Quantification of temperature persistence over the Northern Hemisphere land-area

Abstract

Extreme weather events such as heat waves and floods are damaging to society and their contribution to future climate impacts is expected to be large. Such extremes are often related to persistent local weather conditions. Weather persistence is linked to sea surface temperatures, soil-moisture (especially in summer) and large-scale circulation patterns and these factors can alter under past and future climate change. Though persistence is a key characteristic for extreme weather events, to date the climatology and potential changes in persistence have only been poorly documented. Here, we present a systematic analysis of temperature persistence for the northern hemisphere land area. We define persistence as the length of consecutive warm or cold days and use spatial clustering techniques to create regional persistence distributions. We find that persistence is longest in the Arctic and shortest in the mid-latitudes. Parameterizations of the regional persistence distributions show that they are characterized by an exponential decay with a drop in the decay rate for very persistent events, implying that feedback mechanisms are important in prolonging these events. For the mid-latitudes, we find that persistence in summer has increased over the past 60 years. The changes are particularly pronounced for prolonged events suggesting a lengthening in the duration of heat waves.

Appendix

1.1 Regression analysis

The correlation between persistence and EKE (SPI) is tested by a regression analysis. For this purpose, the persistence record and the EKE (SPI) time series are linearly detrended. Each persistent event is associated to the monthly EKE (SPI) value of the mid-point of the event. Multiple events occurring within one month lead therefore to the situation that monthly EKE (SPI) values are considered more than once in the regression analysis. Assuming that consecutive warm (cold) persistent events are independent from each other, the individual data points in the regression analysis can still be considered as independent measurements. The significance values obtained by linear regression are adjusted for multiple testing by controlling the false discovery rate as suggest by Wilks et al. (Wilks and Sciences 2016).

1.2 Significance testing for distribution comparisons

We compare regional persistence distributions of the 20 years at the beginning (1954–1974) and at the end (1990–2010) of the observational record (due to detrending, 3 years and 45 days are lost at the beginning and the end of the time series). For significance testing, we treat periods as independent from each other only if their midpoints are separated by at least one week. However persistent periods measured at neighboring grid points cannot be treated as independent. To account for this spatial correlation, the persistence record is divided into blocks of persistence events occurring within one region and within one week. These blocks are shuffled for the years of the studied periods (1954–1974 and 1990–2010) creating 10,000 random pairs of equally large persistence distributions. Using an adapted Kolmogorov–Smirnov test, we consider differences in persistence distributions as significant if the maximum distance between the two cumulative distribution functions (i.e. of the early and late periods) is larger than 95% (90%) of the randomly created distribution pairs. Thus there would be only a 5% (10%) chance that the detected differences could have occurred by chance.

Note that the significance test tells whether the distributions are significantly different or not, which not per se relates to changes in the tail, i.e. the 95th percentile.

See Fig. 9.

Fig. 9

Colors indicate the differences in the mean of persistence distributions of 1990–2010 and 1954–1974 for cold and warm periods evaluated for each season. Black dots (triangles) indicate where distributions have changed with a significance level of 95% (90%). Regions 1 to 19 are the result of regional clustering techniques and the larger regions “NHpo”, “NHml” and “NHst” in the lower table regroup these regions into climate zones. All regions are visualized in Fig. 2