Streamflow simulations of the terrestrial Arctic domain



[1] Predicting riverine discharge to the Arctic Ocean has become increasingly important because of the dominant role that river runoff plays in the freshwater balance of the Arctic Ocean, and the predicted high sensitivity of the region to global warming. The ability of land surface models to represent runoff and streamflow from northern river basins is critical to an understanding of the Arctic hydrologic cycle. A set of simulations with the land surface scheme VIC (Variable Infiltration Capacity) implemented at 100 km EASE-Grid across the pan-Arctic domain was conducted to evaluate the model's representation of various hydrologic processes in the Arctic land region, and to provide a consistent baseline hydroclimatology for the region. The pan-Arctic drainage basin system was partitioned into 12 regions for purposes of model implementation and testing. Streamflow observations at various basin outlets, satellite-based snow cover extent, observed dates of lake freeze-up and break-up, and sited monitored summer permafrost maximum active layer thickness were used to evaluate various simulated hydrologic variables. The results indicate that the VIC model was able to reproduce these hydrologic processes in the Arctic region. A 21-year average river inflow (1979–1999) to the Arctic Ocean from the AORB (Arctic Ocean River Basin) illustrated in Prowse and Flegg (2000), was estimated with the simulated streamflow as 3354 km3/yr, and 3596 km3/yr with the inclusion of the Arctic Archepelago, which are comparable to the previous estimates derived from the observed data.

1. Introduction

[2] Runoff from the Arctic terrestrial drainage system represents approximately 50% of the net flux of freshwater to the Arctic Ocean [Barry and Serreze, 2000], a large number in comparison with other world oceans, for which freshwater inputs are usually dominated by precipitation over the ocean surface. The magnitude and variation of freshwater inflows to the Arctic Ocean have significant influence on the global thermohaline circulation, salinity, and sea ice formation in the oceans [Aagaard and Carmack, 1989; Broecker, 1997; Macdonald, 2000; Peterson et al., 2002]. Seasonal variations in streamflow critically affect the transport of sediment, mineral, and organic substances including pollutants into the Arctic coastal seas [Rouse et al., 1997; Gordeev, 2000]. Arctic soils and wetlands play an important role in the global carbon dioxide and methane budget which is sensitive to the altered soil moisture and temperature [Oechel et al., 1993]. The dynamics of water in soil and wetlands have a direct impact on the spatial and temporal distributions of runoff in the region.

[3] Climate model simulations generally indicate that future global warming will be at a maximum at high northern latitudes, and will be associated with increased precipitation and earlier snowmelt [IPCC, 1995]. Recent research evaluating the potential effects of climate change suggests that runoff from Arctic catchments could be significantly modified in terms of quantity and timing [Miller and Russell, 1992; Shiklomanov, 1994; Georgievsky et al., 1996; Shiklomanov et al., 2000]. An important question that remains to be addressed is how global warming will impact the hydrologic cycle in the Arctic region and how an altered Arctic hydrologic cycle will further affect the atmospheric dynamics given that pronounced changes have been observed already [Serreze et al., 2000; SEARCH SCC, 2001].

[4] Arctic hydrology differs from that of more temperate regions in several important ways, mostly having to do with unique conditions associated with cold temperatures. These include dominance of snow cover and spring snowmelt floods, the presence of permafrost, and prevalence of lakes and wetlands. All of these considerations motivate the improvement of high latitude land surface representations in land surface schemes that can be coupled with global circulation models (GCMs). A good review of recent progress in representation of land surface arctic processes can be found in the work of Bowling [2002] and Stieglitz et al. [1999]. PILPS, the Project for Intercomparison of Land Surface Parameterization Schemes, devoted its Phase 2(e) to evaluation of the performance of land surface models in an arctic environment [Bowling et al., 2003]. The project tested 21 land surface models with respect to their ability to represent snow accumulation and ablation, soil freeze/thaw and permafrost, and runoff generation. All participating models captured the broad dynamics of snowmelt and runoff, but large differences in snow accumulation and ablation, turbulent heat fluxes, and streamflow were found due to the different model parameterizations in fractional snow coverage and land-surface roughness [Nijssen et al., 2003; Bowling et al., 2003]. The PILPS-2e intercomparison provided valuable opportunities to identify problems in representations of snow cover, surface runoff, and other physical processes in cold region, and to assess the enhanced representations [Essery and Clark, 2003; Van den hurk and Viterbo, 2003].

[5] Numerous hydrologic studies have been conducted in the Arctic region at the plot, hillslope, and small watershed scales [see, e.g., Rovansek et al., 1996; Tilley and Lynch, 1998; Stieglitz et al., 1999; Young and Woo, 2000; Z. Zhang et al., 2000]. A nested set of three instrumented watersheds (four if the nearby Putuligayuk catchment is included) has been studied since 1992 on the North Slope of Alaska [Kane et al., 2000]. The resulting data have facilitated comparison of the variability of runoff-related processes from the headwater foothills to the low gradient, wetland dominated coastal area. Woo et al. [2000] reviewed recent Canadian research on snow, frozen soils, and permafrost hydrology. Despite the development of new process-based algorithms, they found few advances in comprehensive mathematical models that represent cold-region hydrologic processes in coupled land-atmosphere systems. Prowse and Flegg [2000] argues that the likelihood of climate-induced changes in the runoff regime of northern Arctic rivers demands fundamental improvements in understanding of cold-region hydrologic processes and their incorporation into meso- and large-scale hydrologic models.

[6] The work reported here draws from several earlier implementations of land surface models over part or the entire pan-Arctic domain. Bowling et al. [2000] applied the VIC model to the Mackenzie and Ob River basins at a 2° spatial resolution and examined the space-time structure of the predicted hydrologic variables (i.e., runoff, evaporation, soil moisture, and snow water equivalent). Arora [2001] assessed the performance of the Canadian land surface parameterization scheme (CLASS), when operated within a GCM at 3.75° latitude-longitude resolution, for three Siberian river basins (Yenisei, Lena, and Amur). The CLASS model simulated the annual streamflow cycle reasonably well for the three rivers. Rawlins et al. [2003] applied a simple water balance model (P/WBM) to estimate runoff across the pan-Arctic drainage basin at a 25 km spatial resolution with the temperature and precipitation inputs from the NECP (National Centers for Environmental Prediction) reanalysis. These works suggest the potential of hydrologic models to provide seasonal variations in freshwater discharge to the Arctic Ocean.

[7] In this study, a set of simulations were performed with the VIC model implemented at 100 km EASE-Grid across the pan-Arctic domain. The 100 km resolution was selected based on recent work by Serreze et al. [2003] which found that 100 km was about the highest resolution that can be supported by available observations of precipitation, the key forcing variable for land surface hydrology models. The implementation of the model, including the development of forcing data, and soil and vegetation parameters, as well as the procedure used to calibrate and evaluate the model, are described in the first half of this paper. Subsequently, simulations of hydrologic processes for the period 1979–1999 are analyzed and compared to observations of streamflow, snow cover extent, dates of lake freeze-up and break-up, and permafrost active layer depth (ALD). Finally, the 21-year estimate of the annual mean river inflow to the Arctic Ocean and its seasonality are evaluated in comparison with previous estimates.

2. Model Implementation and Data Sources

2.1. Model Implementation

[8] The VIC model [Liang et al., 1994, 1996] is a grid-based land surface scheme which parameterizes the dominant hydrometeorological processes taking place at the land surface-atmosphere interface. The model was designed both for inclusion in GCMs as a land-atmosphere transfer scheme, and for use as a stand-alone macroscale hydrologic model. The model solves both surface energy flux and water balances over a grid mesh. The model is characterized by a mosaic representation of land surface cover and a subgrid parameterization for infiltration, which accounts for subgrid scale heterogeneities in land surface hydrologic processes. The soil column is comprised of three soil layers, which allows the representation of the rapid dynamics of soil moisture movement during storm events and the slower deep inter-storm response in the bottom layer. The VIC model uses the variable infiltration curve [Zhao et al., 1980] to account for the spatial heterogeneity of runoff generation. It assumes that surface runoff from the upper two soil layers is generated by those areas for which precipitation, when added to soil moisture storage at the end of the previous time step, exceeds the storage capacity of the soil. Evaporation from bare soil also accounts for the subgrid variability in soil moisture. The formulation of subsurface runoff follows the Arno model conceptualization [Todini, 1996].

[9] The critical elements in the model that are particularly relevant to high latitude implementations include: (1) a two-layer energy balance snow model [Cherkauer and Lettenmaier, 1999; Storck and Lettenmaier, 1999] which represents snow accumulation and ablation in a ground pack and an overlying forest canopy, where present; (2) a frozen soil/permafrost algorithm [Cherkauer and Lettenmaier, 1999, 2003] that solves for soil ice contents within each vegetation type and represents the effects of frozen soils on the surface energy balance and runoff generation; (3) a lake and wetland model [Bowling, 2002] which represents lakes and wetlands as a separate land cover class, and explicitly determines their effects on moisture and energy fluxes within a grid cell; and (4) a blowing snow model [Bowling, 2002; Bowling et al., 2004] which parameterizes the effects of wind on snow redistribution and sublimation. All of these elements have been tested at points or catchment scales, but none of them have previously been applied over an area nearly as large as the entire pan-Arctic domain. The VIC model can be run in both water balance (assumes that surface temperature is equal to surface air temperature) and full energy balance (iterates for surface temperature to close the surface energy balance) modes. We have found in previous work that a 24-hour time step is adequate for streamflow simulations using the water balance version of the model, however some aspects of snow simulation in particular require a shorter time step (Hourly or 3-hourly) that resolves the diurnal cycle. For this reason, we chose a 3-hour time step for the simulations reported here.

2.2. Data Sources

[10] The meteorological data consist of daily time-series of precipitation (for the period 1979 through 1999), maximum temperature, minimum temperature, and wind speed gridded to a resolution of one-half-degree over all global land areas (excluding Antarctica). Monthly time-series precipitation was estimated by adjusting the Willmott and Matsuura [2001] grids for gauge undercatch, as described by Adam and Lettenmaier [2003], and for orographic effects, as described by J. C. Adam et al. (Correction of global precipitation products for orographic effects, submitted to Journal of Climate, 2005) (hereinafter referred to as Adam et al., submitted manuscript, 2005). Monthly maximum and minimum temperatures were created mainly from the Climate Research Unit (CRU) Version 2 data set of T. D. Mitchell et al. (A comprehensive set of high-resolution grids of monthly climate for Europe and the globe: The observed record (1901–2000) and 16 scenarios (2001–2100), submitted to Journal of Climate, 2004), which includes monthly mean temperature and diurnal temperature range from 1901 through 2000. The two-degree daily time-series grids of Sheffield et al. [2004] were used to create the daily distribution of precipitation, maximum temperature, minimum temperature, and wind speed for 1979 through 1995. For 1996 through 1999, the two-degree daily grids described by Nijssen et al. [2001a], which have recently been extended through the year 2000, were used to create the precipitation and temperature daily distributions. The values of each of the two-degree grids were assigned to each of the half-degree grids contained within the same spatial domain. These daily time-series were then scaled to have the desired monthly precipitation and temperature means. For 1996 through 1999, the daily 10-m wind fields were obtained from NCEP-NCAR reanalysis [Kalnay et al., 1996] and regridded to half-degree resolution through linear interpolation. Finally, the half-degree global forcings were regridded to the 100 km × 100 km EASE grid using an inverse distance interpolation. Daily precipitation was apportioned equally to a 3-hour time step. Temperatures at each time step were interpolated by fitting an asymmetric spline function through the daily maximal and minimal. The 3-hourly wind speeds were made identical to the daily values. The vapor pressure, incoming shortwave radiation, and net longwave radiation were calculated based on daily temperature and precipitation using algorithms of Kimball et al. [1997], Thornton and Running [1999], and Bras [1990] which have also been utilized in many earlier implementations of the VIC model [e.g., Maurer et al., 2002].

[11] The VIC model contains two primary variables upon which other secondary parameters are determined: soil texture and vegetation type. Soil texture information and soil bulk densities were derived from the 5-min Food and Agriculture Organization data set [FAO, 1998]. The soil parameters fall into two general categories. The first category is not adjusted once determined from the FAO data. These include porosity θs (m3m−3), saturated soil potential ψs (m), saturated hydraulic conductivity Ks (ms−1), and the exponent B for unsaturated flow (which was based on Cosby et al. [1984]). Another category of soil parameters is subject to calibration based on the agreement between simulated and observed hydrographs. Parameters in this category include the thickness of each soil layer di, the exponent bi of the infiltration capacity curve, and the three parameters in the baseflow scheme: Dm, Ds, and Ws.

[12] Vegetation types were obtained from the University of Maryland's (UMD) 1 km Global Land Cover product [Hansen et al., 2000], which has a total of 14 different land cover classes. Leaf area index (LAI) was derived based on the gridded (1/4°) monthly global LAI database of Myneni et al. [1997], and was used to derive monthly mean LAI values for each vegetation class for each grid cell as described in Nijssen et al. [2001a]. Other vegetation parameters such as architectural resistance, minimum stomatal resistance, albedo, roughness length, zero-plane displacement, and rooting depth and fraction were specified for each individual vegetation class. In addition to these standard vegetation parameters, the VIC model requires specification of three other parameters for blowing snow: the standard deviation of terrain slope, the standard deviation of terrain elevations, and the lag-one gradient autocorrelation for each model grid cell and each vegetation type [Bowling, 2002; Bowling et al., 2004]. These were calculated using the UMD 1 km land cover classification and the USGS 30 arc-second (about 1 km) digital elevation model (DEM) of the world (GTOPO30) (available from

[13] The lake algorithm in the VIC model is intended to represent the effects of lakes and wetlands by creating a surface wetland class that can be added to the grid cell mosaic, in addition to the vegetation and bare surface land classes [Bowling, 2002]. The wetland class can represent seasonally flooded ground as well as permanent water bodies. A lake geometry input file is needed if the model is run with the lake algorithm. The data included in this file are (for each grid cell) the lake area fraction, the maximum runoff rate for a stage-based curve, and the specification of a variable depth-area relationship. Lake area was calculated from the UMD 1-km land classification and assumed constant with depth.

[14] River discharge data were obtained from R-ArcticNet, a database of pan-Arctic monthly river discharge (available at Human activities, such as fragmentation of the river channels by dams, inter-basin diversion, and irrigation, are not accounted for in the data set [Lammers et al., 2001], nor are they represented in the version of the VIC model used here.

3. Calibration and Parameter Transfer

3.1. Arctic Drainage Basin

[15] The pan-Arctic drainage basin is defined here as all land areas draining to the Arctic Ocean, as well as those regions draining into the Hudson Bay, Hudson Strait, and the Bering Strait (Figure 1). The model domain is comprised of 2,834 computational grid cells at the 100 km resolution across the entire pan-Arctic land region. The model region spans 37° of latitude, from 46°N in the Yenisei basin to 83°N at the tip of Greenland (Figure 1), and several distinct vegetation and climate zones. Annual mean precipitation ranges from 100 to 1200 mm; annual mean temperature spans 30°C (−20 to 10°C) (based on surface data); the median snow-cover period ranges from 90 to 300 days (based on the remotely sensed data described in section 4.3). The variety of the landscape and climate places enormous variability in the hydrological retention characteristics of the diverse drainage systems. Runoff from Arctic rivers is strongly seasonal and is dominated by the spring snowmelt. According to T. Zhang et al. [2000], approximately 13 to 18 percent of the land area in the Northern Hemisphere is underlain by permafrost. The dominance of surface flow is typical of both permafrost and seasonally frozen regimes during snowmelt. Lakes and wetlands are particularly prevalent in large areas of the Arctic region. The relatively large storage capacity of the lakes and wetlands tend to dampen the volume and lag the timing of runoff.

Figure 1.

Definition of the Arctic drainage with twelve modeling sub-regions: 1, Anadyr Kolyma; 2, Lena; 3, Yenisei; 4, Ob; 5, Barents, Norwegin Sea; 6, Yukon; 7, Mackenzie; 8, Northwest Hudson Bay; 9, South and East Hudson Bay; 10, Nelson; 11, Arctic Archipelago; 12, Greenland.

3.2. Calibration and Parameter Transfer

[16] The VIC model uses physically based formulations for the calculation of energy fluxes while uses a conceptual scheme to represent runoff production. Several model parameters in this scheme must be estimated from observations, via a trial and error procedure that leads to an acceptable match of model-predicted discharge with observations. The most sensitive parameters in the water balance components are the infiltration parameter (bi) which controls the amount of water that can infiltrate into the soil, the first two soil layer thicknesses (d1, d2) which affect the maximum storage available in the soil layers and consequently the water available for transpiration, and three baseflow parameters which determine how quickly the water stored in the third layer is evacuated, including the maximum velocity of baseflow (Dm), the fraction of maximum baseflow (Ds) and the fraction of maximum soil moisture content of the third layer (Ws) at which a nonlinear baseflow response is initiated. In this study, the top soil layer depth (d1) was set to 5–10 cm for every grid cell, based on results of Liang et al. [1996] that indicated that a thin top layer significantly improved evapotranspiration predictions in arid climates. The three baseflow parameters and the third layer depth (d3) described in Nijssen et al. [2001b] were used with minor adjustment during the calibration. Therefore only the infiltration parameter (bi) and the second soil depth (d2) were targeted for intensive calibration. In the VIC model, an increase in the infiltration parameter (bi) tends to enhance runoff production, while a decrease of bi tends to reduce runoff. The maximum soil moisture storage capacity (θs*di) is dynamically determined by the change of soil thickness (di). The thicker the soil depths are (resulting in more soil moisture stored in the soil layers), the less the runoff is generated. Parameters bi and d2 were calibrated independently.

[17] A river network was derived for the 100 km × 100 km EASE grid (Figure 2) from a 1 km DEM using the algorithm of O'Donnell et al. [1999]. A routing scheme [Lohmann et al., 1996, 1998] was run offline using daily VIC surface and subsurface runoff as inputs to obtain simulated streamflows at the outlets of selected study basins. The daily runoff was accumulated from the 3-hourly VIC outputs. The calibration generally relies upon three criteria: 1) visual comparison of simulated and observed monthly hydrographs; 2) Nash-Sutcliffe efficiency (Ef) describing the prediction skill of the modeled streamflow as compared to observed value; and 3) relative error (Er) between simulated and observed mean annual runoff. Ef and Er are calculated as:

equation image
equation image

where Qi,o and Qi,c are the observed and simulated streamflow in month i, and equation imageo and equation imagec are the observed and simulated mean annual runoff, respectively.

Figure 2.

Digital river networks for the pan-Arctic drainage basins at the 100 km resolution, showing the watershed boundaries of the Kolyma, Lena, Yukon, Yenisei, Ob, Mackenzie, and Nelson. Dots represent 200 the farthest downstream outlets with the outflows to the Arctic Ocean based on the river networks.

[18] The pan-Arctic drainage basin was partitioned into twelve regions for model calibration and parameter transfer according to geographical definitions and hydroclimatology (Figure 1). Twenty-seven individual and sub-basins (Table 1) within different regions were chosen for model calibration based on the degree of anthropogenic regulation and length of historical streamflow records. Three regions were not calibrated due to intensive regulation in the Nelson (Region 10 in Figure 1) and the absence of observed discharge data for the Arctic Archipelago and Greenland (Regions 11 and 12 in Figure 1). Parameters were transferred to the uncalibrated areas from the nearest calibrated basins within the same region. Table 2 shows the values of the infiltration parameter (bi), the second soil depth (d2), Nash-Sutcliffe model efficiency (Ef), and relative error (Er) for each basin. Here, Nash-Sutcliffe statistics are determined for monthly streamflow.

Table 1. Selected River Basins for Calibration
RegionNameBasin for CalibrationGauge LocationDrainage Area, km2
1Anadry KolymaKolyma at Kolymskoye68.73158.72526,000
Anadyr at Snezhnoye65.45172.97106,000
2LenaAldan, Lena63.32132.02696,000
Lena at Tabaga, Lena61.83129.60897,000
Lena at Kusur70.68127.392,430,000
Anabar at Saskulak71.97114.0878,800
3YeniseiNizhnyaya Tunguska, Yenisei65.6390.02447,000
Podkamennaya Tunguska, Yenisei62.3292.12218,000
Yenisei at Igarka67.4386.482,440,000
4ObOb at Prokhorkino, Ob59.5279.47738,000
Irtish at Tobolsk, Ob58.268.23969,000
Ob at Salekhard66.6366.602,430,000
5Barents, Norwegin seaSevernaya Dvina at Ust'-Pinega64.1341.92348,000
Mezen' at Malonisogorskaya65.0045.6256,400
Pechora at Oksino67.6352.18312,000
6YukonYukon R NR Stevens Village AK65.87−149.71508,417
Yukon at Pilot Station AK61.93−162.88831,000
7MackenzieLiard River Near the Mouth61.74−121.22275,000
Mackenzie at Arc Red Riv67.45−133.741,680,000
8Northwest Hudson BayThelon64.77−97.05152,000
Back Riv above Hermann Riv66.08−96.5193,900
9South and East Hudson BaySevern55.37−88.3294,300
10Nelson 56.03−96.521,010,000
11Arctic Archipelago    
Table 2. Soil Parameters and Calibration Statistics
RegionNameInfiltration Parameter biSoil Layer Depth d2, mBasin for CalibrationNash-Sutcliffe Efficiency (Ef)Relative Error Er, %
Lena at Kusur0.91570.5
Yenisei at Igarka0.78000.91
Irtish at Tobolsk0.83832.13
Ob at Salekhard0.7781−0.92
5Barents, Norwegin sea0.3–0.50.8–1.3Seve.Dvina0.8484−0.5
Pechora at Oksino0.7705−15.41
6Yukon0.50.5–0.6Stevens Village0.83886.28
Pilot Station0.76712.45
7Mackenzie0.30.5–1.2Liard River0.8481−14.16
8Northwest Hudson Bay0.2–0.50.5–1.5Thelon0.3430−9.85
Back River0.6828−13.57
9South and East Hudson Bay0.051.0–2.0Severn<080.17
11Arctic Archipelago0.30.2   

[19] Nash-Sutcliffe efficiencies below zero indicate that the variance of the observed streamflow is smaller than the error variance; while an efficiency of one corresponds to a perfect model prediction. For 19 basins out of 27 Ef exceeds 0.75, and for 13 it exceeds 0.8. Three basins in the South and East Hudson Bay and the Yana in the Lena region have negative model efficiencies and the highest relative error (10.9%–80.82%). High relative error also occurred in the region of Kolyma (14.34%–25.28%), and in the downstream reaches of the Yenisei (9.22%–12.66%). The overestimation of streamflow in the Kolyma, and Yenisei may be in part due to too much precipitation in these regions. According to J. Adam (personal communication, 2004) the precipitation increased 20.64%, 21.65%, and 57.26% in the Lena, Yenisei, and Kolyma after applying the orographic correction of Adam et al. (submitted manuscript, 2005).

[20] Like all hydrologic models, VIC is very sensitive to the meteorological forcing data, particular precipitation. If the forcing data changes, the soil parameters that result in reproduction of observed streamflow by the model will change accordingly. Recognizing that the inputs are subject to error, some care must be taken not to overfit the parameters, which could compromise the model's representation of physical processes [Yu et al., 1999, 2002]. While research has been conducted to optimize parameters in the face of uncertain inputs [e.g., Kuczera, 1982; Sorooshian and Gupta, 1983], these schemes are difficult to apply in practice. Therefore, we used a manual calibration procedure, and attempted to estimate those parameters to which various aspects of model dynamics are most sensitive, within reasonable physical bounds. Our experience was that the VIC model did a good job of reproducing observed streamfow in the coldest areas of the domain that were mostly underlain by permafrost (e.g., Lena, Yenisei), while problems remained in areas of discontinuous permafrost (e.g., the northern Ob and Hudson Bay), where simulated streamflow was mostly oversimulated. It is not entirely apparent whether these persistent errors are due to input errors, model deficiencies, or both. We believe, however, that the oversimulation of streamflow in the Kolyma and the north of the Yenisei is mostly due to precipitation inputs that are too high, and that the problems in Hudson Bay and the northern Ob are more related to the model parameterizations.

[21] The VIC model uses conceptual schemes to represent the spatial variability in infiltration capacity and the baseflow generation. Determination of conceptual model parameters is always a practical problem in the application of the model to continental or global scales [Nijssen et al., 2001b]. Although there are no exact physical equations to determine the conceptual parameters, some empirical relationships exist between the parameters and surface and climate conditions. In this study, relatively thin soil depths (0.15–0.6 m) are observed in regions of Kolyma, Lena, Yenisei, and Yukon while thicker soil depths (0.5–2.0 m) are assigned to Ob, Mackenzie, Hudson Bay, and Barents regions (Table 2). The distribution of parameter d2 can be inferred from the permafrost extent, land cover, surface air temperature, and runoff ratio. Table 3 lists the permafrost distribution, wetlands extent, and runoff ratios for seven Arctic drainage basins. Basin average mean monthly temperatures are shown in Figure 3. The permafrost extent was based on the 25 km EASE grid Circum-Arctic permafrost and ground ice map [Brown et al., 1998; Zhang et al., 1999; T. Zhang et al., 2000], and mean temperature and runoff ratio were calculated from the observed temperature and streamflow data sets used in this study. Permafrost for each EASE grid is classified as continuous, discontinuous, sporadic or isolated. Wetlands extent, defined here as bogs, marshes, lakes, seasonal and permanent freshwater, tidal, mangroves, and lagoons, were from Revenga et al. [1998].

Figure 3.

Long-term mean monthly temperatures over seven watersheds.

Table 3. Permafrost Distribution, Wetland Extent, and Runoff Ratio for the Seven Arctic Drainage Basins
BasinPermafrost, All Classes, %Continuous Percent, %Discontinuous Percent, %Isolated Plus Sporadic Percent, %Wetland Extent, %Runoff Ratio

[22] On the basis of estimates in Table 3, the Kolyma, Lena, Yenisei, and Yukon River basins tend to have the highest runoff ratios in the domain (0.44–0.49). These basins are underlain by permafrost (all classes) 89% to 100% for Yenisei, Kolyma, Lena, and Yukon. In contrast, the Mackenzie, Ob, and Nelson, with 14 to 80% permafrost coverage, have the lowest runoff ratios (0.11–0.33). Generally basins with high permafrost coverage have a low subsurface storage capacity due to the presence of frozen water which limits infiltration into the soil, and thus a low winter baseflow and a high summer peak flow [Kane and Stein, 1983; Granger et al., 1984]. The regions dominated by permafrost have low values of soil depth (d2). There is a clear progression from relatively mild temperatures in the Ob and Nelson to very cold conditions in the Lena and Kolyma (Figure 3). Monthly temperatures in the Nelson are warmer by 17 to 21°C from October to April, and by 7 to 13°C from May to September, than those in the Kolyma. The warmer temperatures in the Ob and Nelson foster stronger evapotranspiration in these basins, and reduce runoff ratios relative to the other basins [Serreze et al., 2003]. During the calibration, we found that changes in model soil parameters had little effects on the annual runoff volumes of the southwestern Ob and Nelson basins. This lack of sensitivity of the annual runoff to model parameters was observed in results of PILPS-2e [Bowling et al., 2003] as well. One explanation may be that evapotranspiration is primarily energy-limited in this environment.

[23] A large part of the Mackenzie (48.9%), Nelson (86.8%) and Ob (11.2%) are wetlands. Warmer temperatures and the larger extent of wetlands enhance the development of the active soil layer, and result in an increased water-holding capacity, higher evaporation rates, and less runoff. The areas with higher temperatures and large wetland fractions (e.g., Ob and Mackenzie) have relatively thicker soil depths (0.5–2.0 m).

[24] Understanding the hydroclimatological features of different Arctic regions provides the basis for parameter transfer and makes the calibration process more efficient. For instance, the parameters for the Ob region were also used for the Nelson because of the similar hydroclimatological features in these two regions. The parameters for the Arctic Archipelago and Greenland were adapted from these for the Lena and Kolyma regions.

4. Model Results and Analysis

[25] We present five comparisons here to diagnose model-simulated variables and key processes using available observations including streamflow at outlets of selected watersheds, derived fields of annual runoff across the pan-Arctic land area, remotely sensed snow cover extent, site observations of ice freeze-up and break-up dates in major lakes, and point observations of active layer thickness in permafrost regions.

4.1. Streamflow Simulations

[26] Streamflow is a spatially integrated response of hydrologic processes within a watershed, and therefore is a useful diagnostic variable for assessing land surface schemes at large spatial scales. In this section, we present the streamflow simulations for the selected river basins in Table 1 and focus particularly on the Lena, Yenisei, and Ob River basins. Figure 4 compares the routed monthly and mean monthly hydrographs at three locations in the Lena basin with the observed for Aldan at Verkhoyanskiy Perevoz, Lena at Tabaga, and Lena at Kusur for the period of 1979 to 1999 (Table 1). The Aldan tributary, with a drainage area of 696,000 km2, occupies the southeast corner of the Lena basin. The upper Lena region, above the Tabaga station with a drainage area of 897,000 km2, covers the mountainous regions in the southwest corner of the Lena basin. There are no major dams in these two tributaries. The seasonal cycle of streamflow for both of these two sub-basins show very low flows during November to April and high flows during May to October, with the maximum discharge in June due to snowmelt (Figure 4). The 21-year model simulations capture the time of the peak flows induced by snowmelt and interannual variations for the Aldan basin, but overestimate the volume of the flood peak in June (Figure 4a). Figure 4b shows the streamflow simulations for the Lena above the Tabaga station. The seasonal variations show consistence with the observations, while the model underestimates the flows in July and August, and overestimates the flows in October.

Figure 4.

Observed versus simulated hydrographs at three locations within the Lena River basin: (a) Aldan at Verkhoyanskiy Perevoz, (b) Lena at Tabaga, and (c) Lena at Kusur (the mouth of the Lena River). The left column of the figure presents time series of monthly streamflow for 1979–1999, and the right column of the figure displays mean monthly streamflow for 1979–1999.

[27] There is a reservoir (with the maximum capacity 35.9 km3) located at the upper Vilui tributary of the western Lena basin. Ye et al. [2003] has documented the significant streamflow hydrology changes induced by the reservoir regulations in the Vilui valley. A comparison of the pre- and post-damming periods shows that reservoir operation maintains a large winter flow but eliminates the spring peak. For this reason, the observed streamflow at the outlet of Vilui tributary is not used to validate the VIC model, which does not take into account the reservoir effects. However the streamflow records at the Lena outlet reflect basin integration of the both natural variation and changes induced by dam regulations within the watershed. Monthly streamflow at the Lena basin outlet was reconstructed with a regression method to reduce reservoir impacts in Ye et al. [2003]. Figure 4c shows the streamflow simulations at the river outlet, Lena at Kusur (2,430,000 km2). Long-term mean monthly reconstructed discharge is also shown in Figure 4c (dotted line). Reconstructed long-term mean streamflows are lower than observed during November to May, and are higher from June to October. The simulated flow in October is much closer to the reconstructed data compared to the observed one. Due to the large basin size of the Lena, the fluctuations caused by dam regulations are moderated and smoothed out. The simulated monthly hydrographs show good consistency with both the observations and reconstructions.

[28] It is evident from Figure 4 that almost all the baseflow from January to April is underestimated. This is due in part to the nature of the frozen soil algorithm in the VIC model. A new soil frost algorithm which represents the spatial variability in soil freezing is described in Cherkauer and Lettenmaier [2003], and was used in this study. Comparison studies indicated that the model tended to increase the production of surface runoff and total runoff, and decrease the baseflow amount when the VIC model was run with the most recent version of the frozen soil model. Although representation of a spatial variation in soil frost increases the baseflow amount in winter and spring as compared to simulations using the previous point version of the frozen soil algorithm [Cherkauer and Lettenmaier, 2003], the simulated baseflow is still much lower than observed. Further work is needed to understand the reason for these errors, and to improve the model accordingly.

[29] The routed monthly and mean monthly hydrographs at three locations within the Yenisei basin are plotted in Figure 5 along with the observations for Nizhnyaya Tunguska at Bol'shoy Porog, Podkamennaya Tunguska at Kuz'movka, and Yenisey at Igarka (Table 1). The Yenisei has the world's largest reservoirs with a total capacity of constructed reservoirs of 482 km3 in the basin. Yang et al. [2004a] found that discharge at the outlet of Yenisei has increased by 30–110% from November to April, and decreased by 50% in May since the mid 1930s mainly as a result of regulation.

Figure 5.

Observed versus simulated hydrographs at three locations within the Yenisei River basin: (a) Nizhnyaya Tunguska at Bol'shoy Porog, (b) Podkamennaya Tunguska at Kuz'movka, and (c) Yenisey at Igarka.

[30] The Nizjnyaya Tunguska tributary occupying the eastern-central section of the Yenisei basin, with a drainage area of 447,000 km2, was selected for calibration because no major dams exist in this tributary. Monthly streamflows at the outlet are characterized by low flows (200–600 m3/s) from November to April and a sharp peak flow in June (19,700 m3/s). For this sub-basin, the peak flow in June is about 120 times greater than the lowest discharge in April. Low cold season baseflows in this basin are consistent with large permafrost coverage over this region. Subsurface hydrologic processes are limited to the shallow active layer owing to the ice-rich permafrost. Subsurface flow is somewhat simple as compared to those in temperate regions where the movement of groundwater is much more complicated. There is a good agreement in the pattern of both the monthly hydrograph and the mean seasonal hydrograph between simulated and observed streamflows for Nizhnyaya Tunguska (Figure 5a), but there is an apparent overestimation of peak flows in June.

[31] Another sub-basin Podkamennaya Tunguska (218,000 km2) above the Koz'moveka station (Table 1), covers the eastern portion of the Yenisei watershed. This watershed is unregulated. Mean monthly streamflow for the Podkamennaya Tunguska sub-basin shows a different regime when compared with the Nizjnyaya Tunguska sub-basin (Figure 5b). Peak flows occur in May and June, with May having slightly higher flows on average (Figure 5b). Simulated peak flows also occur in May and June, but with a slight overestimation. This shift of the highest peak flow from June to May reflects the response of the river system to a warmer winter/spring climate and an earlier snowmelt in the southern parts of the Yenisei basin [Yang et al., 2004a], which is captured well by the VIC model.

[32] Major dams exist in the southeast corner and upper portions of the Yenisei River basin. Results in Yang et al. [2004a] show that the reservoir regulations have significantly altered the monthly discharge regimes in these regulated areas. Discharge records observed at the Yenisei basin outlet tend to underestimate the natural flows in summer and overestimate the flows in winter and fall seasons. Figure 5c shows the streamflow simulations at the basin outlet. Reconstructed monthly mean streamflow in Yang et al. [2004a] is also plotted in Figure 5c (dotted line). Because of the large runoff contribution from the unregulated areas, the reconstructed and observed mean flows have similar seasonal patterns, i.e. peak flow appearing in June owning to snowmelt, high flow in summer due to rainfall floods, and low flows in winter. It is still evident that the overall effect of reservoir regulations is to enhance winter season flow and reduce summer month discharge. Low flows (from January to April) are underestimated in the VIC model. This can be attributed to the effects of reservoir regulation and the low baseflow simulations coincident with issues associated with the VIC soil frost algorithm noted above.

[33] The routed monthly and mean monthly hydrographs at three locations within the Ob River basin are compared with observations at Prokhorkino (Ob main stem), the Irtish at Tobolsk, and the Ob at Salwkhare. The upper Ob tributary, above the Prokhorkino station (738,000 km2), occupies the southeast section of the Ob basin. Annual discharge at the Prokhokino station is about 4,816 m3/s, and accounts for 38% of the Ob total flow. Monthly streamflow and the seasonal cycle are reasonably simulated for the upper Ob basin, even though the winter baseflow is underestimated (Figure 6a). The Irtish sub-basin (969,000 km2) above the Tobolsk station covers the southwest parts of the Ob basin and has an annual discharge of 2,198 m3/s, 17% of the total flow in the Ob basin. The drainage area of the Irtish sub-basin is 31% larger than that of the upper Ob tributary, but the annual discharge at the Prokhokino is only 52% of that at the Tobolsk. The runoff ratio in the Irtish sub-basin is only 0.15. There are several internal drainage basins within the Irtish that do not connect to the river (Figure 2). The Irtish is the major spring-wheat production region in Russia. Water withdrawals for irrigation (about 2.6–3.1 km3 per year) [State Hydrology Institute, 2001] and high evaporation from the cropland, and the internal drainage basins may be responsible for the low streamflow. Three reservoirs exist in the upper Irtish basin, which affect the seasonal streamflow regime along the upper Irtish valley [Yang et al., 2004b]. Streamflow records at Tobolsk reflect the combined effects of reservoir regulation and water withdrawals for irrigation. Despite these effects, the seasonal cycle of streamflow at the outlet of Irtish (Tobolsk) is well represented by the model, apparently because of runoff contributions from unregulated areas and large basin size. The timing and the volume of snowmelt peak are well simulated for the Tobolsk (Figure 6b).

Figure 6.

Observed versus simulated hydrographs at three locations within the Ob River basin: (a) Ob at Prokhorkino, (b) Irtish at Tobolsk, and (c) Ob at Salwkhare.

[34] As compared with the other large Eurasian rivers flowing to the Arctic, runoff ratios in the Ob are lower, and evapotranspiration is higher. Higher temperatures, less permafrost and larger coverage of wetlands, irrigation, and reservoirs may all contribute to the higher evaporation and smaller runoff in the Ob basin. Both the upper Ob and the Irtish sub-basins have their highest peaks in May and June, while the peak flow at the outlet of the Ob basin appears in June (Figure 6c). The simulations capture these features with a slight overestimation of spring flows (Figure 6c). The delayed response of the peak flow at the outlet of the Ob in June and July (compared with the two upper sub-basins) may be due to the following: (1) time delay of runoff routing from the upstream through the lower basin to the outlet (due to flat topography in the lower Ob); (2) higher surface storage capacity due to a high percentage of lakes and wetlands in the lower Ob [Revenga et al., 1998]; (3) late onset of snowmelt in the northern parts of the Ob region [Yang et al., 2003]; and (4) river ice jams and the delay of river ice break-up from south to north [Vuglinsky, 2002].

[35] Streamflow simulations for some basins in the other regions, such as Kolyma (Region 1), Mackenzie (Region 7), Yukon (Region 6), Barents (Region 5), Hudson Bay (Regions 8), and some independent basins with inflows to the Arctic Ocean directly, are included in Appendix A.

[36] In almost all cases, the baseflows are underestimated by the model, and in 10 of 15 cases the peak flows are overestimated (Figures 4, 5, 6, and Appendix A). This is a problem with most land surface models, and occurs mostly because the effects of groundwater-surface water interactions are not represented (VIC, like most land surface models, conceptualizes the subsurface as a soil column of depth typically 1–2 meters). Table 4 shows monthly relative error statistics based on simulated and observed streamflows of the Lena, Yenisei, Ob, Mackenzie, and Yukon. The statistics confirm that low flows are negatively biased (by 49%–78%) during the winter (November to March), and positively biased (6%–20%) during the spring (April to June). Baseflow is the main contribution to streamflow during the winter, and flood peaks usually appear in the spring. The model shows relatively smaller negative biases of 3%–13% during the summer and autumn. In general, for long-term means, the VIC model considerably underestimates the baseflow in cold season, and the errors in other seasons are relatively smaller (within 20%).

Table 4. Monthly Error Statistics Based on the Simulated and Observed Streamflows of the Lena, Yenisei, Ob, Mackenzie, and Yukon
Relative Error, %−0.66−0.72−0.780.200.120.06
Relative Error, %−0.03−0.08−0.13−0.06−0.10−0.49

4.2. Pan-Arctic Runoff

[37] Lammers et al. [2001] produced long-term annual runoff based on observed discharge for the 68% of the pan-Arctic region that is gauged for the reference period 1960 to 1989. Fields of VIC simulated mean annual runoff, gridded annual runoff from Lammers et al. [2001], and observed mean annual precipitation and temperature for years between 1979 to 1999 are shown in Figure 7. The VIC simulated annual runoff (Figure 7a) is high across southern Alaska, Barents and coastal Norway, southern and eastern Hudson Bay, mountain area of eastern Lena, and the northern portions of the Ob and Yenisei basins, where precipitation is also high (Figure 7c). Low-runoff regions include the southwestern Ob basin, the Nelson River system, the center and south parts of the Yenisei, and the center of the Lena basin (Figure 7a). The VIC simulated annual runoff is highly correlated with precipitation across many regions (Figure 7c), and the correlation with precipitation is 0.9 across all the 2,834 grid cells in the pan-arctic domain. On the other hand, in the Nelson and southwestern Ob, relatively high precipitation did not result in high runoff (Figures 7a and 7c), which is consistent with the low runoff ratio (Table 3) and high temperature (Figure 7d) and evaporation (not shown) in these two areas. Lammers et al. [2001] also indicated low runoff based on observed data (Figure 7b). The distribution of simulated high and low annual runoff fields (Figure 7a) is somewhat consistent with those of Lammers et al. (Figure 7b) in the areas such as the southwestern Ob, Nelson, eastern Hudson, and Selenga basin in northern Mongolia. Overestimates of the annual runoff are also evident in the northern Ob and Yenisei, the edge of eastern Lena, and the very southern Kolyma, which are probably due to the extremely high precipitation in those areas. Large regional variability in annual runoff is evident in Figures 7a and 7b, which are likely to be related to the differences in precipitation, temperature, land cover, and the spatial distribution of permafrost.

Figure 7.

Fields of (a) VIC simulated annual mean runoff (mm) for 1979–1999, (b) observation-based annual mean runoff for 1960–1989 [Lammers et al., 2001], (c) observation-based annual mean precipitation (mm) for 1979–1999, and (d) observation-based annual mean temperature (°C) for 1979–1999.

[38] Figure 8 shows mean monthly fields of VIC simulated runoff over the Arctic terrestrial domain. Cold-season runoff (Figure 8, November through March) is relatively low over much of the terrestrial Arctic during winter owing to a limited groundwater contribution, with the exception of southern Alaska and the Scandinavian coast, which are associated with relatively high precipitation and temperature (not shown). In Eurasian basins, the peak flow, which is mainly caused by snowmelt, progresses northward toward the higher latitudes as the temperature increases in spring. In most cases the peak values occur during the period of May to June. There is a precipitation peak in summertime (June through August) over most land areas (not shown), which contributes to the rainfall induced runoff especially in Hudson Bay, southern Alaska, and Eurasia. High evaporation, which occurs in the same season as the peak precipitation and temperature, tends to produce relatively dry conditions (e.g., Ob and Nelson basins). The inter-annual variation of monthly runoff is generally small in the cold season and large in spring and summer months due to floods associated with snowmelt which can be exacerbated by warm season precipitation. The Siberian basins show the largest seasonality due to the strong continental climate.

Figure 8.

VIC simulated long-term monthly runoff (mm) for 1979–1999.

4.3. Snow Cover Extent

[39] Snow cover accumulation, redistribution, and ablation dominate the land surface hydrology of the Arctic. The snow cover extent simulated by VIC in the pan-Arctic was compared with estimates from the weekly snow cover and sea ice extent version 2 product for the Northern Hemisphere [Armstrong and Brodzik, 2002]. This data set combines snow cover and sea ice extent at weekly intervals for the period from October 1978 through June 2001. The data were derived from digitized versions of manual interpretations of Advanced Very High Resolution Radiometer (AVHRR), Geostationary Operational Environmental Satellite (GOES), and other visible band satellite data, which were gridded to a spatial resolution of 25 km. Figure 9 compares the spatial distribution of mean number of days with snow cover for the 1980–1999 period for the two estimates. The number of days with snow progresses with the increase of latitude in both Eurasia and North America, indicative of latitudinal changes in surface air temperature (Figure 7d). A low number of snow cover days is seen in the southwestern Ob, Nelson, the southern Mackenzie, and Selenga in northern Mongolia, all of which are areas of relatively low annual runoff and comparatively warm temperatures (Figure 7d). A high number of snow cover days (270–300) is observed in the Arctic Archepelago, Greenland, and southern Alaska. In general, there is a good agreement in the pattern of spatial distribution of snow cover days between the VIC simulated and the remote sensing estimate.

Figure 9.

Remotely sensed (left) and model simulated (right) annual mean number of days with snow cover (1980–1999).

[40] Figure 10a shows the seasonal cycle of monthly snow cover extent (fraction of basin area with snow) for the Lena, Yenisei, Mackenzie, Ob, and Nelson River basins for the 1980–1999 period. Figure 10b provides the difference between the simulated and remote sensing estimates in Figure 10a. Snow cover begins to form around late September in the Lena, Yenisei, and Mackenzie, while the accumulation begins later in the Ob and Nelson basins. The warmer temperatures during September and October in the Ob and Nelson (Figure 3) lead to the late formation of snow cover in these basins. Snow cover reaches 59% of the Lena basin at the end of October, and only 20% and 9% of the Ob and Nelson basins, respectively. The VIC model simulation captures the general features of snow accumulation over the five basins. The VIC model underestimates snow cover extent over the Mackenzie and Nelson for the entire accumulation period; and a slight underestimation of snow cover extent is apparent during the period from November to January over the Ob and Yenisei.

Figure 10a.

Mean monthly basin snow cover fraction over the Lena, Yenisei, Mackenzie, Ob, and Nelson River basins (1980–1999).

Figure 10b.

The difference between the simulated and remote sensing estimates of basin snow cover fraction in Figure 10a.

[41] Snowmelt starts around late February in the Nelson basin (with a snow cover fraction of 0.91) due to much warmer winter and spring temperatures (Figure 3), and the rate of snow cover depletion is very high during February to May. Snowmelt starts in the Ob in March, the same as the Lena, Yenisei, and Mackenzie, but the snow depletion during April to May in the Ob is much faster than in the other three basins. Snow cover disappears around late June and early July over the five basins, with the Ob and Nelson ending a little earlier. The processes of accumulation and ablation are over simulated in the Lena basin, coincident with the overestimation of snow cover days over the eastern Lena. For the other basins, the model tends to underestimate the snow cover extent during the accumulation period and overestimate the rates of snow depletion in the melting season. These characteristics are limited to these five basins; the simulations for the Yukon and Kolyma (not shown) do not show the same features.

[42] Snowmelt and associated floods are the most important hydrologic event of the year in Arctic river basins. Snow accumulation and ablation processes are somewhat different for the different Arctic basins mainly owing to the spatial variation in climatology (particularly temperature and precipitation) [Yang et al., 2003]. Overall, the snow cover extent was reasonably simulated as compared to the remotely sensed snow cover data. Temperature and precipitation are two main variables controlling snow accumulation and ablation processes in the model, which may account for some of the difference between observed and simulated snow cover extents.

4.4. Dates of Lake Freeze-Up and Break-Up

[43] Information on the calendar dates of freezing and thawing of lakes is essential to understanding short-term climate phenomena in cold regions. Magnuson et al. [2000] indicated that freeze and break-up dates of ice on lakes and rivers have noticeably changed over the past 150 years in the Northern Hemisphere, with break-up advanced an average of 5.8 days per 100 years and freeze-up delayed 5.8 days. These changes correspond to an increase in air temperatures of about 1.2°C per 100 years. Freezing and thawing of lake ice are represented in the VIC model [Bowling, 2002]. The “freeze date” is defined in the model as the date on which the ice cover becomes continuous and the “breakup date” is the date on which the ice cover becomes fractional.

[44] The VIC simulated dates of lake freeze-up and break-up were compared to the records derived from the Global Lake and River Ice Phenologh Database [Benson and Magnuson, 2000] which contains freeze and break-up dates of ice on rivers and lakes across the Northern Hemisphere. Figure 11 compares observed (dark column) and VIC simulated (shaded column) time series of ice freeze-up and break-up dates for Primrose Lake (54.75°N, −110.05°W) associated with average air temperature for each year (solid line). The “effective lake” in the 100 km VIC grid cell is centered at 55.0685°N, −110.9279°W. The average absolute error of the VIC algorithm versus observations is 3 (days) for freeze-up, and 1 (day) for break-up during the period of 1979–1994 for Primrose Lake. In the ice freeze-up time series (Figure 11a), 1987 stands out with the latest observed freeze-up date (Julian day 352, December 18) coinciding with the highest annual average temperature (2.8°C), and the simulated freeze-up date is 7 days earlier for this year. Correspondingly, in the ice break-up time series (Figure 11b), 1987 stands out with the earliest observed break-up date (Julian day 125, May 5). The simulated break-up date is 4 days later than the observation. On average, the VIC algorithm is able to reasonably represent the duration of the ice-covered period for the Primrose Lake.

Figure 11.

Comparison between observed and simulated julian day of (a) lake freeze-up and (b) lake break-up for the Primrose Lake and annual average temperature. Station position: 54.75N, −110.05W; VIC grid: 55.0685N, 111.0375W.

[45] Figures 12a and 12b show scatterplots of the VIC simulated dates of lake freeze-up and break-up against the observed dates for the period of 1979–1994 for 14 lakes in Canada. The slope of the linear regression lines through the origin and their R2 coefficients are shown on the plots. The linear correlation is quite high for the break-up dates (Figure 12b), while the plot for the simulated freeze-up dates against the observed ones is much scattered (Figure 12a). Magnuson et al. [2000] suggests that the strong influence of spring runoff on the break-up dates of rivers and lakes may account for the insignificant trends in the break-up dates for Russian rives. In this study, Figure 11 and Figure 12 indicate that the difference between VIC simulated dates and observed values is larger for freeze-up than for break-up on these Canadian lakes. Overall, the performance of the VIC algorithm on the lake ice duration is quite encouraging. River ice effects (e.g., ice jams and dams) are not taken into account in the VIC model. River ice is actually one of objectives of our further model developments. For this reason, the discussion in this section has been restricted to lake ice freeze-up and break-up.

Figure 12.

Observed versus simulated julian day of (a) lake freeze-up and (b) lake break-up for the period 1979–1994 for 14 lakes in Canada.

4.5. Active Layer Thickness

[46] The soil active layer is the portion of the soil column that thaws and refreezes each year. The active layer depth (ALD) is the maximum thaw depth each year. Seasonal freezing and thawing processes in cold regions play an important role in hydrology and in the exchange of energy between the land surface and the atmosphere [Hinzman et al., 1991; Kane et al., 1991; Rouse et al., 1997]. Warming and thawing of the permafrost may result in changes in hydrology, biology, and topography [Nelson, 2003; Jorgenson et al., 2001; Zhang et al., 2004]. Results from numerical modeling, observations, and estimates from soil temperature show that active layer thickness has increased and freeze depths on seasonally frozen ground have decreased in the Northern Hemisphere in response to a general warming of high latitude climate [Stendel and Christensen, 2002; Frauenfeld et al., 2004].

[47] Soil freezing and thawing processes are represented in the VIC model [Cherkauer and Lettenmaier, 1999, 2003]. Figure 13 compares the VIC simulated ALD values on the 100 km EASE-Grid with a set of observed site data from the Circumpolar Active Layer Monitoring network (CALM [Brown et al., 2000]) for 1990 through 1999. The CALM site data were collected at 10 m or 100 m grid intervals. In this comparison, we use measurements from Canada, Greenland, Mongolia, Norway, Poland, Russia, Sweden, and Alaska. The VIC simulated ALD in Figure 13 represents the maximum value within that year, so the date is not the exact day when the CALM measurement was made. The simulation overestimates ALD in Alaska and Russia, and underestimates ALD in Norway and Poland. The observed ALD values in Mongolia are extremely low (<7 cm, red dots), and the quality of the data in this area seems questionable. The values from Canada and Sweden are more evenly distributed along the 1:1 line. The Root Mean Square (RMS) error between measured and modeled active layer depth is 98.45 cm for all the regions except for Mongolia. The comparison between modeled thaw depth based on model results for 10,000 km2 grid cells (based on averages of forcing data and soil characteristics over this area) and a measured point values should be viewed more as a rough indication of model performance rather than as a true validation [Oelke et al., 2003]. ALD is known to be highly variable both temporally and spatially, in response to meteorological conditions, topographic features, and soil moisture regime [Nelson et al., 1998]. Results from Cherkauer and Lettenmaier [1999, 2003] showed reasonable simulated ALD from the VIC model at point scale. We found in this study that the simulated ALD is very sensitive to the prescribed temperature at damping depth which is not well known (damping depth temperatures were observed in the simulations reported by Cherkauer and Lettenmaier [1999, 2003]). We are continuing to investigate the parameterization of the VIC frozen soil model, and we hope will eventually provide better simulations in cases where less observational data exist.

Figure 13.

VIC simulated versus CALM measured active layer depth (ALD) from Canada, Greenland, Mongolia, Norway, Poland, Russia, Sweden, and Alaska.

5. Estimation of River Flows to the Arctic Ocean

[48] Numerous estimates have been made of the fresh water inflow to the Arctic Ocean based on available observed streamflow data. The estimates differ considerably mainly because of different contributing land areas [Prowse and Flegg, 2000]. In order to make the estimates more comparable, Lewis [2000] recommended that only those areas that actually drained into the Arctic Ocean should be included. The recommended definition of Arctic Ocean drainage basin was illustrated in Prowse and Flegg [2000], and is designated AORB (Arctic Ocean River Basin). The water flow from the Canadian Arctic Archepelago, Hudson Bay, and Yukon are not included in the AORB. Here we use the Prowse et al. definition of the AORB to calculate the river inflow to the Arctic Ocean based on the simulated discharge from the VIC model. In this calculation, we do not include the northern tip of Greenland, primarily because VIC does not represent ice sheet and glacier water balances.

[49] Two hundred outlets along the Arctic Ocean coast were selected from the 100 km river network (solid cycles in Figure 2). In order to provide the best estimate of river inflow to the Arctic Ocean, a bias correction approach described in Snover et al. [2003] was used to remove the simulation biases for these gauged areas (no attempt was made to correct for biases, which in any event are unknown, for the ungauged areas). The resulting simulated average inflow to the Arctic Ocean for the years of 1979–1999 is 3,354 km3/yr from a contributing area of 15.017 × 106 km2. If the Canadian Arctic Archepelago is included, the estimate becomes 3,596 km3/yr for a total area of 16.387 × 106 km2. Without bias correction, the simulated annual average river inflow to the Arctic Ocean (not including the Canadian Archepelago) is 3,547 km3/yr. Including the Canadian Archipeligo, the estimate prior to bias correction is 3779 km3/yr. Uncertainties in forcing inputs, model parameters, and model parameterization may all contribute to the simulation biases.

[50] Table 5 compares these estimates of annual freshwater inflow to the Arctic Ocean with some previous estimates. For instance, Prowse and Flegg [2000] estimated the 10-year (1975–1984) average inflow to the Arctic Ocean to be 3,299 km3/yr from the AORB based on an extrapolation from gauged to ungauged areas. Grabs et al. [2000] obtained an annual inflow of 2,603 km3 from 35 gauged stations (including the Yukon River) archived at the Global Runoff Data Centre (GRDC). To estimate the streamflow at the river mouth from the unmonitored coastal area, a river transport model was used to scale up the observed station flow as the river mouth outflow in Dai and Trenberth [2002]. Two other estimates in Shiklomanov et al. [2000] from different contributing areas are also included in Table 5.

Table 5. Estimates of Annual Continental Freshwater Into the Arctic Ocean
Basin DefinitionContribution Area, ×1000 km2Volume, km3 yr−1Periods
  • a

    The first area is the gauged area; the second area is the total contributing area in the definition; the second volume is the extrapolation over the total area.

“Arctic Ocean River Basin” in Prowse and Flegg [2000]a11045/155042338/32991975–1984
“All Arctic Regions” in Shiklomanov et al. [2000]1887543001921–1996
“Arctic Ocean Basin” in Shiklomanov et al. [2000]2373252501921–1996
“Arctic Climate System” in Grabs et al. [2000]a12868/181472603/3671 
AORB − Northern Greenland + Arctic Archipelago in Lammers et al. [2001]1619233021960–1989
The largest Arctic Rivers in Dai and Trenberth [2002]168503658 
AORB − Northern Greenland, VIC11501733541979–1999
AORB − Northern Greenland + Arctic Archipelago, VIC21639735961979–1999

[51] Figure 14 shows the relationship between the inflow volume and contributing area resulting from various data sources and VIC simulations (Table 5). The inflow-area relationship is remarkably linear (with an r2 = 0.97). The VIC model is comparable to the estimates derived from the observed streamflow, although the lengths of the data series on which the various estimates are based differ. Shiklomanov et al. [2000] indicated that no noticeable changes occurred in both the annual and seasonal distribution of total river inflow to the Arctic Ocean during 1921 to 1996. The VIC simulated mean monthly discharge into the Arctic Ocean during the period of 1979 to 1999 is presented in Figure 15. Similar to the estimates of Dai and Trenberth [2002] and Grabs et al. [2000], the discharge into the Arctic Ocean has a sharp peak in June arising from snowmelt in late spring and gradually declines from August through April. The months from May to September contribute 79% of the annual discharge into the Arctic Ocean and the month of June alone accounts for 30% of the annual discharge.

Figure 14.

Basin area-annual flow volume relationship for different estimations in Table 4.

Figure 15.

Simulated mean monthly streamflow discharge into the Arctic Ocean (1979–1999).

6. Conclusions

[52] The ability of land surface models to simulate runoff and streamflow from northern river basins is critical to understanding the Arctic hydrologic cycle and to assessing runoff with different variants of climate change. In this study, a set of simulations was conducted with the land surface scheme VIC at 100 km EASE-Grid across the pan-Arctic domain to evaluate the representation of Arctic hydrologic processes in the model. The main conclusions are as follows:

[53] 1. The distribution of soil parameters that control the partitioning of precipitation and snowmelt into “fast” and “slow” response, and indirectly the partitioning of precipitation into runoff and evapotranspiration, is related to the regional climatological surface air temperature, land cover, and permafrost extent. Understanding these relationships provides the basis for model parameter transfer over the domain and makes the calibration process more efficient.

[54] 2. The 21-year simulations over the pan-Arctic land area indicated that the VIC model was able to reproduce the seasonal and interannual variations in streamflow quite well. However, for long-term means, the VIC model considerably underestimated the baseflow in the winter. Further research is needed to understand the bias and improve the model parameterization.

[55] 3. As for all land surface models run off-line, the model is highly sensitive to the precipitation forcings, and the quality of the forcing data probably is the primary source of uncertainty in the model simulations. The effects of human activities on the streamflow records are another source of prediction uncertainty. Reservoir regulations (which occurs mostly in headwater areas) tends to enhance winter season flow and reduce summer month discharge. Reservoir effects could be represented explicitly (e.g., via post-processing of VIC outputs in a manner similar to that reported by Hamlet and Lettenmaier [1999]), however this requires knowledge of reservoir operating rules, which are site specific and difficult to obtain. Nonetheless, research in progress will report on methods of representing reservoir regulation effects on the discharge of global rivers.

[56] 4. Good correspondence was observed between the simulated and satellite derived snow cover extent. The spatial variation in temperature and precipitation is the main reason for the variability in snow accumulation and ablation processes in different Arctic basins.

[57] 5. Good correspondence between the simulated and observed dates of lake freeze-up and break-up suggests that the VIC algorithm has the potential to provide information about long-term trends in lake ice duration for climate change studies.

[58] 6. Considerable error was found (RMS error = 98.45 cm) when comparing the VIC simulated active layer depth on the 100 km EASE-Grid with the CALM data collected at 10 m or 100 m intervals. Due to the large spatial and temporal variability in the active layer depth, this comparison should be viewed as an indication of model performance rather than a true validation. The sensitivity of the frozen soil model in the VIC to different scales and boundary conditions needs to be further investigated.

[59] 7. The VIC model simulated an average rive inflow of 3,354 km3/yr to the Arctic Ocean based on the drainage area definition from Prowse and Flegg [2000]. A slightly larger number (3,596 km3/yr) results when the Canadian Archepelago is included in the drainage area. Given the significant portion of ungauged areas (30%) contributing to the Arctic Ocean, the model-based estimate (which is corrected for bias over the gauged areas) offers new insights into the sources of freshwater discharge to the Arctic. On the other hand, the 183 km3/yr decrease of river inflow to the Arctic Ocean after the bias correction suggests that at the scale of the pan-Arctic domain, and for annual averages, the effects of uncertainties in the model simulations are modest.

[60] Simulation over the entire pan-Arctic drainage basin in representing Arctic hydrologic processes provides a useful approach for assessing hydrologic processes in response to the environmental changes in the pan-Arctic region. Improved parameterizations of cold hydrologic processes in coupled land-atmosphere models are expected to provide better prediction of the impacts of climate change on the northern river flow. The estimation of river flows to the Arctic Ocean can be used in climate and ocean model evaluations and offers the freshwater forcing in coupled ocean models.

Appendix A

[61] Figures A1 and A2 show streamflow simulations in the Kolyma, Mackenzie, and Yukon, and in some independent basins with inflows directly to the Arctic Ocean.

Figure A1.

Streamflow simulations for the Kolyma, Yukon, and Mackenzie River basins.

Figure A2.

Streamflow simulations for the individual basins in the region of Lena, Barents, and Hudson.


[62] This work was supported by grants 0230372 and 0230327 to the University of Washington from the National Science Foundation, Office of Polar Programs.