Previous - 4 Presentation of Uncertainty Index Next - 6 Satellites
Alternative approaches to using the SST data in a way that is less sensitive to biases and other data errors have been made. The following approaches make use of knowledge concerning the types of errors and uncertainties found in SST data and have been adapted to account for them. They highlight the importance of combining understanding of the measurements and their potential errors, as well as understanding of the phenomenon being analyzed. Perhaps the simplest example is Schell [1959] who suggested discarding grid-box averages (in that case Marsden squares) based on small numbers of observations.
Thompson et al. [2008] identified an abrupt drop in the observed global average SST anomaly in late 1945, which they attributed to a rapid change in the composition of ICOADS 2.0 [Worley et al., 2005] from mostly US ships immediately before the 1945 drop to mostly UK ships immediately afterwards. This hypothesis was lent further weight by Kennedy et al. [2011c]. In a follow-up paper [Thompson et al., 2010], a drop in northern-hemisphere SSTs was identified. In order to show that the drop was not an artifact of the change in measurement method, they divided the ICOADS data into distinct subsets based on the country of the ships making the measurements, considered a range of different SST analyses, and looked at related variables such as NMAT and land surface air temperatures. The probability of a drop being due to a coincident change in the way that all countries measured SST, simultaneous with a sudden change in NMAT and land temperature bias, is small. The fact that the drop was seen in all the different data sets implied that the drop was real. Tokinaga et al. [2012] took a similar approach, using bucket measurements from ICOADS as a quasi-homogeneous estimate of SST change over the period 1950 to 2009.
In detection and attribution studies it is common to reduce the coverage of the models to match that of the data. Doing so reduces the exposure of the study to uncertainties associated with interpolation techniques, but it does not avoid the problem of systematic biases. Recent studies [Jones and Stott, 2011] have explicitly used a range of data sets to start to map out the effects of structural uncertainties on detection and attribution studies.
SST data sets are routinely compared to the output of climate simulations. Bearing in mind the discussion in section 2 on the definition of SST it might be necessary to ensure that the modeled output and the measured SST correspond to the same quantity. Many climate models employ a surface mixed layer that is several meters thick. However, models have been run with greater resolution in the near-surface ocean [e.g., Bernie et al., 2008] in order to simulate diurnal variability.
Another common use of SST data for which an understanding of the limitations of the data is important is in the calculation and interpretation of EOFs. In many studies EOFs are calculated from globally complete SST analyses because the lack of missing data makes calculating EOFs easy. However, it seems wise to bear in mind that a good deal of statistical processing has already been applied to the SST analyses to make them globally complete. Extracting EOFs from (or applying any other analysis technique to) what are in some cases EOF analyses already, could lead to difficulties of interpretation on top of the more general problems [Hannachi et al., 2007; Dommenget, 2007; Karnauskas, 2013]. Techniques exist for estimating EOFs from gridded data sets with missing data and these can also incorporate uncertainty information though many assume that the errors are uncorrelated and will tend to underestimate uncertainty in the EOFs and their principal components. See for example, Roweis [1998], Schneider [2001], Beckers and Rixen [2003], Rutherford et al. [2004], Houseago-Stokes and Challenor [2004], Kondrashov and Ghil [2006], Ilin and Kaplan [2009] and Luttinen and Ilin [2009].
Previous - 4 Presentation of Uncertainty Index Next - 6 Satellites