result of instrument degradation. The drifting buoy error decreases initially before reaching a more stable value after 2005.

The target accuracy for ARC SSTs is for the bias to be less than 0.1 K, for all regions (assessed on ~1000-km scales). Overall, the statistical analysis has shown the ATSR-2/AATSR period to have a low and stable bias within the limit of 0.1 K while the three-way error analysis has shown that the ARC SSTs from the AATSR era (2003–09) are of the required accuracy, leading to the dataset being of clear benefit to climate trend users. This demonstrates the ARC project has achieved its aim of producing a climate-quality SST dataset that has a consistent processing and no discontinuities between successive satellites. However, the statistical analysis revealed differences between the biases of the early (ATSR-1) and later years. In some cases it was difficult to judge the quality of the ATSR-1 data because of the lack of collocations resulting in larger uncertainties. Some regional and zonal biases also remained throughout as well as slight dependences on wind speed and TCWV.

Acknowledgments

The ARC project is funded jointly by the Natural Environment Research Council (NERC), the Ministry of Defence, and the Department of Environment, Food and Rural Affairs. This support is gratefully acknowledged. The ARC SST data were acquired from the NERC Earth Observation Data Centre and the latest versions can be found at http://neodc.nerc.ac.uk/browse/neodc/arc. AMSR-E data are produced by Remote Sensing Systems and sponsored by the NASA Earth Science MEaSUREs DISCOVER Project and the AMSR-E Science Team. Data are available at www.remss.com.

REFERENCES

  • Balmaseda, M. A., M. K. Davey, and D. L. Anderson, 1995: Decadal and seasonal dependence of ENSO prediction skill. J. Climate, 8, 27052715.

    • Search Google Scholar
    • Export Citation
  • Casey, K. S., and P. Cornillon, 2001: Global and regional sea surface temperature trends. J. Climate, 14, 38013818.

  • Dee, D. P., and Coauthors, 2011: The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553597, doi:10.1002/qj.828.

    • Search Google Scholar
    • Export Citation
  • Donlon, C. J., M. Martin, J. D. Stark, J. Roberts-Jones, E. Fiedler, and W. Wimmer, 2012: The Operational Sea Surface Temperature and Sea Ice Analysis (OSTIA). Remote Sens. Environ., 116, 140158.

    • Search Google Scholar
    • Export Citation
  • Elsner, J. B., J. P. Kossin, and T. H. Jagger, 2008: The increasing intensity of the strongest tropical cyclones. Nature, 455, 9295, doi:10.1038/nature07234.

    • Search Google Scholar
    • Export Citation
  • Embury, O., and C. J. Merchant, 2012: A reprocessing for climate of sea surface temperature from the along-track scanning radiometers: A new retrieval scheme. Remote Sens. Environ., 116, 4761.

    • Search Google Scholar
    • Export Citation
  • Embury, O., C. J. Merchant, D. Berry, and E. Kent, 2011: A reprocessing for climate of sea surface temperature from the along-track scanning radiometers. Proc. 2011 EUMETSAT Meteorological Satellite Conf., Oslo, Norway, EUMETSAT, P.59.

  • Embury, O., C. J. Merchant, and G. K. Corlett, 2012: A reprocessing for climate of sea surface temperature from the along-track scanning radiometers: Initial validation, accounting for skin and diurnal variability. Remote Sens. Environ., 116, 6278, doi:10.1016/j.rse.2011.02.028.

    • Search Google Scholar
    • Export Citation
  • Horrocks, L. A., A. R. Harris, and R. W. Saunders, 2003: Modelling the diurnal thermocline for daytime bulk SST from AATSR. Met Office Forecasting Research Tech. Rep. 418, 27 pp. [Available online at http://www.metoffice.gov.uk/media/pdf/i/k/FRTR418.pdf.]

  • Kantha, L. H., and C. A. Clayson, 1994: An improved mixed layer model for geophysical applications. J. Geophys. Res., 99 (C12), 25 23525 266.

    • Search Google Scholar
    • Export Citation
  • Kennedy, J. J., N. A. Rayner, R. O. Smith, D. E. Parker, and M. Saunby, 2011a: Reassessing biases and other uncertainties in sea surface temperature observations measured in situ since 1850: 1. Measurement and sampling uncertainties. J. Geophys. Res., 116, D14103, doi:10.1029/2010JD015218.

    • Search Google Scholar
    • Export Citation
  • Kennedy, J. J., N. A. Rayner, R. O. Smith, D. E. Parker, and M. Saunby, 2011b: Reassessing biases and other uncertainties in sea-surface temperature observations measured in situ since 1850: 2. Biases and homogenisation. J. Geophys. Res., 116, D14104, doi:10.1029/2010JD015220.

    • Search Google Scholar
    • Export Citation
  • Kilpatrick, K. A., G. P. Podestá, and R. Evans, 2001: Overview of the NOAA/NASA advanced very high resolution radiometer Pathfinder algorithm for sea surface temperature and associated matchup database. J. Geophys. Res., 106 (C5), 91799197.

    • Search Google Scholar
    • Export Citation
  • Lea, D. W., 2004: The 100 000-yr cycle in tropical SST, greenhouse forcing, and climate sensitivity. J. Climate, 17, 21702179.

  • Llewellyn-Jones, D., and J. Remedios, 2012: The Advanced Along Track Scanning Radiometer (AATSR) and its predecessors ATSR-1 and ATSR-2: An introduction to the special issue. Remote Sens. Environ., 116, 13.

    • Search Google Scholar
    • Export Citation
  • Llewellyn-Jones, D., M. C. Edwards, C. T. Mutlow, A. R. Birks, I. J. Barton, and H. Tait, 2001: AATSR: Global-change and surface-temperature measurements from Envisat. ESA Bulletin, No. 105, Ecological Society of America, Washington, DC, 10–21. [Available online at http://www.esa.int/esapub/bulletin/bullet105/bul105_1.pdf.]

  • Mackie, S., C. J. Merchant, O. Embury, and P. Francis, 2010: Generalized Bayesian cloud detection for satellite imagery. Part 2: Technique and validation for daytime imagery. Int. J. Remote Sens., 31, 25952621, doi:10.1080/01431160903051711.

    • Search Google Scholar
    • Export Citation
  • Merchant, C. J., A. R. Harris, E. Maturi, and S. Maccallum, 2005: Probabilistic physically based cloud screening of satellite infrared imagery for operational sea surface temperature retrieval. Quart. J. Roy. Meteor. Soc., 131, 27352755.

    • Search Google Scholar
    • Export Citation
  • Merchant, C. J., and Coauthors, 2008: Deriving a sea surface temperature record suitable for climate research from the along-track scanning radiometers. J. Adv. Space Res., 41, 111, doi:10.1016/j.asr.2007.07.041.

    • Search Google Scholar
    • Export Citation
  • Merchant, C. J., and Coauthors, 2012: A twenty-year independent record of sea surface temperature for climate from Along-Track Scanning Radiometers. J. Geophys. Res., 117, C12013, doi:10.1029/2012JC008400.

    • Search Google Scholar
    • Export Citation
  • O'Carroll, A. G., R. W. Saunders, and J. G. Watts, 2006a: The measurement of the sea surface temperature by satellites from 1991 to 2005. J. Atmos. Oceanic Technol., 23, 15731582.

    • Search Google Scholar
    • Export Citation
  • O'Carroll, A. G., J. G. Watts, L. A. Horrocks, R. W. Saunders, and N. A. Rayner, 2006b: Validation of the AATSR Meteo Product sea surface temperature. J. Atmos. Oceanic Technol., 23, 711726.

    • Search Google Scholar
    • Export Citation
  • O'Carroll, A. G., J. R. Eyre, and R. W. Saunders, 2008: Three-way error analysis between AATSR, AMSR-E, and in situ sea surface temperature observations. J. Atmos. Oceanic Technol., 25, 11971207.

    • Search Google Scholar
    • Export Citation
  • Rayner, N. A., D. E. Parker, E. B. Horton, C. K. Folland, L. V. Alexander, D. P. Rowell, E. C. Kent, and A. Kaplan, 2003: Global analyses of sea surface temperature, sea ice, and night marine air temperature since the late nineteenth century. J. Geophys. Res., 108, 4407, doi:10.1029/2002JD002670.

    • Search Google Scholar
    • Export Citation
  • Smith, D. L., J. Delderfield, D. Drummond, T. Edwards, C. T. Mutlow, P. D. Read, and G. M. Toplis, 2001: Calibration of the AATSR instrument. Adv. Space Res., 28, 3139.

    • Search Google Scholar
    • Export Citation
  • Stoffelen, A., 1998: Toward the true near-surface wind speed: Error modeling and calibration using triple collocation. J. Geophys. Res., 103 (C4), 77557766.

    • Search Google Scholar
    • Export Citation
  • Trenberth, K. E., and Coauthors, 2007: Observations: Surface and atmospheric climate change. Climate Change 2007: The Physical Science Basis, S. Solomon et al., Eds., Cambridge University Press, 235–336.

  • Veal, K. L., G. K. Corlett, D. Ghent, D. T. Llewellyn-Jones, and J. J. Remedios, 2013: A time series of mean global skin-SST anomaly using data from ATSR-2 and AATSR. Remote Sens. Environ., 135, 6476.

    • Search Google Scholar
    • Export Citation
  • Wentz, F., and T. Meissner, 2000: Algorithm Theoretical Basis Document (ATBD), version 2. Remote Sensing Systems, 66 pp.

  • Wentz, F., C. Gentemann, and P. Ashcroft, 2003: On-orbit calibration of AMSR-E and the retrieval of ocean products. Preprints, 12th Conf. on Satellite Meteorology and Oceanography, Long Beach, CA, Amer. Meteor. Soc., P59. [Available online at http://ams.confex.com/ams/pdfpapers/56760.pdf.]

  • Woodruff, S. D., and Coauthors, 2011: ICOADS release 2.5: Extensions and enhancements to the surface marine meteorological archive. Int. J. Climatol., 31, 951967, doi:10.1002/joc.2103.

    • Search Google Scholar
    • Export Citation
  • Závody, A. M., C. T. Mutlow, and D. T. Llewellyn-Jones, 1995: A radiative transfer model for sea surface temperature retrieval from the along-track scanning radiometer. J. Geophys. Res., 100 (C1), 937952.

    • Search Google Scholar
    • Export Citation
Save