Causes of recent changes in western North American snowpack
- First Online:
- Cite this article as:
- Kapnick, S. & Hall, A. Clim Dyn (2012) 38: 1885. doi:10.1007/s00382-011-1089-y
- 534 Views
Monthly snow water equivalent (SWE) station observations and gridded temperature data are used to identify mechanisms by which warming affects the temporal and geographical structure of changes in western North American mountain snowpack. We first exploit interannual variability to demonstrate the sensitivity of snowpack to temperature during the various phases of the snow season. We show that mechanisms whereby temperature affects snowpack emerge in the mid to late portion of the snow season (March through May), but are nearly absent during the earliest phase (February), when temperatures are generally well below freezing. The mid to late snow season is precisely when significant loss of snowpack is seen at nearly all locations over the past few decades, both through decreases in snow accumulation and increases in snowmelt. At locations where April 1st SWE has been increasing over the past few decades, the increase is entirely due to a significant enhancement of accumulation during the earliest phase of the snow season, when the sensitivity analysis indicates that temperature is not expected to affect snowpack. Later in the snow season, these stations exhibit significant snowpack loss comparable to the other stations. Based on this analysis, it is difficult to escape the conclusion that recent snowpack changes in western North America are caused by regional-scale warming. Given predictions of future warming, a further reduction in late season snowpack and advancement in the onset of snowmelt should be expected in the coming decades throughout the region.
KeywordsSnow water equivalentClimate changeClimate sensitivityTrendsSurface observations
The yearly recurrence of cold season mountain snowfall is a critical aspect of water resources in western North America. The water supply from spring and summer snowmelt runoff supports agriculture, urban areas, and the ecological health of coastal ocean and riverine environments. The mountain snowpack provides natural storage of precipitation from the cold season months until it gradually melts during the warm season and subsequently flows to sea level through rivers and aqueducts. Water supply and springtime runoff predictions are necessary for intra-annual water use planning and flood management, as well as for longer term planning, design, and operation of water resource systems (Brekke et al. 2009). Recent studies have documented the precarious balance between current water supply and a growing demand (Gleick and Palaniappan 2010; Rajagopalan et al. 2009) with the potential for climate change to require significant changes to water resource management (Brekke et al. 2009; Barnett et al. 2005). An understanding of the mechanisms causing changes in the mountain snowpack over the last half-century is of paramount importance to understanding future changes, and developing credible adaptation strategies to avert water shortages or flooding.
Given that global temperature has been increasing over the past century and that the warming is expected to accelerate due to anthropogenic forcing (Solomon et al. 2007), understanding how temperature affects snowpack is of the utmost importance if we hope to project future snowpack changes in western North America. As temperatures vary across the freezing point of water, snowpack in a mountainous region such as western North America would be directly affected through: (1) changes to the percentage of precipitation falling as rain or snow, and (2) melt anomalies.
These two direct temperature influences on snowpack would vary in importance over the snow season. The first mechanism, where temperature modulates precipitation partitioning between rain and snow, should be detected through anti-correlations between changes in snow water equivalent (SWE) and temperature anomalies. This should mainly be observed during the accumulation phase of the snow season. Several papers examine snowpack as a fraction of accumulated precipitation to assess this mechanism. Barnett et al. (2008) forced a hydrologic model with output from two general circulation models over the Cascades, Northern Rockies, and Northern Sierras to show that simulated January through March minimum temperatures are anti-correlated with April 1st SWE as a fraction of accumulated winter precipitation. Feng and Hu (2007) similarly showed that there are strong observed trends in seasonal mean SWE/precipitation ratios and November–March wet-day temperature over the Pacific Northwest and central to eastern United States with almost uniform negative correlations between the two metrics. Knowles et al. (2006) complements this work by exploring the phase of the seasonal cycle when temperature has its greatest influence on observed precipitation partitioning over western North America.
The second mechanism, where temperature modulates snowmelt, should also be detected through anti-correlations between changes in SWE and temperature anomalies, particularly during the melting phase of the snow season when temperatures rise above freezing during the day. However, unlike the first mechanism, this should mainly be observed during the later, melting phase of the snow season. Previous studies have utilized submonthly SWE measurements to ascertain the relationship between snowmelt and temperature. This work focuses on data available since the late 1970s; these include daily snowpack measurements from automated snow telemetry sites (Serreze et al. 1999) and snow cover products from the more recent satellite record (McCabe and Wolock 2010; Leathers and Robinson 1997). Streamflow measurements have also been used as a proxy for SWE to explore the relationship between temperature and snowmelt over the longer streamflow record, and show a trend towards earlier streamflow (Stewart et al. 2004, 2005; Lundquist et al. 2004). Studies using the available SWE record (from 1930 to present) have used linear regressions to connect mean temperatures with SWE for observations (Mote 2006; Cayan 1996) and variables derived from a hydrologic model (Mote et al. 2005). Kapnick and Hall (2010) used four monthly SWE measurements during the snow season to calculate a proxy for melt timing in California. They examined its relationship to temperature and found strong anti-correlations.
A few studies have briefly explored the intraseasonal variation in the influence of temperature on snowpack as it relates to long term climate variability. Brown and Mote (2009) directly compared average December through February temperature with February 1st and April 1st SWE trends in western North America to compare snow depth sensitivity calculations in observations against simulated conditions in Canada. Similarly, Hamlet et al. (2005) compared average December through February temperature with trends in simulated 1st of the month SWE for March, April and May. Mote (2003) calculated regionally averaged monthly SWE values across the Pacific Northwest to document interdecadal variability in snowpack and provide a qualitative discussion of its relationship with temperature and precipitation trends.
In spite of all these studies addressing the two main temperature-driven mechanisms affecting snowpack, it is not clear whether and how they have manifested themselves in snowpack trends. Increasing temperatures should lead to a lower percentage of precipitation falling as snow and enhanced snowmelt; however, these mechanisms are likely strongest during different phases (accumulation versus melt) of the snow season. Moreover, internally-generated or anthropogenic changes in accumulation at any phase of the snow season may also affect snowpack trends. For all these reasons it is difficult to interpret trends in time-integrated snowpack variables physically. This makes it all the more challenging to ascribe such trends to a clearly warming climate.
Studies of April 1st SWE provide a vivid illustration of this problem. They show that April 1st SWE has been decreasing over the historical record in most places. However, it is not clear that this decrease is due to simultaneous warming. Moreover, April 1st SWE has been increasing at a few locations, in particular over the southern California Sierras and southern Rockies (Mote 2006; Mote et al. 2005; Regonda et al. 2005). This has raised further questions about the robustness of the warming signal in the snowpack (Christy and Hnilo 2010; Howat and Tulaczyk 2005). All of this evidence could be consistent with a warming climate if strong evidence for the two mechanisms by which temperature can affect snowpack were found at all locations, including those where April 1st SWE is increasing.
In this study, we exploit interannual variability to demonstrate the sensitivity of snowpack to temperature during the various phases of the snow season. We show that the two mechanisms whereby temperature affects snowpack emerge in the mid to late portion of the snow season (March through May), but are nearly absent during the earliest phase (February). The mid to late snow season is precisely when significant loss of snowpack is seen at nearly all locations over the past few decades. At locations where April 1st SWE is increasing, the increase is entirely due to a significant enhancement of accumulation during the earliest phase of the snow season, when the sensitivity analysis indicates that temperature is not expected to affect snowpack. Later in the snow season, when temperature is expected to affect snowpack according to the sensitivity analysis, these stations also exhibit significant snowpack loss comparable to the other stations. Based on this analysis, it is difficult to escape the conclusion that recent snowpack changes in western North America are caused by regional-scale warming.
We present our study by first providing a description of observational data in Sect. 2 and the methods for analyzing the data in Sect. 3. We discuss the links between snowpack and temperature in Sect. 4. The trends in snowpack, accumulation, snowmelt, and temperature are discussed in Sect. 5. We summarize our main findings in Sect. 6. The physical interpretations of our findings and implications of our work are found in our concluding remarks in Sect. 7.
Temperature data are also used to diagnose changes in SWE and melt timing. Maximum and minimum daily gridded temperature data from 1930 to 2003 were downloaded from the web site of the Surface Water Modeling group at the University of Washington (http://www.hydro.washington.edu/Lettenmaier/Data/gridded/). The development of this data set is described by Hamlet and Lettenmaier (2005). It was chosen for its long temporal coverage (1915–2003) and high spatial resolution (1/8 degree) relative to other temperature products that include the pre-satellite era. It is also useful for our study, as it was developed to facilitate long-term trend analysis. For example, various methods were employed to adjust for the inhomogeneous temperature record, since stations dropping in and out of availability can lead to spurious local trends if not properly accounted for. Moreover, for the time period explored in our analysis (post-1950), the number and spatial distribution of stations used to develop the Hamlet and Lettenmaier (2005) product are fairly consistent (Bonfils et al. 2007). This temperature data has also been used in previous snowpack studies (Kapnick and Hall 2010; Casola et al. 2009; Das et al. 2009; Mote et al. 2005; Hamlet et al. 2005) and temperature studies (Bonfils et al. 2007, 2008). It covers four main regions on the western side of the Continental Divide: the Pacific Northwest and Columbia River, California, the Great Basin, and the Colorado River. (Fig. 7 shows the geographical coverage.)
To provide a physical interpretation of SCD, we tested the relationship between SCD and the actual peak date of SWE at a limited number of stations with daily SWE data. These stations are in California and were analyzed in Kapnick and Hall (2010). We first calculated the date of maximum SWE from February 1st to May 1st using daily data. For years with days of sustained maximum SWE, we averaged the measurement days of sustained maximum SWE to find the peak SWE date. We then aggregated the data from all the years and all available stations for correlation analysis. The monthly SCD, calculated with 1st-of-the-month snapshots, and date of maximum SWE from daily data have a correlation coefficient of r = 0.71 (p < 0.05) and regression coefficient of 1.33 (peak days per SCD days). Thus it appears that SCD anomalies are roughly proportional to anomalies in peak date, and that peak date variability can be assessed without the need for daily data through the SCD technique. Because of this close relationship between peak date and SCD, we use the terms interchangeably in the rest of this paper.
3.2 Temperature adjustments
The daily gridded temperature data were adapted for comparison to snow station data by averaging in time and adjusting temperatures for station elevations. Given that our snow station data is generally only available at monthly intervals, we converted the daily temperature data into monthly means of daily maximum and minimum temperatures. In analyses involving both temperature and snow station data, the local gridcell temperature is adjusted for elevation differences between the gridcell and snow station by assuming a constant lapse rate of 6.1°C/km. This lapse rate was used in the construction of the original gridded temperature product (Hamlet and Lettenmaier 2005) and is used here for consistency. Using observations from the National Weather Service Cooperative weather stations over the Pacific Northwest, Minder et al. (2010) showed that this value is too low for annual mean daily minimum temperatures, but is consistent for annual mean daily maximum temperatures. If a station is equidistant from two or more gridcells, an average lapse-rate adjusted temperature is calculated over the relevant gridcells. Unfortunately, some Wyoming and Colorado snow stations used in Sect. 5 fall outside of the region covered by the temperature product, and are thus left out of Sect. 4 where snow observations are compared to temperature data. Therefore, only stations within 2 grid cells (roughly 28 km) of available temperature data are included in these analyses.
3.3 Linear trends
To understand how the western North American snow pack has changed over the past few decades, we calculated linear trends from observations. Individual snow station and temperature gridcell linear trends (found in Figs. 4, 5, 6, 7) were calculated for each station or gridcell using all years with available data over the specified time period. Snow trends were only calculated at snow stations with records covering at least 75% of the time period being considered. Trend analysis was conducted using: (a) monthly SWE values, (b) month-to-month changes in SWE, (c) the centroid date of SWE over the snow season (discussed in more detail in Sect. 3.1), and (d) monthly maximum daily temperature. The calendar dates associated with the time dimension of the trend calculations are different for each of these variables: (a) monthly SWE trends use each observation’s measurement date, (b) month-to-month changes in SWE trends use the midpoint of the measurement dates of the adjacent months, (c) trends in the centroid dates of SWE use the mean of the measurement dates of the four observations over each snow season from mid-January to mid-May, (d) temperature trends are calculated based simply on the year of measurement, as the measurement dates do not change. Tables 2 and 3 provide western North American climate trends using two different methods to combine individual observations and aggregate them to create regional trends. These are quantitative means of assessing station and gridcell trends shown in Figs. 4, 5, 6, and 7. Individual stations and gridcells may exhibit trends that are counter to those found in the tables; discussion of these differences is provided in the text where the figures are introduced (Sect. 5).
For snowpack variables, the linear trend corresponding to all available snowpack data over a given time period was calculated. Rather than average snowpack trends of individual stations found in Figs. 4, 5, and 6 to create a regional trend, we treated the entire set of snowpack observations as one variable for trend analysis. This method gives equal weight to each snowpack measurement and is well-suited for a data set like snowpack that is inhomogeneous in space and time. A similar technique was used in the regional trend calculations of Kapnick and Hall (2010) and Cayan et al. (2001). There are a variety of other available methodologies for aggregating snowpack trends. For example, averages have also been taken over specific mountainous sub-regions over western North America to distinguish trends that may be driven by different local meteorological conditions (Pierce et al. 2008); we considered this approach, however, find that to the extent that there are distinct trends, they are easily discerned upon examination of the figures. Another popular approach, employed by Clark et al. (2001) and Mote (2003), first converts the time series of station measurements into z-scores by subtracting the mean and normalizing by the standard deviation; then these individual time series are averaged and converted back into SWE using the mean and standard deviation averaged over all of the time series. In Mote et al. (2005) this approach was compared with taking simple regional means, and the interannual and interdecadal variations were found to agree quite well. In light of this, we have chosen to adopt a simple approach for quantifying total western North American snowpack trends. We provide the geographic distributions of individual station trends in the figures to show the extent to which these means are representative of individual locations.
In contrast to snow variable trends, regional western North American temperature trends (found in Table 3) were calculated by finding the mean over the entire gridded product domain (see Fig. 7) to allow for a qualitative discussion of the influence of large-scale temperature trends on snowpack. Equal weight is given to each gridcell in determining the combined regional trend in temperature. This technique is more appropriate for a data set like temperature that is spatially and temporally homogenous. Trend analysis can be affected by the choice of start date due to availability of data (for snowpack variables) and interdecadal and inter-annual variability (for all climate variables). To determine the consistency of trends and to allow for comparison against studies using other time periods, we have also provided information on sensitivity of trend calculations to start date in these tables.
Previous studies have noted that historical monthly snow course measurement dates may vary (Cayan 1996), with a potentially systematic shift in the actual date of measurement towards later in the season (Mote et al. 2005). Thus if snow measurement is assumed to take place on a specific and unvarying date, typically the first of the month, errors could result. In reality the measurement can be taken within 2 weeks of the first of the month (Cayan 1996; Mote et al. 2005), resulting in approximately 29 potential measurement dates. To avoid this source of error in our analyses, we use the actual measurement date rather than assuming the measurement date is the first of the month. However, even if measurement dates are accounted for correctly, spurious positive or negative trends in SWE could still occur. For example, a systematic shift in measurement dates could result in a corresponding shift in the sampling of SWE to a significantly different point in the seasonal cycle (e.g. to earlier or later in the melt phase). To assess this potential error, we calculated the SWE-related analyses presented in this study using subsets of observations that were taken within either 5, 10, or 14 days from the first of the month (not shown in analysis). We found that including the measurements taken more than 5 days away from the first of the month does not alter our results. Since this error source appears insignificant, we have chosen to utilize measurements taken within 14 days of the first of the month to maximize the number of stations available for analysis.
4 Temperature sensitivity
We explore the relationship between snowpack and temperature in this section. We exploit interannual variability to identify the phases of the snow season when SWE is most sensitive to temperature. All results rely on monthly means of daily maximum temperatures.
Correlation between SWE metrics and temperature from 1950 to 2003
SWE metric observation
Month of temperature
Total no. of stations
No. of significant stations
Feb to Mar
Mar to Apr
Apr to May
Mean Mar & Apr
The distribution of station correlations between monthly temperature and changes in SWE and the steadily increasing correlations between temperatures and SWE changes as the snow season progresses can be explained by the relationship of the mean station temperature to the freezing point of water. In Fig. 2d–f the same correlations displayed in Fig. 2a–c are plotted against the mean maximum station temperature (i.e. the average for the entire record of monthly mean maximum daily temperatures at each individual station). During the colder months of February and March (panels d and e) mean temperatures cluster near the freezing point, and nearly all stations experience accumulation (green circles). For these accumulating stations, there is a strong linear relationship between mean temperature and the correlations between temperature anomalies and SWE changes. As mean station temperatures increase, the correlations depart from zero, becoming more and more negative. This is because the mean maximum temperature is a surrogate for the frequency with which maximum temperatures breach the freezing point. The more frequently maximum daily temperature rises above freezing, the more likely it is that precipitation will fall as rain rather than snow and the stronger the negative influence of temperature on SWE changes. If the mean maximum temperature falls below about −2°C, precipitation almost always falls as snow. The changes in SWE are controlled almost exclusively by precipitation anomalies and the anti-correlation with temperature disappears. This pattern may be interpreted as a confirmation of the emergence of the first temperature mechanism, as temperatures increase and move beyond the freezing point. Snowmelt also influences SWE, but only when temperatures rise significantly above freezing; the partitioning of stations exhibiting melting or accumulating SWE on average allows us to identify the dominant mechanism in each month. The geographical pattern shown in Fig. 2a, b is similar to previous work exploring the geographic distribution of the influence of daily snowmelt on April 1st SWE. Using daily SWE data for about 20 years, Mote (2006) showed that total melt from October through March is anti-correlated with April 1st SWE across western North America, but with the weakest influences found in the Rocky Mountains due to daily melt events being insignificant. We therefore conclude that in Fig. 2a, d, and to a lesser extent Fig. 2b, e, the neutral and weak anti-correlations are caused by the dominance of precipitation in driving SWE anomalies.
Most stations generally experiencing melt are found in April, and to a lesser extent March (grey circles in Fig. 2e–f). At these stations and during these months, mean temperatures are enough above freezing that changes in SWE and temperature anomalies are consistently anti-correlated, no matter what the mean temperature. (See the tight cluster of grey circles on the left side of Fig. 2f). Warmer temperatures therefore consistently result in more melt. This may be interpreted as the emergence of the second temperature mechanism as the snow season progresses to the melt phase. There is also a small population of grey circles generally associated with very high mean temperatures and relatively low correlations. These may correspond to stations where the change in SWE simply represents the final loss of snowpack through melt and thus is determined by the amount of SWE remaining rather than any other factor such as temperature.
As we showed in Sect. 4.1, the relationship between temperature and SWE changes becomes progressively stronger over the course of the snow season, as the first, and then the second temperature mechanism (snowmelt and precipitation partitioning respectively) become increasingly important. This is also reflected in the relationship between monthly temperature anomalies and SCD. When individual station SCD time series are correlated with the various monthly temperature time series (Table 1), it becomes apparent that mean spring temperatures have the greatest influence on peak snowmass timing. Interestingly, mean spring temperatures (mean March and April) have a greater influence on SCD (r = −0.58 for all stations) than either month separately (r = −0.40 and r = −0.50 for all stations in March and April respectively).
The distributions of correlation and regression coefficients seen in Fig. 3 are expected from our analysis of the relationship between temperature and monthly changes in SWE from Fig. 2. In contrast to the late season correlations, February correlations (Fig. 3; circled line) have a nearly bell-shaped station distribution centered over 0. This is consistent with the relatively weak influence of February temperature anomalies on inter-month SWE changes (Fig. 2a, d). The generally greater anti-correlations between mean record temperature and monthly change in SWE over March and April (Fig. 2e, f; Table 1) is similarly mirrored in Fig. 3 with greater negative distributions between mean spring temperatures and SCD.
In this section we explore the simultaneous trends in snowpack variables and temperature. Combined with the results of Sect. 4, we can identify the mechanisms underpinning the relationships between trends in snowpack and temperature.
Trends in monthly SWE and delta SWE with variations in start date
February to March change in SWE
March to April change in SWE
April to May change in SWE
Together Figs. 4a and 5a allow us to assess how snow accumulation has changed over the time period ending with March. The positive trends prior to February found mainly in California and the southern Rocky Mountains (Fig. 4a) were discussed above. Positive trends are also seen from February to March at many more stations, and snow accumulation seems to be increasing at more locations than it is decreasing during this time period. Positive accumulation trends continue to be most pronounced in the California Sierras and southern Rocky Mountains, with most other stations exhibiting weakly positive accumulation trends.
A dramatic reversal occurs during the month of March at nearly all stations, countering any increased snowfall occurring earlier in the season. Nearly all accumulating stations exhibit large negative trends indicating strong reductions in snowfall (Fig. 5b). Meanwhile, at melting stations, trends are also very negative during this time period, indicating large increases in melting (Fig. 5e). Since April SWE trends are more consistently negative than the trends in accumulation from February to March (compare Figs. 4c and 5a), the large and significant decreases in snowfall and increases in snowmelt during the month of March (Fig. 5b, e) must overwhelm any positive trends seen earlier in the snow season. The unique locations in southern California and the southern Rocky Mountains that exhibit positive trends in April SWE (Fig. 4c) are thus driven by increases in early season accumulation overwhelming March to April reductions in SWE (Fig. 5b, e) as exhibited by changes in integrated pre-February accumulation (Fig. 4a) and February to March inter-month accumulation (Fig. 5a).
During April, many stations also show reduced accumulation and increased melt (Fig. 5c, f), though the trends are generally smaller than during March and exhibit less statistical significance. There is one station in California, “Lower Lassen Peak”, with a significant increase in accumulation from April to May (Fig. 5c) over the record (6 cm/decade). This value is driven in large part by an extreme snowfall event from April to May in 2003. (More than 250 cm fell between April and May; the average is only 9 cm.) The trend is reduced to 2 cm/decade if this one extreme year is removed from the station data. This station is also anomalous because it experiences snow accumulation later in the season than other stations in the Feather River basin due to its unique combination of topographic variables (Jost et al. 2007; Regonda et al. 2005). This is not at odds with our findings in Sect. 4.1 and Fig. 2 however. A few unique stations at extremely high elevations have climatologically cooler temperatures than the rest of the region and can thus still accumulate snow during the month of April. Large late season storms that would otherwise result in significant melt during rain on snow events at lower elevations therefore lead to enhanced accumulation at higher altitudes. Accumulating stations still exhibit negative correlations between SWE and temperature (green circles in Fig. 2c). But warming at some of these high elevations has not increased enough to reduce accumulation (as exhibited by the single positive SWE point and neutral changes in Fig. 2c). Neglecting this unique station, accumulating stations (Fig. 5c) either experience reduced or neutral accumulation from April to May. Melt is either enhanced, neutral, or slightly reduced during this time period (Fig. 5f). Overall, Fig. 5 shows that the dominant and most geographically uniform change to the snowpack occurs during the month of March. The strong and widespread decreases in snowfall and increases in snowmelt in March to April SWE, even at stations showing significant positive trends in April 1st SWE, are highly suggestive of a coherent and strong forcing during this part of the season.
The last three rows of Table 2 provide a quantitative aggregation of individual trends shown in Fig. 5. The region exhibits some positive statistically significant trends from February to March and from April to May over different time periods. In later decades, February to March trends in inter-month changes in SWE become more positive, consistent with the general predominance of stations with accumulation increases in Fig. 5a. April to May trends become consistently negative for later periods. However, these trends are generally small in magnitude compared to the March to April figures, which are on the order of −1 to −2 cm/decade, and increase with later analyses periods. And, the only inter-month interval with consistent statistically significant trends for every start year is March to April. The information in this table is consistent with the idea that a modest increase in accumulation generally occurred during February followed by a strong and coherent forcing during March causing a much more significant snowpack loss.
Trends in SCD and temperature with variations in start date
All station SCD (−2008)
Number of stations
Select station SCD (−2003)
Number of stations
Averaged March & April temperature
Temperatures in western North America have predominantly increased over the available record. However, there are seasonal and geographical variations in temperature trends. Figure 7 displays monthly-averaged maximum daily temperature trends from 1950 to 2003 for: (a) February, (b) March, (c) April, and (d) mean spring temperature (mean March and April). The month of March generally exhibits the largest warming—on the order of 0.5°C/decade at almost all locations, creating a similar pattern in the mean March and April temperature trend. We hypothesize that this has been due to uniform snowpack loss during March resulting in lower surface albedo as snow melts across the region, causing further warming. The surface albedo feedback should be the strongest when snow cover coincides with rising amounts of solar radiation (i.e. during the spring) (Hall 2004). A few small isolated areas exhibit negative trends, especially in February, which also exhibits the weakest warming trends. Given internal climate variability, it is possible—even likely—that isolated locations will exhibit negative trends without contradicting the ultimate conclusion that there has been a uniform warming signal across the region. The main regions with cooling in Fig. 7 coincide with the main regions of increased SWE (the California Sierras and southern Rockies). It is possible the SWE increases have played a role in the associated cooling trends.
Table 3 gives the mean trends in temperature for the entire gridded product domain. Sensitivities for various start years are also provided. For March, April, and mean March and April temperatures, the warming trends are larger when start years are chosen from later decades. There are no large scale cooling trends, and all but 2 entries show warming trends, with March generally exhibiting the strongest trends and February the weakest, consistent with Fig. 7. We discuss the relationship between trends in SWE metrics and temperature in the final two sections.
In this paper, we examined the interannual variability in snowpack and explored its relationship to monthly temperature during different phases of the snow season. This analysis allows us to interpret the causes of recent changes in mountain snowpack over western North America. The strength of the temperature influence on SWE is directly proportional to mean station temperatures as they approach the freezing point of water in the early part of the snow season (Fig. 2d, e). In particular, during the month of February, the mean daily maximum temperatures have had an almost equal probability of being below or above freezing. For stations with mean subfreezing temperatures during February, temperature has little influence on snowpack. As a result, snow accumulation has increased at these unique locations in California and the southern Rocky Mountains until March while total precipitation has increased.
After February, there is a greater direct influence of temperature on snowpack (as measured by absolute correlations in Table 1). The region is dramatically warmer; mean maximum daily temperatures are almost always above freezing. Above-freezing temperatures lead to strong anti-correlations with SWE as temperatures rise significantly above freezing in March (Fig. 2). The regression coefficients corresponding to the correlations shown in Fig. 3b support the strong influence of late season temperature on the timing of the snowpack peak. Individual stations exhibit an SCD sensitivity of up to −6 days/°C, which translates into a peak SWE date or snowmelt onset sensitivity of up to −8 days/°C (using the relationship between SCD and peak SWE found in Sect. 3.1).
The analysis of the relationship of interannual variability in snowpack to temperature anomalies allows us to interpret the relationship between trends in SWE metrics and temperature. During February, any snowpack trends are unlikely to be attributable to the weak February warming trend, given the lack of sensitivity of snowpack variables to temperature during this phase of the snow season. Instead, February snowpack trends are likely attributable to accumulation trends. However, after February significant warming has been observed over the last few decades. And given the high degree of sensitivity of snowpack to temperature during the mid to late snow season seen in the interannual variability analysis, it seems very unlikely that the simultaneous loss of snowpack is not closely related to this warming. During the month of March, uniform warming has probably resulted in reduced accumulation and enhanced snowmelt at all stations (Fig. 5b, e). With the exception of a few high elevation stations in California and the southern Rocky Mountains, these secular, temperature-driven changes in snowpack combine to form a mainly negative trend in April SWE across western North America in all locations that is enhanced through May (Fig. 4c, d). As springtime temperatures continue to warm, the onset of snowmelt will continue to shift earlier and post-February snowpack values will continue to decline.
The majority of warming has thus far happened during months when maximum temperatures are above freezing (March and April, Table 3). These are also the months when the snowpack shifts into the melt phase of the season (as measured by the distribution of stations exhibiting melt in Fig. 5d–f). The warming in these months may have been enhanced through surface albedo feedback due to the loss of snowpack and increased patchiness of existing snow cover. In the future, the much larger warming signal currently seen in March and April may shift to earlier in the season as snowpack decreases also shift earlier through increased snowmelt and increased percentage of precipitation falling as rain.
As a result of the intraseasonal changes in the influences of precipitation and temperature on snowpack and variable elevations across western North America, certain geographical and temporal trends in SWE measurements may at first appear to be at odds with the idea that a warming climate is influencing snowpack. These include the positive trends in April SWE seen in Fig. 4c and noted in previous literature (Kapnick and Hall 2010; Barnett et al. 2008; Mote 2006; Howat and Tulaczyk 2005; Mote et al. 2005). Here we show that these trends can be attributed to increased accumulation prior to March. These positive snowpack trends do not prove that temperature has little influence on snowpack however. In fact, during the months when snowpack is expected to be sensitive to temperature, these stations show comparable snowpack loss to other stations in association with warming. For example, stations with strong April SWE trends generally show reduced accumulation or enhanced melt trends from March to April (Fig. 5b, e) when there have also been significant warming trends (Fig. 7). The peak date of SWE is also shown to shift earlier in the season at these locations. Given the sensitivity of peak date to temperature at all locations (Fig. 3), this is likely the result of warming. By analyzing monthly SWE measurements using inter-month changes in SWE and the SCD metric, we can isolate various phases in the seasonal cycle to show that the direct influences of temperature increases on snowpack can be detected in the observational record at nearly all locations.
A few locations appear to be experiencing hydrologic intensification, particularly prior to March. For example, high altitude stations in California and the southern Rocky Mountains have experienced enhanced accumulation in the early part of the snow season (Figs. 4a, 5a). These precipitation increases could be the result of anthropogenic intensification of the global hydrologic cycle. However, the increases in accumulation and their expression in the positive April 1st SWE trends shown in Fig. 4c may also arise from internally-generated accumulation anomalies in the early part of the season. In some cases these may be large enough to counter the significant loss of snowpack during the later part of the season due to a warming climate. These unique locations may continue to experience increased snowfall during months of sub-freezing temperatures particularly given the lower warming rates in February. However, given our analysis of the influence of temperature on snowpack at the same stations later in the season and lower elevation stations in February and March, we can conclude that as temperatures rise above freezing more often, these locations will eventually exhibit statistically significant negative correlations between inter-month SWE and temperature, and rising temperatures would therefore eventually cause snowpack loss.
The analyses presented in this paper rely on the assumption that SWE station observations allow for meaningful interpretation of snowpack variability over the region. Given that stations are single point locations and do not measure the entire snowpack mass covering western North America, it is important to note that some changes may not be captured in this analysis. Snow cover located near or below the freezing line will be more susceptible to the minor warming trends than locations at higher elevations with colder temperatures. To the extent these locations are undersampled, the real snowpack may be experiencing a greater extent of snowpack loss than suggested by the station trends presented in Sect. 5. Figure 2d, e support this by providing evidence that locations with temperatures close to and above freezing should be more susceptible to temperature fluctuations than those with sub-freezing climatologies. Once temperatures are significantly above freezing, this systematic relationship between the sensitivity to temperature and mean temperatures breaks down as the snowpack simply melts regardless of temperature. Therefore, locations at lower elevations may be experiencing more pronounced and earlier season snowpack loss loss, particularly in February and March, than those covered by the analyses presented in this study.
The message of a declining late season western North America snowpack parallels findings of other studies using observations (Barnett et al. 2008; Mote 2006; Mote et al. 2005; Cayan 1996), hydrologic models (Barnett et al. 2008; Hamlet et al. 2005), regional models (Kim et al. 2009; Leung et al. 2004), and global climate models (Das et al. 2009; Pierce et al. 2008; Maurer 2007). They underscore a fundamental shift in the availability of water coming from late spring and summer snowmelt towards earlier in the season. This shift has various implications, including: changes to the statistics of flood risk from rain on snow events (McCabe et al. 2007), less warm season water supply in regions without sufficient water reservoir capacity, reduced effectiveness of hydroloelectric dams in the absence of pumping water upstream (Vicuna et al. 2008; Christensen et al. 2004), and increased risk of dry season wild fires (Westerling et al. 2006).
The potential negative impacts of increased runoff during February and March and reduced warm season snowmelt runoff highlight the need for confident predictions of future snowmelt timing and snowpack values. Future changes to total precipitation and thus early snowpack values when temperatures are below freezing (February and to some extent March in our study) however, remain unclear and do not show as consistent or significant of a signal in this observation study or in other global climate modeling studies (Solomon et al. 2007). Even if snow accumulation in the early cold season increases or remains unchanged, continued warming in March and April will very likely significantly shift the onset of melt earlier and reduce late season snowpack and summer snowmelt runoff.
The authors would like to thank Philip Mote and one anonymous reviewer for their comments that greatly improved this manuscript. We are grateful to Christopher Cham for his help in compiling SWE observations for all stations used in this analysis. Sarah Kapnick’s work was supported by NASA headquarters under the NASA Earth and Space Science Fellowship Program—Grant #07-Earth07F-0232 and a 2010 Switzer Environmental Fellowship. Alex Hall is supported by NSF grant AGS-0735056. The writers would also like to gratefully acknowledge funding provided by JPL (#1312546) under a project entitled: “Evaluating key uncertainties in IPCC Climate Change projection of California snowpack: Topography, snow, physics and aerosol deposition.”