lizkent
commented
6 years ago One major source of "mistakes" (I think by the definitions above) is the mislocation of observations in space and time. Ship tracking (Carella et al. 2017) has proved valuable in identifying mispositioned data, as has the preliminary inventory produced under Copernicus C3S_311a, Lot 2 which shows maps of observations by ICOADS "DCK" and "SID".
ET-NCMP
The text doesn't talk much about quality control and currently mixes up "mistakes" with "errors".
Mistakes might be characterised as values that are unrelated to the true SST (a missing value flag like -99, say), or related to it in some abstract way like the transposition of two digits in the value (27.8 become 72.8 for example) or a missing minus sign.
Errors as "defined" in the general classification of uncertainties are perturbations away from the true SST caused by the physical circumstances and limitations of the measurement. There are errors in all measurements.
Quality Control (QC) aims to minimise the presence of mistakes, while preserving errors. There are all sorts of complications here that relate to uncertainty. First, it's not always possible to separate mistakes and errors so there will always be the possibility that the distribution you are looking at is a mixture of data that have and don't have a relationship with the true SST. Second, QC can affect long-term trends (Huang et al. ERSSTv4/5) so it needs to be figured in separately to capture that aspect. Third, QC affects the assessed uncertainties. if you trim off the tails of a distribution, you limit the standard deviation and hence the estimated uncertainty (if you are estimating it that way). In satellite validation, standard deviations have sometimes been calculated using trimmed distributions, robust fitting method, best-estimate gaussian etc. which discard outliers and hence give a too-small estimate of the uncertainty.