Process-based hydrologic models require extensive meteorological forcing data, including data on precipitation, temperature, shortwave and longwave radiation, humidity, surface pressure, and wind speed. Observations of precipitation and temperature are more common than other variables; consequently, radiation, humidity, pressure, and wind speed often must be either estimated using empirical relationships with precipitation and temperature or obtained from numerical weather prediction models. This study examines two climate forcing datasets using different methods to estimate radiative energy fluxes and humidity and investigates the effects of the choice of forcing data on hydrologic simulations over the mountainous upper Colorado River basin (293 472 km2). Comparisons of model simulations forced by two climate datasets illustrate that the methods used to estimate shortwave radiation impact hydrologic states and fluxes, particularly at high elevation (e.g., ~20% difference in runoff above 3000-m elevation), substantially altering the timing of snowmelt and runoff (~20 days difference) and the partitioning of precipitation between evapotranspiration and runoff. The different forcing datasets also exhibit differences in hydrologic sensitivity to interannual temperature at high elevation. The results suggest that the choice of forcing dataset is an important consideration when conducting climate impact assessments and the subsequent applications of these assessments for water resources planning and management.
Given observed changes in streamflow (e.g., Milly et al. 2005; Regonda et al. 2005; Stewart et al. 2005; Hidalgo et al. 2009), water management agencies now put a considerable amount of effort into planning for potential water availability in the future (e.g., Brekke et al. 2009). While a growing number of modeling studies have attempted to quantify hydrologic impacts under climate change (e.g., Christensen et al. 2004; Christensen and Lettenmaier 2007; Cayan et al. 2010; Das et al. 2011), more recent modeling studies illustrate that the magnitude of hydrologic sensitivity to climate variability depends on the specific modeling approach used (Hoerling et al. 2009; Vano et al. 2012).
Selection of climate forcing data for hydrologic modeling introduces one type of uncertainty, among many other uncertainties related to modeling strategies in hydrologic simulations, including the choice of hydrologic models (e.g., Vano et al. 2012) and parameter estimation strategies (Wilby 2005). Several past studies have evaluated how different climate forcing datasets affect the hydrologic simulations (Guo et al. 2006; Materia et al. 2010; Mo et al. 2012; Wayand et al. 2013). The results from the previous studies show that differences in precipitation affect the magnitude of runoff while other forcing variables such as radiative fluxes can affect estimates of evaporation and rates of snowmelt, changing both the overall partitioning of precipitation to evapotranspiration (ET) and runoff and changing the timing of spring snowmelt runoff (Materia et al. 2010; Gao et al. 2011; Nasonova et al. 2011; Haddeland et al. 2012; Wayand et al. 2013).
Data requirements for hydrologic models depend on the model complexity. Among the many types of hydrologic models available, process-based hydrologic models are valuable for improving our understanding of physical mechanisms of long-term hydroclimatological variability, including the potential impact of future climate change on hydrologic processes. However, these models have more complex data requirements, including radiative fluxes, humidity, winds, and so on, and some of these variables are a challenge to obtain, particularly in remote mountainous areas where surface observations are sparse and spatial variability in near-surface meteorology can be extreme.
Two methods are commonly used to provide radiation and humidity data required by process-based hydrologic models: 1) radiation and humidity estimated using an empirical algorithm with temperature (T) and precipitation (P) and 2) radiation, humidity, and wind speeds as obtained from numerical weather prediction (NWP) models. One of the empirical methods, the Mountain-Microclimate Simulation Model (MT-CLIM) (Hungerford et al. 1989; Kimball et al. 1997; Thornton and Running 1999; Thornton et al. 2000), has been widely used to empirically estimate daily downward shortwave radiation and vapor pressure for many hydrologic modeling applications, including drought assessment over the contiguous United States (CONUS) (e.g., Andreadis et al. 2005; Wang et al. 2009); studies on climate change impacts on hydrology in the western United States (Christensen and Lettenmaier 2007; Adam et al. 2009; Elsner et al. 2010); seasonal hydrologic forecasting (Wood et al. 2002); and hydroclimatological evaluation over the La Plata basin in South America (Su and Lettenmaier 2009), among many others.
Because of the large number of hydrologic modeling studies that have relied on empirically derived radiation and humidity estimations, Bohn et al. (2013) evaluated the radiative fluxes and humidity values estimated with MT-CLIM based on in situ measurements at a global scale. They found that, although the MT-CLIM algorithm produced biases in shortwave radiation less than 3% under most climate conditions, larger bias existed: >−20% in coastal regions and −15% over snow-covered areas. Pierce et al. (2013) examined dewpoint temperature estimates from MT-CLIM over the western United States and found larger negative biases up to −10°C in Southern California compared to continental semiarid regions.
NWP models are an alternative data source for process-based hydrologic models. These models can assimilate available ground-based and satellite observations to define the best “analysis” of three-dimensional atmospheric fields at the start of each forecast period, including data for near-surface radiative fluxes, winds, and humidity. Several efforts have rerun current-generation NWP models over a historical period to give “reanalysis” products with a multidecadal time series of the full suite of meteorological variables required to drive process-based hydrologic models (e.g., Uppala et al. 2005; Mesinger et al. 2006; Saha et al. 2010; Rienecker et al. 2011). However, observations assimilated into the NWP model are typically inconsistent in time over the analysis periods (Rienecker et al. 2011), and some persistent problems from the NWP models can negatively affect surface meteorology, particularly precipitation estimates (e.g., Reichle et al. 2011). Another drawback of these reanalysis datasets is their most common grid spacing of >0.3° (~30 km at midlatitudes), which is still too coarse for hydrologic applications in mountainous areas. This means that some type of downscaling is required to resolve effects of complex physiographic features on the spatial variability of meteorological fields.
This study examines the effects, on hydrologic simulations over mountainous regions, of differences in methods for estimating climate variables where the estimations of model forcings are particularly challenging. Two historical climate datasets are examined in this study.
Datasets developed by Maurer et al. (2002, hereafter M02) extrapolate observed temperature and precipitation from low-elevation stations on a ⅛° grid together with estimates of radiative fluxes and humidity using the MT-CLIM algorithm.
Datasets developed by Xia et al. (2012, hereafter X12) also extrapolate observed precipitation to the same ⅛° grid but use temperature, radiation, and humidity derived from a regional-scale NWP model with the National Centers for Environmental Prediction (NCEP) North American Land Data Assimilation (NLDAS) project phase 2.
These are the same datasets used by Mo et al. (2012), who show large differences in solar radiation over the upper Colorado River basin (see Fig. 10 in Mo et al. 2012). Therefore, this study builds on the Mo et al. study by focusing attention on the mountainous regions, illustrating large differences in simulations of hydrologic fluxes in these regions and, importantly, attributing differences in model simulations to specific methodological decisions made when constructing these forcing datasets. Our broader goals are to document some of the difficulties in estimating forcing data for process-based hydrologic models in complex terrain and to illustrate how decisions made when constructing forcing datasets can affect the portrayal of climate change impacts on hydrologic processes.
The remainder of this paper is organized as follows. Reviews of methodologies for constructing both datasets and descriptions of model simulation procedures are given in section 2. Results and discussion of the analyses performed are given in section 3: the climate variables (i.e., temperature, precipitation, humidity, wind speeds, radiation, and pressure) used to force the hydrologic model in section 3a; the hydrologic simulations forced by the two forcing datasets in 3b; differences in shortwave radiation estimations in 3c; and an illustration how different forcing datasets may lead to differences in the portrayal of the hydrologic sensitivity to climate variability in 3d. Section 4 summarizes the findings from this work and suggests some conclusions.
2. Datasets and model
a. Climate forcing data
This study evaluates the two climate datasets most commonly used in land surface modeling studies: the NCEP NLDAS phase 2 derived data (X12) and the data produced by M02. Because both datasets use the same grid spacing (i.e., ⅛°) and CONUS extent, the effect of spatial resolution differences on the model outputs for the comparison is eliminated. As described below, the most significant differences between the two datasets are in their estimates of shortwave (SW) and longwave (LW) radiative energy and humidity. M02 uses the set of empirical algorithms in MT-CLIM together with daily temperature and precipitation, while radiation and humidity data in X12 are based on the North American Regional Reanalysis (NARR) (Mesinger et al. 2006), thus not derived from temperature and precipitation.
1) The M02 dataset
M02 consists of four daily variables—precipitation, minimum and maximum temperature (Tmin and Tmax), and wind speed—as a gridded product derived (excepting wind speed) from observations and interpolated to the ⅛° grids using a consistent set of stations (M02). The dataset covers the CONUS and portions of Canada and Mexico from 1950 through the present. M02 has been used as a forcing input to the Variable Infiltration Capacity (VIC) model (Liang et al. 1994), among others, for a great variety of hydrologic studies of climate effects (e.g., Elsner et al. 2010).
M02 contains derived daily precipitation and temperature grids using ground stations in the NOAA daily Cooperative Observer (COOP) station network in the United States, which are located mostly in low elevations. The daily precipitation gauge data were gridded to the ⅛° resolution using the synergraphic mapping system (SYMAP) algorithm (Shepard 1984). The gridded daily precipitation data were then scaled to match the long-term average of P climatology from the Parameter-Elevation Regressions on Independent Slopes Model (PRISM) (1961–90 monthly climatology; Daly et al. 1994). For mountainous regions in wintertime, precipitation may be underestimated in snow-dominated areas because of undercatch of snowfall estimates (e.g., Goodison et al. 1998), which were not corrected in M02. The Tmin and Tmax data, also obtained from COOP stations, were gridded using the same algorithm as for precipitation and were elevation adjusted with an average lapse rate of 6.5°C km−1 applied to the pixel mean elevation. Wind speed from the 2.5° × 2.5° NCEP–National Center for Atmospheric Research (NCAR) reanalysis (Kalnay et al. 1996) was linearly interpolated to the ⅛° grid.
Derivations of the other forcing variables in M02, shortwave and longwave radiation (SW and LW) and humidity (RH), use the MT-CLIM algorithm (Hungerford et al. 1989) in which theoretical daily extraterrestrial insolation is reduced by daily transmittance estimated empirically with diurnal temperature range (DTR) and humidity and then with a further 25% reduction applied on a day with precipitation (Thornton and Running 1999). Dewpoint temperature is also empirically estimated as a function of Tmin and precipitation (Kimball et al. 1997). Since daily shortwave radiation and humidity are interrelated in the algorithm (i.e., shortwave radiation is a function of vapor pressure, and vapor pressure is a function of shortwave radiation), both variables were iteratively derived in the MT-CLIM algorithm. Estimation of longwave radiation uses an empirical equation developed by Idso (1981), which uses air temperature and vapor pressure to estimate atmospheric emissivity.
Daily precipitation is uniformly distributed throughout the day, and temperature is disaggregated using spline interpolation of the daily maximum and minimum values. Daily total shortwave radiation is disaggregated using the temporal pattern of solar zenith angle for each calendar day. Daily dewpoint is linearly disaggregated between successive days and then converted to specific humidity with subdaily temperature estimates.
2) The X12 dataset
The X12 precipitation data are a merged product of the following: NOAA/Climate Prediction Center (CPC) analysis of daily gauge precipitation, a national mosaic 4-km NOAA/National Weather Service (NWS) Stage II radar, 8-km hourly precipitation analyses using the NOAA/Climate Prediction Center morphing technique (CMORPH) (Joyce et al. 2004), and output from the NARR. CPC PRISM-adjusted gauge-only daily precipitation analyses are temporally disaggregated into hourly fields by multiplying daily CPC precipitation by hourly disaggregation weights derived from the national mosaic of 4-km Stage II radar and 8-km hourly CMORPH precipitation analyses. The mosaic of the Stage II product is interpolated to ⅛° grids, and any gaps in radar coverage (which total, on average, 13% of the area of the CONUS and are due to lack of radar coverage or equipment failure) are filled in with nearest-neighbor Stage II data from within a 2° radius. Where Stage II data are unavailable, CMORPH data are used; where CMORPH data are unavailable, NARR data are used. If the daily Stage II or CMORPH total is zero in an area of nonzero CPC P, the CPC precipitation estimate is spread evenly over the entire day. Since the Stage II and CMORPH data are used only for the hourly disaggregation weights, a daily summation of the precipitation fields will exactly reproduce the original CPC daily precipitation analysis.
All variables in X12 except precipitation are generated from the NARR using spatial and temporal disaggregation with the vertical adjustment methods described by Cosgrove et al. (2003). The 3-h NARR fields are spatially interpolated to the ⅛° grid with bilinear interpolation and then temporally disaggregated to hourly values using linear interpolation, excepting precipitation which uses the temporal disaggregation methods described above. Additionally, the fields of surface pressure, surface downward longwave radiation, near-surface air temperature, and near-surface specific humidity are adjusted vertically to account for the vertical difference between the NARR and the terrain height on the ⅛° grid. For example, this vertical adjustment uses the vertical lapse rate 6.5°C km−1 for air temperature.
The surface downward shortwave radiation field from NARR is known to contain a positive bias, that is, approximately +35 W m−2 in mountainous areas in the northwestern United States; Schroeder et al. 2009), thus is bias corrected based on surface downward shortwave radiation estimated with Geostationary Operational Environmental Satellite-8 (GOES-8) (Pinker et al. 2003). Monthly bias correction factor fields are derived based on comparison between NARR shortwave radiation and GOES-8 derived shortwave radiation for 1996–2000 (Berg et al. 2003). These bias correction factors are applied to the NARR downward shortwave radiation field.
b. Model simulations
Two hydrologic simulations were performed over the upper Colorado River basin (Fig. 1) from 1 October 1980 through 30 September 2008 using the Community Land Model, version 4 (CLM4) (Oleson et al. 2010) with identical model parameters and 1-h time outputs at ⅛° resolution, except for forcing data that we evaluated here.
The model was initialized by cycling through the entire period five times. The analysis used the 27-yr period from 1 October 1981 through 30 September 2008. The bulk of this analysis evaluates spatial averages of model outputs in the three elevation bands shown in Fig. 1.
CLM includes a full spectrum of land surface and subsurface processes, such as interactions of solar and longwave radiation with vegetation canopy and soil; momentum and turbulent fluxes from canopy and soil; heat transfer in soil and snow; and hydrology of canopy, soil, and snow. CLM characterizes the land surface within a grid box by fractions of five different land covers (vegetation, wetland, lake, glacier, and urban) with vegetated cover further split into 16 possible vegetation/forest classes called plant functional types (PFTs). Vegetation structures (e.g., leaf area index, stem area index, canopy top and bottom heights), canopy optical properties (e.g., reflectance and transmissivity), and aerodynamic properties (e.g., roughness length and displacement height) are defined for each PFT. Water and energy fluxes are computed for each PFT and the fluxes from and to a grid box are weighted averages over all PFTs. The PFT distribution used for this study is based on a global 0.5° PFT grid developed by Lawrence and Chase (2007). With the global land surface datasets used here, forest coverage is 4%, 19%, and 48% in the low (<2000 m), middle (2000–3000 m), and high (>3000 m) elevation bands, respectively.
3. Results and discussion
a. Comparison of climate forcing variables
Figure 2 displays scatterplots of monthly spatially averaged values in the elevation bands from M02 and X12; statistical measures of the difference at annual scale are shown in Table 1. Differences between the two are largest in the highest elevation band. For example, while shortwave radiation in M02 is consistently higher than in X12 and this difference is seen in all elevation bands, it is the largest (~68 W m−2) in the highest elevation band. Differences in precipitation are mixed, but, on a monthly scale, M02 has ~3 mm more in the higher elevation band than X12. Temperature from X12 is consistently 1~1.5 K higher than M02 for all elevation bands. Relative humidity exhibits the least similarity between the two datasets with M02 having a larger water vapor deficit, that is, lower humidity, than the reanalysis-based humidity data from X12. The higher temperature and humidity lead to higher longwave radiation in X12, though differences are less than 3%. Wind speed comparisons show that X12 has higher wind speeds overall; this is not surprising, however, because the M02 wind data come from the coarser-resolution NCEP–NCAR reanalysis (Kalnay et al. 1996) while X12 wind speeds come from the 32-km NARR and wind speed diminution at a coarse spatial resolution is well known.
These forcing data differences should be kept in mind to understand how the forcing data differences are propagated into the CLM simulations presented in the next subsection.
b. Comparison of CLM simulations
For this analysis, the CLM simulations forced by X12 (CLM-X12) and M02 (CLM-M02) are focused on seasonal snowpack and water fluxes. Figure 3 shows a 27-yr climatological annual cycle of spatially averaged snow water equivalent (SWE), soil moisture, evapotranspiration (ET), and total runoff in each elevation band at a daily time step. Clearly, CLM-X12 produces more SWE in the highest elevation band, particularly in later spring and early summer, while differences in snow accumulation and ablation patterns are much smaller in the lowest elevation band (Fig. 3a). The timing and magnitude of soil moisture recharge from snowmelt correspond to snowmelt timing and peak SWE, respectively, in the highest elevation band. Despite having more available water from snowpack in CLM-X12 than in CLM-M02, CLM-M02 produces higher ET in the highest elevation band (Fig. 3b) possibly because of larger shortwave radiation in M02, as discussed in the following sections. Furthermore, the seasonal difference in ET is unique to each elevation band with the largest difference occurring throughout the year in the highest elevation band. However, the difference is seen through the end of April over the middle elevation band and there is a very small difference in ET only during a short period in winter in the lower elevation band. Differences in SWE and ET result in an obvious difference in runoff between CLM-X12 and CLM-M02 in the highest elevation band (Fig. 3c). CLM-M02 tends to produce more runoff during midwinter (Fig. 3c), indicating more snowmelt during midwinter. This snow mass loss during midwinter may explain the lower snow accumulation in CLM-M02.
Differences in simulated snow accumulation and ablation processes were further examined for each water year. For this analysis, a set of snow signatures was defined to characterize the snowpack processes and two runoff signatures, temporal centroid of runoff (RO_CT; Stewart et al. 2004) and runoff ratio (RR), were also computed to examine the runoff timing and the partitioning of precipitation between runoff and ET respectively; see Table 2 for the definitions of each signature. As shown in Fig. 4, the CLM-X12 simulation produced more snow accumulation (Peak_SWE), longer snow cover duration (SNOW_LENGTH), later snowmelt timing (SWE_CT), and therefore later runoff timing (RO_CT). Differences in the RR suggest that CLM-X12 contributes a higher percentage of total annual precipitation to runoff than CLM-M02 does.
c. Diagnosing shortwave radiation difference
The hydrologic simulation results perhaps appear counterintuitive since X12 has less precipitation, particularly at the highest elevation; higher air temperature; and greater wind speed over the basin. All of these would contribute to less snow accumulation and more rapid snowmelt. One possible reason for the large differences in the hydrologic simulations is the ~30% differences in shortwave radiation in the highest elevation band, as shown in Table 1 and Fig. 2. A detailed comparison of shortwave radiation is provided here to explore this shortwave radiation difference.
Figure 5 shows the climatological annual cycle of shortwave radiation from the two climate datasets. For each month, daily values from 27 years were used to display interannual variability for each month. As shown in Fig. 5, the difference in shortwave radiation tends to be greater in summer than winter, with the largest differences at higher elevations. The large differences in CLM simulations presented in Fig. 4 can be attributed, at least in part, to the large differences in shortwave radiation. For example, a mean difference in shortwave radiation of 85 W m−2 in May over the highest elevation band, as shown in Fig. 5, translates to a difference in 273 mm in monthly snowmelt assuming an isothermal snowpack and albedo of 0.6 using
where h is the snowmelt (kg m−2), SW is the shortwave radiation (W m−2), α is albedo, and λ is the latent heat of fusion (334 000 J kg−1). Note that differences in incident shortwave radiation on the ground will be smaller than this estimate because 48% of the highest elevation band is covered by forest in the CLM model surface. Nevertheless, this estimate helps understand the impacts of differences in shortwave radiation estimates.
Figure 6 compares shortwave radiation estimates with observations available in the upper Colorado River basin; see Fig. 1 for the locations of the 10 shortwave radiation measurement sites. The meteorological sites selected provided daily incident shortwave radiation measurements from 2004 through 2008. Daily shortwave radiation measurements were quality controlled to eliminate erroneous values (Slater 2012). Figure 6 shows the range of differences in daily values between observations and X12 (top panel) and M02 (bottom panel) for each month in the 5 years. As shown in Fig. 6, M02 clearly overestimates shortwave radiation throughout the year and its biases (about +40 W m−2) tend to be greater during spring and summer. On the other hand, X12 underestimates shortwave radiation in spring and summer, but the bias in X12 is smaller than in M02.
Evaluating the climatological characteristics of the extrapolated precipitation and temperature fields from M02, which are subsequently used in the empirical MT-CLIM radiation algorithms, can help explain why shortwave radiation estimates in M02 are systematically higher. The Tmin, Tmax, DTR, and monthly wet day frequency (where a wet day is defined to have daily precipitation greater than 0 mm) from M02 were compared with data from the 164 U.S. Department of Agriculture Natural Resource Conservation Service (NRCS) Snow Telemetry (SNOTEL) sites (Serreze et al. 1999) around the upper Colorado River basin (see Fig. 1). This comparison restricts consideration to the grid cells having a SNOTEL site. Because SNOTEL measurements are automated and unattended during winter, SNOTEL temperature data can be erroneous (Pepin et al. 2005). Therefore, erroneous values were eliminated following the quality control procedure of Serreze et al. (1999). Values removed in this quality control procedure were filled using the regression-based gap-filling technique developed by Slater et al. (2013). Filling of missing data is necessary because MT-CLIM requires continuous daily time series of temperature and precipitation.
Figure 7 shows that M02 ,Tmin is nearly uniformly lower than values from the SNOTEL measurements, although the Tmax are more comparable. Differences in Tmin translate to higher DTR in M02 than in the SNOTEL observations (see Fig. 7). M02 monthly wet day frequency is overall comparable to SNOTEL observations. However, this unbiased wet day frequency could be due to compensatory effects of lower wet day frequency at the lower-elevation COOP sites used to construct the dataset, which offset increased wet day frequency due to spatial interpolation of precipitation measurements from multiple sites.
Given the difference in DTR, shortwave radiation was estimated using the MT-CLIM algorithm with SNOTEL observations and compared to M02 shortwave radiation. Figure 8 shows M02 shortwave radiation is consistently higher than radiation estimated with SNOTEL data. This suggests that the systematic biases in DTR at high elevations cause large differences in radiation estimates.
These differences in DTR occur because M02 used only low-elevation COOP stations in order to construct a consistent long-term dataset; SNOTEL data were not used because those data became available in the 1980s and so could not create a continuous time series. Use of a constant lapse rate of 6.5°C km−1 for temperature extrapolations has been shown to be inappropriate for Tmin (e.g., Blandford et al. 2008; Minder et al. 2010). Using observations in Idaho, Blandford et al. (2008) showed that the largest diurnal range of a lapse rate is seen in dry conditions and that the commonly used lapse rate of 6.5°C km−1 could be applied only to Tmax since it often produces an overestimate for Tmin.
d. Comparison of hydrologic sensitivity to climate variability
The paper now turns attention to how choice of forcing datasets affects the portrayal of the hydrologic sensitivity to climate change. Offline hydrologic model experiments offers more flexibility to define future climate conditions for synthetic hydrologic sensitivity analysis and many previous assessments of hydrologic sensitivity to climate change have been performed using this approach (e.g., Adam et al. 2009; Elsner et al. 2010; Das et al. 2011; Vano et al. 2012). However, this is not a fully coupled climate model experiment that uses more realistic land–atmospheric feedback processes to simulate current and future climate conditions. This study evaluates hydrologic sensitivity to interannual climate variability, that is, the change in evapotranspiration and runoff per degree of temperature or millimeter of precipitation, by looking at the historical climate datasets over the 27 water years (WY) 1982–2008 to support inferences about the effects of climate change on hydrologic processes.
Understanding the hydrologic sensitivity to interannual climate variability requires consideration of how M02 and X12 represent hydroclimatic regimes using, for example, a simple characterization of ET energy-limited or water-limited regimes. The energy and water controls on precipitation partitioning are shown using a Budyko plot (Budyko 1974; Fig. 9). Where values of the aridity index [potential ET (PET)/P] are >1, actual ET is controlled by precipitation, signifying a water-limited regime; where PET/P values are <1, actual ET is controlled by energy input rather than precipitation, signifying an energy-limited regime. PET was computed here with the Penman–Monteith formula (Shuttleworth 1993). Figure 9 shows that, in the highest elevation band, hydroclimatic regimes represented by both M02 and X12 are closer to the boundary region between water- and energy-limited regimes where PET/P is approximately 1 and that M02 is characterized by being more water limited than X12. This is construed as resulting from the greater energy input from more shortwave radiation in M02. In the middle and lowest elevation bands M02 is drier (i.e., higher PET/P) than X12, but both datasets represent a strongly water-limited regime there.
Figure 10 shows relationships between temperature and precipitation and simulated hydrologic fluxes on an annual scale from both CLM-M02 and CLM-X12. All annual variables were averaged in each elevation band; therefore, Fig. 10 shows not only relationships between these climate and hydrologic variables within individual elevation bands but also overall elevation effects on the relationships.
These plots reveal a number of interesting features. First, the interannual variability in evapotranspiration is more closely tied to precipitation than to temperature for the lowest and middle elevation bands; compare Figs. 10a and 10b. This is consistent with the results shown in Fig. 9, where the lowest and middle elevation bands are in a water-limited regime; that is, the amount of available water controls ET. Put differently, because a high fraction of water input evaporates in both simulations in the lowest and middle elevation bands (~80% and >90% of precipitation in the lowest and middle elevation bands, respectively), interannual variability in ET is directly related to interannual variability in precipitation. Recall that partitioning of precipitation between ET and runoff is similar in CLM-X12 and CLM-M02 (Fig. 9); therefore, sensitivity of ET to precipitation is also similar. A weak, slightly negative within-band relationship between temperature and ET is seen in two lower elevation bands, though this is not statistically significant at the 95% confidence level when computed for individual elevation bands. These negative relationships between ET and temperature are possibly due to the weak anticorrelation (r ≈ −0.5) between precipitation and temperature at annual scale for both datasets.
In the highest elevation band, the within-band temperature–ET relationships in CLM-X12 and CLM-M02 are different from the ones in the lowest and middle elevation bands (see Fig. 10a). CLM-X12 exhibits a slightly positive temperature–ET relationship within the highest elevation band, with CLM-M02 showing no discernible relationship with temperature. The clearer positive relationship seen in CLM-X12 suggests that its hydroclimatic regime is more strongly energy limited in the highest elevation band than is CLM-M02 (see Fig. 9); for this reason, an interannual variability of temperature, which is an indicator of energy input, explains the evapotranspiration variability. It is likely that the temperature–ET relationship in CLM-M02 is weak because CLM-M02 is characterized more by a mix of water and energy limitations depending on water years (Fig. 9).
For the precipitation–ET relationship in the highest elevation band, both CLM-M02 and CLM-X12 likewise show less influence of precipitation on ET variability. Note the nearly zero slope of the precipitation–ET relationship in Fig. 10b.
This mechanism of ET variability is propagated into the relationship between runoff and climate variables (Figs. 10c,d). In the highest elevation band, CLM-M02 has ~100 mm less runoff than CLM-X12 (see Fig. 10d). This is because X12 is strongly energy limited: that is, less energy is available in X12 for evaporation than in M02, leading to more residual water, which is either stored in soil or runoff. The sensitivity of runoff to precipitation is similar between CLM-M02 and CLM-X12 perhaps because the little sensitivity of ET to precipitation shown in Fig. 10b means that interannual variability in precipitation is directly tied to interannual runoff variability. Runoff is less sensitive to precipitation in the lower elevation bands in both CLM-M02 and CLM-X12 because both simulations are strongly water limited. In the lowest and middle elevation bands, runoff variability is suppressed by limited water availability and a higher percentage of water input is lost as evapotranspiration. Since there is no difference in sensitivity of ET to precipitation between the two simulations, runoff sensitivity to precipitation between the two climate datasets is also similar.
In summary, despite differences as large as ~100 mm in annual ET and runoff in the highest elevation band between CLM-M02 and CLM-X12, there is no substantial difference between the two simulations in their sensitivity of ET and runoff to interannual climate variability measured as changes in T and P. The most substantial difference in hydrologic sensitivity to interannual climate variability appears between temperature and ET in the highest elevation band where for individual years X12 climate data are more likely to be in an energy-limited regime than are the data from M02. This difference in sensitivity is likely caused by the large difference in shortwave radiation between the two datasets.
4. Summary and conclusions
This paper examined two climate forcing datasets, the surface observations assembled by M02 and NWP model output reanalysis as described by X12, for process-based hydrologic model forcing over the upper Colorado River basin. Given that the choice of climate forcing dataset affects hydrologic simulations (e.g., ~20% difference in annual runoff at high elevation), careful derivation of hydrologic forcing is required to produce useful hydrologic modeling applications. In particular, our study illustrated the large differences in estimates of shortwave radiation from the different methods used in M02 and X12, most likely leading to different hydrologic simulations in the higher elevation areas of the basin.
Uncertainty in spatial estimates of precipitation and temperature is propagated into other radiative and humidity estimates when using empirical algorithms such as MT-CLIM. This study has shown that biases in daily minimum temperature affect shortwave radiation estimates obtained from the MT-CLIM. One important conclusion from this is that spatially and temporally varying lapse rates should be used to develop temperature grids via spatial interpolation of station data.
Estimates of precipitation in the two datasets are based on a similar set of stations and for this reason are not considerably different; however, agreement of precipitation estimates between these two datasets does not necessarily imply lower errors since both datasets could be wrong for the same reason. Since the precipitation is a primary driver of the hydrologic cycle, careful gauge data quality control and interpolation scheme are important. For example, gauge catch deficiencies may affect gridded precipitation data (Adam and Lettenmaier 2003). Using PRISM to adjust precipitation for producing the grids might cause errors in interpolated precipitation at higher elevations because the PRISM fields potentially misrepresent the spatial pattern of precipitation there, particularly around mountain ridges (Gutmann et al. 2011).
An important consideration is that all of our conclusions on the effects of different forcing datasets on hydrologic simulation are based on use of a single hydrologic model. Vano et al. (2012) showed that quantification of water balance partitioning over the Colorado River basin depends on choices of hydrologic model structures, which can affect quantitative assessments of hydrologic sensitivity to climate change. Differences in model structures could affect assessments of model sensitivities to choice of model forcing data as well. For example, how biases in shortwave radiation affect simulated hydrologic fluxes depends on the specific radiative transfer scheme used in the selected model. In particular, interaction of radiation with forest canopy is particularly complex. In forest environments, shortwave radiation absorbed by the vegetation canopy in turn affects longwave radiation and rate of below-canopy snowmelt. Shortwave radiation biases also affect evaporation and snow sublimation from the canopy-intercepted water/snow and other processes as well. Model representations of such radiative transfer processes differ across hydrologic models. A more complete and systematic quantitative assessment of the effects of model forcing on hydrologic simulations will require multiple simulations using different models.
Finally, explicitly considering uncertainty in climate forcing data is important in the offline (uncoupled) hydrologic assessments used for many water resources management decisions ranging from short-term flood forecasts to long-term hydrologic shifts due to climate change. This point is especially important when automatic calibration methods are used to infer hydrologic model parameters based on differences between simulated and observed streamflow because inferred model parameters can inappropriately compensate for biases in the model forcing data. It is possible that choice of forcing datasets will affect the simulated hydrologic portrayal through the different inferences gained from different model parameter sets and that ultimately this will have an adverse effect on decisions concerning adaptation to the hydrologic impacts of climate change.
The authors thank Dr. Jessica Lundquist and two anonymous reviewers for critical and careful reviews that helped improve the manuscript. The authors also thank Ethan Gutmann, Roy Rasmussen, David Gochis, Kyoko Ikeda, and Pablo Mendoza for discussions at the earlier stage of this research. This work was financially supported by the U.S Bureau of Reclamation and the U.S Army Corps of Engineers.