The linear trends of observed SWEs, presented in Fig. 3a for the February, March, and April months, display large spatial variability and also show differences among the defined sub-regions. Grids located in the eastern part of the ‘NAS’ and ‘NAN’ regions show statistically significant decreasing trends for SWE, which are mainly due to increasing temperatures for the same months (Fig. 3b) and/or decreasing cumulative precipitation from the December to spring months (Fig. 3c). Grids located in the ‘WES’ and ‘WEN’ regions generally exhibit decreasing trends for SWE, which are again generally due to the increasing temperatures for the same months, although some increases in cumulative precipitation from the previous December to spring months can be noted. On the contrary, many grids located in the region ‘EES’ and few grids located in the region ‘EEN’ produce statistically significant increasing trends for SWE, which could be due to significant increases in the cumulative precipitation from the previous winter to spring months (Fig. 3c). Peng et al. (2010) and Cohen et al. (2012; 2014) reported an increase in mean winter (December–February) snow depth and a decrease in winter temperature for the ‘EE’ region during the recent two or three decades. Changes in storm track, jet stream, and planetary waves, and their associated energy propagation induced by recent Arctic warming have been suspected as a main cause for widespread winter cooling for this region (Cohen et al. 2012, 2014).
Observed SWE for the hemispherical to the sub-continental scale regions shows overall decreasing trends for the 1980–2012 period, except for the regions ‘EE’ and ‘EES’ (Fig. 4). Statistically significant decreasing trends are observed for the regions ‘A’, ‘N’, ‘WE’, ‘NA’, ‘WEN’, and ‘NAN’ for February, for regions ‘N’ and ‘NAN’ for March, and for regions ‘N’, ‘NA’ and ‘NAS’ for April, at the 90 % confidence level (i.e., p value of the two-tailed test is smaller than 0.1). Spring SWE for the region ‘EES’, however, displays statistically significant positive trends at the 90 % confidence level, which also results in weak upward trends for the region ‘EE’, particularly for March and April. The general decrease in the spring month SWE noted here is consistent with previous studies (Takala et al. 2011; Luojus et al. 2011; Li et al. 2014) at both hemispheric and sub-continental scales. The increase in spring SWE, for the southern part of East Eurasia, during the last few decades is also consistent with the results of Bulygina et al. (2009, 2011) based on 820 meteorological station data and those of Wu et al. (2014) based on satellite-based passive radiometer data SSM/I.
Linear trends of the spring SWE estimated from the ensemble mean signal of multi-model and each CMIP5 model for the ALL forcing case are presented in Fig. 5. The CMIP5 models generally reproduce the observed decreasing trends for the ‘NA’ and ‘WE’ regions and the observed increasing trends in some eastern parts of the ‘EES’ and western parts of the ‘WEN’ regions. The models, however, indicate significant decreasing trends for the western part of the region ‘EES’, where the observation yielded some increasing trends. They also show increasing trends for the ‘EEN’ and ‘NAN’ regions, where the observation generally yielded decreasing trends. All six CMIP5 models generally show large mean biases for the ‘EES’, ‘WEN’, and ‘NAN’ regions (Fig. 2b). However, NorESM1-M reproduces well the observed temporal tendency of the spring month SWEs, especially for regions ‘EEN’ and ‘NAN’, although it yielded the lowest spatial correlation coefficient and large mean biases when compared with observations for the spring SWE amounts as shown in Fig. 2.
Prior to applying the D–A analysis for the spring (February–April average) SWE changes, the simulated signals for the three different forcings (i.e., ALL, GHG, and NAT) are compared with observations. Figure 6 presents annual series of spring SWE anomalies for observation and simulations for all three forcing cases, for region ‘A’. The anomalies observed and modelled for ALL and GHG cases show a clear decreasing trend for the last three decades. However, the multi-model ensemble mean displays smaller inter-annual variability than observation and also individual ensemble member signals. Low temporal coherence among the individual runs, induced by the diversity of the CMIP5 model physics and structures and initial conditions, can result in small inter-annual variability of the multi-model ensemble mean (Jeong and Kim 2009). The spring SWE anomalies simulated by the CMIP5 models based on the NAT forcing do not show any decreasing trend (Fig. 6c). However, effects of volcanic eruptions are reflected in the SWE anomalies. For instance, the spring SWE anomalies of both observation and simulations based on NAT forcing show a small increase for few years after 1992 as a response to the Pinatubo eruption, which was also noted by observed Rupp et al. (2013) in their study, but for SCE.
Linear trends and their 5–95 % confidence intervals of the spring SWE anomalies estimated from multi-model ensemble mean signals, for the three different cases (ALL, NAT, and GHG) are compared to those from observations for the 12 different regions for the 1980–2012 period in Fig. 7. The multi-model ensemble mean signal for ALL forcing reproduces the observed decreasing trend for the hemispheric-scale region ‘A’, while larger differences are noted for the sub-hemispheric scale regions ‘S’ and ‘N’. As shown in Figs. 2 and 5, the multi-model ensemble mean with ALL forcing fails to reproduce the statistically significant decreasing trends for regions ‘EEN’ and ‘NAN’, which results in the misrepresentation of the observed decreasing trend for region ‘N’. The multi-model ensemble mean also fails to reproduce the observed statistically significant increasing trend for the region ‘EES’, which results in an overestimation of the observed decreasing trend for region ‘S’. However, the multi-model ensemble mean reproduces well the observed decreasing trends for other regions. The multi-model ensemble mean with GHG forcing displays a very similar pattern to that with ALL forcing, implying that the GHG forcing is the main external forcing of the spring SWE changes in the CMIP5 model simulations. The 5–95 % confidence intervals of the linear trends of the multi-model ensemble mean based on ALL and GHG forcings are much smaller than those for the observation for all regions, implying that the inter-annual variability of the multi-model ensemble mean is smaller than that observed. The multi-model ensemble mean based on NAT forcing produces statistically insignificant small trends for all regions. The internal variability estimated from the CTL simulations (column 5 of Fig. 7) is higher for the region ‘N’ compared to the region ‘S’ and similarly increase with decreasing area. The regions, which have large internal variability, generally yield large confidence intervals for the linear trends for both observation and the multi-model ensemble.
Figure 8 shows the temporally smoothed time series of spring SWE anomalies observed and modelled (for ALL), for the hemisphere scale region ‘A’. The smoothed observation shows a significant downward trend particularly after 1996 and an important decreasing for the last 6 years. The individual signals obtained from the six CMIP5 models also suggest a decreasing trend for the SWE anomalies. The multi-model ensemble mean of the 33 ALL forcing runs reproduces the observed decreasing trend, though it displays smaller temporal variability compared to both observation and individual ensemble members. The individual CMIP5 models reproduce the observed decreasing trend in the smoothed observation to varying degrees. BCC_CSM1-1, CNRM-CM5, GISS-E2-H, and NorESM1-M reproduce the temporal variability of the smoothed observation relatively well, compared to CanESM2 and GISS-E2-R in terms of Pearson’s linear correlation coefficient. Ensemble means of the individual models tend to exhibit larger variance than the multi-model ensemble mean. However, ensemble means of individual CMIP5 models yield lower linear correlation coefficients with observations than the multi-model ensemble mean.
Figure 9 shows scaling factors and their 5–95 % confidence ranges estimated by the D–A analysis of the multi-model ensemble mean signals for the three forcing cases for the 12 different regions. The scaling factors of both ALL and GHG forcings are significantly greater than zero for the hemispheric scale region ‘A’, indicating that the combined effects of external anthropogenic and natural forcings or the effect of GHG forcing alone are detected in spring SWE decline for the landmass north of 45°N. Moreover, the 5–95 % confidence ranges of the scaling factors include unity, indicating that the simulated signals of the spring SWE decreases under the ALL and GHG forcings are consistent with the observed spring SWE decrease at the hemisphere-scale. However, detection and attribution of ALL and GHG forcings to spring SWE changes are less clear for the smaller regions. Scaling factors of both ALL and GHG forcings are statistically significant for the regions ‘N’, ‘WE’, ‘NA’, and ‘NAS’. However, some of them (i.e., scaling factors for regions ‘N’, ‘NA’, and ‘NAS’ for ALL and regions ‘N’ and ‘NA’ for GHG) indicate that the simulated signals of the CMIP5 models, as a group, are not consistent with observations and tend to underestimate the spring SWE response to external forcings due to improper reproduction of temporal variability of individual models, as their 5–95 % confidence ranges do not include unity. The observed decreasing trends of spring SWE were reproduced properly by multi-model ensemble signals at the sub-continental regions ‘WES’, ‘WEN’, and ‘NAS’ (Fig. 7). However, the combined effect of ALL forcing or GHG forcing alone are not detected clearly by the D–A analysis, indicating that the observed and simulated spring SWE changes have large noise range, which makes it difficult to detect signals outside the noise range for the small sub-continental scale regions. Scaling factors and their 5–95 % confidence ranges of the NAT forcing only are much larger than unity for the regions ‘A’, ‘WE’, ‘NA’, and ‘NAS’, include zero for regions ‘EE’, ‘WES’, ‘WEN’, and ‘NAN’, and is negative for regions ‘N’, and ‘EEN’, indicating that the effect of NAT forcing alone are generally undetectable or inconsistent with the observed spring SWE changes for the considered regions.
Figure 10 compares best estimates of scaling factors and their 5–95 % confidence ranges for the multi-model ensemble mean and that of single-model ensemble means for the hemispheric scale region ‘A’. The effects of ALL and GHG forcings are generally detected by most individual model simulations, except for BCC_CSM1-1 for GHG forcing and CanESM2 for ALL forcing. The single-model ensembles, however, tend to produce larger 5–95 % confidence ranges of scaling factors for the ALL and GHG forcings compared to the multi-model ensemble, except NorESM1-M, supporting robustness of the detection results from multi-model ensemble mean. This result roughly implies that individual models usually have larger uncertainty in the estimation of SWE responses (fingerprints) (as shown in Fig. 8) and, therefore, the detection and attribution analysis based on the multi-model could provide more robust results than individual models, which has also been reported through the D–A analyses for the global surface temperature (Gillett et al. 2002) and Northern Hemisphere spring SCE (Rupp et al. 2013).