A Review of Temperature Reconstructions

By Andy May

This is an update to a 2016 post; the original post is here.

We often hear that the planet is warming faster than ever before, or at the fastest rate since the beginning of the industrial era! Is it true? We haven’t had thermometers for very long. How do thermometer readings compare to temperature proxies like ice cores and tree rings? Greenland is a good place to start, we see the high-resolution Greenland ice core temperatures all the time. How accurate are they? How do Greenland temperatures compare to temperatures elsewhere?

In previous posts (here and here), I’ve compared historical events to the Alley and Kobashi GISP2 Central Greenland Temperature reconstructions for the past 4,000 years. Unfortunately, these two reconstructions are very different. Steve McIntyre suggested I consider a third reconstruction by Bo Vinther. Vinther’s data can be found here. Unfortunately, Vinther is significantly different from the other two. Nothing agrees very well.

The Alley data has been smoothed, but the details of the smoothing algorithm are unknown. I smoothed the other datasets, so they visually have the same resolution as the Alley dataset. Both datasets (Kobashi and Vinther) were first smoothed with a 100-year moving average filter. Then 20-year averages of the smoothed data were taken from the one-year Kobashi dataset to match the Vinther 20-year samples. The Alley data is irregularly sampled, but I manually averaged 20-year samples where the data existed. If a gap greater than 20 years was found that sample was skipped (given a null value).

All three reconstructions are shown in Figure 1. There is no reason to prefer one of the three reconstructions over the other two, so I simply averaged them. The average is the blue line. I’m not presenting this average as a new or better reconstruction; it is merely a vehicle for comparing the three reconstructions to one another and to other temperature reconstructions. I’m trying to show the variability in common temperature reconstructions for the past 2,000 to 4,000 years. This post is less about the actual temperatures, than the temperature reconstructions and how they compare.

There are some notable outliers apparent in the comparison. In particular, we see the odd 700AD 2.5 degree C Kobashi spike and the scatter in the interval from 700BC to 100BC. The Minoan Warm Period (1600BC to 1300BC) is completely missing in the Vinther reconstruction. The estimates agree better from 900 AD to the present than they do prior to 900 AD. Perhaps as the ice gets older, accuracy and repeatability are lost? Figure 2 shows the same average and the maximum and minimum value for each 20-year sample.

The average temperature for the 4,000-year period is -30.8°C. The average minimum and maximum suggest this value is plus or minus 0.3°C. Perhaps we are simply seeing the error in these methods and nothing more. For those who want to see the messy details of the average temperature calculation, the spreadsheet can be downloaded here. As noted in my previous post, the error in the time axis is probably at least ±50 years. Loehle has suggested a time error of ±100 years based on ¹⁴C laboratory errors. These values give us some perspective in interpreting the reconstructions. Below is a comparison of the average to the same historical events we have used before.

This average temperature reconstruction shows a steady decline in temperature since the Minoan Warm Period, interrupted by ±120-year cycles of warm and cold. Don’t take the apparent 120-year cyclicity too seriously all the data was smoothed with a 100-year moving average filter. After the end of the Little Ice Age, the Modern Warm Period begins, and temperatures rise to those seen in the Roman Warm Period. The Modern Warm Period is equivalent to the Medieval Warm Period within the margin of error. We need to be careful because we are comparing actual measurements to averaged proxies.

When proxies are averaged, all high and low temperatures are dampened. In particular, the Medieval Warm Period is somewhat smeared and dampened due to the Vinther record. The Vinther Medieval Warm Period peak is earlier than the Kobashi and Alley peaks. Major volcanic eruptions fit this timeline reasonably well. Rabaul is dated to 540AD. Thera-Santorini occurred in 1600BC and Tambora in 1815. The HadCRUT 4.4 point shown with a red star is an average of several HADCRUT4 surface temperature grid points in the Greenland area thought to be comparable to the Greenland average temperature.

Comparisons to broader temperature reconstructions
Dr. Craig Loehle published a global composite temperature reconstruction in 2007 and a corrected version in 2008. This reconstruction has been widely reviewed and appears to have stood the test of time. Subsequent work seems to support the reconstruction. In Figure 4 we show his global reconstruction compared to the Greenland average and the recent temperature reconstruction of the extratropical (90° to 30°N) Northern Hemisphere by Christiansen and Ljungqvist. The graph in Figure 4 shows temperatures as anomalies from 1600 to 1800 since each line represents a different area.

All the reconstructions show a trend of decreasing temperature from 900AD to the Little Ice Age, roughly from 1400 to 1880 AD. They also show a temperature peak around 1000 AD. The Northern Hemisphere peak is much higher than the Greenland or global peaks. The most striking thing about Figure 4 is that the temperature swings seen in the extra tropical Northern Hemisphere are larger than in the other reconstructions. There are likely three reasons for this, the first is the global reconstruction averages more proxies. The proxies have different resolutions, anything from one-year to several hundred years and the dating uncertainty is high in each proxy sample. These issues dampen any extremes. The second issue is the Southern Hemisphere and the Northern Hemisphere warm and cool at different times. Since the hemispheres are out of phase, global extremes are further dampened. Finally, the Northern Hemisphere is always anomalous, it has more extreme warming and cooling than the other large regions of the earth. The reason for this is unclear, but it could be because most of the land is in the Northern Hemisphere, between 30N and 60N. The Arctic is mostly ocean, so it escapes some of the extremes seen in the Northern Hemisphere between 30N and 60N.

Temperature changes are amplified in the Arctic and Antarctic, independently of where the land masses are, over geological time, as Chris Scotese explains in his climate history. This is also suggested by Flannery, et al. So, a combination of polar amplification and extra land causes the Northern Hemisphere to be anomalous. Finally, Figure 5 shows the Greenland average compared to two Arctic reconstructions. One is the ice core component of the Arctic reconstruction by Kaufman, 2009 (data here) and the other is the Sunqvist, 2014 “PAGES2K” Arctic reconstruction (data can be found here). In Figure 5, the Arctic and PAGES2K temperature anomalies have been shifted to the average central Greenland temperature for the period for comparison purposes.

These two multi-proxy Arctic reconstructions agree fairly well with the Greenland average if we assume a ±0.3°C temperature error and ±50-year time error. It is interesting that the peak about 400AD is seen in the Arctic and Greenland reconstructions but not in the Loehle global reconstruction in Figure 4. The Sunqvist, 2014 reconstruction was used as presented in his paper, it seemed to be carefully constructed. Sunqvist,et al. did include some tree ring data (fewer than 1% of the proxies), but they used it carefully and tree ring data did not dominate his reconstruction.

Kaufman’s Arctic reconstruction used a lot of tree ring data (4 of 23 proxy records) as can be seen in his Figure 3 and in his dataset. Tree ring data does provide an accurate chronology, but it provides a poor temperature proxy due primarily to what has been called the “divergence” problem. Tree rings may correlate well to temperatures in a “training” period but show little correlation to temperature longer term. There are several possible reasons for this. First, as Keith Briffa explained, the twentieth century tree ring data diverges strongly from historical tree ring data. This could be a result of increasing CO₂ in the atmosphere causing the assumed temperature to tree ring correlation to change.

A second likely factor is that forests adapt to long-term climatic changes by adjusting tree density and tree size, this diminishes their usefulness in measuring long-term changes. Tree ring width and density tend to reflect summer temperatures, precipitation, and many other factors, but extracting the average annual air temperature from them is problematic. Loehle discusses this problem and other problems with tree ring data here and here. For this reason, only the ice core records (seven proxy records of 23) from Kaufman’s reconstruction were used to plot the “Arctic” line in Figure 5.

Figure 6 compares Kaufman’s “All proxies” reconstruction to his ice core, sediment (including lake varves), and tree ring proxies. The ice core, sediment, and tree ring proxies only agree reasonably well for the last 500 years, before that the tree ring proxies diverge dramatically downwards. The lake and marine sediment proxies (12 of the 23) are also lower than the ice core proxies, but not so dramatically. We all know of another paleoclimatologist who took advantage of this divergence.

When all the proxies are used the earlier temperatures are much lower and the modern warm period has a higher peak. The recent peak in Figure 6 is the twenty-year average around 1945. The proxy reconstruction then drops, the last point is centered on 1985. I didn’t bother to “hide the decline.” Except for the sharp drop from 1945 to 1985 Kaufman’s ice core proxies fit the rest of the reconstructions shown above reasonably well.

Discussion
There are many Greenland area temperature reconstructions, they use ice core data, lake and marine sediment core data and other proxies, mainly tree rings. They are not perfect and contain errors in the temperature estimates and dating errors. The exact error is unknown, but by comparing reconstructions we can see that they generally, except for the tree ring proxies, agree to within 0.3°C and in time, to within 50 years or so. Why is this important? Natural climate cycles are poorly understood. Some, like the very irregular, but short-term (~3-7 years) ENSO cycle (La Nina and El Nino) we can identify, but because they are irregular, and the cause is unknown, we cannot model them. The same is true of the ~60 to 80-year Atlantic Multidecadal Oscillation (AMO) and the ~50-70-year Pacific Decadal Oscillation (PDO). These events affect the weather and climate all over the world, but they are not accurately included in the GCM’s (General Circulation Models) used by the IPCC, and other organizations, to compute man’s influence on climate. Thus, some portion of the Modern Warm Period attributed to humans may, in fact, be attributable to these or other natural climate cycles.

During the 1980’s and 1990’s the PDO was mostly positive (warming). From the mid 1990’s to today the AMO has been mostly positive and undoubtedly contributing to warming. There have been numerous attempts to see a pattern in these multidecadal natural climate cycles. Most notably, Wyatt and Curry identified a low-frequency natural climate signal that they call a “stadium wave.” This model is based on a statistical analysis of observed events (especially the AMO) and not on the physical origins of these long-term climate cycles. But it does allow predictions to be made and the veracity and accuracy of the stadium wave hypothesis can and will be tested in the future.

Another recent paper by Craig Loehle discusses how the AMO signal can be removed from recent warming, leaving a residual warming trend that may be related to carbon dioxide. He notes that when the AMO pattern is removed from the Hadley Center HADCRUT4 surface temperature data the oscillations are dampened and a more linear increase in temperature is seen. This trend compares better to the increase in carbon dioxide in the atmosphere and allows the computation of the effect of carbon dioxide on temperature. The calculation, for the post-1970 period, results in a temperature trend of 0.83°C per century. This is roughly half of the observed trend of 1.63°C. Loehle suggests that the AMO may be the best indicator of natural trends. If this is true, then half of recent warming is natural and half is man-made. It also suggests that the equilibrium climate sensitivity to carbon dioxide is about 1.5°C per doubling of CO₂, half the IPCC preferred value. This value also compares well to other recent research.

Conclusions
The use of temperature proxies to determine surface air temperatures prior to the instrument era is very important. It is the only way to determine natural long-term natural climate cycles. Currently, in the instrument record, we can see shorter cycles like the PDO, AMO, and ENSO. When these are incorporated into models, we see that half or more of recent warming is likely natural, belying the IPCC idea that “most” of recent warming is man-made. Yet, these shorter cycles are clearly not the only cycles. When we look at longer temperature reconstructions, we see 40,000, 80,000 or 120,000-year glacial periods interrupted by brief 10,000 to 18,000-year interglacial periods. These longer periods will probably only be fully understood with more accurate reconstructions. Intermediate ~1,500-year cycles, called “Bond events,” the 2,450-year Bray, and the 1,000-year Eddy cycles have also been identified and need to be better understood.

Tree ring proxies older than 500 years and younger than 100 years are anomalous. This anomaly is large enough to cast doubt on any temperature reconstruction that uses tree rings. Between lake and marine sediment proxies and ice core proxies it is hard to tell which is more likely to be closer to the truth. They agree well enough to be within expected error. All proxies diverge from the mean with age, none are accurate (or more precisely in good agreement) prior to 1100 AD. It does appear that all proxies except for tree ring proxies, could be used for analytical work back to 1100 AD. Unfortunately, the uncertainty in dating the samples, the poor temporal resolution, and the lack of temperature accuracy preclude comparing these proxy records to modern temperatures. Statements like “warming today is the fastest in the last 1,000 years [or more]” have no credibility. Computing the speed of warming requires accurate dating and an accurate temperature measurement.

The Northern and Southern Hemispheres are out of phase with one another and cool or warm separately. This makes global temperature averages problematic. This is probably due to Earth’s orbital eccentricity and regular changes in the Earth’s orientation, relative to its orbital plane. For these reasons, the Sun affects Earth’s climate differently by latitude. CO₂ may be a global effect, but solar forcing is not, it is hemisphere specific. This must be considered when calculating natural warming, it discourages focusing on a global average temperature as a climate change metric. The unique climate of the Northern Hemisphere, which is more extreme than the other regions of the world, must also be considered.

The other very important use for temperature reconstructions is to study the impact of climate changes on humans and the Earth at large. Historical events are often known to the day and hour, only when we have reconstructions with more accurate time scales can we properly match them to major events in history. In addition, this post makes it clear that combining multiple proxies causes severe dampening of the temperature response because proxy sample dating errors cause peaks and valleys in the record to be mismatched, reducing apparent variability and accuracy. Multiple proxy reconstructions show less temperature variation than occurred. The National Academy of Sciences/National Research Council and Wegman Committee investigations into the “Hockey Stick” documented this problem well. They also showed that dating accuracy will probably never be better than 50-100 years with current dating tools. Tree ring dates are more accurate than this, but tree rings are a poor and inconsistent indicator of temperature. A proper comparison of warming or cooling rates between the twentieth and twenty-first centuries and any period prior to 1900 will probably never be possible, short of dramatic improvements in proxies or proxy technology.