The vast majority of historical (and modern) weather observations were not made explicitly for climate monitoring purposes. Measurements have changed in nature as demands on the data, observing practices and technologies have evolved. These changes almost always alter the characteristics of observational records, changing their mean, their variability or both, such that it is necessary to process the raw measurements before they can be considered useful for assessing the true climate evolution. This is true of all observing techniques that measure physical atmospheric quantities. The uncertainty in observational records encompasses instrumental/ recording errors, effects of representation (e.g., exposure, observing frequency or timing), as well as effects due to physical changes in the instrumentation (such as station relocations or new satellites). All further processing steps (transmission, storage, gridding, interpolating, averaging) also have their own particular uncertainties. Because there is no unique, unambiguous, way to identify and account for non-climatic artefacts in the vast majority of records, there must be a degree of uncertainty as to how the climate system has changed. The only exceptions are certain atmospheric composition and flux measurements whose measurements and uncertainties are rigorously tied through an unbroken chain to internationally recognized absolute measurement standards (e.g., the CO2 record at Mauna Loa; Keeling et al., 1976a).
Uncertainty in data set production can result either from the choice of parameters within a particular analytical framework—parametric uncertainty, or from the choice of overall analytical framework— structural uncertainty. Structural uncertainty is best estimated by having multiple independent groups assess the same data using distinct approaches. More analyses assessed now than in AR4 include published estimates of parametric or structural uncertainty. It is important to note that the literature includes a very broad range of approaches. Great care has been taken in comparing the published uncertainty ranges as they almost always do not constitute a likefor- like comparison. In general, studies that account for multiple potential error sources in a rigorous manner yield larger uncertainty ranges. This yields an apparent paradox in interpretation as one might think that smaller uncertainty ranges should indicate a better product. However, in many cases this would be an incorrect inference as the smaller uncertainty range may instead reflect that the published estimate considered only a subset of the plausible sources of uncertainty. Within the timeseries figures, where this issue would be most acute, such parametric uncertainty estimates are therefore not generally included. Consistent with AR4 HadCRUT4 uncertainties in GMST are included in Figure 2.19, which in addition includes structural uncertainties in GMST.
To conclude, the vast majority of the raw observations used to monitor the state of the climate contain residual non-climatic influences. Removal of these influences cannot be done definitively and neither can the uncertainties be unambiguously assessed. Therefore, care is required in interpreting both data products and their stated uncertainty estimates. Confidence can be built from: redundancy in efforts to create products; data set heritage; and cross-comparisons of variables that would be expected to co-vary for physical reasons, such as LSATs and SSTs around coastlines. Finally, trends are often quoted as a way to synthesize the data into a single number. Uncertainties that arise from such a process and the choice of technique used within this chapter are described in more detail in Box 2.2.