Importance of error sources in climate data on different analysis scales
This figure is an example of the relative importance of different classes of errors to the uncertainty in the analysis of sea surface temperature changes when that analysis is performed on different space-time scales. The sea surface temperature (SST) data in this scenario are a climate data record (CDR) obtained from a series of meteorological sensors.
When looking at SST changes over short time scales and small areas, noise in the CDR is that largest contribution to uncertainty. Noise errors are random in nature, and are independent between individual SST data. This means that over larger areas and/or longer times, the noise averages down fastest.
The second most important term on the left side of the graph are "locally correlated" errors. These typically arise from ambiguity in SST retrieval connected to the state of the atmosphere. The atmospheric state evolves over the course of days and is often similar across distances much longer than a satellite pixel, hence the local correlation in time and space of associated errors. Although in the long run these errors are random, the local correlations mean they average less than random noise when going to larger scales of analysis.
The calibration of a good instrument degrades in quality slowly over many years, and in time this introduces increasing uncertainty, since we don't know how the calibration errors are changing. Calibration errors tend to be systematic. For a long data set, usually several sensors have to be used. Even if harmonisation steps are taken to make the different sensors' SSTs consistent, there is still an uncertainty in harmonisation which increases as more sensors are used in sequence. This source of uncertainty becomes significant when looking at SST evolution over a decade or more.
Although SST has been used as an example here, the general principle at play applies to most, perhaps all, CDRs: the apparently subtle systematic effects become the dominant sources of uncertainty in long-term climate analyses of the observations.
To create this figure, reasonable assumptions for the case of SST are made. These are detailed in the python script used to create the figure, which is also available on figshare.
