An evaluation of the statistical homogeneity of the Twentieth Century Reanalysis

Abstract

The Twentieth Century Reanalysis (20CR) holds the distinction of having the longest record length (140-year; 1871–2010) of any existing global atmospheric reanalysis. If the record can be shown to be homogenous, then it would be the first reanalysis suitable for long-term trend assessments, including those of the regional hydrologic cycle. On the other hand, if discontinuities exist, then their detection and attribution—either to artificial observational shocks or climate change—is critical to their proper treatment. Previous research suggested that the quintupling of 20CR’s assimilated observation counts over the central United States was the primary cause of inhomogeneities for that region. The same work also revealed that, depending on the season, the complete record could be considered homogenous. In this study, we apply the Bai-Perron structural change point test to extend these analyses globally. A rigorous evaluation of 20CR’s (in)homogeneity is performed, composed of detailed quantitative analyses on regional, seasonal, inter-variable, and intra-ensemble bases. The 20CR record is shown to be homogenous (natural) for 69 (89) years at 50 % of land grids, based on analysis of the July 2 m air temperature. On average 54 % (41 %) of the grids between 60°S and 60°N are free from artificial inhomogenetites in their February (July) time series. Of the more than 853,376 abrupt shifts detected in 26 variable fields over two monthly time series, approximately 72 % are non-climate in origin; 25 % exceed 1.8 standard deviations of the preceding time series. The knock-on effect of inhomogeneities in 20CR’s boundary forcing and surface pressure data inputs to its surface analysis fields is implicated. In the future, reassessing these inhomogeneities will be imperative to achieving a more definitive attribution of 20CR’s abrupt shifts.

Appendices

Appendix 1: Every member analyses for Geneva and Rondonia

In this section, we evaluate variability among the 56 ensemble members from which we assess only the ensemble mean (i.e., official 20CR product) in our study. We focus on two grid points: Geneva, Switzerland (46.20°N, 6.15°E) and Rondonia, Brazil (24.0°S, 51.0°W). The significance of these locations is the dichotomy of their supporting observational record. The record at Geneva benefits from a steady stream of approx. 300 observations per month over the period of availability. In contrast, not a single observation was ever assimilated at Rondonia.

Figures 14 and 15 present results of the Bai-Perron test performed on the 56-member monthly T _a and P data at Geneva and Rondonia, respectively. In general, but particularly for Geneva, the confidence interval for the ensemble mean change point encompasses much of the cumulative 56-member confidence interval range. The figures show that a change point in the ensemble mean could be triggered by abrupt shifts in as few as five ensemble members. On the other hand, there are cases in which clusters of 20 ensemble member change points over 9 years (Fig. 15a, September) do not cause an equally significant shift in the ensemble mean. At Geneva, most of the variable-months are homogenous. This suggests, to us, that if a climate shift did occur, it was weak. At Rondonia, on the other hand, change points manifest in all variable-months (except July for T _a), which we interpret as evidence of a strong shift.

Fig. 14

Summary of the Bai-Perron test results for 20CR a T _a and b P at Geneva, Switzerland. The horizontal magenta lines at the top of each subpanel (where detected) are the change points and associated 95 % confidence intervals computed from the 20CR ensemble mean field. The gray-filled area is the histogram of cumulative confidence intervals summed over 20CR’s 56 ensemble members. Vertical magenta bars show the number of ensemble members for which a change point is detected, multiplied by three for emphasis. Results from CRU T _a (blue) and GPCC P (orange) are also included (where detected). Notice that no change points are detected for the latter. Also, note that the vertical scale in (b) varies

Fig. 15

As in Fig. 14 but for Rondonia, Brazil. The tripling of cumulative change point counts for visual effect places the representative magenta bars outside plotting extents in May 1936 (38/56 ensemble members) for T _a and the following cases for P: January 1988 (26/56 ensemble members); February 1985 (38/56 ensemble members); March 1951 (33/56 ensemble members) and 1973 (37/56 ensemble members); October 1972 (24/56 ensemble members); November 1965 (19/56 ensemble members); December 1976 (36/56 ensemble members)

Figure 15 exposes limitations of the Bai-Perron detection algorithm. Specifically, evidence of competing models (in a Bai-Perron sense) and detection limits can be seen [e.g., Fig. 15a: April, October, and November; Fig. 15b: October and December]. For example, at Rondonia, 17, 15, and 16 members of the T _a ensemble had change points in November 1962, 1969, and 1978, respectively. In other words, three potential shifts were identified within a 17-year time period. Because of the imposed 21-year minimum segment length (Sect. 2.3.2), only one change point was resolved.

Appendix 2: Temperature spread fields over ocean

See Fig. 16.

Fig. 16

For ocean domains, the area-averaged (at 2°) time series of 20CR July a assimilated observations, b T _a ensemble spread, and c TMIN ensemble spread. Notice the y-axis scales of (b) and (c) vary substantially