A Cooling Bias in Global SSTs in the Early 20th Century

A new (currently not paywalled) Nature paper[1] was published this week with some really interesting findings. The authors examined potential biases in sea surface temperatures (SSTs) and found evidence of a cooling bias affecting SSTs between roughly 1900 and 1930 that, if corrected, would warm SSTS during that time frame and also coincidentally make the instrumental record conform more closely with model simulations for the early 20th century. Since this study did not discover a significant bias between 1850-1900, these corrections would not have a significant impact on the amount of global warming above the 1850-1900 mean, but it would have a significant impact on our understanding of multi-decadal variability in temperatures in the instrumental record.

However, some on X have taken this to mean that scientists have overestimated the amount of global warming the earth has experienced. Ryan Maue called this a "bombshell climate paper" and found it disconcerting that it was published in the "prestigious Nature journal was published today without any press coverage." Presumably this is problematic because the paper, in his view, says, "Early 20th Century global temperatures were much warmer than initially thought!" He goes on to complain that if this bias is discovered in the early 20th century, it should make the instrumental record in the 19th century more suspect: 

Furthermore, do you think we know more about the ocean temperatures in 1850-1900 than 1900-1940? 1850-1900 is used as the baseline pre-industrial period for Paris (1.5°) and climate attribution science. So, what are we going to do here, folks?

I'd like to answer his question. That is, I think it would be helpful to investigate whether there is any evidence in this paper that would cause us to view the mean temperature of the IPCC's 1850-1900 baseline as suspect. Robert Rohde, one of the authors of this paper[1], has shared about this study on X and another author, Reto Knutti, has on BlueSky. Tim Osborn, who wrote a (currently paywalled) Nature News and Views article[2] related to this study, has also shared about it. I've consulted these sources as well as my own reading of the paper to write what follows. To understand this paper and what its significance really is, we need a little background.

Background

Scientists currently have a series of Argo floats that can be used alongside buoy and ship measurements, but before WWII, many SST measurements were dependent on ship measurements. Buckets would be tossed over the side of a boat and then lifted out of the water for temperature measurements to be collected. Depending on how well insulated the bucket was, there could be slight cooling of the water by evaporation from the time the water is collected and its temperature measured. The modern rubber buckets are well-insulated, but other buckets were not as much, and canvas buckets would lose more heat to evaporation than wooden buckets.

As technologies and instrumental collection methods change, therefore, a bias can be introduced into the SST record. Temperature collection methods changed from wooden buckets to canvas buckets in the late 19th century. Later in the 20th century, buckets were largely replaced by engine-room intake measurements. Bias correction analysis to date has typically emphasized the transition from buckets to engine-room intake, with the transition from wood to canvas under appreciated. For instance, HadSST4 featured only a small correction at 1906-1910 to account for the change from wooden to canvas buckets. Sippel et al write,

Early SST records, collected before the Second World War (1939–1945), primarily utilized ship-based bucket measurements. The transition within the early record from wooden to canvas buckets around the late nineteenth century, combined with the shifting patterns of shipping routes and shipping fleets, complicates systematic bias adjustments and adds to their uncertainty. Furthermore, essential metadata are often incomplete or missing. Engine-room intake measurements replaced buckets over time, and much work has focused on understanding the biases of buckets relative to this more modern measuring technique.

Scientists have known about this bias for a long time. This was a big part of the discussion in a few of the hacked CRU emails where scientists were trying to figure out how to quantify ocean temperature biases due to changes in SST collection methods. In one email they referred to an "ocean blip."  The discussion in that email revolved around the fact that scientists knew there was a cooling bias prior to 1940, but it was difficult to quantify how much, and it was likely under-corrected. Consequently the "blip" was never fully removed, and it still shows up in all the major GMST datasets, which you an see as a peak in global temperatures in the 1940s during WWII. But since corrections to SSTs prior to 1940 were correcting a cooling bias, the correction necessarily warmed GMST estimates prior to 1940. This correction was larger than the bias correction to land-based temperatures, which suffered from more of a warming bias. The net result of bias correction on land + ocean global temperatures has always been to decrease the amount of global warming above the 1850-1900 mean.

Chan et al 2024

Another paper[3] published in August of this year, found an ingenious way to better quantify the SST bias. They looked at ships taking SST measurement near shore or at port and compared these SST measurements to nearby coastal and island weather stations. Reasoning that these should show similar temperature evaluations, this allowed them to better constrain SST biases from ship-based measurements. Their bias correction method greatly diminished the cold dip in the early 20th century and also made the warm "blip" in the 1940s smaller (see the red DCENT line below).

Chan et al 2024[3]

Chan's results show a more steady warming since 1850 through 1940 than what is found in GMST data sets. While the 1850-1900 mean doesn't change much at all, there is a significant change to the 1900-1930 time period.

Sippel et al 2024

Sippel's paper examined data inconsistencies between land and ocean data from 1850-2020, which  can becomes apparent with ridge regression of SST and land temperature anomalies. From 1850 to 2020 land and SST measurements are largely in agreement, but ridge regression identified an exception in the period roughly between 1900 and 1930, where SSTs are colder than would be predicted. They then consulted proxy data from Pages 2K at locations on land and the sea surface that were near coastlines to confirm their findings. They found that proxy data broadly aligned with land-based instrumental data, and they don't show the cooling dip found in SST measurements. 

Sippel et al 2024[1]

This allowed them to quantify bias affecting SSTs relative to land based thermometers: "existing estimates of ocean temperatures in the early twentieth century (1900–1930) are too cold, based on independent statistical reconstructions of the global mean surface temperature from either ocean or land data. The ocean-based reconstruction is on average about 0.26 °C colder than the land-based one, despite very high agreement in all other periods" (italics mine). This means that two studies have now essentially concluded that scientists have under-corrected for the cooling bias affecting SSTs in the early 20th century. The SST cooling bias affecting 1900-1930 can be seen in panel a below. The agreement between ocean-based and land-based reconstructions between 1850-1900 suggests that the instrumental record for the mean of that baseline is substantially unchanged by these corrections.

Sippel et al 2024[1]

The fact that the bias is detected in roughly 1900-1930 but not in 1850-1900 appears to be largely due to the fact that during the earlier period, collection methods were largely with wooden buckets, and the bias that was introduced (and under-corrected) came from the transition from wooden to canvas buckets. So this study does not provide reason to think that the IPCC's baseline of 1850-1900 was significantly warmer than is represented in GMST datasets. And incidentally, correcting for these biases brings the instrumental record closer to model simulations. This would suggest that internal variability on multidecadal time scales is smaller that would otherwise be indicated, and the Earth's climate system may be a little more predictable than we once thought.

Conclusion

I suspect this is not the last paper to be written on this subject, and more studies may improve on the quantification of the cooing bias affecting SSTs. But a few things should be clear:

1. This paper does not indicate that scientists have underestimated the amount of global warming has occurred above preindustrial levels,
2. This paper does indicate that scientists have been too conservative in applying bias correction to 1900-1930, at least partially because they underestimated the cooling bias introduced by the transition from wooden to canvas buckets near the end of the 19th century,
3. This paper does indicate that scientists can discover new tools to examine hard-to-quantify biases, and
4. It turns out that model simulations appear to be doing even better than we thought at simulating historic temperatures.

I think this is a fascinating paper. It's certainly not a bombshell that disrupts our understanding of the amount of warming that has occurred since the latter half of the 19th century, but it does seem to have some promising conclusions that will resolve some question marks about internal variability in the climate system on multidecadal time scales.

References:

[1] Sippel, S., Kent, E.C., Meinshausen, N. et al. Early-twentieth-century cold bias in ocean surface temperature observations. Nature 635, 618–624 (2024). https://doi.org/10.1038/s41586-024-08230-1

[2] Tim Osborn, John Kennedy. Revised historical record sharpens perspective on global warming. Nature News and Views. November 20, 2024. https://www.nature.com/articles/d41586-024-03551-7

[3] Chan, D., Gebbie, G., Huybers, P. et al. A Dynamically Consistent ENsemble of Temperature at the Earth surface since 1850 from the DCENT dataset. Sci Data 11, 953 (2024). https://doi.org/10.1038/s41597-024-03742-x