On the Monumental Differences in Warming Rates between Global Sea Surface Temperature Datasets during the NOAA-Picked Global-Warming Hiatus Period of 2000 to 2014

UPDATE:  See the 2 updates under the heading of A QUICK OVERVIEW OF SHIP-BUOY BIAS ADJUSTMENTS.

# # #

Alternate Title:  NOAA Has a Sea Surface Temperature Dataset with an EVEN HIGHER Warming Rate than Their Full-of-Problems ERSST.v4 “Pause-Buster” Data

Whether or not there had been a slowdown in global surface warming before the El Niño of 2015/16 depends on which sea surface temperature dataset researchers elect to use in studies. Even over the full term of the satellite-era of sea surface temperature data, the differences in warming rates can be quite large.

Figure 1 is a time-series graph of six global sea surface temperature datasets for the satellite-era of November 1981 to November 2015. (November 1981 is the start month of the original version of NOAA’s Reynolds OI.v2 satellite-enhanced data, and, as of this writing, the HadISST data from the UKMO have only been updated through November 2015.) I’ve also shown the trend lines for the datasets with the highest and lowest warming rates. The HadISST dataset from the UK Met Office (UKMO) is the sea surface temperature dataset that’s used most often in research papers. Of the 6 datasets presented, it has the lowest warming rate over the past 34 years.  At the other end of the spectrum is NOAA’s high-resolution (1/4 deg), daily version of NOAA’s Optimum Interpolation sea surface temperature data (a.k.a. Reynolds OI.v2).  It is presented at websites like the University of Maine’s Climate Reanalyzer and used in products where daily sea surface temperatures are needed. (That version of the Reynolds OI.v2 is NOT the dataset I present in my monthly sea surface temperature updates. More on the two versions of Reynolds OI.v2 SST data in a moment.)  And as you’ll see shortly, the differences in warming rates of those 6 datasets are even greater (slightly) during the NOAA-selected global-warming hiatus period of 2000 to 2014.

Figure 1

But first, Figure 2 shows the spread between those 6 sea surface temperature datasets. The anomalies are all referenced to the WMO-preferred period of 1981-2010, almost the full term, so not to bias the results. And the “global” data were limited to the latitudes of 60S-60N, excluding the polar oceans, because the data suppliers account for sea ice differently.  The monthly minimum and maximum values for the 6 datasets were first determined.  Then the spread was calculated by subtracting the monthly minimums from the monthly maximums. Also shown in maroon is the spread smoothed with a 12-month running-mean filter.

Figure 2

Even before the early 2000s, when the number of buoy-based measurements skyrocketed, the spreads between sea surface temperature datasets are quite large. Keying off the smoothed data, the spread varied between 0.05 deg C and 0.1 deg C from the start until 2004. Then there was an upward shift in 2005, and after that shift, the spread cycled near 0.1 deg C.  There was another obvious shift in the spread in 2013. The spread between sea surface temperature datasets is now cycling as high as 0.15 deg C.

For a global-warming hiatus period, NOAA used the period of 2000 to 2014 in two recent papers:

• Karl et al. (2015) Possible artifacts of data biases in the recent global surface warming hiatus, and
• Huang et al (2015) Further Exploring and Quantifying Uncertainties for Extended Reconstructed Sea Surface Temperature (ERSST) Version 4 (v4)

As noted earlier, the trend difference is slightly greater for the NOAA-selected global-warming hiatus period of 2000 to 2014.  See Figure 3. Once again, NOAA’s high-resolution (1/4 deg) version of NOAA’s Optimum Interpolation (Reynolds OI.v2) sea surface temperature data is showing the highest warming rate.  (Does anyone wonder why alarmists love that dataset?) But this time, the dataset with the lowest warming rate is NOAA’s ERSST.v3b, which, oddly enough, is still being updated by NOAA even though it was replaced by NOAA’s ERSST.v4 “pause-buster” data.

Figure 3

I suspect some readers are imagining that the differences between warming rates have to do with the ship-buoy bias adjustments—that some of the datasets include the ship-buoy bias adjustment and others don’t. You may be surprised to discover two of the sea surface temperature datasets, one with and one without ship-buoy bias adjustments, have basically the same warming rates for the period of 2000 to 2014.  More on that later.

A QUICK OVERVIEW OF SEA SURFACE TEMPERATURE BIAS ADJUSTMENTS

Sea surface temperatures have been measured using a number of different technologies over the years. At first (and continuing to this day), buckets of different types were tossed over the sides of ships, then hauled back aboard, where sailors would place thermometers in the water-filled buckets. Depending on the air temperatures on deck, the bucket-based temperature measurements could be biased cool.  Buckets were the sole method used before the 1930s. Then ship-based engine room inlets (ERI) began to be used to sample sea surface temperatures, so there was a mix of measurements from buckets and ship-inlets from the 1930s to the 1970s. Buoys have also been used to sample ocean surface temperatures since the 1970s, with a very large increase in buoy-based observations starting in the early 2000s from drifting buoys (they are not ARGO floats). So even today there is a mix of sampling methods from buoys, ship inlets and buckets.  The dominant sampling method has varied with time.  See Figure 2 here from Kennedy et al. (2011) Reassessing biases and other uncertainties in sea-surface temperature observations measured in situ since 1850, part 2: biases and homogenization. Because those sampling methods have biases toward one another, data suppliers adjust the source data. The impacts of those adjustments vary with time depending on the mix of measurement technologies.

For the period discussed in this post (2000 to 2014), the ship-buoy bias adjustments are said to play a major role.  Unfortunately, the uncertainties of the ship-buoy bias are extremely high.

My Table 1 is Table 5 from Kennedy et al. (2011) Reassessing biases and other uncertainties in sea-surface temperature observations measured in situ since 1850, part 2: biases and homogenization.

Table 1 – Table 5 from Kennedy et al. (2011)

As listed, for the global oceans, researchers have found that there is a ship-buoy bias of 0.12 deg C with a standard deviation of 0.85 deg C…the buoys reading cooler than the ship inlets.  Let’s rewrite that bias in terms that you may be more familiar with: It’s 0.12 deg C +/- 1.7 deg C.  In other words, the uncertainty of the global ship-buoy bias is an order of magnitude greater than the observed bias.

I’ve seen a climate scientist reframe those uncertainties in a blog comment.  Unfortunately, I can’t recall when or where.  (If you know of that comment, please link it on this thread and I will include it here.) Regardless of how they are framed, the uncertainties in the ship-buoy bias still exist and they are quite large.

UPDATE 1:  Nick Stokes writes in the comment here on the thread at WUWT:

This is Ross McKitrick’s bungle. It is just wrong. 0.85 is the standard deviation (SD) of the pairings that go to make up the average. SE is the standard error of the mean – the figure 0.12 that you are quoting. That is basic stats. It’s 0.12 +/- 0.01.

[End Update 1]

The uncertainties of the ship-buoy bias are so great that researchers in the early 2000s didn’t bother to account for them.  But as the 21^st Century unfolded and the slowdown in global warming became more and more evident as time passed, the researchers began to search for excuses for the slowdown and began to search for ways to increase the warming rate for the post-1997/98 El Niño period. So they blamed the ship-buoy bias and began to adjust sea surface temperature datasets for it.  The most recent sea surface temperature data supplier to do so was NOAA with their ERSST.v4 “pause-buster” data.

UPDATE 2: To confirm my statement The uncertainties of the ship-buoy bias are so great that researchers in the early 2000s didn’t bother to account for them… see Reynolds et al. (2002), where they write:

We have not corrected the OI.v2 in situ data by the factors in Table 2 because of the uncertainties of the biases in the table. However, any correction of satellite data is further complicated by in situ biases and their uncertainties.

[End Update 2.]

DATASETS PRESENTED

Of the 6 sea surface temperature datasets presented in this post, 3 have been adjusted for ship-buoy bias and 3 have not.     Let’s start with the datasets that have been adjusted. They include:

• NOAA ERSST.v4 (NOAA’s recently introduced “pause buster” data, infilled, in situ only)
• UKMO HADSST3 (UK Met Office product, not infilled, in situ only)
• NOAA Optimum Interpolation/Reynolds OI.v2 – High Resolution (NOAA’s satellite-enhanced data/AVHRR-only, ¼-deg resolution, infilled, presented daily) KNMI recently added it to their Climate Explorer.

Notes: The notation “in situ only” means the dataset includes only observations from ships (buckets and ship inlets) and from buoys (moored and drifting). The “satellite-enhanced” datasets also include in situ observations and the satellite-based data are also bias adjusted with the in situ data. “Infilled” means that data suppliers use statistical devices to create data for ocean grids without observations, providing a dataset that, seemingly, is spatially complete. [End notes.]

The datasets that have not been adjusted for ship-buoy biases are:

• NOAA ERSST.v3b (NOAA’s former dataset, infilled, in situ only)
• NOAA Optimum Interpolation/Reynolds OI.v2 – “Original” (NOAA’s original satellite-enhanced data, 1-deg resolution, infilled, presented weekly and monthly) I’ve listed it as “Std” for standard resolution and standard timeframes in the title blocks on the graphs. This is the Reynolds OI.v2 data I use in my monthly sea surface temperature updates.
• UKMO HADISST (UKMO’s interpolated/infilled product, satellite-enhanced)

A REFERENCE ILLUSTRATION

Last month a paper was published about the uncertainties of the new NOAA ERSST.v4 “pause-buster” sea surface temperature data.  That paper is Huang et al. (2015b) Further Exploring and Quantifying Uncertainties for Extended Reconstructed Sea Surface Temperature (ERSST) Version 4 (v4). (Preliminary accepted version is here.)

My Figure 4 is Figure 6 from Huang et al. (2015b).  It includes histograms of trend uncertainties that were determined from the model used to calculate the new NOAA ERSST.v4 data for three periods: 1901 to 2014, 1951 to 2012, and 2000 to 2014.

Figure 4 – (Figure 6 from Huang et al. (2015b))

The trend uncertainties presented in their Figure 6 are “parametric uncertainties”. More on that topic later in the post.

A BRIEF EXCURSION TO THE PERIOD OF 1951-2012

One of the curiosities illustrated and discussed in the recent posts (here and here) was how the trends of NOAA’s new ERSST.v4 “pause-buster” sea surface temperature data resided at or toward the high ends of the uncertainty ranges for the periods of 1951 to 2012 and 2000 to 2014. See the illustration here from the post The Oddities in NOAA’s New “Pause-Buster” Sea Surface Temperature Product – An Overview of Past Posts.

But there’s another curiosity in that illustration. Note how Figure 6 from Huang et al. (2015b) includes a histogram for the UKMO HADSST3 data, but only for the period of 1901 to 2014, Cell a. The obvious intent was to show the similarities between the two datasets for that time period.  That raises a question: Why did NOAA exclude the histograms for the HADSST3 data during the other two periods?  They can’t include it in one and not the others and not draw attention to the fact that it’s missing from others.

We illustrated and discussed in the post Busting (or not) the mid-20th century global-warming hiatus how the UKMO adjusted their HADSST3 data for the 1945 discontinuity presented in Thompson et al (2008) and for the trailing biases, while NOAA had not.

Figure 5

As a result of NOAA’s failure to make those adjustments, the new global NOAA ERSST.v4 data have a noticeably higher warming rate (+0.099 deg C/decade) than the UKMO HADSST3 data (+0.076 deg C/decade) for the period of 1951-2012.  I’ll let you speculate about why NOAA did not include the histogram for the trends of the HADSST3 data for that period in Figure 6, Cell b from Huang et al. (2015b).

Note:  My Figure 5 above is similar to the bottom graph in Figure 13 from the post Busting (or not) the mid-20th century global-warming hiatus.  I’ve ended the data in 2012 in Figure 5 above to agree with the timeframe used by NOAA in Huang et al. (2015b). [End note.]

But what about the period of 2000 to 2014 shown in Cell c of Figure 6 from Huang et al.?  Where does the UKMO HADSST3 data fit in that period? I’ve included Cell c from Huang et al. (2015b) in the next two illustrations for illustration and discussion purposes.

COMPARISON OF DATASETS WITH SHIP-BUOY BIAS ADJUSTMENTS

The top illustration in Figure 6 is a time-series graph that includes the global sea surface temperature anomalies for the NOAA ERSST.v4 “pause-buster” data, the NOAA Reynolds OI.v2 (high resolution, daily version) satellite-enhanced dataset, and the UKMO HADSST3 data. It covers the NOAA-selected hiatus period of 2000 to 2014.  All three datasets have been adjusted for ship-buoy biases.  Quite remarkably, the trends range from +0.054 deg C/decade for the HADSST3 data to +0.131 deg C/decade for the NOAA Reynolds OI.v2 (high resolution version) data, with the NOAA “pause-buster” data coming in at +0.097 deg C/decade.

Figure 6

Referring to the trend histogram, we can see that the +0.131 deg C/decade trend for the NOAA Reynolds OI.v2 (high resolution version) data is so high it’s out of the range of trend uncertainties for NOAA’s latest and greatest ERSST.v4 data. (Off the chart, not even close.)

We can also see that the trend of the HADSST3 data for the period of 2000 to 2014 resides in the lower half of the ERSST.v4 trend uncertainty range. Once again, I’ll let you speculate about why NOAA did not include the histogram of the HADSST3 trend uncertainties in Figure 6, Cell c from Huang et al. (2015b).

COMPARISON OF DATASETS WITHOUT SHIP-BUOY BIAS ADJUSTMENTS

Figure 7 includes NOAA’s original Reynolds OI.v2 satellite-enhanced data, the NOAA ERSST.v3b in situ-only data, and the UKMO HadISST satellite-enhanced data.  These 3 datasets have not been adjusted for ship-buoy biases. Their trends for the period of 2000 to 2014 are clustered much more closely together than the datasets that have been adjusted for ship-buoy biases. The trends of the datasets that haven’t been adjusted for ship-buoy bias range from +0.039 deg C/decade for the NOAA ERSST.v3b data to +0.052 deg C/decade for the original NOAA Reynolds OI.v2 data, with the UKMO HadISST data between them at 0.046 deg C/decade.

Figure 7

For all three of the datasets without the ship-buoy bias adjustments, we can also see that the trends for the period of 2000 to 2014 fit within the range of trend uncertainties that NOAA determined for their “pause-buster” ERSST.v4 data. But instead of the trends residing up at the high end of range like NOAA’s “pause-buster” data, these three datasets without the ship-buoy bias adjustment have trends below the average.  Some readers might believe the datasets without the ship-buoy bias adjustments provide conservative estimates of the warming rate from 2000 to 2014, where the trend of the NOAA “pause-buster” data is far from conservative, more in the realm of extremism.

Referring to the histograms in Figures 6 and 7, there is only one sea surface temperature dataset that falls outside of the range that NOAA determined for their ERSST.v4 data, and it is the high-resolution version of the NOAA Reynolds OI.v2 data. I believe we can treat that version of NOAA’s Reynolds OI.v2 data as an outlier and also treat it as an unrealistic product for global warming presentations.

We can also see in Figures 6 and 7 that the 2000-2014 trend of the UKMO HADSST3 (+0.054 deg C/decade), which has been adjusted for ship-buoy bias, is basically the same as the trend of the standard NOAA Reynolds OI.v2 satellite-enhanced data (+0.052 deg C/decade), which has not been adjusted for ship-buoy bias.

THE OUTLIER’S IMPACT ON THE SPREAD

Figure 8 is the same as Figure 2, except that I’ve excluded the outlying high-trend, high-resolution, daily version of the Reynolds OI.v2 data from NOAA.

Figure 8

The upward shifts in the spread in 2005 and 2013 no longer exist when we exclude the outlying version of NOAA’s Reynolds OI.v2 data (that’s favored by alarmists).  Makes one wonder where those shifts come from. Excluding that outlying dataset also reduces the spread before 2005. See Animation 1.

Animation 1

PARAMETRIC UNCERTAINTY

Based on observations from ships, buoys and in some cases satellites, data suppliers (NOAA and UKMO) use computer models (not the same as climate models) to determine the monthly, weekly and daily values of sea surface temperatures in the ice-free oceans.  There are a number of factors called parameters that data suppliers can adjust in the computer models in order to produce their sea surface temperature end products.  Parameters are commonly thought of as tuning knobs.  The uncertainties shown in the histograms from Huang et al. (2015) are parametric uncertainties.  That is, they are the uncertainties associated with the 24 parameters NOAA uses to “tune” the ERSST.v4 “pause-buster” sea surface temperature data.

As shown in this post and in the post The Oddities in NOAA’s New “Pause-Buster” Sea Surface Temperature Product – An Overview of Past Posts, NOAA has selected those parameters so that the warming rates of their ERSST.v4 data reside at or near the extreme high ends of the ranges of parametric uncertainties for the periods of 1951 to 2012 and 2000 to 2014.

CLOSING

Please understand that I am not saying the high resolution/daily version of NOAA’s Reynolds OI.v2 data doesn’t serve a purpose.  There are many applications that require daily sea surface temperature data. And there are studies where higher resolution data are preferred, like research into western boundary currents and their relationship to local sea surface temperatures.   Caution has to be exercised, though, when using that version of NOAA’s Reynolds OI.v2 data as a reference for global ocean warming.  There are some seemingly unjustifiable warm biases in that dataset.

In many respects, the new NOAA “pause-buster” ERSST.v4 is also an outlier.  It is the only sea surface temperature dataset (with or without ship-buoy bias adjustments) whose 2000-2014 trend resides near the high end of the trend uncertainty histogram NOAA created for that dataset.

The ERSST.v4 trend is so unusual, so high it might make one wonder if it would fall outside of a trend histogram created from HADSST3 data during the NOAA-selected hiatus period of 2000-2014.  Unfortunately, NOAA did not present the histogram with the uncertainty range for the HADSST3 in Huang et al. (2015b) for that period.

SOURCE

The sea surface temperature data presented in this post are available from the KNMI Climate Explorer.