Guest Post by Willis Eschenbach
A while back, I noticed an oddity about the Hadley Centre’s HadISST sea ice dataset for the Arctic. There’s a big change in variation from the pre- to the post-satellite era. Satellite measurements of ice areas began in 1979. Here is the full HadISST record, with the monthly variations removed.
Figure 1. Anomaly in the monthly sea ice coverage as reported by the HadISST, GISST, and Reynolds datasets. All data are from KNMI. Monthly average variations from the overlap period (1981-1994) have been subtracted from each dataset. All data are from KNMI (see Monthly Observations).
There’s a few points of note. First, the pre-1953 data is pretty useless, much of it is obviously not changing from year to year. Second, although the variation in the GISST dataset is doesn’t change in 1979, the variation in the HadISST dataset changes pretty radically at that point. Third, there is a large difference between the variability of the Reynolds and the GISST datasets during the period of their overlap.
I had filed this under unexplained curiosities and forgotten about it … until the recent publication of a paper called Observations reveal external driver for Arctic sea-ice retreat, by Notz and Marotzke, hereinafter N&M2012
Why did their paper bring this issue to the fore?
Well, the problem is that the observations they use to establish their case are the difference in variability of the HadISST during period 1953-1979, compared to the HadISST variations since that time. They look at the early variations, and they use them as “a good estimate of internal variability”. I have problems with this assumption in general due to the short length of time (25 years), which is way too little data to establish “internal variability” even if the data were good … but it’s not good, it has problems.
To their credit, the authors recognized the problems in N&M2012, saying:
Second, from 1979 onwards the HadISST data set is primarily based on satellite observations. We find across the 1978/1979 boundary an unusually large increase in sea-ice extent in March and an unusually large decrease in sea-ice extent in September (Figures 1b and 1d). This indicates a possible inconsistency in the data set across this boundary.
Ya think? I love these guys, “possible inconsistency”. The use of this kind of weasel words. like “may” and “might” and “could” and “possible”, is Cain’s mark on the post-normal scientist. Let me remove the GISST and Reynolds datasets and plot just the modern period that they use, to see if you can spot their “possible inconsistency” between the 1953-1979 and the post 1979 periods…
The inconsistency is clearly visible, with the variability of the pre- and post-1979 periods being very different.
As a result, what they are doing is comparing apples and oranges. They are assuming the 1953-1979 record is the “natural variability”, and then they are comparing that to the variability of the post-1979 period … I’m sorry, but you just can’t do that. You can’t compare one dataset with another when they are based on two totally different types of measurements, satellite and ground, especially when there is an obvious inconsistency between the two.
In addition, since the GISST dataset doesn’t contain the large change in variability seen in the HadISST dataset, it is at least a working assumption that there is some structural error in the HadISST dataset … but the authors just ignore that and move forwards.
Finally, we have a problematic underlying assumption that involves something called “stationarity”. The stationarity assumption says that the various statistical measures (average, standard deviation, variation) are “stationary”, meaning that they don’t change over time.
They nod their heads to the stationarity problem, saying (emphasis mine):
For the long-term memory process, we estimate the Hurst coefficient H of the pre-satellite time series using detrended fluctuation analysis (DFA) [Peng et al., 1994]. Only a rough estimate of 0.8 < H < 0.9 is possible both because of the short length of the time series and because DFA shows non-stationarity even after removal of the seasonal cycle.
Unfortunately, they don’t follow the problem of non-stationarity to its logical conclusion. Look, for example, at the variability in the satellite record in the period 1990-2000 versus the period 2000-2005. They are quite different. In their analysis, they claim that a difference in variability pre- to post-1979 establishes that human actions are the “external driver” … but they don’t deal with the differences pre- and post-2000, or with the fact that their own analysis shows that even the variability of the pre-1979 data is not stationary.
Finally, look at the large change in variability in the most recent part of the record. The authors don’t mention that … but the HadISST folks do.
03/DECEMBER/2010. The SSM/I satellite that was used to provide the data for the sea ice analysis in HadISST suffered a significant degradation in performance through January and February 2009. The problem affected HadISST fields from January 2009 and probably causes an underestimate of ice extent and concentration. It also affected sea surface temperatures in sea ice areas because the SSTs are estimated from the sea ice concentration (see Rayner et al. 2003). As of 3rd December 2010 we have reprocessed the data from January 2009 to the present using a different sea ice data source. This is an improvement on the previous situation, but users should still note that the switch of data source at the start of 2009 might introduce a discontinuity into the record. The reprocessed files are available from the main data page. The older version of the data set is archived here.
08/MARCH/2011. The switch of satellite source data at the start of 2009 introduced a discontinuity in the fields of sea ice in both the Arctic and Antarctic.
Curious … the degradation in the recent satellite data “probably causes an underestimate of ice extent and concentration,” and yet it is precisely that low recent ice concentration that they claim “reveals an external driver” …
In any case, when I put all of those problems together, the changes in variability in 1979, in 2000, and in 2009, plus the demonstrated non-stationarity pre-1979, plus the indirect evidence from the GISST and Reynolds datasets, plus the problems with the satellites affecting the critical recent period, the period they claim is statistically significant in their analysis … well, given all that I’d say that the N&M2012 method (comparing variability pre- and post-1979) is totally inappropriate given the available data. There are far too many changes and differences in variability, both internal to and between the datasets, to claim that the 1979 change in variability means anything at all … much less that it reveals an “external driver” for the changes in Arctic sea ice.