Improved hydrological modeling for remote regions using a combination of observed and simulated precipitation data



[1] Precipitation, as simulated by climate models, can be used as input in hydrological models, despite possible biases both in the total annual amount simulated as well as the seasonal variation. Here we elaborated on a new technique, which adjusted precipitation data generated by a high-resolution regional climate model (HIRHAM4) with a mean-field bias correction using observed precipitation. A hydrological model (USAFLOW) was applied to simulate runoff using observed precipitation and a combination of observed and simulated precipitation as input. The method was illustrated for the remote Usa basin, situated in the European part of Arctic Russia, close to the Ural Mountains. It was shown that runoff simulations agree better with observations when the combined precipitation data set was used than when only observed precipitation was used. This appeared to be because the HIRHAM4 model data compensated for the absence of observed data from mountainous areas where precipitation is orographically enhanced. In both cases, the runoff simulated by USAFLOW was superior to the runoff simulated within the HIRHAM4 model itself. This was attributed to the rather simplistic description of the water balance in the HIRHAM4 model compared to a more complete representation in USAFLOW.

1. Introduction

[2] The water balance in large river basins can be simulated in a general way using climate models—general circulation (or global climate) models (GCMs) as well as regional climate models (RCMs) [Russell and Miller, 1990], or using more advanced hydrological models [Wood et al., 1992; Kwadijk, 1993; Bergström, 1995; Kite, 1995; Arnell, 1999]. The application of either model type in remote areas gives rise to several problems. The spatial resolution of current GCMs is too coarse to accurately represent hydrological processes at regional scales [Christensen et al., 1998; Bergström et al., 2001; Intergovernmental Panel on Climate Change (IPCC), 2001]. A problem encountered with macro scale hydrological models is the limited availability of (meteorological) data that are necessary as input.

[3] Precipitation is responsible for runoff generation and is consequently the most important input variable in hydrological models. Therefore, it is necessary that precipitation data represent an accurate distribution over space and time. In remote areas, precipitation is measured only at few meteorological stations. This is further complicated in mountainous regions, as these are subject to modified precipitation amounts due to the orographic effects, but often no stations are present at all. Precipitation data generated by climate models may better represent the spatial variation in remote areas than estimates based on very restricted observed precipitation data do, although there may be apparent discrepancies between the amount simulated and available observations [Christensen et al., 1998; Christensen and Kuhry, 2000; Dethloff et al., 2002].

[4] Many studies use raw GCM or RCM output for the use in hydrological models. Kite and Haberlandt [1999] indicated that raw GCM output could give good results when used in macro scale hydrological models. Wilby et al. [1999, 2000] found that applying a statistical downscaling technique to the GCM data gave better results than the input of raw GCM data in a hydrological model. Leung et al. [1996] used an RCM with and without a parameterization of subgrid effects of topography on clouds and precipitation updated by analyses at the lateral boundaries to provide driving conditions for a hydrological watershed model in a mountainous region within the US. They showed that the hydrological model driven by the RCM version including parameterization was superior. Besides data derived from climate models, radar data can also be used to give an areal estimate of precipitation. Fulton et al. [1998] adjusted radar-derived rainfall products using a mean field hourly gauge-radar bias.

[5] The present work dealt with an approach aiming at maximizing the usage of all available precipitation data. This approach is comparable with the approach followed by Fulton et al. [1998]. The method proposed here is especially valuable in remote areas where only a few meteorological stations are present. It adjusted long-term average precipitation data generated by a 15 yearlong simulation by a high-resolution RCM (HIRHAM4) using a monthly mean-field bias correction using observed precipitation data. Runoff was simulated with a hydrological model (USAFLOW) using 1) gauge measured precipitation only and 2) a combination of long-term mean HIRHAM4 precipitation with monthly gauge measured precipitation. Runoff simulated with USAFLOW was also compared to runoff simulated directly within HIRHAM4. This comparison highlights some of the differences associated with the representation of the water balance by the climate model, and the more complete representation of the water balance represented in the hydrological model. HIRHAM4 has a somewhat simplified treatment of surface water, particularly in a subarctic environment. The approach was illustrated for the subarctic Usa basin in the northeastern part of European Russia (Figure 1). The Usa river catchment comprises the core study area for the Tundra Degradation in the Russian Arctic (TUNDRA) project funded by the European Commission (EC). TUNDRA studied the effects of global change in the East European Arctic. The main focus of TUNDRA was to assess feedback processes to the global climate system that originate in the Arctic. The present study is a key contribution to hydrological efforts within TUNDRA.

Figure 1.

The Usa basin and two subcatchments with hydrometeorological stations in the area.

[6] In section 2, the environmental setting of this work is described. Section 3 contains a description of the climate and hydrological models used. The method to combine RCM generated precipitation and observed precipitation is given in section 4, followed by the verification of this method (section 5). In section 6, runoff simulated with USAFLOW as well as HIRHAM4 is compared to observed runoff. An interpretation of the results and conclusions are given in sections 7 and 8.

2. Environmental Setting

[7] The study area was the Usa basin, in the East-European Russian Arctic (Figure 1). The catchment area measures about 100,000 km2. The Ural Mountains in the east (occupying approximately 20% of the catchment area) have an elevation up to 1800 m above sea level. The remainder of the area comprises lowland terrain with elevations between 50 and 200 m. Hypsometric curves for the Usa and two subcatchments are presented in Figure 2. Sporadic and discontinuous permafrost is present in the Usa basin. The northeastern part of the basin is underlain by continuous permafrost. A detailed map of the permafrost conditions in the region is provided by Christensen and Kuhry [2000]. In the south, taiga forests with large wetland complexes are found. The northern part of the catchment is covered with tundra vegetation. The mean annual temperature ranges from −3°C in the south to −7°C in the northernmost regions. Annual precipitation ranges from 400 mm in the lowland to 950 mm within the mountains [Christensen and Kuhry, 2000] with maximum values in summer; see also Figure 3. The discharge in the basin is characterized by a typical (sub)Arctic flow regime. In autumn and winter (October until April) only a small base flow is sustained due to the discharge of groundwater. A major runoff peak occurs in May or June due to snowmelt when temperature rises above 0°C. After the snowmelt period, runoff decreases but strong peaks occur due to heavy rainstorms in summer (July to September).

Figure 2.

Hypsometric curves for the Usa, Kosyu and Khosedayu basins and elevation of the meteorological stations (projected on the curve for the Usa basin).

Figure 3.

Mean monthly precipitation (a) and temperature (b) at Kozhim Rudnik and Khoseda-Khard and runoff (c) of the Usa at the Adzva station.

[8] Two subcatchments were studied in more detail. Both catchments are representative for a specific part of the basin (Figure 1). The Khosedayu catchment is located in the northern tundra lowland area. This area has little relief and is underlain by discontinuous permafrost. Mean annual temperature is −6.5°C and observed annual precipitation at the nearby Khoseda-Khard climate station is 430 mm. The Kosyu River is located in the southern taiga zone. The headwaters of the catchment are located in the Ural Mountains (elevation up to 1600 m). Sporadic permafrost is present, while in the Ural Mountains discontinuous permafrost exists. Mean annual temperature is −3°C and precipitation is 600 mm on annual basis at the nearby station Kozhim Rudnik.

3. Model Descriptions and Data Availability

3.1. The Regional Climate Model HIRHAM4

[9] The dynamical regional climate model HIRHAM4 applied in this study is described in detail by Christensen et al. [1996], while the model set-up used in the present study is further discussed by Christensen and Kuhry [2000]. The dynamical part of the model is based on the hydrostatic limited area model HIRLAM, documented by Machenhauer [1988] and Källén [1996]. Prognostic equations exist for the horizontal wind components, temperature, specific humidity, liquid water content, and surface pressure. HIRHAM4 uses the physical parameterization package of the general circulation model ECHAM4, developed by Roeckner et al. [1996]. These parameterizations include radiation, land-surface processes, sea surface sea-ice processes, planetary boundary layer, gravity wave drag, cumulus convection and stratiform clouds. An advanced parameterization for precipitation processes was adopted. Convection is described in the Tiedtke [1989] mass-flux formulation with modifications to the formulation of deep convection. Liquid water in stratified clouds is a prognostic variable and was treated according to Sundqvist [1988]. The land-surface parameterization uses five prognostic temperature layers and one bucket moisture layer. Runoff was calculated using the Arno scheme [Dümenil and Todini, 1992]. This is essentially a bucket formulation assuming an algebraic spectrum of field capacities in each grid point taking into account subgrid scale orographically driven variations by increasing the runoff with increased slope. The soil module in the model does not treat freezing or melting processes within the soils. Moreover when snowmelts, it goes directly into runoff, whereby it is lost from the system. Both simplifications will affect the simulation of runoff generation, but the effect may only be serious during the main snowmelt period. In winter and summer the effect will vanish. In winter, all precipitation falls as snow and is stored. In summer the active layer is thawed and infiltration is not restricted.

[10] The computational grid utilized a rotated latitude-longitude coordinate system, with the coordinates of the rotated South Pole at 55°E, 23°S, whereby the rotated equator crosses the middle of the computational domain minimizing projection effects. The model grid had a mesh of 130 by 101 points with a horizontal resolution of 0.15° by 0.15° (approximately 16 km). The vertical discretization consisted of 19 irregularly spaced levels in hybrid sigma-p coordinates from the surface up to 10 hPa with 5 vertical layers in the planetary boundary layer. The model time step was 120 s. At the lateral and lower boundaries HIRHAM4 was forced by the data from the European Centre for Medium-Range Weather Forecast (ECMWF) Reanalysis (ERA) [Gibson et al., 1997]. A simulation for the entire period 1979–1993 was carried out. The lateral forcing included all prognostic variables except the liquid water content, which is not available in ERA. The information from the lateral boundaries was transferred to the model interior by a boundary relaxation in a 10 grid point wide boundary zone with boundary data updated four times per day. At the lower boundary ERA sea surface temperature and sea-ice fraction was used. Results of model simulations in the Arctic and validation against ECMWF analyses and station data are described by Dethloff et al. [1996], Rinke et al. [1997, 1999], Kattsov et al. [2000], and Dethloff et al. [2002]. For a thorough discussion about the performance of the present model set-up, the reader is referred to Christensen and Kuhry [2000].

3.2. The Hydrological Model USAFLOW

[11] USAFLOW is a simple Geographical Information System (GIS) based water balance model that estimates the components of the water balance on a monthly basis. The model was adapted from the RHINEFLOW model that was developed for large river basins and was applied successfully in the Rhine and Meuse basins, Europe, the Ganges-Brahmaputra basin, India and the Yangtze basin, China [Kwadijk, 1993; Kwadijk and Middelkoop, 1994; Kwadijk and Rotmans, 1995; Van Deursen, 1995]. USAFLOW uses the following input variables: monthly temperature, monthly precipitation, and a Digital Elevation Map (DEM). Discharge data are necessary for verification of the model. The data are stored in a GIS using a grid of 1 km. A flow diagram of the model is presented in Figure 4.

Figure 4.

Flow diagram of the USAFLOW model.

[12] First, the water balance is calculated for each grid cell. Precipitation is treated as rain or snow depending on a critical temperature (0°C). A temperature above the critical temperature results in rainfall, a temperature below the critical temperature results in snowfall, which is stored in the snow cover. Snowmelt takes place from the snow cover and is calculated with a degree-day snowmelt model that uses a critical temperature (0°C) as well as a snowmelt factor. Evapotranspiration equations that are commonly used in modeling either need input data that are not available for the Usa basin [e.g., Penman, 1948; Priestley and Taylor, 1972; Spittlehouse, 1989] or do not give reliable estimates in cold environments [e.g., Thornthwaite and Mather, 1957]. Therefore, evapotranspiration is based on evapotranspiration simulated by the HIRHAM4 model. The spatial evapotranspiration distribution simulated by HIRHAM4 is in agreement with evapotranspiration estimates based on field data from the Usa basin [Taskaev, 1997], but the annual amount is about 20% higher than annual observed evapotranspiration estimates found in the work of Taskaev [1997]. Therefore, simulated evapotranspiration was decreased by 20%. A sensitivity analysis showed that a 20% change in evapotranspiration results in a change in runoff of only about 8%, so runoff is relatively insensitive to evapotranspiration and possible errors in the evapotranspiration estimation have little implication for the results in this paper. Actual evapotranspiration is the minimum of the available water at the surface (rain and snowmelt) and the HIRHAM4 evapotranspiration. The partition between water that is discharged directly and water that percolates to the groundwater is made with a separation coefficient, which depends on slope, soil physical properties, land use variation and soil moisture conditions. Groundwater remains stored for a longer period, but in each time step part of it is discharged as delayed runoff. The proportion of the groundwater that is discharged is defined by a recession coefficient, which depends on the geohydrological properties of a catchment. Finally, runoff is generated at the catchment outlet by summing the runoff generated in all upstream cells. The generated runoff for each grid cell reaches the basin outlet within a time step and no water remains stored in the channels. This procedure is repeated for each monthly time step.

[13] The model parameters (snowmelt factor, separation coefficient and recession coefficient) were determined by calibration. The value of the separation coefficient varies with time and was set to zero during the winter period with limited infiltration due to the seasonal freezing of soils. Runoff simulation on a monthly basis is not negatively influenced by a fixed value for the recession coefficient [Van der Linden and Woo, 2002], so a single values was chosen to reduce calibration efforts. The model was calibrated with observed meteorological data and runoff for a relatively wet 5-year period (1980–1984) and validated for a relatively dry 5-year period (1969–1973). Even though observed precipitation is underestimated, the use of these data for model calibration will not effect the choice of the parameter values, because the parameters mainly determine the timing of runoff, while the amount is determined by the precipitation input. Modeled runoff appears to be in close agreement with observed runoff, for both verification periods the R2 [Nash and Suthcliffe, 1970] is 0.9. The total annual discharge is underestimated, which can possibly be attributed to an underestimation of the actual precipitation in the observational data—one of the major points, which has motivated the present work. Nevertheless, the timing simulated by the model is correct. These outcomes indicate that the model can be used to simulate runoff in the Usa basin.

3.3. Data Used by the USAFLOW Model

[14] A DEM of the Usa basin was derived from 1:200.000 digital topographic maps from the State GIS Centre (GOSGISCentre) Moscow, Russia. From these data the Finnish Forest Research Institute (FFRI), Rovaniemi, Finland constructed a DEM with a grid size of 1 km.

[15] Time series of observed monthly temperature and precipitation were obtained from the Komi Science Centre, Russia and the Russian Meteorological Survey, Russia for the period 1950–1987. Temperature and precipitation records are available for only 13 meteorological stations (Figure 1). All stations are located in the lowland area or at the foothills of the Ural Mountains; while in the mountain areas no stations are present (Figure 2). Most stations only have meteorological records for a few months per year in the early part of the record. After 1970, records are more complete. The observed temperature data were distributed over the area using simple Thiessen polygons. The temperature in a GIS grid cell was obtained from the temperature of the representative Thiessen polygon and was corrected for the altitude of that cell (with elevation data from the DEM), assuming a lapse rate of 0.6°C per 100 m. The spatially variable temperature created this way was used as input for USAFLOW. For simulations using observed precipitation only, observed precipitation data were also spatially distributed using Thiessen Polygons. No attempt however, was made to correct for altitudinal differences, as the data sample did not permit any statistical treatment for the full altitudinal range present in the Usa catchment.

[16] Monthly discharge data of the Usa basin are available for the Adzva hydrological station, for the Kosyu River at the Kosyu station and for the Khosedayu River at the Khoseda-Khard station. Discharge records were obtained from the Komi Science Centre, Russia and the Russian Research Institute of Hydrometeorological Information, and cover the period 1950–1990.

4. Methods

4.1. Combining Observed and Simulated Precipitation

[17] The absence of observed precipitation data from mountainous regions as well as biases in the amount of precipitation simulated by HIRHAM4 were expected to cause incorrect estimates of runoff when these data sets were used separately as input in the hydrological model. To overcome these problems, a combined data set was generated, in which the spatial distribution of the precipitation calculated by the climate model was adjusted with a mean field bias correction using the absolute values of the observed precipitation data. This way, the simulated precipitation was bias-adjusted with observations. The combined data set (HIRHAM4 adjusted with observed precipitation) was generated for 1981–1986. In Figure 5, a schematic representation of the method used is shown.

Figure 5.

Flow diagram of the method used for combining observed and HIRHAM precipitation data.

[18] The spatial distribution of the precipitation simulated by HIRHAM4 is consistent with the distribution of observed precipitation at the resolution of the model. The climatology at the monthly scale as simulated with HIRHAM4 was thoroughly evaluated by Christensen and Kuhry [2000]. They found that the model was depicting the areal distribution quite realistically when compared to the limited amount of station data available in the vicinity of the Usa river catchment. Basically the same systematic errors were found for all stations, namely a tendency for too much precipitation during summer and quite agreeable amounts in the (very long) winter season. A particularly important feature of RCMs is that they take into account orographic effects (at the model resolution). In a set of simulations over Scandinavia, Christensen et al. [1998] assessed the hydrological cycle as simulated by a GCM and HIRHAM4 at two different resolutions including one at 19 km, comparable to the 16 km adopted in this work. One major conclusion from that work was that at the highest resolution (19 km) it was found that in mountainous regions the high-resolution RCM simulation shows an improved performance in simulating the various components of the hydrological cycle, compared to the GCM simulation. Moreover, when compared with observed data from Sweden, the simulated runoff indicated that a precipitation analysis based on observations is underestimating true precipitation severely, which accordingly should be almost doubled on an annual basis in the most mountainous regions of Sweden. These conclusions referred specifically to regions where the rain gauge network is particularly sparse. Due to the absence of rain gauges at high elevation in the Ural Mountains, it is not possible in the present case to assess the accuracy of the simulated precipitation amount (but see also the study by Christensen and Kuhry [2000]). The only way to assess it, therefore, is via indirect methods. One possibility is to follow Christensen et al. [1998] and try to validate the other components of the hydrological cycle and this way get at least a qualitative feeling for the accuracy. Here we used the HIRHAM4 precipitation to drive the calibrated USAFLOW hydrological model and by analyzing the resulting runoff for various subcatchments of the Usa basin we obtained such a qualitative assessment of the modeled precipitation at least in a climatological sense.

[19] A mean annual HIRHAM4 precipitation distribution was obtained by averaging the precipitation distribution over all available model years, which are 1979 to 1993. The resulting spatial distribution is represented as a GIS-layer, with high precipitation values in front of and within the Ural Mountains and low values in the lowlands. As it is not obvious how the precipitation should be distributed within the 16 km grid, the grid point values representing an area mean were used also at the 1 km grid (Figure 6).

Figure 6.

Annual precipitation distribution calculated from the HIRHAM model.

[20] Annual observed precipitation was calculated for each station and averaged for each year between 1981 and 1986. The average HIRHAM4 precipitation at the same locations was derived from the HIRHAM4 precipitation distribution calculated above. For each year, the mean-field bias was calculated, i.e., the difference between the average annual observed precipitation at the location of the stations and the corresponding average HIRHAM4 precipitation for those locations:

equation image

where: ΔP(t) is the mean-field bias for year t (mm/year); Pa(xs,ys,t) is annual observed precipitation averaged over all stations s in year t (mm/year); Pa,m(xs,ys) is average annual HIRHAM4 precipitation averaged over all stations s (mm/year).

[21] As a first step we chose to use the long-term annual mean HIRHAM4 precipitation pattern, mainly because this is a robust pattern in the sense that the annual precipitation does not vary substantially between the years. At the monthly scale, systematic biases between simulated and observed precipitation are known to vary during the year, with a relative overestimation during summer and more realistic values during winter [Christensen and Kuhry, 2000]. Sources of bias are the relative proportion of convective and stratiform precipitation during warm seasons and the gauge undercatch of snow. Evidently, gauge undercatch of snow is partly responsible for the wet bias of simulated precipitation by HIRHAM4 during the winter months [Christensen and Kuhry, 2000]. Consequently, negative rain gauge corrections in these months may be excessive, since they are based on biased observations. The biases in the warm season and in winter would suggest the use of a variable bias during the year. However, the interannual variation of monthly—and even seasonal—precipitation is much larger than for the annual mean. Therefore, the monthly HIRHAM4 precipitation bias is likely to be more variable than at the annual scale, which will require use of a different bias for each month in the entire model period. This will not only significantly increase the degree of freedom, but also prevents the application of the method to periods outside the range of HIRHAM4 simulation, while the method can be applied to other time periods when a long-term mean precipitation pattern is used. Given the desire to apply this technique outside the HIRHAM4 simulation period, the negative effect of using a long-term annual precipitation bias instead of a monthly bias on the accuracy of the results is accepted, especially when compared to using only raingauge data. The use of an annual average HIRHAM4 precipitation field instead of a HIRHAM4 precipitation field for each individual year is acceptable in relatively small geographic regions where spatial patterns of precipitation do not vary significantly from year to year, or in regions orographic modulation of precipitation is significantly greater in magnitude than interannual variations in the spatial distribution.

[22] For each year, the grid cells with the average HIRHAM4 precipitation were altered with ΔP to create combined annual precipitation for each cell:

equation image

where: Pout(x, y, t) is combined precipitation at grid cell x, y, in year t (mm/year); Pm(x,y) is average annual HIRHAM4 precipitation at grid cell x, y (mm/year).

[23] The combined precipitation was not allowed to be less than zero, because the total amount of precipitation is always larger than the overestimation of precipitation by the HIRHAM4 model. The biases in HIRHAM4 precipitation were nearly constant over the whole area where comparison with observations is possible [e.g., Christensen and Kuhry, 2000, Figure 5]. While, it seems likely that the bias should be different in the mountains than in the lowlands, it was not possible to assess how much in a quantitative sense. In a recent study, Dethloff et al. [2002] assessed the ability of HIRHAM4 to simulate net accumulation over Greenland. It was demonstrated that HIRHAM4 depicts the general patterns well, particularly the areas with very low accumulation rates over the interior parts of the Ice Sheet (annual accumulation of below 150 mm/yr). But there is also considerable evidence for a good representation of the precipitation/accumulation in the complex coastal mountain ranges, particularly in the southern part of Greenland (annual values exceed 2500 mm/yr). Given the caveat that the observations within the coastal ranges are heavily biased due to the undercatch of solid precipitation in winter and that the gauge stations are located near the sea surface and thus not representative for the complex terrain, this study suggests that even with this large dynamical range of precipitation values, there is no reason to assume that there is a simple multiplicative relation between the HIRHAM4 precipitation bias in regions with low precipitation when compared to regions with high precipitation. Therefore, the correction for the Usa basin was made in an additive way instead of using a ratio, which was done by Fulton et al. [1998]. The use of a ratio would in fact imply that biases are always higher (or lower) in the mountains. This would not depict any geographical skewness, which could easily be introduced from systematic regional errors in the mean flow conditions.

[24] The combined precipitation was converted to monthly values by using the distribution of the annual observed precipitation over the months:

equation image

where: Pm(x, y, t) is monthly combined precipitation at grid cell x, y, in month t (mm/month); d is the percentage of the annual observed precipitation that falls in month t (%).

[25] This procedure was repeated for the entire model period to create time series of monthly precipitation.

4.2. Runoff Calculation

[26] To determine the effect of using combined precipitation on runoff compared to that based on observed precipitation only, runoff was simulated with USAFLOW for 1981–1986 using 1) observed precipitation, and 2) combined observed and HIRHAM4 precipitation. The runoff generated by USAFLOW was furthermore compared to runoff directly estimated by HIRHAM4. The HIRHAM4 model provided values for runoff generated at each 16 km grid node for each time step. To generate HIRHAM4 monthly runoff values at the catchment outlet, runoff of all upstream cells was summed for each month.

5. Verification

[27] The method of adjusting the simulated precipitation pattern by HIRHAM4 with a mean-field bias correction was validated to obtain an indication of the performance of the method used. To obtain an independent validation, combined monthly precipitation was generated for each meteorological station separately, while the observed precipitation from the station under consideration was not used to calculate ΔP. In Figure 7, the observed and combined precipitation for the Adzva meteorological station are given. The R2 [Nash and Suthcliffe, 1970] between combined and observed precipitation for all meteorological stations is given in Table 1. Most stations show a very high R2, up to 0.98, while a few stations have a poor skill (Kozhim Rudnik and Sivaya Maska). By comparing Figures 1 and 5 it is seen that these two stations are situated at locations at the foothills of the Ural Mountains, where relatively large spatial variation in the simulated precipitation is found. Hence the R2 values are subject to how well the individual grid point precipitation is simulated. This is supported by the R2 values calculated with precipitation simulated by HIRHAM4 and observed precipitation (Table 1), which are also very low for both Kozhim Rudnik and Sivaya Maska. Most of the remaining stations are located far away from such variations, although Polar Urals could be subject to a similar problem. For this station, however, the grid point value actually being used seems to agree well with the observations, which could obviously be an incident of good luck. To compensate for such effects Christensen and Kuhry [2000] applied a nine point smoother to the model data before comparing with observations. They demonstrated that this was only important very near to the Ural Mountains. The present discussion supports this conclusion.

Figure 7.

Observed and combined precipitation at Adzva meteorological station.

Table 1. Nash and Suthcliffe Criterion for Combined Precipitation Versus Observed Precipitation and HIRHAM4 Precipitation Versus Observed Precipitation
Meteorological StationR2 Combined PrecipitationR2 HIRHAM4 Precipitation
Khorei Ver0.730.05
Adzva Vom0.890.01
Kozhim Rudnik0.61−1.47
Sivaya Maska0.61−0.74
Khalmer Yu0.910.55
Polar Urals0.970.70

5.1. Runoff-Precipitation Ratios

[28] Runoff-precipitation ratios (R/P ratios) give an indication of the proportion of precipitation that is discharged. Woo et al. [1994] and McNamara et al. [1998] found that R/P ratios in Arctic regions with a permafrost-dominated hydrology are usually high. High R/P ratios can result from processes acting in cold regions: low evaporation and limited subsurface storage due to frozen soils. Furthermore, high R/P values can result from errors in the data or incomplete data; for example, mountainous areas with high precipitation values are usually ungauged. R/P ratios were calculated for the entire Usa basin and the two subcatchments using observed precipitation and observed runoff, using combined precipitation and observed runoff, and using precipitation from the HIRHAM4 model and observed runoff data for the years 1981–1986.

[29] R/P ratios derived from observed precipitation and observed runoff are 0.7 for the northern and 1.2 for the southern subcatchment (Table 2). For the entire Usa an intermediate value of 0.9 is obtained. Annual R/P ratios that are higher than 1 during a sequence of years are physically impossible, because they suggest that more water leaves the system than had entered it. R/P ratios for the HIRHAM4 precipitation and observed runoff data range from 0.5 and 0.7, and ratios for the combined precipitation are between 0.6 and 1.0. The R/P ratios for both HIRHAM4 precipitation and combined precipitation are more plausible than ratios based on observed precipitation, because all values are lower than 1. Also, they are in better agreement with values found by Woo et al. [1994] and McNamara et al. [1998] in other permafrost dominated regions, i.e., 0.6–0.8.

Table 2. Runoff-Precipitation Ratios for Khoseda-Khard, Adzva, and Kosyu With Observed Runoff and Precipitation From HIRHAM4, Observed Precipitation, and Combined Precipitation
Hydrological StationR/P Ratio
HIRHAM4 PObserved PCombined P

6. Model Results

[30] Runoff simulation with the different data sets shows the effect of using different data types (observed or combined precipitation) and model types (RCM or hydrological model). In Figure 8, the R2 [Nash and Suthcliffe, 1970] as well as the deviation of annual volume between observed runoff and simulated runoff are shown for the entire Usa basin at the Adzva station as well as the Khosedayu and Kosyu subcatchments for each year in the model period (1981–1986). Discharge data for the years 1985 and 1986 were not available for the Kosyu catchment and consequently R2 values could not be calculated for these years. Hydrographs for the three catchments for 1983, a relatively wet year, are illustrated in Figure 9.

Figure 8.

Nash and Suthcliffe criterion and percentage deviation of simulated runoff (with USAFLOW using observed precipitation, with USAFLOW using combined precipitation and with HIRHAM4) for the Usa (a,b), the Kosyu (c,d) and the Khosedayu basin (e,f).

Figure 9.

Modeled runoff and observed runoff for the Usa (a), the Kosyu (b) and the Khosedayu basin (c).

[31] A comparison of the R2 of the runoff simulations with different precipitation data (Figure 8) shows that for all catchments and years, runoff simulated with USAFLOW is in better agreement with observed runoff (R2 is 0.6–0.9 for the Usa, 0.4–0.6 for Kosyu and 0.5–0.7 for Khosedayu) than runoff simulated with HIRHAM4 (R2 is 0.1–0.8 for the Usa, 0–0.6 for Kosyu and 0.2–0.6 for Khosedayu). An exception is the year 1981 for the Usa and Kosyu catchment, where HIRHAM4 performs better. In general, the use of combined precipitation to simulate runoff with USAFLOW results in a slightly higher R2 than the use of uncorrected observed precipitation. Exceptions are the years 1982, and 1986 for the Usa and 1982, 1983 and 1986 for the Khosedayu catchment. Annual runoff simulated with both observed and combined precipitation is underestimated for the three catchments. The deviation from observed runoff for each year is significantly smaller when combined precipitation is used for modeling (deviation is 0 to −37% for the Usa, for −19 to −45% Kosyu, and 20 to −38% for Khosedayu) than when observed precipitation is used (−18 to −49%, −43 to −76%, and 9 to −43%). Exceptions in the Khosedayu catchment are 1982, 1983 and 1986 when runoff is overestimated. Runoff simulated with HIRHAM4 is overestimated in the Usa (9 to 37%) and underestimated in the Kosyu (−32 to −46%) and the Khosedayu catchment (−58 to −69%). As expected, the most consistent improvements when using the combined versus only the observed precipitation is achieved for the Kosyu basin, which is the most mountainous basin.

[32] The seasonal variation in runoff simulated with the HIRHAM4 model shows major differences from observed runoff (Figure 9). The snowmelt runoff peak starts too early in the season (March instead of April or May). It is very likely that a substantial part of this discrepancy is due to the instantaneous conversion into runoff, when the snowpack starts to melt. If refreezing was considered the water would be withheld for a longer period. Runoff simulated with USAFLOW and observed precipitation produce better results, although deviations from the observed runoff exist in summer as well as in fall, especially for the Khosedayu catchment. The annual variation in runoff simulated with observed precipitation is similar to the variation in runoff simulated with combined precipitation, but total volume and peak flow simulation are improved.

7. Interpretation of the Results and Discussion

[33] Precipitation data derived by combining observed and RCM data seem to provide a spatially distributed precipitation field that is superior to observations or RCM rainfall fields alone. Runoff simulations can benefit from the use of such data. A comparison between observed precipitation and combined precipitation at the location of the meteorological stations (Table 1) shows that not all stations have a high R2 value. This cannot be ascribed to the effect of elevation or mean annual precipitation only, i.e., low R2 values do not necessarily occur at high elevation or with high precipitation values. Instead the low correlation appears to be caused by the geographical location of the station sites within a relatively large precipitation gradient, whereby the exact representation of the location in the HIRHAM4 model becomes important. Therefore, methods such as the adjustment of precipitation data with a relationship between observed precipitation and elevation to correct precipitation values of ungauged pixels or the use of a multiplicative bias field do not seem advantageous. Furthermore, the rain gauges in the area do not allow for any statistical treatment of the data, which could take elevation variation into account in a GIS based application of filling out the 1 km grid pixels. All stations are situated in a very narrow zone between 150 m and 200 m above sea level.

[34] Runoff simulation can be significantly improved (as measured by R2 and deviation from annual observed runoff) by using precipitation data derived from the spatial distribution of the HIRHAM4 model adjusted with observed precipitation instead of observed precipitation directly. In the present case, we have demonstrated that it is possible to adjust HIRHAM4 precipitation somewhat downward, with the excess precipitation mainly being a summer time phenomenon. The mean annual HIRHAM4 precipitation pattern used to adjust observed precipitation is spatially variable and takes into account the high precipitation amounts in the mountains, which are not available from the observed data set. The error resulting from the use of a long-term mean precipitation pattern instead of a interannually variable pattern is not great enough to cancel out the benefits of this approach.

[35] The largest improvement is seen in the mountainous Kosyu catchment, where the use of only observed precipitation result in an underestimation of runoff ranging from −43 to −76% in a year, while the underestimation of runoff simulated with the adjusted precipitation data ranges from −19 to −45%. In lowland areas, such as the Khosedayu catchment, the effect is smaller because precipitation is less variable over the area and therefore it is thought to be better represented by the limited observed data. The volume of simulated peak flow (summed over the months May and June) is for all catchments lower than the volume of observed peak flow. The deviation in volume simulated peak flow from observed peak flow is only 4% for the entire Usa basin, but can be as high as 20% for Kosyu and Khosedayu. This may be the result of gauge undercatch of snow during the winter months, causing the combined precipitation fields to be excessively dry. Simulated runoff is overestimated for 1982, 1983 and 1986 in the Khosedayu catchment. Figure 9 illustrates that this can be attributed to an overestimation of flow in August, caused by overestimation of precipitation, underestimation of evapotranspiration, or too little storage. As was indicated in a section 3, runoff is relatively insensitive to possible errors in the evapotranspiration estimation, which leaves the overestimation of precipitation or the underestimation of water storage as possible causes.

[36] In all areas, runoff simulated with USAFLOW is in better agreement with observed runoff than is the runoff simulated directly with the HIRHAM4 model. The deviation of HIRHAM4 runoff from observed runoff is not only caused by the error in timing of the runoff, but also by the relatively coarse spatial resolution of the HIRHAM4 model. This makes it difficult to delimit the catchments and to estimate the model parameters in the hydrological processes accurately. Another cause may be the simplistic treatment of surface water and soil processes in connection with thawing, i.e., neglecting refreeze of meltwater, but there are also indications from other studies [Hagemann et al., 2001], which point toward an excessive evaporation during snowmelt, which could lead to a local positive feedback of precipitation falling as rain, which would subsequently be directed into runoff, since the soils are completely saturated. As previously noted, HIRHAM4 tends to overestimate evaporation on an annual basis. For the present study area, this is only important during spring and summer (April to August). However, one of the consequences is an intensified recycling of the surface water, which was also argued by Christensen and Kuhry [2000] to be the main reason for the summer time wet bias of the model. We do note that the runoff simulated by HIRHAM4 is surprisingly realistic at the full basin scale. But it is also taken ad notem that in order to treat the hydrological cycle in detail for such an area as the Usa basin additional and more advanced hydrological modeling tools such as provided by the USAFLOW seem necessary.

8. Conclusions

[37] The method of combining precipitation data from an RCM and observed data was successfully applied in the Usa basin. Runoff simulated with observed precipitation data in combination with the spatial distribution deduced from simulated precipitation patterns by the HIRHAM4 climate model is in better agreement with observed runoff than when using observed precipitation only. Even though the modeled runoff does not perfectly match the observations, the total annual amount is in better agreement with the observed amount of runoff as is the seasonal cycle. Combining observed precipitation data with data from an RCM could significantly improve hydrological model simulations with respect to annual volume especially in remote areas and mountainous regions such as the Usa basin. We also find that the use of a hydrological model allowing for lateral discharge instead of a climate model describing runoff only locally offers a better simulation of runoff variation within the year as well as the total annual amount of runoff.

[38] The lack of spatial resolution in meteorological inputs has long deterred the reliable use of area-distributed hydrological models for impact study of environmental change of water resources. Through a combination of USAFLOW with HIRHAM4 used as an advanced data interpolator, we have demonstrated that more consistent meteorological information can be constructed and used to drive a simple GIS based water balance model. The very fact that this procedure seems to work out, suggests that the HIRHAM4 simulated precipitation amounts and geographical distribution are realistic, at least in a climatological sense. The approach adopted in this study should be further explored by possibly refining the algorithm used to scale the HIRHAM4 simulated precipitation to the observational records.


[39] We would like to thank Peter Kuhry for many fruitful discussions on the topic of this paper. We also thank the anonymous reviewers for their useful comments. This study was supported by the EC Environment and Climate Research Programme (TUNDRA project, contract ENV4-CT97-0522, climate and natural hazards). Some field observations and consultations with Russian colleagues were made available with support from INTAS (PERUSA project, contract 97-10984).