Observed and simulated temperature extremes during the recent warming hiatus

We start with an assessment of the recent trends in the HadEX2 temperature extremes. Here we are particularly interested in question (a), whether the trends are significantly different from zero, and in the spatial trend patterns. The warmest day in a year (TXx), usually occurring in summer, shows significant positive trends in the last 40 years, particularly in central Europe, eastern parts of Asia (figure 1(a)) and northeastern North America. The trends of the recent 15 years have greater amplitude in many regions, but also show greater spatial variability, reflecting their higher statistical uncertainty, which is due to the small sample size of 15 years. This also makes it difficult to evaluate whether the recent short term trends are significantly different from zero (see also Nicholls 2001). Coherent regions with cooling trends, while not statistically significant, emerge in the interior of Canada and in central parts of northern Asia during this period.

Zoom In Zoom Out Reset image size

The coldest night in a year (TNn), usually occurring in winter, show stronger significant positive trends than TXx over the last 40 years in large parts of the Northern Hemisphere, except in Southern and Central Europe (figure 1(b)). In the recent 15 years, the northern latitudes show strong warming trends (∼1 °C per decade), but a coherent zonal band of cooling trends (although again not statistically significant) emerges in the mid-latitudes including western North America and Southern Europe stretching all the way to East Asia. Cooling trends are also prevalent in some areas with sufficient observational data in the Southern Hemisphere (i.e., Southern South America, South Africa and parts of Australia). Note that even if we see a cooling trend regionally (as in figure 1), globally averaged TNn and TXx in the recent 15 years remain on average warmer as observed in the preceding 25 years (i.e., 1971–1995) (see supplementary figure S1). Thus, short-term regional cooling trends do not undermine the global long-term warming trend.

Looking at the median of trends estimated from the ensemble of CMIP5 model simulations in figure 2, we see that significant positive trends in both TXx and TNn dominate across the globe during the past 40 years. In the recent 15 years, some small areas with no or slight cooling trend become apparent; however, these patterns are much less pronounced and not as coherent as in HadEX2. More notable is the similarity between trend patterns over the past 40 years compared to the past 15 years in the model simulations. The ensemble median pattern indicates the dominance of positive trends as found in the bulk of model simulations and masks cooling trends that are apparent in some individual ensemble members. For instance, globally averaged cooling trends in TNn during 1996–2010 occur in individual ensemble realizations of four CMIP5 models (see online supplementary table S1). The observed spatial TNn trend pattern is reasonably reproduced by two of these models (CMCC-CMS and CSIRO-Mk3-6-0, see supplementary figure S2). All models, except MRI-CGCM3, simulated a globally averaged warming trend in TXx in the recent 15 years.

Zoom In Zoom Out Reset image size

From this analysis, regional and seasonal features in TNn and TXx become apparent that are not seen when studying globally averaged Tmean alone. Spatial and seasonal features in Tmean have been discussed in Kosaka and Xie et al (2013) and Cohen et al (2012), but those patterns cannot be directly compared to the TXx and TNn patterns shown here because different mechanisms are involved in generating extreme minimum night-time or maximum day-time temperature conditions compared to Tmean (Seneviratne et al 2014). We should therefore not expect that they could be inferred from a simple first-order shift in the temperature distribution.

An interesting feature in figure 1(b) is the zonal structure of the cooling trends in observed cold extremes in the last 15 years. We investigate this in more detail by looking at zonally averaged time series of 15-year running trends in HadEX2 and the models for the period 1971–2010. We distinguish between zonal bands including the high latitudes in the Northern Hemisphere (45 °N–90 °N) and the mid-latitudes in both Hemispheres (20 °S–45 °S and 20 °N–45 °N). We exclude the low latitudes (i.e., 20 °S–20 °N) due to sparse observational data coverage.

As we have seen, individual trend estimates based on short (i.e., 15-year) records are highly uncertain. A technique such as that of Fyfe et al (2013) can nevertheless be used to determine whether a collection of model simulated trends as obtained from an ensemble of climate simulations is consistent with an observed trend. Thus, our objective in this section is to investigate question (b) as to whether model simulated trends in temperature extremes are consistent with observed trends.

Running 15-year trends in globally averaged TXx and TNn from HadEX2 lie within the 90% range of model trends (i.e., 5–95% of the models in the following) for both TXx and TNn as can be seen in figure 3. For TXx, the ensemble median and range of the running 15-year model trends follows the evolution of observed 15-year running trends in HadEX2 particularly at the global scale, and also for the zonal averaged latitudinal bands, except in the higher northern latitudes (45 °N–90 °N). In the latter region, the observed trend falls outside the model range in the mid-1980's, where also the null hypothesis (i.e., observed and simulated trends are equal) is rejected for both assumptions discussed in Fyfe et al (2013). This single departure is, however, within the expected rejection rate of a statistical test that has a nominal significance level of 10%. It appears that rejections are actually very rare, suggesting that intermodel differences are probably larger than internal variability (i.e., evidence of a discrepancy between observed and simulated trends has to be strong to reject).

Zoom In Zoom Out Reset image size

For TNn, the temporal evolution of the HadEX2 running 15-year trends is more variable than the corresponding evolution of TXx in HadEX2. Observed global features in TNn, such as the larger trends in the 1980s and comparably smaller trends in the recent 15 years, are neither found in the model ensemble median nor range. This feature manifests itself in the mid-latitudes, and particularly in the Northern Hemisphere mid-latitudes (20 °N–45 °N), where observed trends fall outside the model range in a sequence of years between the late 1990s and early 2000s. This again is supported by the rejection of the null hypothesis of the applied test statistic. Similar results can also be found when considering 20-year running trends (see supplementary figure S3). The steep increasing trend in observed TNn in the higher northern latitudes (45 °N–90 °N) is embedded in the comparably wide model range for this zonal band, but is not reproduced in its full extend by the ensemble median.

To pin down regional patterns of recent observed and simulated trends, we compare trends in globally and regional averaged TXx and TNn for the two time periods. We choose seven continental-scale regions (see figure 4(a)) according to the HadEX2 data coverage as well as climatological features of observed trend patterns, as discussed in section 3.1. Note, that the trends in globally averaged TXx and TNn do not represent the sum of the averages over the seven regions as grid-boxes in South America and South Africa (gray shading in figure 4(a)) are not included in any of the defined regions.

Zoom In Zoom Out Reset image size

As can be seen in figure 4(b), the observed trends in globally averaged TNn and TXx fall within the 5–95% model  range. Compared to the test results (i.e., using the method of Fyfe et al 2013) presented in the previous section 3.2., the model range represents a rather conservative estimate of whether the observed trends are consistent with the simulated trends. Thus, in the following, we consider only the model range as the basis for our comparison of regional trends from models and observations. Note that the global results shown in figure 4(b) for the 1996–2010 period are identical to the last time point shown in the panels of the upper row of figure 3. Note also that the model spread is substantially wider when a shorter period is considered (i.e., 1971–2010 versus 1996–2010), as is expected due to the greater uncertainty in trend estimates from the shorter records. In general, the observed trends of globally averaged TNn and TXx are comparable between the recent 15-year and the longer 40-year period with the differences between the trends of the two periods centering around zero (figure 4(c)).

Recent detection and attribution studies by Zwiers et al (2011) and Christidis et al (2011) as well as a detection study by Fischer and Knutti (2014) argue that models generally underestimate observed trends in globally averaged TNn and some models overestimate observed trends in globally averaged TXx. If we repeat our analyses using the set of CMIP5 models and ensemble members as in online supplementary table S1 for similar periods as considered in these studies (e.g., 1960–2000, 1971–2000, 1960–2010), we generally confirm these results (see supplementary figure S4). This comparison, however, reveals that there is some sensitivity to the time period chosen for the comparison of simulated and observed trends, particularly when limiting the model information to one number (i.e., the ensemble median or mean).

We look now at regional scales and as at the global scale we find again that the trends in simulated TXx are generally in agreement with the observed trends (figure 4(b), left), with particularly close agreement in Australia (AUS) and North Asia (NAsia) for both periods. Differences between the longer and shorter period generally center on zero (figure 4(c), left) for both observed and simulated trends in all regions, which indicates comparable trends between these two periods.

Larger regional discrepancies are found between simulated and observed trends in TNn (figure 4(b), right). The observations show moderate to pronounced negative trends in the recent 15 years in several regions (particularly in Southern Asia (SAsia)), whereas the majority of models simulate a positive trend. Only one model (i.e., CCSM4) simulated a negative trend as large as the observed in SAsia (see also supplementary figure S2(b), (c)). For this region, the observed difference in TNn trends between the shorter and longer period falls outside the 5–95% model range of differences (figure 4(c), right). This indicates a discrepancy between the observed and simulated trends for the recent 15 years as reflected also in the statistically significant departure of observed running trends from the range of simulated trends in the northern mid-latitudes as shown in figure 3.

For the other regions, observed trends in TNn fall within the range of simulated trends (figure 4(b)). Differences between the two periods deviate further from zero than for trends in TXx (figure 4(c)), indicated by the larger model spread, but still center around zero. Slight cooling trends in HadEX2 in the recent 15 years can be found in the Mediterranean region (MED) and AUS, which however fall within the model range. In the other regions, located primarily in the mid-to-high northern latitudes (i.e., Eastern North America (ENA), Northern Europe (NEUR) and Northern Asia (NAsia)), the observations show an increased warming trend for TNn in the recent 15 years compared to the longer period, which is also reflected in the ensemble median. In general, it becomes apparent that the warming trend in the recent 15  years is somewhat more pronounced in cold extremes (i.e., TNn) in high northern latitudes, and exceeds the warming observed in the hot extremes (i.e., TXx).