Addressing the World War 2 Warm Anomaly in HadSST.4.2.0.0

We present an update to the Hadley Centre Sea-Surface Temperature dataset (HadSST.4.2.0.0) that addresses residual warm bias during the Second World War (WW2). Using an existing quantitative definition of the WW2 warm anomaly we identify Engine Room Intake (ERI) bias corrections as the dominant factor in this warm bias in HadSST4, and use this to propose new constraints on ERI bias estimates prior to 1950. In addition, we implement corrections for truncation bias in observations from the Japanese Kobe Collection, spanning the period from 1933 to 1961. We evaluate the effects of these changes with respect to the previous version of HadSST and compare with the most recent iterations of other SST datasets including ERSSTv6, COBE-SST3 and DCENT. We show that it is possible to remove the WW2 warm anomaly using a physically-based approach that maintains the independence of HadSST from land surface temperature records, and preserves structural diversity within the range of available global SST datasets.

💡 Research Summary

The paper presents a comprehensive update to the Hadley Centre Sea‑Surface Temperature dataset, resulting in version HadSST.4.2.0.0, with the explicit goal of eliminating the residual warm bias that has persisted in the World War II (WW2) period (1941‑1945). The authors begin by quantifying the WW2 warm anomaly using the definition of Chan and Huybers (2021), which calculates the difference between the mean SST anomaly for the war years and the midpoint of the five‑year periods immediately before and after the war. Applying this metric to the previous HadSST.4.0.0.0 ensemble yields an average anomaly of 0.19 °C (range ‑0.09 °C to 0.45 °C), while the ensemble median for HadSST.4.1.1.0 is 0.21 °C (range 0.11 °C to 0.48 °C). These values are clearly larger than the “true” WW2 anomaly estimated by Chan and Huybers (‑0.01 °C to 0.18 °C), indicating under‑correction in the existing dataset.

The authors identify Engine‑Room Intake (ERI) measurements as the dominant source of the bias. During the war, ERI observations became the primary SST source; ERI sensors tend to read warm, whereas bucket measurements are generally cool, creating a discontinuity. In the current HadSST framework, ERI bias is estimated regionally only after 1950, while a single global bias is applied to all pre‑1950 data, sampled from a uniform distribution of 0.1‑0.5 °C (0‑0.5 °C during WW2). By examining the strong linear relationship (correlation ≈ 0.93) between the assumed ERI bias and the WW2 anomaly, the authors argue that the pre‑1950 ERI bias range should be tightened. A literature review reveals that early ERI bias estimates are sparse, highly uncertain, and often based on very small ship samples. Studies such as James & Fox (1972) report a mean bias of 0.3 °C with a 1 σ spread of ±0.9 °C, while Kent & Kaplan (2006) find random‑error variances of 2.7‑3.0 °C² for the period 1975‑1994. Given this uncertainty, the authors propose a revised pre‑1950 ERI bias interval of 0.2‑0.5 °C, which aligns with the magnitude of the WW2 warm anomaly and is consistent with the broader literature. This new constraint is then implemented in the HadSST.4.2.0.0 ensemble.

In parallel, the paper addresses a separate source of systematic error: truncation bias in the Japanese Kobe Collection (ICoADS decks 118 and 119) covering 1933‑1961. Deck 118 records SST to the nearest whole degree, effectively dropping the decimal part, while deck 119 applies a similar truncation after computing SST from air temperature and air‑sea temperature differences. The authors calculate an average truncation bias of 0.45 °C for deck 118 (consistent with a 0‑1 °C uniform distribution assumption) and an additional 0.25 °C bias for deck 119. These corrections are added to the raw observations, mirroring the approach taken by COBE‑SST3, which adjusts deck 118 upward by 0.5 °C.

A technical re‑implementation of the HadSST processing pipeline is also described. The legacy IDL codebase has been rewritten in modular Python, and the monthly update workflow is now automated using Cylc 8, reducing human error. The re‑implementation uncovered a subtle bug in the handling of canvas‑to‑rubber bucket transition dates. The original HadSST.4.0.1.0 allowed a 29 % probability that the transition began before 1941, whereas the new implementation restricts this to 1.1 %—a result that aligns with historical documentation indicating that rubber buckets were not widely adopted until the mid‑1950s.

The authors evaluate the new dataset against three contemporary SST products: ERSSTv6, COBE‑SST3, and DCENT. Metrics include global mean anomalies, regional pattern differences, ensemble spread, and spectral characteristics. HadSST.4.2.0.0 successfully reduces the WW2 warm anomaly to less than 0.02 °C while preserving the overall climatology and variability of the earlier version. Ensemble spread is modestly reduced, indicating increased confidence in the corrected values. Importantly, the update maintains the independence of HadSST from land‑surface temperature records, a core design principle of the dataset.

In conclusion, by tightening the pre‑1950 ERI bias range to 0.2‑0.5 °C and applying rigorously quantified truncation corrections to the Kobe Collection, the authors demonstrate that the WW2 warm anomaly can be removed from HadSST without compromising the dataset’s structural diversity or its physical basis. This work enhances the reliability of historical SST records, which are essential for climate attribution studies, model evaluation, and the detection of long‑term oceanic warming trends. Future work is suggested to further refine ERI bias estimates using ship‑log metadata and to explore other wartime observational changes that may affect early‑20th‑century SST reconstructions.