To improve the representation of the land surface in their operational numerical weather prediction (NWP) models, the Meteorological Research Division of Environment Canada (EC) is developing an external hydrometeorological modeling and data assimilation system. The objective of this study is to verify the improvement in simulating snow cover extent (SCE) and snow water equivalent (SWE) over the Canadian Rockies with this new modeling system. This study will be an important first step in determining the optimal configuration of the land surface model and atmospheric forcing for a future operational implementation. Simulated SCE is compared with the Interactive Multisensor Snow and Ice Mapping System (IMS) analysis, while simulated SWE values are verified against a series of manual snow survey sites located within the Canadian Rockies. Results show that land surface model simulations of SCE and SWE were sensitive to precipitation forcing. Simulations at both low and high resolution forced with EC’s experimental precipitation analysis were found to underestimate SCE and SWE values. Mountain snowpack retreated too early during the spring melt period. Precipitation forcing derived from EC’s short-range NWP model resulted in improved values for both SCE and SWE, which also contributed to higher contributions to streamflow. Terrain adjusting the atmospheric forcing data was found to be important for properly modeling local extreme SWE values. A comparison with available precipitation observations over the Canadian Rockies region found EC’s experimental precipitation analysis to possess a negative precipitation bias that increases with increasing elevation.
Land surface models and assimilation systems have been used for a few decades to provide surface conditions and fluxes for numerical weather prediction (NWP) systems (e.g., Bélair et al. 2003a). The aim of land surface models is to simulate the exchanges of mass, energy, and momentum between the soil and vegetated surfaces and the atmosphere (Rodell et al. 2004). Driven by advances in computing power and an increased understanding of the underlying physical processes that influence the coupled earth–atmosphere system, land surface models have continued to evolve in sophistication and currently demonstrate an improved ability to simulate complex interactions (Bélair et al. 2003a,b; Mitchell et al. 2004; Rodell et al. 2004; Gottschalck et al. 2005).
Snow and ice are key components of the hydrologic cycle over North America, acting as storage reservoirs for moisture during the winter, influencing the incoming energy fluxes and the subsequent hydrologic conditions in the spring melt season (Cayan 1996; Brown 2000; Pan et al. 2003; Leung and Qian 2003; Mote et al. 2005; Mote 2006). An increasing number of studies have demonstrated the importance that winter snowpack has upon seasonal forecasting owing to the important linkages between snowpack and the dominant atmospheric circulation regimes in the spring and summer seasons (Cohen and Entekhabi 1999; Gong et al. 2002).
Two important variables related to cold season processes are the snow water equivalent (SWE) and the snow cover extent (SCE). The accurate representation of SCE in land surface models is important for the correct partitioning of the incoming radiation owing to the impacts of snow cover on surface albedo (Slater et al. 2001; Bélair et al. 2003b; Sheffield et al. 2003). Several studies have documented the important influence that SCE has upon the correct modeling of snow processes (Sheffield et al. 2003; Rodell and Houser 2004; Clark et al. 2006; Zaitchik and Rodell 2009) and recent efforts to improve both land surface and hydrologic predictions have focused upon the assimilation of SCE data (Rodell and Houser 2004; Clark et al. 2006; Kumar et al. 2008). SWE on the other hand is important as it quantifies the amount of frozen moisture storage and controls the amount of spring melt and subsequent runoff and flooding (Pan et al. 2003; Mote et al. 2003). Basin-scale estimates of water stored as snow is essential for managing water resources, forecasting runoff, and understanding climate change. For many river basins located throughout western North America, snowpack is the largest component of water storage (Cayan 1996; Mote et al. 2005; Hamlet et al. 2005; Lapp et al. 2005; Mote 2006).
One of the most important atmospheric forcing to land surface models is precipitation (Mote et al. 2003; Pan et al. 2003; Cosgrove et al. 2003; Gottschalck et al. 2005; Feng et al. 2008). Using simulations of the Global Land Data Assimilation System (GLDAS), Gottschalck et al. (2005) examined the sensitivity of land surface states to several external precipitation forcing datasets and noted important sensitivities, up to 100–400 mm in monthly averaged SWE, for regions located over the interior west and the northeast of the United States. In a comparison of simulated SWE values from four different land surface models with automated snow telemetry (SNOTEL) measurements over the western United States, Pan et al. (2003) noted a reduction in negative SWE bias when precipitation forcing values were adjusted both locally and regionally. The authors found that biases in precipitation forcing were more important than temperature forcing biases in determining simulated SWE errors at these high-elevation stations. Temperature forcing becomes more important in explaining SWE trends with decreasing elevation as noted by Mote (2006).
In an effort to improve the representation of the land surface in their operational NWP models, the Meteorological Research Division (MRD) of Environment Canada (EC) is currently developing a hydrometeorological modeling and data assimilation system. The essential component of the hydrometeorological modeling system is the land surface model, driven by three principal sources of information: geophysical land surface characteristics, atmospheric forcing, and initial conditions. Coupled to the land surface model are a subsurface lateral flow model and a river routing model (Soulis et al. 2000). The hydrometeorological modeling system currently runs in an external mode whereby the atmospheric forcings interact with the land surface without feedback on the atmosphere.
A collaborative project between EC and Agriculture and Agri-Food Canada (AAFC), the National Agri-Environmental Standards Initiative (NAESI), was recently completed (Pellerin et al. 2009). The water availability subcomponent of the NAESI water theme had as a goal to produce water-balance indicators using a coupled hydrometeorological modeling system for the South Saskatchewan River basin (SSRB) located in western Canada. The NAESI study provided an excellent opportunity to evaluate the external hydrometeorological modeling system currently under development as part of the extensive prevalidation work needed prior to an operational implementation.
Owing to the importance of mountain snowpack as moisture storage for subsequent spring runoff over the SSRB, the objective of this study is to verify the improvement in simulating SCE and SWE over the Canadian Rockies with this new modeling system. In particular, the sensitivity of the simulated SCE and SWE to changes in land surface model resolution and precipitation forcing will be examined. The land surface model will provide the very important first-guess component to the assimilation system and evaluation studies such as this are an important first step in determining the optimal configuration of both the land surface model and the atmospheric forcing for future operational implementations.
The simulation of mountain snowpack is a very challenging task (Essery et al. 1999; Slater et al. 2001; Boone and Etchevers 2001; Feng et al. 2008) as SCE and SWE are affected by a wide variety of processes and land surface characteristics such as elevation and shading effects, vegetation cover, and wind redistribution. Numerous land surface models of varying complexity have been developed, however; more complex physics does not necessarily mean that coupled atmospheric simulations will be improved. The choice of the Canadian Rockies domain for this study was somewhat dictated by the objectives of the NAESI project, nonetheless this study will provide an opportunity to highlight problem areas and suggest future directions for development work.
The organization of the paper is as follows. In section 2 the hydrometeorological modeling system is described and the atmospheric forcing data and datasets used for verification are described. In section 3 the experimental design is discussed. Verification of model-simulated SCE and SWE is presented in section 4 along with a quantitative assessment of the impacts of terrain adjusting the atmospheric forcing data. A comparison of precipitation-forcing datasets is also presented in section 4. Discussion of problems identified and future directions in model system development is presented in section 5.
2. Modeling system and datasets
a. Hydrometeorological modeling system
At EC’s MRD, a hydrometeorological modeling system was developed by combining model components taken from the NWP, climate prediction, and hydrological prediction infrastructures. Currently, atmospheric forcing data from EC’s operational NWP model, the 15-km version of the regional Global Environmental Multiscale (GEM) model (Mailhot et al. 2006), is used to drive the evolution of the land surface state in external mode. Air temperature, specific humidity, and wind forcing data are taken from the regional GEM’s lowest vertical level, which is at a height of roughly 40 m. This forcing configuration was found to be important in allowing for an interactive evolution of the atmospheric surface layer, which provides a land surface feedback at screen level (Balsamo et al. 2007). For more details on the regional GEM model the reader is referred to Mailhot et al. (2006).
The land surface modeling component of the hydrometeorological modeling system in this study is the Canadian implementation of the Interactions between Soil, Biosphere, and Atmosphere (ISBA) model (Noilhan and Planton 1989), which is described in detail in Bélair et al. 2003a,b. The Canadian implementation of ISBA consists of two layers, a superficial layer and a rooting-depth layer, in which the temperature and soil water contents evolve following the force-restore equations (Bélair et al. 2003a). A single energy budget calculation is performed over the land portions of the model grid with the surface thermal coefficient accounting for the area-weighted contributions of bare soil, vegetation, and snow within the grid cell.
A single layer snow scheme simulates the evolution of the prognostic variables related to snow and accounts for the liquid water reservoir in the snowpack, the melting effects of incident rainfall on the snowpack, and includes a new formulation for snow density (Bélair et al. 2003b). The prognostic variables include the snow mass, liquid water in the snowpack, relative snow density, and snow albedo. Snow density evolution accounts for decreases of density due to fresh snow, increases owing to snow aging, and increases due to the freezing of liquid water in the snowpack. The evolution of snow albedo depends upon whether the snow is melting or not, decreasing linearly if no melting is occurring or exponentially if melting is occurring. Comparisons of the ISBA modified snow scheme with ISBA’s original snow scheme (Douville et al. 1995) in stand-alone mode simulation experiments showed considerable improvements in the representation of important snowpack characteristics such as SWE and snow density values (Bélair et al. 2003b). For more details on the snow scheme please refer to Bélair et al. (2003b).
The WATROF parameterization (Soulis et al. 2000) for overland flow, subsurface flow, and groundwater recharge was coupled to ISBA to improve the vertical water balance calculations. The WATROUTE parameterization, a gridded river routing model (Kouwen 2010), was used to route the water fluxes generated by ISBA. The overland flow parameterization employs a kinematic wave-routing model to delay the surface runoff produced by a variable infiltration capacity model (VIC; Liang et al. 1996) in ISBA. To account for the importance of sloping terrain for runoff production, the hillslope model of Soulis et al. (2000) was used to generate interflow at the seepage face. Soil suction is accounted for by ensuring that the water content stays above bulk field capacity as defined by Soulis et al. (2010).
b. Atmospheric forcing
In external mode the hydrometeorological modeling system is driven by atmospheric forcing data, which come from 6–18-h regional GEM model forecasts. As noted by Mahfouf et al. (2007), model spinup issues act to severely underestimate precipitation amounts in the 0–6-h regional GEM model forecasts, precluding their usage as atmospheric forcing. The atmospheric forcing variables taken from the regional GEM model are shortwave and longwave radiation incident at the surface, temperature, specific humidity, wind, surface pressure, and precipitation.
In this study precipitation forcing data comes from two related sources. Regional GEM model forecasts of hourly precipitation accumulation are used as one type of input precipitation forcing. Mailhot et al. (2006) document significant improvements in regional GEM model precipitation forecasts in terms of sharper structures, better representations of trace precipitation amounts, and overall better quantitative precipitation verification scores, when compared with earlier operational versions of the regional GEM model.
The other source of precipitation information comes from the Canadian Precipitation Analysis (CaPA; Mahfouf et al. 2007). CaPA is constructed using the optimum interpolation (OI) procedure, which is a statistical interpolation technique currently used at the Canadian Meteorological Centre (CMC) to generate screen-level analyses of several meteorological variables (Brasnett 1999). The CaPA analysis begins with a 6-h regional GEM model precipitation forecast which serves as a first guess or background field. The use of model precipitation information within CaPA is necessitated by the low density of the observational network in the northern regions of Canada.
The observational database used by CaPA consists of surface synoptic reports of 6-h precipitation accumulation from the SYNOP network. The OI technique performs the analysis on innovations (difference between an observation and the corresponding background value) and requires the specification of error statistics (Mahfouf et al. 2007). To bring the probability distribution of the innovations as close as possible to a normal distribution a cubic root transformation is performed on the innovations and the OI analysis is performed in this transformed space. The observation and model error along with the correlation length scale are calculated at each analysis time based upon a theoretical fit to the experimental semivariogram (Pannatier 1996). The 6-h CaPA precipitation product is disaggregated into hourly accumulation totals using the temporal structure provided by the regional GEM hourly precipitation.
The analysis of precipitation in mountainous terrain poses very unique problems (Daly et al. 1994; Schultz et al. 2002; Fortin et al. 2007). Surface observing stations are typically located in valleys (not at the top of the mountains) and thus are not able to capture the very important precipitation–elevation relationships found in areas of complex terrain (Daly et al. 1994; Schultz et al. 2002). Moreover, snowfall measurements are systematically underestimated for windy conditions (Fortin et al. 2007). CaPA does not explicitly account for topography when estimating precipitation amounts; however, the regional GEM first-guess precipitation field used in CaPA does account for topographic variations in precipitation. The CaPA precipitation product is used in this study because it may soon become operational and this study provides an opportunity to examine the impacts of using surface gauge precipitation observations in constructing a precipitation analysis in complex terrain.
c. Geophysical fields
To describe the land surface characteristics several databases from different sources are used in this study. For the lower-resolution experiments (i.e., 15 km; see section 3a) a global U.S. Geological Survey (USGS) database at 1-km resolution is used to characterize the orography. For the high-resolution simulations (1 km), the Canadian Digital Elevation Dataset (CDED) at 20-m resolution is restricted to the Canadian territory and was thus combined with the Shuttle Radar Topography Mission (SRTM) digital elevation model with a 90-m pixel size to cover the entire integration domain. Vegetation characteristics, including vegetation type and fractional coverage, for both the high- and low-resolution experiments are derived from the global USGS database. Soil texture information is provided by combining a U.S. Department of Agriculture (USDA) database (STATSGO) at 1-km resolution, restricted to the U.S. territory, and an AAFC database at 10-km resolution.
d. Verification datasets
Snow cover data is obtained from the Interactive Multisensor Snow and Ice Mapping System (IMS) analysis from the National Oceanic and Atmospheric Administration (NOAA; Ramsay 1998; Helfrich et al. 2007). The IMS analysis incorporates data from various satellite platforms along with derived mapped products and surface observations (Ramsay 1998). Owing to difficulties in deriving operational passive microwave algorithms, the IMS analysis is still largely based upon manual analyses of visible satellite imagery (Romanov et al. 2000; Brubaker et al. 2005; Brown et al. 2007; Helfrich et al. 2007). IMS daily snow cover maps are available at a nominal resolution of 24 km covering the Northern Hemisphere. Beginning in February 2004, the IMS snow cover maps have been produced at 4-km resolution largely in response to the needs of the operational weather forecasting community (Helfrich et al. 2007).
Several studies have evaluated the IMS snow cover product against other satellite-based snow products, including the Moderate Resolution Imaging Spectroradiometer (MODIS) and QuikSCAT platforms (Romanov et al. 2000; Sheffield et al. 2003; Pan et al. 2003; Brubaker et al. 2005; Brown et al. 2007). Additionally, the IMS analysis has been verified against in situ ground observations and automated SNOTEL sites (Brubaker et al. 2005). The consensus from the above studies is that the IMS product performs well for the core winter months of December, January, and February. Where the IMS product, along with other satellite-based products have difficulty is in the early winter period and during the spring melt from March to June (Brubaker et al. 2005; Brown et al. 2007). When the IMS analysis is compared to station data, snow detection rates varied from a low value of roughly 20% in November to a maximum of 95% in December, with a value near 70% for the month of March (Brubaker et al. 2005). Brown et al. (2007) in their study of snow cover duration over the Canadian Arctic region found the IMS product to be superior in capturing the spatial variability in spring snow cover duration but was found to possess a positive bias of 22–26 days most likely related to frequent cloud cover.
SWE is also difficult to measure from satellite platforms, especially in areas overlain by dense forest cover (Derksen et al. 2003, 2005; Dong et al. 2005). It is well known that the accuracy of satellite SWE retrievals diminishes in areas of complex terrain owing to the apparent increase in microwave emission resulting from multiple reflections off of the rough surfaces. On the other hand, direct comparisons of modeled and observed in situ SWE values suffer from issues of representativeness and scale (Pan et al. 2003; Mote et al. 2003), as simulated SWE values represent a grid-averaged value while the observed in situ SWE values represent point measurements. Differences in elevation between the model grid box and the in situ SWE observation site can introduce systematic biases (Pan et al. 2003; Leung and Qian 2003).
Within the Canadian Rockies manual snow surveys are conducted on a monthly basis over the winter/spring season by the British Columbia Ministry of Environment, Water Stewardship Division (more information available online at http://www.env.gov.bc.ca/rfc/). These data archives extend back to 1935 and are used to provide warnings and forecasts of hydrological conditions throughout the province of British Columbia. Measurements are typically taken at the beginning of the month from January through April and twice monthly in May and June. For each snow survey site, the recorded snow depth, SWE, and snow density values represent averages from a sample of sites measured. These data have been used in previous studies of western North American mountain snowpack (Mote et al. 2005; Mote 2006). The locations of the snow survey sites utilized in this study are shown in Fig. 1b.
3. Experimental design
To examine the sensitivity of the simulated mountain snowpack four simulations with the external hydrometeorological modeling system over the SSRB domain for the period 20 May 2004 to 31 October 2007 are performed. Figure 1a shows the geographical domains for the four simulations. Experiment CaPA-15 consists of running the external system at 15-km resolution with CaPA as the precipitation forcing. In experiment GEM-15 the precipitation forcing is taken from the regional GEM model while the model resolution remains at 15 km. For the final two sensitivity experiments the horizontal resolution is increased to 1 km and both the CaPA (CaPA-1) and the regional GEM (GEM-1) are used as precipitation forcing. Each model simulation consists of uninterrupted 24-h integrations of the ISBA land surface model with daily updates of land surface characteristics and atmospheric forcing data. The model time step in all 4 simulations is 30 min.
When running the external system at a higher resolution than that of the forcing data, a downscaling of the atmospheric forcing is performed to take into account the change in model resolution (Cosgrove et al. 2003; Mitchell et al. 2004). Owing to differences in the topographic fields between the 1-km simulations and the smoothed version of the regional GEM forcing model, temperature, humidity, and surface pressure forcing data were downscaled to the higher-resolution topography for experiments CaPA-1 and GEM-1. A lapse rate of 6°C km−1 was used to downscale the temperature forcing, which are then used to hydrostatically adjust the surface pressures. Conservation of relative humidity is assumed to downscale the specific humidity values to the 1-km grid. Offline tests comparing different lapse rate values found that the 6°C km−1 value provided the best overall fit to surface observations. The difference in topography between the 15-km regional GEM forcing model and the 1-km simulations is shown in Fig. 1b. Elevation differences are on the order of several hundreds of meters, which translate into temperature adjustments of several degrees Celsius. The general trend for the high-resolution topography is for elevations to be increased over the mountains and reduced in the valleys as evidenced by the distribution of elevations in Fig. 1c.
The downscaling of the temperature forcing data has an important influence upon the precipitation phase. In the hydrometeorological modeling system, the air temperature forcing data at 40 m is used to distinguish between snowfall and rainfall. Precipitation occurring in association with air temperature forcing greater or equal to 0°C is classified as rain, while precipitation occurring in association with air temperature forcing less than 0°C is classified as snow.
4. Model simulation verifications
a. Snow cover extent
The calculation of the macroscale distribution of snow cover within a model grid, referred to as the SCE, is difficult owing to the lack of a precise definition of the term SCE (Slater et al. 2001; Sheffield et al. 2003; Bélair et al. 2003b). The weaknesses and somewhat crude nature in which SCE is calculated in the CMC operational version of ISBA are documented in Bélair et al. (2003b). After a careful review of the literature and the testing of several SCE algorithms, the vegetation-dependent SWE formulas used in the Noah land surface model, which are outlined in Koren et al. (1999) and Sheffield et al. (2003), were chosen.
In this algorithm, maximum values of SWE are specified above which the grid cell is considered snow covered. If the modeled SWE is less than the maximum SWE value then the SCE is calculated following the snow depletion curve given in Sheffield et al. (2003). Three cover types are considered: forest, bare soil, and other. Consistent with the release of version 2.6 of the Noah land surface model the maximum SWE values for each of the three cover types are reduced by half from those of Sheffield et al. (2003), while the curve shape parameter is increased from 2.6 to 4.0.
The IMS determination of snow/no snow is somewhat qualitative and subjective (Romanov et al. 2000) and only a binary classification is used to indicate if a grid cell is snow covered or not. To evaluate model simulations of SCE against the IMS product, a quantitative measure was needed to transform the model-simulated SCE into a binary product similar to the IMS. In the absence of any specific information an assumption was made that if the simulated grid cell SCE is greater or equal to 0.5 or 50% then it is considered snow covered for the purposes of comparing with the IMS. Brubaker et al. (2005) and Livneh et al. (2010), in their evaluations of IMS snow cover estimates, used a similar quantitative criterion.
For the mountain domain (Fig. 1a) a comparison of the time evolution of the area snow cover fraction (i.e., the fraction of grid points considered snow covered) is shown in Fig. 2. The SCE from the various model simulations is calculated based upon daily SWE values at 1800 UTC (1100 h local time). For the purposes of comparison, the IMS high-resolution product available at 4-km resolution was interpolated to the high-resolution grid using the nearest-neighbor algorithm to preserve the binary nature of the IMS data. For the two low-resolution simulations (CaPA-15 and GEM-15), values of SCE for each grid cell were calculated at low resolution and then interpolated linearly to the high-resolution grid prior to calculating the area snow cover fraction. During the simulation study period a total of 12 IMS maps were missing.
The comparisons of the simulated SCE from the various models with the IMS data are intended to provide a broadscale measure of performance among the land surface model simulations. Quantitative results presented in Tables 1 and 2 are sensitive to the choice of the assumed snow cover threshold of 0.5 or 50%. However, the sensitivity is systematic with the differences among the models at both high and low resolution remaining robust. Because the high-resolution simulations are forced with higher-resolution topographic databases and undergo a terrain adjustment to the atmospheric forcing data the spatial variance generated in the snow fields is much greater when compared with the low-resolution integrations (e.g., see Figs. 8a,b). The IMS snow cover product is generated at a lower resolution and does not possess the same degree of spatial variance when interpolated to 1 km. The smoother IMS product will act to mask some of the finer structures seen in the high-resolution integrations introducing a caveat when directly comparing the low- and high-resolution simulations.
All model simulations represent a distinct annual cycle consistent with the IMS signal (Fig. 2). Table 1 summarizes the area snow cover fraction biases, with respect to the IMS product, for each model. For the period 1 October 2004 to 15 July 2007 CaPA-15 underestimates the fractional snow coverage, while the GEM-15 simulation overestimates fractional snow coverage by a small amount. The high-resolution simulation results are also shown in Table 1 and again a similar sensitivity to precipitation forcing is noted with the GEM-1 (CaPA-1) simulation overestimating (underestimating) area snow cover fraction.
Additional insight can be gained by examining the core winter months and the spring melt period separately. See Table 1 for definitions of spring melt and core winter month periods. For the core winter months both CaPA simulations exhibit an increase in negative area snow cover fraction bias. The GEM-15 simulation has virtually the same bias, but the overall positive GEM-1 bias is reduced. Compared to the core winter months, in the spring period the low-resolution (high resolution) simulations exhibit a decrease (increase) in area snow cover fraction when compared with IMS. This spring period result is most likely related to the terrain adjustment of the atmospheric forcing data that is present in the high-resolution simulations. A study of simulated snow cover extent among four land surface models run at 12-km resolution by Sheffield et al. (2003), verified against the lower-resolution IMS data (i.e., 24 km), noted higher model variability in simulated snow cover extent during the spring melt period. One has to be cautious in drawing firm conclusions given the documented problems with the IMS snow cover product during the spring period (Brown et al. 2007).
The area snow cover fraction for the same mountain region as a function of elevation is shown in Fig. 3. Owing to the usage of the high-resolution topographic databases for the two high-resolution simulations, the comparison with the IMS data is restricted to experiments CaPA-1 and GEM-1. Elevation bands are chosen at intervals of 400 m. With increasing elevation the intraseasonal winter variability in area snow cover fraction is reduced in both the IMS data and the GEM-1 and CaPA-1 simulations (see Figs. 3a–e). Nonetheless, for the higher-elevation bands (Figs. 3c–e), both model simulations fail to capture the intraseasonal snow cover variability seen in the IMS data during the peak winter months. The onset or transition periods from low values of area snow cover fraction in the summer months to complete snow coverage in the fall period are more abrupt in the models when compared with the IMS data (Fig. 3).
The time-averaged area snow cover fraction biases, when compared with the IMS product, as a function of elevation range for the period 1 October 2004 to 15 July 2007 for the GEM-1 and CaPA-1 simulations are given in Table 2. A trend is evident whereby the GEM-1 area snow cover fraction biases become more positive with increasing elevation, while the CaPA-1 negative biases at low elevations are reduced with increasing elevation. Examining the core winter months separately an increase (decrease) in the CaPA-1 (GEM-1) biases is noted at lower (higher) elevations. This finding suggests that the positive biases in GEM-1 at higher elevations occur during the fall and spring transition periods.
To examine the issue of whether the model simulations are producing snow in the correct locations, a grid cell by grid cell comparison with the IMS data is shown in Fig. 4. For the mountain region time series of the percent correct score for each of the model simulations are displayed in Fig. 4. The percent correct score is defined as the total number of hits (model and IMS classify grid cell as snow covered) and correct rejects (model and IMS classify grid cell as not snow covered) divided by the total number of grid cells. All model simulations represent an annual cycle with varying degrees of intensity (Fig. 4). The percent correct scores for the entire period 1 October 2004–15 July 2007 and for the core winter months for CaPA-15 are (78.03%, 79.4%), GEM-15 (82.9%, 90.9%), CaPA-1 (80.2%, 83.8%), and GEM-1 (82.2%, 91.1%), respectively. All simulations exhibit an improved performance during the core winter months. The greater improvements noted for GEM-15 and GEM-1 is suggestive of a better overall performance during the core winter months. Sheffield et al. (2003) undertook a similar grid cell by grid cell comparison of four land surface model simulations with IMS data over several regions of the United States. Their findings indicated an average percent correct score in the range of 75%–80% for the winter months, with scores increasing to 85%–90% for the spring period.
b. Snow water equivalent
In this section the ability of the different land surface model simulations to represent SWE in the mountain region of the integration domain is examined. Figure 5 shows a time series of SWE, averaged over the mountain region for all four model experiments for the period 20 May 2004–31 October 2007. For the purposes of comparison, SWE values from the 15-km simulations were linearly interpolated to the 1-km high-resolution grid. Note that the SWE values are daily values at 1800 UTC (1100 h local time). For all four model experiments a clear annual cycle is evident with maximum SWE values in the winter season. A comparison of experiment CaPA-15 with GEM-15 (Fig. 5a) and experiment CaPA-1 with GEM-1 (Fig. 5c) highlights the influence of precipitation forcing at both low and high resolutions. A rough calculation indicates that maximum area-averaged modeled SWE values increase by a factor of 2 at both low and high resolutions when the precipitation forcing is changed from CaPA to that of the regional GEM.
Also shown in Fig. 5 is the overall contribution to streamflow, averaged over the mountain domain, from all four experiments. The contribution to streamflow represents the sum of lateral flow, baseflow, and surface runoff for the three hydrological years, defined as the October–September period. Differences in simulated SWE do translate into important differences in surface runoff and streamflow contributions. At both high and low resolutions, the overall contribution to streamflow for the simulations with the regional GEM as precipitation forcing is more than double that of the experiments with CaPA as the precipitation forcing. To compare the contribution to streamflow with observed streamflow values, the annual discharge values for the Bow River at Calgary, Canada, are plotted in Fig. 5 for three hydrological years. The location of the Bow River at Calgary is shown in Fig. 1a. The Bow River is one of the principal rivers routing snowmelt east of the Canadian Rockies. The total streamflow contributions from the simulations with GEM as the precipitation forcing are in better agreement with those at the Bow River at Calgary.
The mean elevation of the snow survey sites used to verify the simulated SWE values is 1850.5 m with the highest (lowest) elevation site located at 2230 (1213) m, respectively. A crude elevation screening was performed to ensure that the elevation differences between each of the 13 manual snow survey sites and the high-resolution simulation topography field is less than 300 m in absolute value. Simulated SWE values at 1800 UTC were used to compare with the observed SWE values.
Figure 6 shows scatterplots of the observed versus simulated SWE for the four model simulation experiments at all 13 snow survey sites for the period December 2004–June 2007. A total of 163 observed–simulated pairs are stratified according to time of year with the respective biases and correlation coefficients shown for each simulation. Examining the two low-resolution experiments (Figs. 6a,b), an important influence of precipitation forcing upon locally simulated SWE values is noted. The negative SWE bias for CaPA-15 is roughly double that for GEM-15 and the GEM-15 SWE values are more strongly correlated with observed SWE values. What is most striking, especially when discussing the CaPA-15 experiment (Fig. 6a), is the large number of nonzero observed SWE values that are associated with simulated SWE values of zero. Examining Fig. 6a more closely, these nonzero simulated SWE values occur predominantly in the melt season during the period April–mid-June (yellow and red colors). A reduced clustering of zero-simulated SWE values associated with nonzero-observed SWE values is found in GEM-15 (Fig. 6b). Additionally, the GEM-15 simulation appears to perform better during early and midseason (black and blue colors). Clearly precipitation forcing has an important impact on simulated SWE values.
The comparison for both high-resolution simulations is shown in Figs. 6c,d. For both CaPA-1 and GEM-1 the correlation coefficients indicate a stronger linear relationship between simulated and observed SWE. A substantial reduction in the number of zero-simulated SWE values associated with observed nonzero SWE values is noted for CaPA-1 (Fig. 6c), which indicates a better representation of SWE during the melt season. Nonetheless, a negative model bias still persists. The overall best performance is found with the GEM-1 simulation (Fig. 6d), which is associated with the highest linear correlation and the smallest model bias.
c. Comparison of the regional GEM and CaPA precipitation
In this section the regional GEM and CaPA precipitation are compared with observations. The Climate Research Division of Environment Canada has developed high-quality precipitation time series for a set of over 400 stations across Canada that have been rehabilitated to account for known inhomogeneities such as change of site location, changes in observing procedure, and instrument deficiencies (Mekis and Hogg 1999). Daily snow measurements are based upon ruler measurements that have always been the standard practice in Canada. The conversion to liquid equivalent is accomplished using a density of fresh snow of 100 kg m−3 (Mekis and Hogg 1999). As such, these bias-corrected precipitation measurements have not been assimilated into CaPA and will be referred to as the Historical Adjusted Precipitation for Canada (HAPC). Additionally, observations of daily precipitation are available from the SNOTEL network for the U.S. portion of the high-resolution domain. For details on the SNOTEL network and methods used to measure precipitation the reader is referred to Serreze et al. (1999).
The precipitation evaluation focuses upon the principal snow accumulation months of December–January–February–March (DJFM) for the three winter seasons between 2004 and 2007. Although the number of stations over the Canadian Rockies is limited, a total of 11 stations from the HAPC dataset were chosen to perform the precipitation validation (Figs. 7a,b). The location of these stations is shown as squares in Figs. 7a,b. Monthly precipitation totals from both CaPA and GEM were linearly interpolated to the location of the observing station and the spatial distribution of the monthly precipitation bias for both CaPA and the regional GEM are shown in Fig. 7. CaPA possesses a negative bias for 9 of the 11 stations (Fig. 7a), while the regional GEM possesses a positive bias for 8 of the 11 stations (Fig. 7b).
Figures 7c,d show scatterplots of the individual monthly CaPA and regional GEM precipitation totals versus observed for the 11 stations. Integrated over all 11 HAPC stations CaPA possesses a negative precipitation bias of 21.35 mm, while the regional GEM has a positive bias of 4.51 mm. The winter precipitation biases for the regional GEM and CaPA are of opposing sign and the CaPA bias is roughly 4 times that of the regional GEM. If all of the precipitation is assumed to have fallen as snow, the difference in precipitation bias can translate into roughly 100 cm more of snow depth over a given winter season at some locations. The results from the precipitation comparison are consistent with the underestimation of SCE and SWE in the simulations with CaPA as the precipitation forcing.
Also shown in Figs. 7a,b are monthly precipitation biases for 10 SNOTEL sites (diamonds) in Montana and Idaho whose mean elevation is 1640.13 m. At these higher-elevation stations, both CaPA and the regional GEM possess negative monthly precipitation biases. The bias in CaPA is negative at 9 of the 10 SNOTEL sites, while the regional GEM bias is negative at 7 of the 10 SNOTEL sites. Similar to the findings for the HAPC data, the CaPA monthly precipitation bias is roughly 3–4 times that of the regional GEM (Figs. 7e,f), and integrated over an entire winter season the CaPA bias can translate into snow depth differences of over 200 cm at some locations. The negative precipitation biases in both CaPA and the regional GEM point to the difficulties in estimating precipitation at high-elevation locations.
d. Impact of terrain adjusting the atmospheric forcing data
Considering only the regional GEM precipitation forcing, an additional experiment is performed, GEM-1_NA, in which the temperature, humidity, and surface pressure forcing data are not downscaled to the higher-resolution grid. Because temperatures are not downscaled in GEM-1_NA the phase of the precipitation forcing will be different from that of GEM-1. Experiment GEM-1_NA does benefit from the higher-resolution geophysical fields; however, the important temperature–elevation relationship is not replicated. A comparison of GEM-1 and GEM-1_NA isolates the impact of the terrain-adjusted atmospheric forcing upon simulated SWE.
Extended seasonal mean SWE from GEM-1 and GEM-1_NA, plotted for the mountain domain, is shown in Figs. 8a,b. It is clear that one of the impacts of terrain adjusting the atmospheric forcing is to increase the spatial structure and variance of the simulated SWE fields. The area-averaged SWE, averaged over an entire season, does not differ substantially between GEM-1 and GEM-1_NA; however, the plots do suggest that local values of SWE will be affected by terrain adjusting the atmospheric forcing. For each of the 13 snow survey sites the simulated SWE values from GEM-1 are plotted against those from GEM-1_NA and are shown in Fig. 8c. The simulated SWE values have been stratified according to the sign of the difference in elevation at the snow survey site between the regional GEM model and that of the high-resolution simulations, where a positive (negative) value implies that the elevation is higher (lower) in the high-resolution simulations.
A clear trend is evident in Fig. 8c, in which SWE values are greater (less) in GEM-1 when compared to GEM-1_NA at snow survey sites for which the elevations in the high-resolution database simulations are greater (less) than the regional GEM forcing data. The largest differences in simulated SWE, between GEM-1 and GEM-1_NA, appear to be concentrated near the end of the cold season (red colors), with values similar during the core winter months of February and March (blue and green colors). Figure 8e shows the differences in simulated dates for 10% SWE accumulation and 90% SWE melt between GEM-1 and GEM-1_NA (GEM-1_NA-GEM-1) for all 13 snow survey sites. For snow survey sites in which elevations in the high-resolution database simulations are greater (less) than the regional GEM forcing data, the 10% SWE accumulation date is generally later (earlier) in GEM-1_NA, while the 90% SWE melt date is typically earlier (later) in GEM-1_NA. The effect of terrain adjusting the atmospheric forcing data is to extend (shorten) the snow season at locations where the elevations in the high-resolution database simulations are greater (less) than the regional GEM forcing data.
A comparison of the simulated SWE values from GEM-1_NA with the SWE observations from the 13 snow survey sites is shown in Fig. 8d. Compared with the results from GEM-1 (Fig. 6d) the overall negative SWE bias increases from −29.60 to −36.39 mm, a 23.0% increase, while the strength of the linear relationship between simulated and observed SWE decreases from 0.70 to 0.49. These results highlight the importance that terrain-adjusted atmospheric forcing can have upon locally simulated SWE values.
5. Discussion and future work
The focus of this paper was an evaluation of the improvements in simulating SCE and SWE with the hydrometeorological modeling system currently under development at EC. This study represents one of the first documented quantitative evaluations of this new system. In general, it has been shown that reasonable estimates of SCE and SWE can be obtained over the Canadian Rockies from our hydrometeorological modeling system by increasing the horizontal resolution and improving atmospheric forcing, including precipitation. Although the study period was limited to 3 winter seasons and the SWE verification dataset was restricted to 13 snow survey sites, the findings have highlighted important areas that need further work.
Comparisons of the simulated SCE with the IMS analysis indicated that at low resolution, the impact of changing the precipitation forcing from CaPA to that of the regional GEM resulted in important decreases in the mean SCE bias. Similarly, a verification of simulated SWE values against a series of snow survey sites within the Canadian Rockies showed important sensitivities to precipitation forcing with the overall SWE bias in GEM-1 lower by a factor of 4 compared with CaPA-1.
The analysis of precipitation based upon surface observations in areas of complex terrain is a challenging task. This study has shown that the CaPA analysis underestimates precipitation amounts over complex terrain, which directly impact simulated SCE and SWE snow parameters. Motivated by the results from this study, several steps are being taken to improve the CaPA analysis. The cubic root data transformation used in CaPA acts to increase the skill for lower precipitation quantities but introduces a bias for large precipitation amounts. As a first step, instead of transforming individually the observations and the model background precipitation, their difference is transformed, which has the advantage that the bias can be more easily estimated. Preliminary evaluation results for the September 2005 time period over the Great Lakes region are encouraging as the skill is improved for both large and small precipitation quantities without introducing a significant bias.
In regions of complex terrain, the most promising approach concerns the development of a sufficiently long climatology of high-resolution simulations (∼1 km) to calculate robust and useful local precipitation–elevation relationships. These relationships can be then be used to downscale the lower-resolution regional GEM precipitation fields that are used as the first-guess field in CaPA. A similar strategy to account for topographical influences is used in the production of the North American Land Data Assimilation System (NLDAS) real-time precipitation forcing (Cosgrove et al. 2003). Furthermore, the correlation length scales, or the horizontal scales of influence for a given precipitation observation, need to account for the “true” distance, sometimes referred to as the walking distance, between the observing station and model grid point (Dubois 2006). In areas of complex terrain this means including the elevation differences in the distance calculations.
Terrain adjusting the atmospheric forcing data when increasing the horizontal resolution was found to be important for properly modeling local extreme SWE values. Several intercomparison studies within the context of the Project for the Intercomparison of Land-Surface Parameterization Schemes (PILPS) discuss the importance of surface roughness parameterization as a source of intermodel differences in snow sublimation and ablation rates (Slater et al. 2001; Nijssen et al. 2003; Samuelsson et al. 2003; van den Hurt and Viterbo 2003). In ISBA the ratio of the roughness length for momentum to that for heat transfer is set to a constant value of 5 over the snow free portion of the model grid. Several studies suggest that this ratio should be stability dependent and can be very different in areas of complex terrain (Beljaars and Viterbo 1994; Samuelsson et al. 2003; van den Hurt and Viterbo 2003). This issue is currently being investigated.
A prototype version of the hydrometeorological modeling system provided high-resolution land surface modeling support for the Vancouver 2010 Winter Olympic and Paralympic Games (Mailhot et al. 2010). Although the ISBA land surface model was used for this study, development work has begun to include the Canadian Land Surface Scheme (CLASS; Verseghy 1991; Verseghy et al. 1993) within the hydrometeorological modeling system.
The authors thank Alain Pietroniro and Brenda Toth of the Hydrometerology and Arctic Laboratory, Meteorological Service of Canada, National Hydrology Research Centre for their efforts and contributions to the NAESI project. The British Columbia Ministry of Environment, Water Stewardship Division, is acknowledged for providing the snow survey data used in this study. The National Snow and Ice Data Center is acknowledged for providing the IMS snow cover data. Seth Bennett kindly extracted the monthly precipitation values from the Adjusted Historical Canadian Climate Database. SNOTEL precipitation data was obtained from the U.S. Department of Agriculture Natural Resources Conservation Service Web site. The authors thank the three anonymous reviewers for their valuable comments and suggestions.
Corresponding author address: Dr. Marco L. Carrera, Meteorological Research Division, Environment Canada, Canadian Meteorological Centre, 5th Floor, 2121 Trans-Canada Highway, Dorval QC H9P 1J3, Canada. Email: firstname.lastname@example.org