Projected changes in regional climate extremes arising from Arctic sea ice loss

We analyse simulations from two independent AGCMs, namely, the UK Met Office Hadley centre global atmospheric model version 2 (HadGAM2) and the National Centre for Atmospheric Research (NCAR) community atmosphere model version 4 (CAM4). Full descriptions are these AGCMs can be found in Collins et al (2011) and Gent et al (2011), respectively. The version of HadGAM2 used here has a horizontal resolution of 1.875° longitude and 1.25° latitude (known as N96) and 38 vertical levels. The utilized configuration of CAM4 has a horizontal resolution of 1.25° longitude by 0.9° latitude and 26 vertical levels.

We performed two core experiments with both AGCMs. In the first experiment, we prescribed a repeating annual-cycle of sea surface temperatures (sst) and sea ice concentrations (sic) representative of the late twentieth century. These sst and sic values were taken from the CMIP5 'historical' simulations with the coupled versions of the models (known as HadGEM2 and CCSM4), averaged for the period 1980–99 and across all available ensemble members. The late twentieth century simulations with HadGAM2 and CAM4 are referred to hereafter as had20c and cam20c, respectively.

In the second experiment, we prescribed the same sst values as just described, but this time with sea ice conditions representative of the late twenty-first century. The sic values were taken from the CMIP5 'rcp8.5' simulations averaged for the period 2080–99 and across all available ensemble members. The 'rcp8.5' simulations are forced by a continuous increase in greenhouse gas concentrations and are often viewed as a 'business-as-usual' scenario, with limited mitigation strategies applied. This scenario was chosen to maximize the signal-to-noise ratio. In grid-boxes that lost ice between the late twentieth and twenty-first centuries, the sst value for the late twenty-first century was used. This procedure accounts for the ocean surface warming in regions where ice is reduced. The late twenty-first century simulations with HadGAM2 and CAM4 are referred to hereafter as had21c and cam21c, respectively. Further details on the experimental setup can be found in Screen et al (2015). Additionally in the supplementary material, we present selected results from a third experiment with mid-twenty-first century sea ice conditions (had21c_mid and cam21c_mid).

The mean March (month of annual maximum) and September (month of annual minimum) sea ice concentrations, and the annual cycle of Arctic sea ice area, from each simulation are shown in figure 1. In September there are ice-free conditions in both had21c and cam21c. In fact, had21c is ice-free (or very nearly) from July to November and cam21c is ice-free from August to October. In March, had21c retains ice along the northern coasts of Greenland, Canada and Siberia, but open water dominates the Atlantic sector. In comparison, cam21c retains more multi-year ice with the largest reductions (compared to cam20c) in the Barents Sea, Sea of Okhotsk and Bering Sea. In terms of mean March sea ice area, had21c has a 50% reduction (compared to had20c) and cam21c has a 20% reduction (compared to cam20c). The CMIP5-mean loss of March sea ice area over this time period (under rcp8.5) is 32% (15%–63%; 5th–95th percentile range). Thus, the prescribed sea ice area loss in HadGAM2 is towards the upper-end of the CMIP5 model spread and in CAM4, is towards the lower end.

Zoom In Zoom Out Reset image size

All simulations (had20c, cam20c, had21c and cam21c) were run for 260 years. In our modelling framework where the boundary forcing repeats annually, each year can be considered a separate ensemble member starting from a different atmospheric initial condition. Thus, we have 260-member ensembles from which to calculate robust statistics of extreme weather. We note that since these are atmosphere-only simulations, we cannot capture possible changes in weather extremes related to ocean circulation changes that may result from Arctic sea ice loss (Deser et al 2015).

In what follows we analyse daily values of maximum near-surface (1.5 m in HadGAM2; 2 m in CAM4) temperature (T_max), minimum near-surface temperature (T_min), daily-mean near-surface temperature (T_ave) and total precipitation (P_tot). We applied the ETCCDI indices (see below) to the full 260 years and then derived long-term means. To approximate the response to Arctic sea ice loss, we subtracted the long-term mean of the had20c (cam20c) simulation from that in the had21c (cam21c) simulation. Statistical significance was calculated using a difference of means test (Student's T-test) where the null hypothesis is that the two samples (n = 260) have the same mean. We report responses where the null hypothesis can be rejected with 95% confidence. In addition to testing for statistical significance, we also assess the robustness of the response between the two models. Robust responses are considered to be those for which there is model agreement on the sign and significance (i.e., the response is significant at the 95% confidence level in both models).

We analyse sixteen of the ETCCDI core indices, which are only briefly introduced here; full definitions can be in Zhang et al (2011). Specifically these are (official identifier in parentheses): frost days (FD; annual count of days when T_min < 0 °C), icing days (ID; annual count of days when T_max < 0 °C), cold nights (TN10p; % of days when T_min < 10th percentile), cold days (TX10p; % of days when T_max < 10th percentile), cold spell duration (CSDI; annual count of days with at least 6 consecutive days when T_min < 10th percentile), warm nights (TN90p; % of days when T_min > 90th percentile), warm days (TX90p; % of days when T_max > 90th percentile), warm spell duration (WSDI; annual count of days with at least 6 consecutive days when T_max > 90th percentile), wet days (R10mm; annual count of days when P_tot > = 10 mm), very wet days (R20mm; annual count of days when P_tot > = 20 mm), wet day precipitation (R95pTOT; annual total precipitation on days when P_tot > 95th percentile), very wet day precipitation (R99pTOT; annual total precipitation on days when P_tot > 99th percentile), precipitation intensity (SDII; average precipitation on days when P_tot > = 1 mm), longest wet spell (CWD; annual maximum number of consecutive days when P_tot > = 1 mm), dry days (Rnnmm; annual count of days when P_tot = 0 mm; see note below) and longest dry spell (CDD; annual maximum number of consecutive days when P_tot < 1 mm). We note that one ETCCDI index is user-defined (Rnnmm; annual count of days when P_tot is above a user-chosen threshold); here we apply a threshold of 0 mm and refer to this index as dry days. For the percentile-based indices, the percentile thresholds were determined from the late twentieth century simulations (had20c and cam20c) and kept the same for the evaluations of the changes in these indices in the late twenty-first century.

The indices were calculated at each model grid-point before regional averaging, using twelve regional domains (shown in figure 3(a)) based on those defined in the IPCC special report on managing the risks of extreme events and disasters to advance climate change adaptation (SREX; Seneviratne et al 2012) and later used in the IPCC fifth assessment report (IPCC 2013). These are: Alaska and western Canada (AWC; 105–168 °W 50–75 °N), eastern Canada and Greenland (ECG; 10–105 °W 50–85 °N), Scandinavia (SCA; 0–40 °E 58–85 °N), Siberia (SIB; 40–180 °E 50–85 °N), western United States (WUS; 105–130 °W 30–50 °N), central US (CUS; 85–105 °W 30–50 °N), eastern US (EUS; 60–85 °W 30–50 °N), central Europe (CEU; 10 °W–40 °E 45–58 °N), Mediterranean (MED; 10 °W–40 °E 30–45 °N, western Asia (WAS; 40–75 °E 30–50 °N), central Asia (CAS; 75–100 °E 30–50 °N) and eastern Asia (EAS; 100–145 °E 30–50 °N). We hereafter refer to these regions by the three letter abbreviations just provided.

To evaluate the models' ability to simulate realistic weather extremes, we compared climatological values of the ETCCDI indices (for the non-percentile-based indices) in the late twentieth century simulations to estimates of these quantities in the real world. Since there is large observational uncertainty, we utilize four different reference data sets: HadEX2 (Donat et al 2013), based solely on in situ observations, and three commonly used reanalysis products, ERA-Interim (Dee et al 2011), NCEP/DOE (Kanamitsu et al 2002) and NCEP/NCAR (Kalnay et al 1996). The two models simulate well the main spatial features of the observed climatologies of the ETCCDI indices (supplementary figure 1), for example, the latitude and altitude dependence of cold extremes and enhanced precipitation extremes in coastal and mountainous regions. Figure 2 shows the mean biases in the models, compared to each reference data set, averaged within our regional domains. More specifically, we area-averaged each index over our twelve regions and compared the regional averages derived from had20c and cam20c to those from the reference data sets. The biases were simply calculated by subtracting the reference value (area and time mean) from the simulated value, such that a positive bias implies the model, on average (in space and time), overestimates the quantity of interest. For most indices and regions, the magnitudes of the biases vary considerably depending on the reference data set used (in some cases the biases are of opposite sign). This prevents a meaningful quantitative assessment of the model biases; however, there are some consistent features of the biases (corroborated by multiple reference data sets) that enable qualitative statements on the ability of the models to simulate particular types, or characteristics, of extreme weather. The models have generally smaller biases for the 'event frequency' indices, for example frost days or wet days (R10mm), than for the 'event duration' indices. Most notably, both models substantially overestimate the mean cold spell duration (CSDI). They also tend to underestimate dry spell duration (CDD), but these biases are smaller than for cold spells. Relatively large biases are also identified for warm spell duration (WSDI), but the sign of these vary between the reference data sets, complicating their interpretation. Given the large observational uncertainty, biases in all the indices likely reflect inadequacies in the observations as well as in the models.

Zoom In Zoom Out Reset image size

Sillman et al (2013b) assessed the ability of the CMIP5 models (including HadGEM2-ES and CCSM4) to simulate observed global climatologies of the ETCCDI indices. HadGEM2-ES and CCSM4 were shown to perform comparatively well in this regard with generally small biases globally (see their figure 10), and to out-perform most other CMIP5 models (especially for the precipitation indices). Thus despite the substantial biases shown in figure 2, we consider these models to be two of the best available for studying future changes in extremes in response to sea ice loss. However, the projected changes in the ETCCDI indices arising from sea ice loss revealed in the following sections need to be viewed with a degree of caution in light of the model climatological biases in these indices.

Both models project significant reductions in frost days (figure 3(b)), icing days (figure 3(c)), cold nights (figure 3(d)) and cold days (figure 3(e)) over the high-latitude regions; namely, AWC, ECG, SCA and SIB. Both models also project significantly fewer frost days, icing days, and cold nights over certain mid-latitude regions; namely, CUS and EUS. Also, cold days decrease in both models over EUS and a small, but statistically significant, reduction in frost days is projected over MED. CAS is the only region for which both models project an increase in any of the aforementioned indices, a small, but significant, increase in cold nights. Projected changes in these indices over WUS, CEU, WAS and EAS are either statistically insignificant in one or both of the models, or are of opposite sign in the two models and thus, cannot be considered to be robust. Regarding the latter, frost days and icing days over WAS both decrease in HadGAM2 but increase in CAM4, and cold nights and cold days over WUS both increase in HadGAM2 but decrease in CAM4.

Zoom In Zoom Out Reset image size

As well as altering the frequency of cold extremes, Arctic sea ice loss may effect the duration of cold extremes. Indeed, a shift towards more prolonged cold extremes has been hypothesized (Francis and Vavrus 2012, 2015). Turning then to cold spell duration (figure 3(f)), both models project significant decreases over AWC, ECG, SCA, SIB, CUS and EUS, but a significant increase over CAS. The shift towards shorter cold spells, especially over the central and eastern US, is opposite to that hypothesized by some in response to Arctic warming.

In summary, both models depict the largest reduction in cold extremes and their duration over the high latitudes. Over mid-latitudes, there are consistent reductions in cold extremes and their duration over central and eastern US, but little change over Europe. Central Asia is unique in being the only region where cold extremes are projected to increase in frequency and length.

Both models depict significant increases in warm days (figure 3(g)) and warm nights (figure 3(h)) over the high latitude regions of AWC, ECG, SCA and SIB. Also, warm nights increase significantly over EUS in both models. In general over mid-latitudes, the changes in warm extremes are less robust than those in cold extremes. For example over WUS, HadGEM2 projects significant decreases in warm nights and warm days, whereas both indices significantly increase in CAM4. Similar model divergence is found for warm nights over CAS and warm days over CUS and EUS.

Next we consider warm spell duration (figure 3(i)), which like cold spell duration, has also been hypothesized to increase in response to Arctic warming (Tang et al 2014, Coumou et al 2015). Warm spells are projected to lengthen over AWC, ECG, SCA and SIB in both models in response to Arctic sea ice loss. However, no robust changes in this index are projected over the mid-latitudes.

In summary, robust increases in warm extremes and their duration are projected over the high-latitudes, but the mid-latitude changes are weak and non robust.

Significant increases in wet days (figure 4(b)) and in very wet days (figure 4(c)) are projected by HadGAM2 for all regions considered. Fewer of the regional changes are significant in CAM4, but those that are, are predominantly positive. CAM4 depicts significant increases in wet days over AWC, ECG, SIB, MED and CAS; and very wet days over the same regions (except AWC, where the change is insignificant). In contrast to HadGAM2, CAM4 projects a significant decrease in wet days and very wet days over CUS, and in wet days over SCA.

Zoom In Zoom Out Reset image size

Total wet day precipitation (figure 4(d)) and total very wet day precipitation (figure 4(e)) also increase significantly across the majority of regions in HadGEM2. Again the changes in these indices are weaker in CAM4. CAM4 projects significant increases in one or both these indices over ECG, SIB, MED and CAS; but a decrease in the former index over CUS.

The precipitation intensity index (figure 4(f)) documents changes in the average amount of precipitation on wet days. Thus, this metric is a measure of the severity of wet extremes rather than their frequency. Precipitation intensity increases significantly in all regions in HadGAM2. Significant changes are also projected in some regions in CAM4, but they are of mixed sign. Precipitation intensity increases significantly over CAS, but decreases significantly over AWC, SCA, SIB, EUS and EAS.

Next, the longest wet spell index (figure 4(g)) provides a measure of the annual average maximum duration of wet spells. The longest wet spell increases significantly in HadGAM2 for all regions except SCA. In CAM4, this index increases significantly for five of the twelve regions, specifically, AWC, ECG, SIB, CAS and WAS; and decreases significantly for CEU. The latter change mirrors decreases in wet day (and very wet day) precipitation and precipitation intensity over this region in CAM4, but is in stark contrast to the increases projected by HadGAM2.

In summary, the two models project contrasting changes in wet extremes. HadGAM2 depicts significant increases in the frequency, severity and duration of wet extremes over all or most regions. In contrast, the changes in CAM4 are of mixed sign and significance.

Dry days (figure 4(h)), i.e. days with no precipitation, are generally projected to decrease, especially over mid-latitude Europe and Asia. Both models depict significant reductions in dry days over SIB, EUS, CEU, WAS, CAS and EAS. The models diverge significantly over ECG (HadGAM2 increases; CAM4 decreases) and CUS (HadGAM2 decreases; CAM4 increases).

Lastly, the longest dry spell (figure 4(i)) is projected by HadGAM2 to decrease in all regions, except SCA. Significant decreases are corroborated by CAM4 over AWC, ECG, SIB and EAS. The only region with a significant increase in dry spell length, in either model, is SCA in HadGAM2.

In summary, there is an overall tendency for fewer dry days and shorter dry spells in both models, although these changes are significant over more regions in HadGAM2 than they are in CAM4.