The role of winter precipitation and temperature on northern Eurasian streamflow trends

Authors


Abstract

[1] Eurasian river discharge into the Arctic Ocean has steadily increased during the 20th century, and many studies have documented the spatial distribution of the trends and hypothesized the causes. There is a large variation in the scope of these studies, including the spatial scale of interest, and they often lack consistency in the time period analyzed. Studies have shown a connection between changes in the seasonal snowpack and discharge, but they have been constrained by the limitations of the snow observational network, which contains few long-term stations. This study overcomes these problems by using both in situ observations and a land surface model to evaluate the role snowpack changes have had on increases in runoff across northern Eurasia from 1936 through 1999. Our analysis shows consistent trends in both observations and model predictions. Increases in cold season precipitation propagate into increases in maximum snow water equivalent, which lead to increases in runoff. A series of model experiments demonstrate that the nonlinear interaction between winter precipitation and temperature has driven changes in the snowpack, which are manifested in the modeled runoff trends. Given that winter precipitation is expected to continue to increase and temperatures to warm during the 21st century in this region, these results point to the importance in understanding how the projected changes will influence the seasonal snowpack, which may have important consequences for streamflow in this region and freshwater export to the Arctic Ocean.

1. Introduction

[2] Streamflow has increased 7% across northern Eurasia during the 20th century [Peterson et al., 2002], and numerous studies have hypothesized different causes of the increase, such as permafrost melt, increased precipitation, and dams and reservoirs regulations [Berezovskaya et al., 2004; Groves and Francis, 2002; McClelland et al., 2006; Pavelsky and Smith, 2006; Rawlins et al., 2009]. Permafrost melt alone cannot explain the trends: 4 m of permafrost would had to have thawed evenly across the entire permafrost zone to explain the discharge trends [McClelland et al., 2004]. This has led to speculation that the increases must, at least in part, come from precipitation [McClelland et al., 2004], but the annual trends in precipitation over the entire region are inconsistent with annual streamflow trends. Adam and Lettenmaier [2008] examined annual precipitation and streamflow trends and found that in the coldest basins streamflow trends exceeded precipitation trends, but in warmer basins the opposite was true. They hypothesize that in the coldest basins, storage changes in the form of melting ground ice explain the difference in trends, whereas in the warmest basins, increasing precipitation leads to increases in both streamflow and evapotranspiration. In the threshold basins, they hypothesize that it is a combination of both permafrost degradation and evapotranspiration effects. Berezovskaya et al. [2004] compared basin precipitation trends with streamflow trends and found little consistency between precipitation and streamflow. The Yenisei had a positive streamflow trend with a negative precipitation trend; the Lena had only a weak precipitation increase that was not sufficient to explain the streamflow increase. Pavelsky and Smith [2006] performed a compatibility analysis between annual streamflow and precipitation for small basins throughout the region, concluding that precipitation plays a role in explaining streamflow trends for many, although not all, of the basins. They also found that permafrost melt could be playing a role in streamflow trends, although they concluded it was not a strong role.

[3] Reservoir regulation has played a role in seasonal trends, causing increased winter streamflow and decreased spring and summer streamflow. Of all the rivers, the Yenisei is the most regulated, with winter flows increased by up to 85% in the upper basin [Yang et al., 2004a]. Adam et al. [2007] also found similar results for the Yenisei, as well as for the Ob and Lena basins. However, they also found that dams cannot explain the increase in annual streamflow [Adam et al., 2007].

[4] Rawlins et al. [2009] hypothesized that the driver of streamflow trends could be a change in the seasonality of precipitation, with more precipitation occurring in the winter and less in the summer, resulting in little or no trends in annual precipitation. Several studies have provided evidence for this change in precipitation seasonality [e.g., Rawlins et al., 2006, 2009]. Bulygina et al. [2009] documented an increase in mean snow depth during the continuous snow cover period in the western half of the domain with similar results for maximum snow depth (study period 1966–2007). Between 1936 and 1983, snow depth increased over most of northern Russia and decreased over southern Russia, with the increase in the north exceeding the decrease in the south [Ye et al., 1998]. Popova [2007] showed, through the use of empirical orthogonal functions (EOFs), that there is snow accumulation across portions of the region. The first principal component of snow depth variability between the White Sea and the Lena in the north corresponded to the NAO. Winter precipitation is expected to increase with warming temperatures [Kattsov et al., 2007] with attendant increases in runoff [Intergovernmental Panel on Climate Change (IPCC), 2007]. Understanding the linkage between snowpack changes and runoff is therefore instructive for understanding future climate changes.

[5] Shiklomanov et al. [2007] demonstrated a correlation between winter precipitation and spring daily maximum streamflow, with decreasing maximum streamflow in southern Siberia and the Kolyma basin and increases in other regions. This is consistent with the individual basin trends calculated by Peterson et al. [2002]. Relationships have also been drawn linking snow cover extent and streamflow. Yang et al. [2003] demonstrated that high snow cover extent led to an increase in streamflow during the snowmelt period. Over the Yukon River, snow cover extent was shown to correspond to streamflow during the melt season, and snow cover extent and snow water equivalent were closely related [Yang et al., 2009]. Although warming temperatures in the region could lead to earlier snowpack melting, which could potentially lead to changes in streamflow, Ye et al. [2004] showed no significant trend in the timing of the last day of snow cover in the region.

[6] These studies have shown a link between the snowpack and runoff trends, but they have been constrained by the limitations of the snow and precipitation observational networks, which are sparse over Siberia and the Russian Far East. Because long-term measurements were not always continuously taken, stations may come in and out of the network, which can affect the calculated climatology [Rawlins et al., 2006] and therefore the trends. In addition to a sparse observational network, many studies have documented trends in streamflow and other hydroclimatic variables for many regions [Yang et al., 2002, 2004a, 2004b; Ye et al., 2003, 2004] with explanations for the trends that differ spatially.

[7] There have been different hypotheses posited to explain the increasing streamflow over the region, from increasing annual precipitation to warming causing round ice melt to changes in the seasonality of precipitation. From the previous studies that document a connection between the snowpack and streamflow, we investigate the role that changes in the seasonal snowpack may have in driving the trends in seasonal and annual streamflow while acknowledging that the dynamics in other seasons will also play a role. We first evaluate the observations, quantifying trends in cold season precipitation (CSP), snow depth, and streamflow for 1936–1999. To complement the observations and fill in gaps in the observational network, we use the Variable Infiltration Capacity (VIC) land surface hydrologic model [Liang et al., 1994] to evaluate the role changes in cold season precipitation and snowpack have played in the streamflow trends. The consistency of modeled and observed trends in streamflow, cold season precipitation, and the snowpack are analyzed. Finally, a series of model experiments are performed to understand the role changes in the snowpack play in annual streamflow trends.

2. Data and Model

[8] This study uses several different data sets to evaluate trends in runoff and streamflow, snow water equivalent and depth, and cold season precipitation. (See Table 1 for a summary.) For the purpose of this study, runoff refers to modeled surface and subsurface flows before reaching the river network, and streamflow refers to measurements of flow once it is in the river network. For runoff and streamflow, streamflow measurements and the VIC-modeled runoff fields are examined. For evaluating the snowpack, a long-term, in situ data set of snow depth and the VIC snow water equivalent (SWE) estimates are used. Precipitation is taken from an in situ, gauge-based data set and from a gridded, gauge-based data set. The analysis is conducted for 1936–1999 (with the exception of snow depth because available observations end in 1995), and all calculations are for water years (September through August) rather than calendar years. A station is included only if fewer than 5 years of data are missing to ensure that as many stations as possible are included in the analysis without affecting the calculations of trends. Sampling experiments using data from stations with continuous records indicated that the 5 year threshold did not affect the calculated trends appreciably.

Table 1. Summary of the In Situ Observations Useda
Data SetVariableTemporal CoverageTemporal ResolutionData CenterReference
  • a

    The temporal coverage listed here only includes coverage between 1950 and 2006; some of the data sets begin earlier than 1950.

TD-9813Precipitation1891–2001DailyNCDCNational Climatic Data Center [2005]
Russian Snow SurveysSWE1966–199010 daysNSIDCKrenke [1998]
Historical Soviet Daily Snow Depth v2Snow Depth1881–1995DailyNSIDCArmstrong [2001]
r-arctic-netStreamflow1950–1999MonthlyUNHLammers et al. [2001]

2.1. In Situ Observations

[9] The snow survey data set, with 1345 stations measuring SWE, is from the National Snow and Ice Data Center's (NSIDC) Former Soviet Union Hydrological Snow Surveys, 1966–1996 [Krenke, 1998]. This data set is used solely to validate the land surface model because it is significantly shorter than the 1936–1999 analysis period. Daily snow depth measurements are from the Historical Soviet Daily Snow Depth Version 2 data set [Armstrong, 2001], available from the NSIDC. The data set contains records for 284 stations, with the earliest observations beginning in 1881 and the last in 1995. For this study, only those stations with fewer than 5 years of missing data between 1936 and 1995 were used: 64 stations meet this criteria. Precipitation measurements are from the National Climatic Data Center's (NCDC) TD-9813 data set; about 300 stations are located within the Pan-Arctic and meet the long-term data requirements.

[10] Streamflow measurements for the Pan-Arctic drainage area were obtained from the r-arctic-net database (http://www.r-arcticnet.sr.unh.edu/). If a station was missing one month of data in a given year, the entire year was discarded, and only stations with fewer than 5 years of missing data were used: 93 stations meet this criterion. Unlike snow depth, SWE, and precipitation measurements, the streamflow measurements potentially do not represent independent data points of observations. Streamflow spatially integrates the basin response upstream of the measurement, and, as such, many of these measurements are nested within larger basins. However, in terms of understanding the spatial pattern of the trends, using all basins regardless of size provides an indication of which regions have trends, either negative or positive.

2.2. Gridded Meteorological Data Set and the Land Surface Model

[11] The VIC land surface model is a large-scale hydrological model that solves for both water and energy budget closure at the land surface [Cherkauer and Lettenmaier, 1999; Liang et al., 1994; Liang et al., 1996]. It has been used in a variety of global and northern high-latitude studies [Adam et al., 2007; Nijssen et al., 2001; Sheffield and Wood, 2007; Su et al., 2005, 2006]. It provides temporally and spatially continuous estimates of each component of the terrestrial water budget while ensuring budget closure. The model requires inputs of land cover distribution, vegetation and soil parameters, and meteorological forcings. Land cover distribution was taken from the work by Bartalev et al. [2003], a 1 km resolution land cover data set from the Global Land Cover 2000 Project (GLC2000) data set based on the SPOT-4 satellite data that contains vegetation classes specific to northern Eurasia, such as peatlands. The leaf area index (LAI) was taken from the Moderate Resolution Imaging Spectroradiometer (MODIS) product (v. MOD15BU C4.1) [Knyazikhin et al., 1998]). An updated parameterization for snow density and heat flux through the snowpack was used in the VIC model that is based on the SNTHERM model [Jordan, 1991]. The frozen soil algorithm has been updated with a different numerical method to calculate heat fluxes, and modeled soil temperatures compare well with observations [Troy, 2010]. The model was run with a 50 m soil column with 50 thermal nodes, exponentially distributed such that there is a higher density of nodes near the surface. To validate the modeled runoff against gauge streamflow observations, runoff was routed through the stream network using the routing model of Lohmann et al. [1996].

[12] The VIC model was run at 100 km resolution on an equal-area (EASE) grid over the Eurasian Pan-Arctic and was forced by the Princeton Global Forcing (PGF) data set [Sheffield et al., 2006]. This is a global, 1°, 3 hourly data set of precipitation, air temperature, downward longwave and shortwave radiation, wind, humidity, and surface pressure, originally for 1948–2000 and recently extended to 1901–2008. It was interpolated to the EASE grid for this study. The PGF data set uses the National Centers for Environmental Prediction-National Center for Atmospheric Research (NCEP-NCAR) reanalysis [Kalnay et al., 1996] to temporally downscale monthly observations of precipitation and temperature from the CRU data set [Mitchell et al., 2004] after bias-correcting the rain day statistics [Sheffield et al., 2004]. Shortwave radiation was adjusted to match the remote-sensing based Surface Radiation Budget (SRB) data set [Stackhouse et al., 2004] for 1984–2007 and regressed to match variations in monthly cloud cover data at other times, as described by Sheffield et al. [2006]. The shortwave radiation was validated by Troy and Wood [2009] for northern Eurasia. This data set was chosen rather than precipitation data sets that have been created specifically for the region, such as the data set of Adam and Lettenmaier [2008], because the current internal algorithm in VIC to generate shortwave radiation as a model input overestimates downward shortwave radiation in the high latitudes. The precipitation and temperature for the PGF data set was validated by Troy [2010].

3. Methods

[13] Trends are calculated using the Mann-Kendall test with the estimate of the slope from the methods of Hirsch [1982]. The slope of the trend is calculated as the median value of Dij, where

display math

for all pairs of data points (Xi, Xj), ij, for which i and j are indices for the year or season. Trends are considered significant at the 90% level throughout this study.

[14] To evaluate the cause of the runoff trends, a series of model experiments was performed in which specific atmospheric forcing variables were held constant while others were allowed to vary, to isolate their impact. In the first experiment (denoted TClim/PVary), temperature was scaled to the monthly climatology for each year, such that any changes in runoff would be due to changes in precipitation or other meteorological variables. In the second experiment, monthly precipitation was set to climatology, such that air temperature is the driving variable (PClim/TVary). In the third experiment (SWEVary), the effect of the snowpack on runoff was isolated as follows: Whenever snow was present in the historical simulation, the historical meteorological data were used; and when snow was not present, the climatological temperature and precipitation were used as inputs to the model. Comparisons of the modeled snowpack between the historical and SWEVary simulations showed that this allowed for the snowpack to be correctly simulated for each year as compared with the historical run, while the warm season was held to climatology. This experiment does not mean that the hydrologic cycle after snowmelt reverts to climatology: For example, if a deeper snowpack was present, it is possible for this to lead to higher soil moisture that is due to more water infiltration and then more base flow generation. However, these effects would all be due to the deeper snowpack, not due to changes in warm season precipitation of temperature.

[15] In all of these experiments, the historical meteorological data were scaled for each month to match the monthly climatological precipitation or temperature value. Thus, the daily variability within the month remains the same. For example, if it rained on the third day of the month, it will still rain on the third day of the month but the amount is scaled such that that month's total precipitation matches the climatology. This method is imperfect in that daily variability may be affected with changes in climate, but it should be able to be used as an approximation of climatology and allows the model to receive consistent meteorological forcings (e.g., less shortwave radiation on rain days, etc.).

4. Results

4.1. Model Validation

[16] The model simulations used in this study are all uncalibrated: Biases therefore exist in the simulation of the streamflow. The assumption that calibration was unnecessary was tested by running 15 grid cells scattered across the domain with 64 different parameter combinations and calculating the modeled runoff trends. Only small differences in the magnitude of modeled annual runoff trends were seen across all the grid cells. Because this study focuses on trends, the bias is less important than the ability of the model to capture the interannual variability and trends. Figure 1 shows the locations of the basins used in this study, and Figure 2 shows the annual streamflow standardized anomalies for 1936–2006, with gauge observations in black and the routed VIC streamflow in gray. For the Northern Dvina, Ob, and Lena Rivers, the model replicates the interannual variability reasonably well. The exception to this is the Yenisei River, as the modeled trend (decreasing) is opposite in sign to the observed trend. This can be traced to the spatial variability that exists in the runoff trends in the Yenisei, which is discussed further below.

Figure 1.

Map of major basins used in this study. Dots are the location of the streamflow gauges used in Figure 2.

Figure 2.

Validation of the VIC model's ability to replicate the interannual variability of streamflow. The mean from the observed and modeled time series is removed from each time series, showing that VIC simulates the interannual variability well.

[17] The ability of the model to capture the magnitude and spatial pattern of the observed streamflow and snow depth trends is critical in order to use the VIC for trend attribution. Figure 3 plots the magnitude of the trend estimated for VIC against the observed trend, with streamflow in the top panel and snow depth in the bottom. For streamflow, this was done for each of the stations for which basin delineations and sufficient observations were available: 73 of the 96 basins with observations were delineated. Lumped routing was assumed as in the work by Lohmann et al. [2004], such that any runoff generated in a grid cell in the basin was assumed to contribute to that year's calculated streamflow without routing the runoff through the river network. Although this assumption may introduce some errors in estimating streamflow for very large basins because travel time is not taken into account, it should not have an influence on trend calculations at the annual scale. The lumped routing will affect only a small percentage of the year's runoff because the majority of the runoff is generated during other times of the year rather than at the end of the water year. Overall, the model shows an underestimation of the magnitude of observed trends, but in general there is good correlation between the observed and modeled annual streamflow trends shown in Figure 3 (r = 0.66, p < 0.01). For negative trends, the VIC tends to overestimate the significance. The majority of these basins occur in the southern Yenisei, where observations also showed negative trends, but of smaller magnitude and not necessarily statistically significant.

Figure 3.

Evaluation of VIC's ability to model (top) observed annual streamflow trends (1936–1999) and (bottom) March snow depth trends (1936–1995), with the magnitude of the observed or modeled trend plotted and the shape of the symbol indicating the agreement in significance. A significance level of 90% is used. The 1:1 line is plotted as a dashed line for comparison.

[18] For trends in March snow depth, the observed station trend was compared with the modeled snow depth trend for the grid cell in which the station was located. To account for any potential bias missing data might introduce, the modeled trend is calculated for only those years in which there is observed data at a particular station. The bottom panel of Figure 3 shows there is good correlation between the modeled and observed snow depth trends (r = 0.66, p < 0.01), with a systematic underestimation of the modeled snow depth trend as compared with the observed trends.

[19] The errors between simulated and observed trends will primarily result from forcing and model errors. The forcings may not have the correct trends in observed precipitation as well as temperature. Errors in precipitation trends can directly propagate into errors in streamflow trends, and errors in temperature trends can affect modeled SWE trends and evapotranspiration. VIC does not have a parameterization to allow for excess ground ice, which may cause it to underestimate streamflow trends in the permafrost regions where ground ice melt may be playing a role. Other periods were also compared to evaluate if the model is able to capture how the trend evolved in time, and results similar to those in the top panel of Figure 3 were found for 1960–2000 and 1980–2000, although with more scatter.

[20] For this study, the model's ability to simulate the snowpack is also critical, and VIC has little bias in the SWE trends across the basins (Figure 4). VIC shows an underestimation for the Ob, Lena, and Yenisei for some months. The largest scatter is in the Kolyma basin, which is probably due to the sparse observational network there. There are few long-term precipitation measurements there, meaning the model forcings will be prone to errors in this region that will propagate into errors in the modeled snowpack. In addition, there are only a few snow stations with observations available, which means errors average out less across the Kolyma as compared with other basins. Errors in the modeled snowpack could also play a role in the differences between modeled and observed simulated streamflow trends in Figure 3.

Figure 4.

Validation of the VIC-simulated snow water equivalent. The observed snow water equivalent is on the x axis, and the modeled snow water equivalent is on the y axis. Units are mm/month. Model data were filtered to include only those grid cells with an observation on a given day. Each dot represents the basin-averaged snow water equivalent for a single month between January 1966 and December 1990. The 1:1 line is plotted as a dashed line for comparison.

4.2. Trends in Observations

[21] Peterson et al. [2002] showed that annual streamflow as averaged over the six largest basins in northern Eurasia has increased between 1936 and 1999 with regional variations in the magnitude of the trends. To understand the spatial pattern of observed streamflow changes, trends were calculated for any stations in the r-arctic-net data set that had fewer than five missing years of streamflow observations between 1936 and 1999. Figure 5 shows the trends for 1936–1999, with statistically significant trends plotted as solid circles and the stations with statistically insignificant trends shown as open circles. The size of the symbol is proportional to the size of the basin. Spatial correlation exists between some of the observations because the stations can be on different reaches of the same river. Despite this, it is the general spatial pattern that is informative. The majority of the region shows an increase in annual streamflow, particularly north of 60°N. The exception to this is the southern Yenisei region and the Kolyma basin, both of which have negative trends. Yang et al. [2004a] also found positive annual streamflow trends overall in the Yenisei with a negative trend in the upper Yenisei, which this study also found. Yang et al. [2004b] documented similar trends as shown in this study in streamflow for the Ob, with negative trends in the southeastern portion of the basin but with an overall positive trend for the basin.

Figure 5.

Trends in streamflow for 1936–1999, according to the Mann-Kendall test with the Hirsch estimate of the slope. Units are mm yr−1 yr−1. The size of the circle indicates the size of the basin (see legend in bottom right of figure), the color is the slope of the trend. Slopes that are significant at the 90% level are plotted with a black outline. Only stations that were missing fewer than 5 years of data were used.

[22] The hypothesis that streamflow trends have been driven by changes in the snowpack cannot be tested directly against SWE measurements because of data availability, but long-term measurements of snow depth are available from 1936 to 1995. Although the relationship between SWE and snow depth depends on precipitation type, snowpack age and melt, and refreeze regimes, the depth measurements can be used as a surrogate. Figure 6 shows the trends in snow depth between 1936 and 1995. A similar pattern to the trends in runoff is seen, with positive trends north of 60°N and some negative trends in the southern Yenisei. The limitations of the snow observational network are apparent in this figure, as large areas in central and western Siberia and the Russian Far East lack long-term measurements. This confirms the results of Ye et al. [1998], which documented increases in snow depth in the north and decreases in the south, with no station data available for the Russian Far East.

Figure 6.

As in Figure 5 for trends in March snow depth for 1936–1995. Units are cm mo−1 yr−1.

[23] The positive trends in snow depth must be driven by precipitation, with temperature playing a role in shaping the snowpack dynamics and its characteristics such as cold content. To evaluate this, trends in cold season precipitation (CSP) were calculated for the TD-9813 gauge observational data set. CSP is defined here as the monthly precipitation that occurs, provided that snow is present any time during the month in the VIC simulation. This was done rather than defining the cold season as specific months, such as November through March, because the length of the cold season varies with latitude. By using the snowpack as an indicator, we ensure that all the precipitation of the cold season is included. Figure 7 shows negative trends in the southern Yenisei, southwestern Ob, and parts of the far eastern domain near the Kolyma, with statistically significant positive trends throughout the rest of the region. Qualitatively, this pattern is generally consistent with the patterns of trends in snow depth and annual streamflow. The overall positive trend in CSP is consistent with the trends of Rawlins et al. [2006] despite slightly different methods of calculation, indicating that the positive trend is robust.

Figure 7.

As in Figure 5, but for trends in cold season precipitation (CSP) in the TD-9813 precipitation gauge data set. Units are mm yr−1 yr−1.

4.3. Trends in the VIC Model Results

[24] In general, Figure 3 shows that the VIC model is able to replicate the observed streamflow trends, and Figure 8 confirms this. Figure 8 plots the trend in modeled annual runoff simulated by the VIC model for 1936–1999. The magnitude and spatial pattern of annual runoff trends in VIC match that seen in the observed streamflow trends, with a decrease in annual streamflow in the southern Yenisei, southwestern Ob, and the Kolyma basin. Increases are seen in the Severnaya Dvina, Ob, and northern Yenisei basins. This is consistent with the spatial pattern seen in Figure 5, both in terms of magnitude of trend and contrasting positive and negative trends, depending on region.

Figure 8.

Trends in runoff for 1936–1999, according to the Mann-Kendall test with the Hirsch estimate of the slope. Units are mm yr−1 yr−1. Trends that are significant at the 90% level or greater are plotted as squares; trends that are not significant are plotted as dots.

[25] As was shown in Figure 6, there is observational evidence that changes in the snowpack are occurring, but long-term SWE measurements are not available to the scientific community. Modeled SWE can therefore act as a surrogate, given that it compared well with the observations that exist for 1966–1991. Figure 9 shows the March SWE trends for 1936 through 1999. Just as the pattern in observed snow depth replicated the pattern in observed annual streamflow, the same is true for modeled SWE and annual runoff trends. The spatial pattern and magnitude of the modeled SWE trends match the observed and modeled trends in annual runoff, indicating that the SWE trends can explain, at least in part, the trends in annual runoff. Trends calculated on the runoff estimated by VIC during the spring show the same pattern and magnitude as the annual trends in observed runoff (as discussed further below).

Figure 9.

As in Figure 8, but for trends in modeled March SWE for 1936–1999.

[26] Figure 10 shows a consistent pattern in trends of CSP in the PGF meteorological forcing data set compared with modeled annual runoff and observed streamflow, as well as modeled SWE and observed snow depth. The magnitude of the trends is also consistent with the increases shown in SWE and annual runoff. Overall, there is a positive trend in the PGF CSP over the Ob and in parts of the Severnaya Dvina and the Yenisei basins. Negative trends exist over the Kolyma and the southern Yenisei, which are consistent with the trends in observed streamflow and modeled runoff and SWE.

Figure 10.

As in Figure 8, but for trends in cold season precipitation in the PGF data set.

[27] Figure 11 plots the trends in modeled March SWE against the trends in modeled annual runoff, where each dot represents a grid cell. Only those grid cells with statistically significant runoff are shown in order to focus on the regions where change is occurring, which is why there are very few points around the zero trend in runoff. Overall, the trends are generally of the same sign. Spatially, the trends have a Pearson's correlation coefficient of 0.72 that is statistically significant (p < 0.01).

Figure 11.

Comparison of trends in modeled March SWE and annual runoff. Only those grid cells with statistically significant runoff trends are included. The 1:1 line is plotted as a dashed line for comparison.

4.4. Trend Attribution

[28] Although visual inspection of the trends shows consistency between the gauge observations of cold season precipitation, snow depth, and streamflow, the evidence is circumstantial as the observational network is sparse, particularly over Siberia and the Russian Far East. The VIC land surface model is able to simulate broad scale patterns for northern Eurasia as a whole that are consistent with the sparse observational data. The consistency in the patterns of trends in modeled CSP, SWE, and runoff adds confidence to the hypothesis that changes in SWE trends are at least partially driving the trends in annual streamflow. In order to test this more rigorously, a series of model experiments, as described in section 3, were performed to test this hypothesis. Figure 12 plots the maps of annual runoff trends calculated from the series of experiments. The first column shows the trends in annual and seasonal runoff for the baseline, historic simulation, the second column shows the trend from the SWEVary experiment, the third column the trends from PClim/TVary, and the right column the trends from TClim/PVary.

Figure 12.

Annual and seasonal runoff trends in the VIC simulations. (left) Historic baseline run, (middle) influence of the snowpack (SWEVary) and influence of air temperature (PClim/TVary), (right) influence of precipitation (TClim/Pvary). See the text for descriptions of the model experiments. Units for the annual trends are mm yr−1 yr−1 and for the seasonal trends are mm season−1 yr−1.

[29] It is readily apparent that only the SWEVary simulation replicates the pattern and magnitude of the trends in the baseline, the historic model run. It is also apparent that spring is the dominant season for modeled runoff trends, with strong positive trends in the northern part of the domain during summer. These positive summer trends may still be caused by melting snow in June at such high latitudes. Overall, the streamflow trends simulated in the PClim/TVary do not match the streamflow trends of the historical model runs nor those of the SWEVary experiment for spring, which is the season that drives the historical annual trends. The TClim/PVary experiment shows larger than modeled (and observed) streamflow trends annually, driven predominantly by increases in spring streamflow from deeper snowpacks.

[30] The interactions of precipitation and temperature control the snowpack dynamics. Warmer atmospheric temperatures allow the atmosphere to hold more water and therefore allow for more precipitation, but they also can cause more precipitation to fall as rain. That seems to be the case during the 20th century. When only precipitation varies, the model overestimates observed runoff trends by allowing too much precipitation to fall as snow. When the cold season precipitation and temperature both increase, the increase in precipitation would lead to an increase in snow, but the increase in temperature would lead to a decrease in the snowpack. The SWEVary simulation shows that the two act together in such a way that overall the snowpack increases, although not as much as it would without the warming temperatures. To evaluate the contribution of the snowpack on the historical modeled trends, we calculated the percentage of contribution of the SWEVary annual runoff trend in historical modeled simulation for every grid cell that had a statistically significant annual runoff trend. The SWEVary trends have a Pearson correlation coefficient of 0.82 with the historical trends, the TClim/PVary experiment correlation coefficient is 0.66, and the PClim/TVary experiment correlation coefficient is 0.01. The SWEVary trends explain 67% of the magnitude of the historical streamflow trends.

[31] The SWEVary experiment also highlights where the snowpack dynamics may not be the primary driver of how streamflow has changed. In those regions with negative annual trends in streamflow, the SWEVary experiment underestimates the magnitude of the annual trend. This is driven by an underestimation of the negative summer runoff trends in those regions with negative trends. The SWEVary experiment holds precipitation and temperature to climatology during the warm season, and the underestimation of the negative summer trend in many parts of the region demonstrates that the warm season precipitation and temperature are leading to decreases in streamflow. This may be done through decreasing summer precipitation, as is seen in the southern Yenisei in the TClim/PVary experiment, or increases in evapotranspiration through warming temperatures.

5. Discussion and Conclusions

[32] The VIC model is able to capture the observed trends in snow water equivalent and runoff sufficiently well to be used in the analysis presented in the paper. This is despite the fact that the model is uncalibrated and the modeled runoff may be biased. Calibration would not affect the modeling of the snowpack, and the majority of the water from snowmelt goes into the stream network in this region. We acknowledge that the lack of calibration could introduce biases into the model results. For example, calibration can affect the amount of water that infiltrates into the soil and is therefore available for summer evapotranspiration. However, tests on grid cells scattered across the domain demonstrated that the effect of calibration on model results is small, and, as such, using an uncalibrated model will not significantly affect the results presented above.

[33] The region where the model did not perform well is the Yenisei basin, where VIC modeled a decrease in streamflow while the observations show a strong increase. This may be due to the lack of modeled anthropogenic influences in the VIC model, specifically large dams with the attendant changes during reservoir filling and hydropower operations. Given that reservoirs do not significantly influence annual streamflow trends [Adam et al., 2007], it is probably due to the spatial variation and nature of the trends within the basin. The southern Yenisei has shown a decrease in observed streamflow and the northern Yenisei an increase [Yang et al., 2004a], with the increase in the north dominating the total basin trend. In VIC, the negative trend in the south dominates, although it does replicate the observed spatial pattern of positive trends in the north and negative in the south. The modeled negative trend in the south may be overestimated because of errors in the meteorological forcings; for example, the trend in warming summer temperature may be leading to a stronger trend in evapotranspiration, leaving less water for runoff generation. This region also is mountainous, and the model may not be capturing the hydrologic processes as well in a topographically complex terrain.

[34] When considering trend attribution, the effects of other climate variables and potential human impacts should not be ignored. The region has several large dams, particularly on the Yenisei, which have the potential to influence the streamflow regime significantly. The dams do alter the seasonal streamflow trends, increasing streamflow in the winter and reducing streamflow during the melt season. However, previous studies have shown that the dams do not alter the annual trends at longer time scales [Adam et al., 2007]. Aside from reservoirs, it is well documented that significant uncertainty exists in the estimates of precipitation over this region because of a sparse gauge network [Serreze et al., 2003] and gauge undercatch of solid precipitation [Adam and Lettenmaier, 2003]. This may induce errors, both in estimates of trends from observations and from the model, which is forced by observation-based gridded data sets. Changes in evapotranspiration may also play a role, affecting the amount of water available for runoff generation. Permafrost melt, which is essentially a long-term storage change in the water budget, may also be contributing to positive streamflow trends in the permafrost-dominated regions, as Adam and Lettenmaier [2008] and Pavelsky and Smith [2006] hypothesized. This study does not address these potential factors, although they may play a role in the observed streamflow trends.

[35] Previous work has shown that snow depth and snow water equivalent have increased across the former Soviet Union [Bulygina et al., 2009; Ye et al., 1998], and changes in streamflow have been linked to changes in winter precipitation [Rawlins et al., 2006, 2009] and snow cover extent [Yang et al., 2003]. These studies have focused on one aspect of the changes that might drive streamflow. This study takes a more holistic approach, showing the link between cold season precipitation, changes in the snowpack, and streamflow in a modeling framework that bridges the spatial and temporal gaps in the observational network.

[36] Many of our results using observed data confirm what has been previously reported in the literature using the same data sets in different ways. Rawlins et al. [2006, 2009] documented increases in winter precipitation with the TD-9813 data set. Although for a shorter time period (1966–1995), Rawlins et al. [2009] showed widespread positive trends in winter precipitation for the Ob, Yenisei, and Lena basins, with a region of negative trends in the southern Yenisei. This confirms the results shown in Figure 7. This was also seen in the snow water equivalent observations used in that study. Bulygina et al. [2009] documented trends for maximum snow depth for 1966–2009 that were positive in the northeastern portion of the domain and negative in the southern, which is largely what was seen in Figure 6, albeit for a longer time series.

[37] VIC shows an increase in snow water equivalent that is consistent with observed snow depth measurements and with the pattern of observed and modeled runoff trends. Previous studies have shown that snow cover extent and duration are changing, particularly in the spring, because of earlier melt [Brown, 2000; Brown and Mote, 2009]. Snow cover extent and SWE depth are linked [Yang et al., 2009], so that using only SWE rather than changes in depth, cover, and duration simplifies the analysis and directly accounts for the water in the snowpack that may contribute to runoff.

[38] Despite the consistency in magnitude and spatial pattern in trends of CSP, SWE, and runoff streamflow in the observations and model, the evidence is circumstantial that changes in CSP and therefore SWE trends are driving the trends in streamflow, given the uncertainty that exists in measurements over this region. To more fully test this hypothesis, we ran a series of model experiments that held constant certain atmospheric conditions, such as temperature or precipitation, in order to determine which variables are driving the trends. These experiments identify changes in the snowpack as the driver of streamflow trends in this region. The use of the hydrological model for the analysis was necessary: Direct trend calculations on cold season precipitation revealed increases, but the actual accumulation and melt of the snowpack is a nonlinear process that depends on the interaction of precipitation and air temperature, whose individual impacts can be isolated only through a modeling approach.

[39] The SWEVary experiment's streamflow trends showed strong correlation with the historical simulation trends, indicating that the snowpack dynamics do play a large role in the historical streamflow trends, but can explain only 67% of the trends. This is consistent with other studies that have shown that annual precipitation can explain streamflow trends [e.g., Adam and Lettenmaier, 2008; Pavelsky and Smith, 2006], as this indicates that it is not only the cold season precipitation that matters in explaining trends. Other processes, whether it is warm season precipitation or evaporation changes or some other factor, must also be playing a role.

[40] This study does not focus on any possible contribution to streamflow trends other than the influence of the snowpack. For example, changes in evapotranspiration that are due to warming temperatures, permafrost melting, and the changes that may have an effect on infiltration, or changes in precipitation outside of the cold season may all play a role in streamflow trends. In fact, changes in summer climate have to play a role in the negative streamflow trends seen in the southern Ob and Yenisei basins. Figure 12 showed that the snowpack could not be solely responsible for the negative trends in this region, and other factors, such as evapotranspiration increases, must be considered. The VIC model is run over large grid cells (100 km EASE grid) that may not be able to capture processes occurring at smaller resolutions, such as melting of ice wedges in the permafrost. If these smaller-resolution processes play larger roles in the runoff trend, this study would not be accounting for them and could potentially be another source of uncertainty and error.

[41] Many in the Pan-Arctic hydrological community have been studying the trends in annual streamflow across Eurasia. Many of the studies linked the trends with potential drivers, but these have generally been at local or short time scales because of the sparse observational network and the short observational record length. This study has built on previous studies to provide a more definitive attribution for the increases in northern Eurasian streamflow. This was done through the use of a validated hydrological model that shows consistency with trends in observed snow and streamflow. Increases in SWE trends appear to be driving trends in annual runoff, and it is the interaction between increasing cold season precipitation and warmer temperatures that produces the increase in SWE trends. This bears further investigation: Climate models predict both warmer and wetter winters in this region, and the evolution of the snowpack may be critical in understanding how annual and seasonal streamflow will change over the 21st century.

Acknowledgments

[42] The research was supported through NSF grant 0629471 (Collaborative Research: Understanding Change in the Climate and Hydrology of the Arctic Land Region: Synthesizing the Results of the ARCSS Fresh Water Initiative Projects) and NASA grant NNX07AR18G (Use of International Polar Year Data to Improve Attribution of Long-Term Hydrologic Changes in Arctic Eurasian Land areas). We thank Arvid Bring for providing the basin delineations and the NSIDC and r-arctic-net for providing data.