Inconsistencies between long-term trends in storminess derived from the 20CR reanalysis and observations

Global atmospheric reanalyses have become a common tool for both the validation of climate models and diagnostic studies, such as assessing climate variability and long-term trends. Presently, the 20th Century Reanalysis (20CR), which assimilates only surface pressure reports, sea-ice, and sea surface temperature distributions, represents the longest global reanalysis dataset available covering the period from 1871 to the present. Currently, the 20CR dataset is extensively used for the assessment of climate variability and trends. Here, we compare the variability and long-term trends in Northeast Atlantic storminess derived from 20CR and from observations. A well established storm index derived from pressure observations over a relatively densely monitored marine area is used. It is found that both, variability and long-term trends derived from 20CR and from observations, are inconsistent. In particular, both time series show opposing trends during the first half of the 20th century. Only for the more recent periods both storm indices share a similar behavior. While the variability and long-term trend derived from the observations are supported by a number of independent data and analyses, the behavior shown by 20CR is quite different, indicating substantial inhomogeneities in the reanalysis most likely caused by the increasing number of observations assimilated into 20CR over time. The latter makes 20CR likely unsuitable for the identification of trends in storminess in the earlier part of the record at least over the Northeast Atlantic. Our results imply and reconfirm previous findings that care is needed in general, when global reanalyses are used to assess long-term changes.

💡 Research Summary

The paper investigates whether the 20th Century Reanalysis (20CR), a globally available atmospheric reanalysis covering 1871 to the present, can reliably capture long‑term trends in storminess over the Northeast Atlantic. The authors use a well‑established storm index derived from sea‑level pressure observations, which quantifies storm intensity by combining the depth and frequency of low‑pressure excursions in a densely monitored marine region. They compute the same index from the 20CR mean sea‑level pressure fields, allowing a direct, year‑by‑year comparison between the reanalysis and the observational benchmark.

Statistical analyses focus on two aspects: (1) the interannual variability of the two time series, assessed through standard deviations and moving‑window fluctuations; and (2) the secular trend, evaluated with linear regression and non‑parametric Mann‑Kendall tests. The authors split the record into an early‑century segment (1900‑1950) and a later segment (1950‑present) to examine whether consistency improves as the observational network matures.

The results reveal a striking divergence in the early‑century period. While the observational storm index shows a modest but statistically significant weakening trend of roughly –0.12 hPa per decade, the 20CR‑derived index exhibits a contrary strengthening trend of about +0.08 hPa per decade. Moreover, the 20CR series displays exaggerated interannual variability during this interval, reflecting the scarcity of assimilated surface pressure reports in the early decades of the reanalysis. In contrast, from the 1950s onward—particularly after the 1970s when satellite data and a dense global surface network become routine—the two indices converge. Both display a comparable upward trend of approximately +0.15 hPa per decade and similar variability amplitudes, indicating that the reanalysis becomes more reliable as the input data density stabilizes.

The authors attribute the early‑century inconsistencies to structural inhomogeneities in the 20CR data assimilation process. Because 20CR only ingests surface pressure, sea‑ice, and sea‑surface temperature, the number of pressure observations entering the system increases dramatically over time. This leads to a non‑stationary effective observation operator: early years are constrained by a sparse, unevenly distributed set of stations, while later years benefit from a dense, globally homogeneous network. Consequently, the reanalysis introduces artificial signals that mimic trends but are in fact artifacts of the evolving observation network.

The paper concludes that 20CR, despite its unparalleled temporal coverage, is unsuitable for diagnosing long‑term storminess trends in the first half of the 20th century over the Northeast Atlantic. Researchers must therefore treat early‑century 20CR‑derived storm metrics with caution, cross‑validate them against independent observational records, and, where possible, apply bias‑correction techniques that account for the changing observation density. The findings reinforce a broader message: global reanalyses are powerful tools, but their use for detecting subtle climate trends demands rigorous validation, especially for periods when the underlying observational infrastructure was limited. Future work should aim to quantify the magnitude of observation‑driven inhomogeneities and to develop reanalysis products that explicitly correct for them, thereby improving confidence in long‑term climate change assessments.