Evaluating climate change over the Colorado River basin using regional climate models

Authors


Abstract

[1] We use regional climate model (RCM) simulations from the North American Regional Climate Change Assessment Program (NARCCAP) to evaluate implications of climate change for the discharge of the Colorado River in the mid-21st century. We compare historical RCM simulations and simulations from their host global general circulation models (GCMs) to one-eighth degree gridded observations of precipitation, surface air temperature, and runoff (generated by the Variable Infiltration Capacity (VIC) land surface model forced with gridded observations) for the historical period 1970–1999. The RCMs capture the primary features of observations better than their host GCMs in part because of their ability to better represent strong upward lift in topographically complex regions. Although the RCMs do not significantly improve the simulation of precipitation, their ability to better represent surface temperature in mountainous regions has important effects on simulations of evapotranspiration, snowpack, and runoff. In the Colorado River basin, projected mid-21st century runoff changes are mostly impacted by the combination of snow cover change in winter, temperature change in spring, and precipitation change in summer. In particular, the response of cold season temperatures in headwater streams is key to determining the basin's susceptibility to a warming climate. Because of the cooler temperature and higher snowline in RCMs relative to GCMs, the RCMs project less warming in the spring and thus generate smaller decreases in runoff, both during spring and annually, as compared with GCMs. Changes in surface air temperature, runoff, and snow water equivalent at high elevations all indicate that headwater streams of the Colorado River are less susceptible to a warming climate in climate change simulations that use RCMs than in simulations that use GCMs. Nonetheless, the 50 km NARCCAP grid resolution has some limitations in resolving orographic effects, which suggests that there may remain residual biases in the climatic sensitivity of the RCM simulations.

1. Introduction

[2] The Colorado River basin (CRB) includes parts of seven U.S. states and Mexico (Figure 1). The headwaters lie in the Rocky Mountains of Wyoming and Colorado, where about 75% of annual streamflow is generated from 25% of the area [Hoerling et al., 2009; Franz et al., 2003; Cayan et al., 2001], while the lower basin contributes to only 8% of annual streamflow. With an annual average discharge of roughly 18.5 BCM/yr, the CRB is not particularly large in terms of discharge, especially when compared to other major continental rivers in the United States, like the Mississippi or Columbia. Still, the Colorado River is the most important source of water for the vast, arid southwestern United States. It provides water to 27 million people in the southwestern United States and Mexico [Barnett and Pierce, 2009]. The CRB is especially vulnerable to shifts in climate due to the sensitivity of its discharge to precipitation and temperature changes (both of which affect snow accumulation and melt patterns as well as evapotranspiration), and these effects are exacerbated by the semiarid nature of the basin [Loaiciga et al., 1996]. Given the warming climate and strong reliance of steadily increasing populations on the river and its water storage system, variations in Colorado River flows across the basin are more important now than ever before.

Figure 1.

(a) Location, (b) topography distribution (meters), and (c) histogram of topography for the Colorado River basin.

[3] Many studies have examined the effects of climate change on components of the hydrologic budget in the CRB [e.g., Seager et al., 2007; Milly et al., 2005; Christensen et al., 2004; Christensen and Lettenmaier, 2007]. Seager et al. [2007] analyzed output from 19 general circulation models (GCMs) archived for the Intergovernmental Panel on Climate Change (IPCC). They documented a general trend in almost all the models toward permanently drier conditions in the southwestern United States as increased humidity in the tropics induces poleward migration of the subtropical dry zones. The U.S. Climate Change Science Program [U. S. Climate Change Science Program and The Subcommittee on Global Change Research, 2007] shows, based on results from Milly et al. [2005] (but using a slightly different subset of the 2007 IPCC model runs), that the greatest decreases in runoff by midcentury will range from 10 to 25 percent, that these changes will occur in the upper and lower Colorado basin, and the Great Basin, and that there is a high degree of agreement among the models as to the directions of these changes. In contrast, Christensen and Lettenmaier [2007] used statistical downscaling of a slightly different subset of the same GCMs used by Milly et al. [2005] and Seager et al. [2007] to force a physically based (and more highly spatially resolved) hydrology model over the CRB. They found the mean change in Colorado River discharge, averaged over all GCM A2 emissions scenarios by midcentury, would be a decline of about six percent. Subsequent reanalysis of their results to adjust the hydrology model precipitation forcing to be more consistent with the results of Milly et al. [2005] doubles the decline (to about 13 percent), but still shows runoff reductions that are considerably less severe than those projected by Seager et al. [2007]. The differences in these estimates of Colorado River discharge to a changing climate have caused considerable concern among western U.S. water managers, and implications go well beyond scientific interest [see, e.g., Gertner, 2007]. There is, therefore, some urgency in resolving the causes of the differences, to the extent possible.

[4] GCMs are the most appropriate and powerful tool for predicting the effects of climate change; however, formulation of adaptation policies in response to climate change impacts requires information at finer spatial scales than can be represented by GCMs, which typically have grid cells with dimensions 200–300 km or greater. Thus, although GCMs can provide useful information about possible future changes in atmospheric circulation at the regional (e.g., continental) scale, they do not provide the detail required for regional and national assessments. This is particularly true for heterogeneous regions, where sub-GCM grid-scale variations in topography, vegetation, soils, and coastlines can strongly affect climate. In addition, extreme events, such as heavy precipitation, are often not captured or their intensity is unrealistically low at coarse resolutions.

[5] Regional climate models (RCMs), which provide finer spatial detail than GCMs, provide one possible solution to these issues. One shortcoming of RCMs has been that the computational requirements have resulted in an inability to represent the range of inferred future climate conditions in the same way that has been possible through the application of statistical downscaling to the full suite of IPCC models. The issue is further complicated by the combinatorial problem associated with multiple RCMs (and even specifics of implementation of a given RCM) and multiple GCMs, which would provide the RCM boundary conditions. The North American Regional Climate Change Assessment Program (NARCCAP) has attempted to address this issue by producing multiple RCM simulations over the continental United States [Mearns et al., 2009] with different GCM “parents” providing the boundary conditions. Given their relatively high spatial resolution, these simulations should offer more insight into the nature of projections of future climate for the CRB, and in particular, for its runoff. The spatial resolution of the NARCCAP models (typically about 50 km) is much higher than that of the GCMs on which past CRB studies [e.g., Seager et al., 2007; Christensen and Lettenmaier, 2007] have been based. The NARCCAP runs should, therefore, be better able to capture the effects of topographic variations in climate, and hence, should be more useful to address questions regarding the sensitivity of future Colorado River runoff to climate change. We report herein a comparison of Colorado River runoff as inferred from the NARCCAP RCMs and their host GCMs.

2. Data Description and Methods

2.1. RCM Output

[6] The general NARCCAP strategy (as in most RCM applications) consists of two phases. In Phase I, six RCMs were forced with global reanalysis from the National Center for Environmental Prediction/Department of Energy (NCEP/DOE) reanalysis [Kanamitsu et al., 2002] as the boundary conditions. Because the reanalysis effectively consists of weather prediction model analysis fields (with “frozen” model versions and analysis systems), it is appropriate to compare the RCM output with observations on a time step basis. In the second phase, GCM output was used to provide boundary conditions for both historic and future climate runs. For the historic run, given the chaotic nature of the atmosphere as represented in the GCM boundary conditions, comparisons with observations is only possible in a statistical context. For future climate runs, no comparison with observations is possible, of course. We primarily used model output from Phase II, although we made some comparisons with Phase I output as well.

[7] The six RCMs participating in NARCCAP are the Hadley Regional Model 3 (HRM3 [Jones et al., 2004]), the Regional Climate Model version 3 (RCM3 [Giorgi et al., 1993]), the Canadian Regional Climate Model (CRCM [Laprise et al., 1998]), the NCEP Experimental Climate Prediction Center Regional Spectral Model (ECPC RSM [Juang et al., 1997]), the MM5- PSU/NCAR mesoscale model (MM5 [Grell et al., 1993]), and the Weather Research and Forecasting model (WRF [Skamarock et al., 2005; Leung et al., 2005]). The overall design of the NARCCAP experiment is described by Mearns et al. [2009].

[8] In Phase I of NARCCAP, 25 year (1980–2004) RCM simulations were implemented using the NCEP/DOE Reanalysis for boundary conditions. In Phase II, each RCM was nested within at least one GCM at 50 km spatial resolution for the periods 1971–2000 and 2041–2070 (the latter using an SRES A2 global emissions scenario; see the work of Nakićenović and Swart [2000] for details). The RCM-GCM combinations are listed in Table 1 and GCM scenarios are listed in Table 2.

Table 1. Status of NARCCAP Archived RCM/GCM Combinations
 Phase IIPhase I NCEP/DOE
GFDLCGCM3HadCM3CCSM3
CRCM-finished-plannedfinished
ECPCrunning-planned-finished
HRM3planned-finished-finished
MM5--plannedrunningfinished
RCM3finishedfinished--finished
WRF-planned-finishedfinished
Table 2. Forcing Global Climate Model (GCMs) Runs Chosen for the NARCCAP Phase II Multiple Regional Climate Model (RCMs) Runs for Current (20C3M; 1970–1999) and Future (A2; 2040–2069) Scenarios
ScenarioPCMDI/CMIP3Notes
CCSM3
Currentncar_ccsm3_0 20c3mRun5
Futurencar_ccsm3_0 sresa2Run5
 
CGCM3
Currentcccma_cgcm3_1 20c3mRun4
Futurecccma_cgcm3_1 sresa2Run4
 
HadCM3
CurrentNot archivedCustom run for NARCCAP
FutureNot archived 
 
GFDL
Currentgfdl_cm2_1 20c3mRun2
Futuregfdl_cm2_1 sresa2Run1

[9] All of the RCM-based analyses reported here used seasonal and annual means derived from 3 hourly NARCCAP output [Mearns et al., 2007]. Five combinations of RCMs and GCMs for Phase II had been archived as of the date of this writing (labeled as “finished” in Table 1). Of those, the MM5/CCSM3 future scenario run had an error, and is being rerun (the current-climate run for this RCM/GCM combination is complete). For the RCM3/CGCM3 and RCM3/GFDL runs, only surface air temperature and precipitation output were archived. Therefore, for evaluation of historical performance, we effectively had access to six RCM/GCM combinations for precipitation (P) and surface air temperature (T), and three combinations for evapotranspiration (ET), runoff (R) and terrestrial storage change (P-ET-R). For the climate change runs, there were five combinations for P and T, and three combinations for ET and R. Because we focus on water budget changes over the CRB, we only analyzed the three model combinations for which P, ET, and R were all available.

2.2. Host GCMs

[10] For comparison, monthly surface air temperature (T), precipitation (P), evapotranspiration (ET), and runoff (R) for CCSM3, CGCM3, and GFDL CM2.1 were obtained from the World Climate Research Programme's (WCRP's) Coupled Model Intercomparison Project phase 3 (CMIP3) multimodel data set [e.g., Meehl et al., 2007]. The HadCM3 run for NARCCAP was different from that in the CMIP3 archive; therefore, output for this GCM was obtained directly from the NARCCAP team.

2.3. Model Evaluation

[11] Land surface variables for the historical period (1970–1999) were taken from the one-eighth degree historical North American Land Data Assimilation System (NLDAS) data set of Maurer et al. [2002]. This is the reference data set for comparisons with the NARCCAP RCMs and the host GCMs. A frequently encountered difficulty in assessing model-predicted land-atmosphere exchanges of moisture and energy is the absence of comprehensive observations to which model predictions can be compared at the spatial and temporal resolutions at which the models operate. The NLDAS data set we used provides both gridded observations (for precipitation and temperature) and model-derived output (e.g., for runoff, snow water equivalent (SWE), and evapotranspiration) for land surface states and fluxes over the conterminous United States and portions of Canada and Mexico. The model-derived output are from the Variable Infiltration Capacity (VIC) land surface model driven by gridded precipitation, temperature and wind time series, as well as downward solar and longwave radiation derived from the daily temperature range and temperature, respectively, following algorithms detailed by Maurer et al. [2002]. VIC is a macroscale terrestrial hydrologic model that balances both surface energy and water over each model grid cell [Liang et al., 1994, 1996]. The VIC model simulates streamflow, and by closure it is constrained to balance other terms in the land-surface water and energy budgets. Although direct comparisons of VIC-derived SWE is complicated by point to area comparison issues, when VIC is forced with local meteorology, it has been shown to reproduce observed SWE well [Cherkauer and Lettenmaier, 2003], and it is an attractive surrogate for observations. The approach of Maurer et al. [2002] to gridding precipitation and temperature observations preserves observed orographic effects. Furthermore, model-derived variables, like runoff and SWE, are effectively constrained by the fact that model streamflow was calibrated to match observed values. The VIC gridded observations and model-derived variables simulations are described and have been evaluated extensively in publications like that of Maurer et al. [2002] and have previously been used for evaluation of other, less constrained, derived data sets like the NCEP/NCAR reanalysis by Maurer et al. [2001]. Similar global derived data sets have been evaluated by Nijssen et al. [2001]. Furthermore, the data have been used extensively in widely cited published assessments of climate change impacts on the surface hydrology of the CRB [Christensen et al., 2004; Christensen and Lettenmaier, 2007].

2.4. Methods

[12] We first show that the mean patterns of the driving GCM are reproduced in the RCM simulations for key variables. We then evaluate the finer-scale features that are more resolved by the RCMs than by their host GCMs. The baseline spatial resolution of the RCMs, as run for NARCCAP, is 50 km, whereas the GCMs typically range from 1 to 4 degrees latitude-longitude. The NLDAS data sets used for evaluation have one-eighth degree spatial resolution, which is considerably finer than either RCMs or GCMs. For comparison, both GCM and RCM output were interpolated to one-eighth degree resolution using an inverse distance squared interpolation, and only points within the CRB were compared. We did not interpolate to coarser spatial resolutions to avoid losing information from the finer resolution. Initially, we evaluated surface air temperature, precipitation, evapotranspiration, runoff, and storage change for the historical period over the CRB. Annual and seasonal changes in the water budget (precipitation, evapotranspiration, and runoff) for the future (2040–2069) period were then compared with the historical period (1970–1999).

3. Results: Evaluation for Historical Period

[13] We first evaluate the historical RCM and GCM simulations through comparisons with NLDAS data (Table 3a and Figures 211). We focus on surface air temperature (T), precipitation (P), evapotranspiration (ET), runoff (R), and terrestrial storage change (P-ET-R).

Figure 2.

Surface air temperature distribution pattern for current period (1970–1999) of the one-eighth degree historical (a) North American Land Data Assimilation System (NLDAS) data set, (b–g) regional climate models/general circulation models (RCMs/GCMs) combinations, (h–m) the host GCMs, and (n–s) difference between the RCMs and its host GCMs (RCMs-GCMs).

Figure 3.

Variations in annual, winter, spring, and summer surface air temperature as elevation from the RCMs/GCMs and the host GCMs comparing to the one-eighth degree historical NLDAS data set (NLDAS) over the Colorado River basin (CRB) (the solid line is the historical period and the dashed line is the future period).

Figure 4.

Same as Figure 2 but for precipitation.

Figure 5.

Same as Figure 3 but for annual, winter, and summer precipitation.

Figure 6.

Same as Figure 2 but for evapotranspiration.

Figure 7.

Same as Figure 5 but for evapotranspiration.

Figure 8.

Same as Figure 2 but for runoff.

Figure 9.

Same as Figure 3 but for runoff.

Figure 10.

(a) Annual snow water equivalent distribution and (b) annual, (c) winter, and (d) spring variations as elevation from the RCMs/GCMs and the host GCMs comparing to the one-eighth degree historical NLDAS data set (NLDAS) over the CRB (the solid line is the historical period, and the dashed line is the future period). (top) WRF/CCSM3, (middle) CRCM/CGCM3, and (bottom) HRM3/HadCM3. Results for HadCM3 not archived.

Figure 11.

Same as Figure 5 but for terrestrial storage change.

Table 3a. Annual (ANN) and Seasonal (DJF, MAM, JJA, and SON) NLDAS and Simulation of the Surface Air Temperature, Precipitation, Evapotranspiration, Runoff, Terrestrial Storage Change, and Snow Water Equivalent for the RCM Driven by GCM (RCM/GCM) and GCM Over the CRB for 1970–1999 and Spatial Correlation Coefficient (CORR) Between RCMs, GCMs, and NLDASa
 ValueCORR
ANNDJFMAMJJASONANNDJFMAMJJASON
  • a

    T is in units of °C and P, ET, R, and P-ET-R are in units of mm.d−1; SWE are in units of mm. NR denotes “not reported.”

Surface Air Temperature
NLDAS10.650.379.7721.3311.2711111
WRF/CCSM39.15−2.118.8620.868.990.950.960.940.930.95
MM5/CCSM39.63−1.429.4421.359.130.950.960.940.920.95
CCSM310.34−1.668.6223.8810.510.880.920.850.820.89
CRCM/CGCM37.52−4.216.6720.467.150.940.960.930.910.94
RCM3/CGCM37.45−3.785.7719.817.980.940.950.930.910.94
CGCM39.5−3.337.5823.4710.290.860.910.840.80.88
RCM3/GFDL6.03−5.283.3919.116.890.940.950.930.920.94
GFDL8.32−3.586.4321.938.490.870.910.840.820.89
HRM3/HadCM312.10.169.7925.612.870.950.960.940.920.95
HadCM311.26−0.929.524.5611.890.880.920.840.830.89
 
Precipitation
NLDAS1.011.040.931.021.0511111
WRF/CCSM30.911.620.980.360.690.720.560.750.430.7
MM5/CCSM31.091.921.10.340.990.730.640.780.420.75
CCSM31.141.911.180.441.050.22−0.180.470.550.33
CRCM/CGCM30.991.11.090.830.950.620.470.680.580.63
RCM3/CGCM31.171.641.240.571.240.60.490.740.470.59
CGCM30.981.191.010.760.970.4−0.080.620.370.49
RCM3/GFDL1.862.351.841.122.110.650.510.650.510.6
GFDL1.612.011.610.971.840.34−0.240.510.20.22
HRM3/HadCM31.351.631.691.180.890.620.530.70.620.64
HadCM31.391.531.781.260.980.31−0.110.540.420.41
 
Evapotranspiration
NLDAS0.870.360.851.590.6911111
WRF/CCSM30.780.51.161.10.340.710.540.680.70.55
CCSM30.650.441.030.570.580.480.680.430.340.47
CRCM/CGCM30.820.391.281.020.590.560.60.340.550.58
CGCM30.810.461.270.880.630.50.530.230.430.54
HRM3/HadCM31.090.61.661.570.550.590.490.510.610.58
HadCM31.290.652.021.760.720.340.360.220.40.48
 
Runoff
NLDAS0.150.090.210.220.0711111
WRF/CCSM30.150.20.330.040.030.160.240.260−0.11
CCSM30.430.720.80.060.13−0.02−0.150.190.1−0.15
CRCM/CGCM30.170.080.60.0100.5−0.250.560.310.32
CGCM30.180.140.56000.26−0.190.30.280.2
HRM3/HadCM30.250.090.670.20.060.470.20.450.410.5
HadCM30.10.040.170.130.050.08−0.220.070.110.03
 
Terrestrial Storage Change
NLDAS−0.010.59−0.13−0.790.2911111
WRF/CCSM3−0.020.92−0.52−0.780.320.010.530.330.650.62
CCSM30.060.75−0.65−0.190.330.10.06−0.370.240.46
CRCM/CGCM300.64−0.78−0.20.36−0.240.47−0.310.540.62
CGCM300.59−0.82−0.120.340.160.16−0.450.340.38
HRM3/HadCM300.95−0.64−0.590.28−0.190.53−0.180.650.61
HadCM300.84−0.41−0.630.210.080.060.50.390.31
 
Snow Water Equivalent
NLDAS14.9524.5128.683.972.6311111
WRF/CCSM35.8216.485.7201.050.470.560.47--
CCSM319.9629.9225.904.520.210.310.21--
CRCM/CGCM316.7238.3425.670.062.830.470.530.59--
CGCM311.2529.3914.2301.40.280.360.30--
HRM3/HadCM313.6631.6821.260.051.630.420.530.44--
HadCM3NRNRNRNRNRNRNRNRNRNR

3.1. Surface Air Temperature

[14] In comparisons of 30 year mean T, all the GCMs and RCMs captured obvious T patterns of lower T in the north and higher T in the south (Figure 2). Spatial T patterns of the RCMs and their driving GCMs were highly correlated (greater than 0.89); however, RCMs had a tendency to be warmer in the southern part of the basin and cooler in the northern part of the basin than GCMs (Figures 2n2s), except for HRM3/HadCM3, which was warmer than HadCM3 over the entire basin (Figure 2 and Table 3a).

[15] Surface air temperature correlations for the RCMs and NLDAS were much higher than for other variables (Table 3a). This implies that regional and global model simulations for T are more reliable than for other variables, which agrees with other studies [Randall et al., 2007; Snyder and Sloan, 2005; Zhang et al., 2009; Leung and Ghan, 1999a; Leung et al., 2003a, 2003b; Plummer et al., 2006]. For annual and seasonal T, the spatial correlations between the RCMs and NLDAS were much higher than for GCMs because of the enhanced representation of the land-surface boundary, especially in spring, summer, and fall.

[16] Generally, the RCMs produced better T lapse rates and smaller biases at high elevations than GCMs (Figure 3). The T lapse rate below 2250 m in both GCM and RCM output is well defined (Figure 3). Annual T from GCMs was overestimated above 2250 m and underestimated below. The RCMs have two obvious T characteristics: (1) they produce more realistic lapse rates than GCMs, especially above 2250 m, (2) RCM performance varied because of different model physics, but the RCMs forced by the same GCMs (WRF/CCSM3 and MM5/CCSM3, CRCM/CGCM3 and RCM3/CGCM3) had similar performance in terms of T (Figure 3). WRF/CCSM3 and MM5/CCSM3 had the best performance above 2250 m.

[17] Seasonally, biases varied for both RCMs and GCMs when averaged over the CRB (Table 3a). In the cold season, GCMs had a cold bias and RCMs amplified this cold bias in the forcing data, whereas in the summer GCMs had a warm bias and RCMs had a cold bias (except for HRMs/HADCM3). As a result, RCMs show a colder bias in annual T than GCMs because in GCMs the cold bias in winter is offset by the warm bias in summer (Table 3a). The cold season behavior we found coincides with several other studies. Duffy et al. [2006] found that RCMs were also too cold in late winter and spring over the western United States. MacKay et al. [2003] concluded that annual mean daily temperatures were 1.8°C colder than observed over the Mackenzie basin. Plummer et al. [2006] found that although CRCM had a warm bias over a large part of the continent, the model had a cold bias between 3° and 4°C over the United States southwest and northern Mexico. Leung et al. [2003a] also found that the WRF simulation was slightly too cool in the southwestern United States, in spite of a warm bias over the coast and mountains; they associated the cold bias to a longer snow covered season and larger peak snow water equivalent than was observed.

[18] Biases arise from the host GCM forcings but also from RCM physics. To further explore origins of the biases, we compared the bias from RCM/NCEP and RCM/GCM simulations. We considered the biases of RCM/NCEP simulations to come from errors in model physics (biasphys) whereas the counterpart for the RCM/GCM simulations represents total bias (biastotal). The absolute of biasphys divided by the absolute of biastotal represents how much physical process errors contribute to the total bias. CRCM had a large cold bias when driven by NCEP (Table 3b), and the bias was amplified when forced by CGCM3 (Table 3a). HRM3 had a large warm bias even when forced by NCEP/DOE (Table 3b) (the annual bias was much larger than simulations forced by HadCM3). Zhang et al. [2009] also found that HRM3 had a warmer bias compared to WRF, even using the same NCEP/DOE forcing over the U.S. Pacific Northwest. The contribution of physical process errors to total bias was 40% and 60% for CRCM and HRM3, respectively. However, WRF, MM5, and RCM3 biases came mostly from their GCMs, with the contribution from physical process errors being 5% or less. For example, RCM3/GFDL had the largest cold bias relative to other RCMs, which was not present when RCM3 was forced by NCEP/DOE (Table 3b). This indicates that the bias largely came from the GFDL GCM forcing. In general, our analysis suggests that depending on the GCM, the bias not only relates to the host GCM, but also to model physical processes, and that the relative contributions vary considerably among the GCM/RCM combinations.

Table 3b. Biases of Surface Air Temperature for the RCM Simulationsa
 Biasphys (°C)Biastotal (°C)Contribution (%)
  • a

    Biasphys is the biases of the RCM simulations forced by NCEP/DOE, which represents the bias from physics processes; Biastotal is from the RCM simulations forced by GCMs, which represents the bias from physics processes and forcing GCM. Contr is the absolute of biasphys divided by the absolute of biastotal, which represents the contribution of bias from physics processes.

  • b

    RCM3 forced by CGCM3.

  • c

    RCM3 forced by GFDL.

WRF0.07−1.54
MM50−1.020
CRCM−1.26−3.1340
RCM3b0.18−3.25
RCM3c0.18−4.624
HRM34.331.4560

3.2. Precipitation

[19] The seasonal cycle of precipitation was not as well represented as temperature by either GCMs or RCMs (Table 3a). This is consistent with many previous studies [Randall et al., 2007; Snyder and Sloan, 2005; Zhang et al., 2009; Leung et al., 2003a, 2003b; Plummer et al., 2006].

[20] Spatial correlation of annual precipitation between GCMs and NLDAS ranges from 0.22 to 0.40. None of the four GCMs was able to capture the higher annual precipitation over the mountainous part of the CRB, and some were unable to represent dry conditions in the lowlands because of their coarse spatial resolution (Figure 5). In winter, precipitation had mostly negative spatial correlations with NLDAS (Table 3a). Figure 5 shows that precipitation in winter mostly occurred at 1800 m for GCMs, not above 2500 m as it occurred in NLDAS, which again is due to the coarse GCM resolution that does not distinguish the considerable topographic changes in the CRB. There is not enough dynamical uplift in GCMs because of their smoothed topography [Leung and Ghan, 1999a]. Castro et al. [2007] also states that GCMs cannot properly represent the diurnal cycle that strongly depends on terrain-induced convection; this is the most important value-added component of the enhanced representation of the land-surface boundary of the RCMs. In summer, GCMs had better lapse rates than winter, so unrealistic annual lapse rates for GCMs were largely from winter differences (Figure 5).

[21] In contrast to GCMs, the RCMs have enhanced variation in precipitation with elevation because of their finer topography (Figure 5). Spatial correlations of annual precipitation between the RCMs and NLDAS ranged from 0.60 to 0.73 (Table 3a) (summer correlations were smaller, not surprisingly as even RCMs were unable to resolve the small-scale processes that dominate precipitation during the warm season). Improvements in RCMs (relative to their host GCMs) in winter and spring are especially apparent (Table 3a). The RCM simulations were closer to NLDAS than GCMs, except for oversimulations in HRM3/HadCM3 and RCM3/GFDL (Figure 5). WRF/CCSM3 and CRCM/CGCM3 were biased low at high elevations despite improvements relative to their host GCMs (Figure 5).

[22] Consistent with their host GCMs, all RCMs produced more precipitation in winter and less (except HRM3/HadCM3 and RCM3/GFDL) in summer, hence the annual values and correlations reflect some compensation of these differences (Table 3a). A robust amplification of cold season precipitation was also found by Wang et al. [2009] based on NARCCAP Phase I analysis, by Leung et al. [2003a, 2003b] based on MM5 simulations, and by Giorgi et al. [1992] based on RCM3 simulations. Summer precipitation simulations had a larger intermodel spread than in other seasons, which is also suggested by the lower spatial correlations noted above. Basic convective research [e.g., Kain and Fritsch, 1990] indicates that warm season precipitation is especially sensitive to differing representations of subgrid convection. Previous studies [Gochis et al., 2002; Liang et al., 2007] also found that RCM simulations of precipitation in the North American Monsoon region (which includes the southern part of the CRB) were quite sensitive to the choice of the convective parameterization. The NARCCAP models use different convective and microphysical parameterizations. In this context, the differences in summer precipitation are not surprising. Furthermore, given the subgrid nature of convective precipitation (and to a lesser extent, representations of cold season processes), the 50 km NARCCAP resolution still has limitations in resolving precipitation in topographically complex regions like the CRB, notwithstanding that the spatial resolution is much finer than that of the GCMs.

[23] Unlike air temperature, the RCMs differed in how well they simulate precipitation even when forced by the same GCMs. For example, WRF/CCSM3 and MM5/CCSM3 had similar precipitation simulations at low elevations, but MM5 simulated precipitation better than WRF at high elevations. This suggests differences in physical processes related to topography. Differences because of how physical processes were simulated are also apparent when host GCMs were coupled with different RCMs. For example, CRCM/CGCM3 simulated precipitation better than RCM3/CGCM3. CRCM/CGCM3 had the smallest biases among the RCMs in both annual mean and monthly averages (Table 3a), whereas RCM3/CGCM3 overestimated precipitation, especially in winter. The wetter bias by RCM3/CGCM3 was partly from model physics parameterizations. RCM3 had wet biases with GFDL too, greatly overestimating precipitation in winter and at high elevations in summer (Figure 5). Because the only difference between RCM3/CGCM3 and RCM3/GFDL was their forcings, this larger bias is also attributable to the host GCMs.

3.3. Evapotranspiration

[24] Evapotranspiration (ET) was archived by NARCCAP for three RCM/GCM combinations: CRCM/CGCM3 (CRCM for short in the following), HRM3/HadCM3 (HRM3), and WRF/CCSM3 (WRF). The three GCMs did not have much spatial variation and did not capture ET distributions at high elevations well, especially CCSM3 (Figures 6h6m). Not surprisingly, the RCMs better resolved inferred regional variations (Figures 6b6g). The RCMs had higher spatial correlation with NLDAS and also captured ET magnitude better than GCMs, especially in summer (Table 3a).

[25] Performance varied for each RCM. WRF matched NLDAS better than the other two RCMs, except for slight underestimations below 1800 m, which occurred mostly in summer. CRCM had almost the same performance as its host CGCM3, which underestimated annual ET at high elevations and overestimated ET at low elevations. HRM3, consistent with its host HadCM3, overestimated annual ET, especially in high elevations (Table 3a and Figure 7).

[26] Because ET is limited by moisture availability over much of the basin, the ET lapse curve follows precipitation patterns to some extent (compare Figures 5 and 6), especially below 2600 m. Generally, the RCMs tended to overestimate ET in winter and underestimate ET in summer, especially for elevations below 2000 m, which is consistent with the modeled precipitation patterns in Figure 5.

3.4. Runoff

[27] As Figure 8 shows in the 30 year mean spatial distribution of runoff over the CRB, most of the runoff comes from the mountains, which cover a small portion of the total basin area in the north. The RCMs better reproduced the basic spatial characteristics of the surface hydrology and topographically induced characteristics of runoff over the CRB than the GCMs as also found in earlier studies [e.g., Giorgi et al., 1994]. The RCMs generally had much higher correlations with NLDAS runoff than their host GCMs, especially HRM3/HadCM3 and CRCM/CGCM3 (Table 3a). Nonetheless, runoff correlations were lower than for P and ET. In WRF the spatial correlation was low, although it captured the runoff magnitude well. The GCMs generally overestimated runoff in winter, whereas RCM runoff was closer in magnitude to NLDAS than GCMs, especially in winter (Table 3a). This is because in the CRB, runoff is driven by springtime snowmelt, which in turn results from winter snow accumulation. As noted in section 3.1, RCMs generally had colder temperatures than GCMs; these colder temperatures promote winter snow accumulation at high elevations, where a large portion of the CRB's runoff is generated [Barnett et al., 2005; Barnett and Pierce, 2009; Cayan et al., 2001].

[28] The maximum runoff for the RCMs occurred in spring, not spring-summer as in NLDAS (Table 3a), and runoff in spring was mostly overestimated, especially in HRM3. This mismatch is linked with underestimated runoff in summer, mainly because of the underestimation of precipitation and the overestimation of snowmelt in spring. This suggests that spring runoff is quite sensitive to the model treatment of snow cover relative to temperature, as also indicated by Plummer et al. [2006] and MacKay et al. [2003].

[29] Similar to P and ET, the RCMs captured runoff variations with elevation better than GCMs (Figure 9). The NLDAS distribution implied that most of the CRB's runoff comes from areas above 2250 m, and decreases sharply at lower elevations. For the three host GCMs, runoff was mainly from elevations that range from 2000 m to 2400 m, especially CCSM3 and CGCM3 in winter, and not from the highest elevations as implied by NLDAS. The large runoff at midelevation in winter was mostly consistent with precipitation. Commonly, in winter, precipitation falls as snow rather than rain because of temperatures below freezing. But snow cover in CCSM3 and CGCM3 was located at lower elevations than in other models (Figure 10) where temperature was near freezing in the diurnal cycle (Figure 3). The accumulation of low-elevation snow cover and the overall higher temperatures of CCSM3 and CGCM3 resulted in increased runoff. Furthermore, snowfall at low elevations is mostly related to the GCM's weak uplift due to the smoothed topography and model physics in the cold season. Also, it is notable that all GCMs and RCMs underestimated runoff production in summer (Figure 9) and fall (not shown).

[30] Of the models considered here, HRM3/HadCM3 simulated the most realistic annual runoff. It produced large runoff from high elevations in summer; although not as large as NLDAS (none of the RCMs reproduced the large observed summertime runoff values). WRF/CCSM3 overestimated winter runoff especially at elevations below 2400 m, underestimated snow water equivalent (SWE) in winter, and missed most summer runoff as a result of the underestimation of high-elevation precipitation (section 3.2 and Table 3a). It therefore had significant negative bias in annual runoff for elevations above 2250 m. CRCM/CGCM3 had slight improvement relative to its host CGCM3 at high elevations but missed the summer runoff because of underestimation of high-elevation precipitation.

3.5. Terrestrial Storage Change

[31] We calculated the change in terrestrial storage as P-ET-R, an index of water budget closure. As mentioned previously, we used an off-line VIC run as the source of surrogates of NLDAS runoff and ET, which we combined with gridded precipitation observations to compute NLDAS storage change. In all RCMs, water balance (change in storage on an annual basis averaged over the historical period) approached zero (Table 3a), which is essentially the same as NLDAS. Although annual mean P-ET-R was zero, these values vary for each season. In winter and fall, P-ET-R was positive, especially above 2250 m, and in summer and spring, it was negative (Table 3a), especially at high elevations (Figure 11); however, none of the GCMs had summer drying to the extent inferred from NLDAS. The RCMs showed conditions that are slightly drier than GCMs in spring and summer, although not as much as NLDAS. Generally, P-ET-R was opposite of the runoff pattern (Figure 11).

4. Results and Discussion: Climate Change Impacts

[32] To evaluate climate change and related impacts on water availability, we focus on temperature and runoff change as well as snow cover and precipitation between future (2040–2069) and historical (1970–1999) simulations for three RCM/GCM combinations: CRCM/CGCM3 (CRCM for short in this section), HRM3/HadCM3 (HRM3), and WRF/CCSM3 (WRF). We selected these GCM/RCM combinations because they are the only ones for which all of the variables of interest were archived.

4.1. Temperature Changes

[33] First, we examine climate change-induced differences in temperature over the CRB (Figure 12). RCMs projected warming signals that are consistent with their host GCMs. All RCMs and GCMs projected more warming in the north than south (Figure 12a) and more warming in the warm season than in the cold season (Figure 12b). The magnitude of changes between RCMs and GCMs, however, differed, depending on the season. In the cold season, the RCMs projected less warming than GCMs (Figure 12b). However, in the warm season, two RCMs (CRCM and HRM3) projected stronger warming than their host GCMs. WRF projected less warming than CCSM3. In the CRB, snowmelt is the source of runoff, and for this reason, warming in the cold season has a very strong influence on water availability. It is worth noting that effects of increased atmospheric CO2 on the surface air temperature from the RCMs show strong elevation dependencies, consistent with the work of Giorgi et al. [1997], Leung and Ghan [1999b], and Kim et al. [2002].

Figure 12.

The (a) distribution and (b) basin average of annual and seasonal surface air temperature (T) change over the CRB for the RCMs and the host GCMs for the future (2040–2069) minus current (1970–1999).

4.2. Climate Change Impacts on Water Availability and Related Variables

[34] Generally, annual runoff changes for the RCMs were consistent with their host GCMs (see ANN, Figure 13). CCSM3 and CGCM3 projected decreases in annual runoff, whereas HadCM3 projected slight increases. Consistent with their host GCMs, CRCM and WRF had decreasing runoff, with the same pattern but different magnitudes (Figure 13). HRM3 had a slight increase in annual runoff, consistent with HadCM3. Runoff changes for RCMs and GCMs mostly occurred in winter (DJF), spring (MAM), and summer (JJA) (Figure 13b). These results from runoff being impacted by the combination of three factors: snow cover change in winter, temperature change in spring, and precipitation change in all seasons. In summer, with temperatures considerably above freezing, the changes in runoff were mostly consistent with those of precipitation. In winter and spring, however, runoff change was complicated because of snowmelt. We highlight changes in winter, spring, and summer to show how RCMs and GCMs differ in how they represent the susceptibility of headwater streams to climate change.

Figure 13.

Same as Figure 12 but for runoff (R) change.

4.2.1. Winter

[35] In winter, all RCMs and GCMs projected runoff increases except for WRF (Figure 13b). HRM3 projected larger increases than its host, HadCM3, whereas CRCM projected smaller increases than its host CGCM3. Runoff changes primarily tracked precipitation and snow cover changes (Figures 13, 14, and 15). The increase in CCSM3 runoff closely tracked the trend in snow cover (Figures 13 and 15). Different temperature simulations resulted in different snow cover in RCMs and GCMs. In the upper basin, most areas of CCSM3 snow cover were located near the freezing level, which is lower than other GCMs and RCMs (Figure 9). Lower-elevation snow was more susceptible to melting in response to varying temperatures; this, in turn, led to larger runoff. WRF projected decreased runoff as a result of the combination of no precipitation change (Table 4 and Figure 5) and increased evapotranspiration (Table 5 and Figure 7) due to increased temperature (Table 8 and Figure 12). Generally, smaller winter precipitation increases (CRCM/CGCM3 and HRM3/HadCM3, Table 4) and reduced snowmelt (WRF/CCSM3, Table 7) were the main reasons for smaller winter runoff increases in the RCMs than in the GCMs (Figure 13).

Figure 14.

The distribution of winter and summer precipitation (P) change for the RCMs and their host GCMs for the future (2040–2069) minus current (1970–1999).

Figure 15.

The distribution of winter and spring snow water equivalent (SWE) change for the RCMs and the host GCMs for the future (2040–2069) minus current (1970–1999). Results for HadCM3 not archived.

Table 4. Annual and Seasonal Precipitation (P) Change (2040–2069)-(1970–1999) for RCMs and GCMs over the CRB (mm d−1 (%))
 ANNDJFMAMJJASON
WRF−0.03 (−3%)0.00 (0%)−0.04 (−4%)−0.04 (−11%)−0.03 (−4%)
CCSM3−0.06 (−5%)−0.12 (−6%)−0.11 (−9%)0.13 (30%)−0.14 (−13%)
CRCM−0.09 (−9%)0.04 (4%)−0.13 (−12%)−0.22 (−27%)−0.03 (−3%)
CGCM3−0.01 (−1%)0.10 (8%)−0.06 (−6%)−0.07 (−9%)−0.02 (−2%)
HRM3−0.06 (−4%)0.08 (5%)−0.14 (−8%)−0.24 (−20%)0.07 (8%)
HadCM30.01 (1%)0.11 (7%)−0.19 (−11%)−0.05 (−4%)0.16 (16%)
Table 5. Annual and Seasonal Evapotranspiration (ET) Change (2040–2069)-(1970–1999) for RCMs and GCMs Over the CRB (mm d−1 (%))
 ANNDJFMAMJJASON
WRF0 (0%)0.04 (8%)0.02 (2%)−0.09 (−8%)0.05 (15%)
CCSM30.06 (9%)0.11 (25%)0.03 (3%)0.09 (16%)0.02 (3%)
CRCM−0.05 (−6%)0.09 (23%)0 (0%)−0.23 (−23%)−0.06 (−10%)
CGCM30.02 (2%)0.09 (20%)0.05 (4%)−0.03 (−3%)−0.03 (−5%)
HRM3−0.07 (−6%)0.04 (7%)−0.05 (−3%)−0.28 (−18%)0.02 (4%)
HadCM3−0.01 (−1%)0.05 (8%)−0.09 (−4%)−0.14 (−8%)0.12 (17%)

4.2.2. Spring

[36] The most significant change and differences in changes across models occurred in spring (Figure 13b). In the western United States, the runoff-change signal tracks temperature and snow cover change. GCM runoff decreased dramatically in CCSM3 and CGCM3 because of reduced snowmelt (Figures 13 and 15); whereas, in HadCM3 runoff slightly increased. In contrast, the RCMs predicted smaller runoff decreases or larger increases than their host GCMs, especially for WRF (Figure 13). The lower-elevation snow cover (Figure 10) and larger warming (Figure 12) caused larger reductions in snow cover for GCMs than for RCMs (Figure 15). Larger snowmelt leads to lower albedo. The reduced albedo absorbs more energy and generates positive feedback for more warming and increased snowmelt as a result of enhanced warming. As a result, GCMs projected larger temperature increases than the RCMs in spring (Figure 12); temperature differences are especially notable between WRF and its host CCSM3. WRF projected 2.2°C increases over the CRB, while CCSM3 projected 3.0°C increases. Similar to its response in winter, HRM3 projected larger runoff increases in spring at high elevations. In contrast, its host GCM (HadCM3) predicted almost no change in spring. Leung and Ghan [1999b] also found a negative signal over the CRB (see their Figure 7), although over the Pacific Northwest they concluded the opposite in that the RCM warming signal is stronger than the CCSM3 signal during winter in most areas of the Pacific Northwest. This indicates that climate change susceptibility as simulated by the RCMs as contrasted with GCMs depends on the particulars of the region being investigated.

[37] The surface temperature, snow cover, and runoff change signals were strongly elevation dependent. Furthermore, most reductions of snow cover were near the snowline in the control simulation, which is consistent with the work of Giorgi et al. [1997], Leung and Ghan [1999b], and Kim et al. [2002]. Leung and Ghan [1999b] and Kim et al. [2002] attributed runoff change in spring to higher freezing levels in the future climate, and Kim et al. [2002] stated that the warming signal between 1000 and 2000 m is larger than areas over 2000 m in the western United States. For the CRB, the snowline is often above 2000 m (Figure 10). In our study, all the RCMs and HadCM3 agreed with NLDAS surrogates for SWE as to how snow accumulation changes with elevation, although they underestimated snow amounts. For CCSM3, however, the snow amount was not only under-predicted, but the snowline was predicted to be much lower than NLDAS (Figure 10). Most of the snow was located between 1700 and 2400 m in CCSM3, which was lower than the location of snow in WRF. As freezing levels become higher in the future, large amounts of snow from below 2000 m melted in CCSM3 but not in the RCMs. Therefore, since CCSM3 had a lower snowline, snow was easier to melt in CCSM3 than in other models. The reduced albedo, due to increased snowmelt, led to greater absorption of radiation, which raised surface temperatures. Raised temperatures accelerated snowmelt further and caused significant changes in runoff (Figure 13 and Table 6) and temperature (Figure 15 and Table 8). Thus, consistent with previous studies, the elevation-dependent runoff and temperature signal is mainly related to reduced winter snowfall and spring-summer snow cover in high altitudes in the altered climate. The snow-albedo feedback plays an important role in determining the climate change signal.

Table 6. Annual and Seasonal Runoff (R) Change (2040–2069)-(1970–1999) for RCMs and GCMs Spatial Averaged Over Area Above 2250 m Over the CRB (mm d−1 (%))
 ANNDJFMAMJJASON
WRF−0.03 (−16%)0 (0%)−0.11 (−19%)0 (0%)0 (0%)
CCSM3−0.07 (−16%)0.27 (61%)−0.56 (−50%)0.01 (17%)0 (0%)
CRCM−0.07 (−16%)0.06 (200%)−0.32 (−19%)−0.03 (−75%)0 (0%)
CGCM3−0.04 (−13%)0.12 (80%)−0.3 (−27%)0.01 (100%)0 (0%)
HRM30.05 (5%)0.15 (88%)0.2 (7%)−0.15 (−20%)0.01 (5%)
HadCM30.01 (6%)0.01 (20%)0.04 (15%)−0.02 (−9%)0 (0%)

4.2.3. Summer

[38] Runoff change differences between RCMs and GCMs also occurred in summer (although they contribute less to annual flow changes). All RCMs projected decreases in summer runoff, whereas their host GCMs projected no change for CGCM3 and HadCM3 and increases for CCSM3 (Figure 13). The runoff change difference between the RCMs and GCMs in summer resulted mostly from differences in precipitation changes; the RCMs projected greater decreases than the GCMs in summer (Figure 14 and Table 4), in agreement with Leung and Ghan [1999b] over the CRB. Runoff changes at low elevations were similar to the host GCMs, but differed at high elevations, which we discuss further in section 4.3.

4.3. Climate Change Impacts at High Elevations

[39] As mentioned above, climate change is highly elevation dependent over the CRB. Much of the runoff comes from areas above 2250 m (Figure 9) [Hoerling et al., 2009; Franz et al., 2003; Cayan et al., 2001]. Areas with elevations higher than 2250 m occupy about 22% of the basin's area (Figure 1c). Table 6 lists annual and seasonal runoff changes above this threshold. Differences between the RCMs and GCMs were most prominent in winter, spring, and summer. In winter, all RCMs and GCMs except WRF showed increases in runoff. Two RCMs (CRCM and HRM3) projected larger increases than their host GCMs (CGCM3 and HadCM3, respectively). In spring, two RCMs projected smaller decreases than their host GCMs, specifically −19% for WRF and CRCM, compared to −50% and −27% for CCSM3 and CGCM3, respectively, whereas HRM3 and its host HadCM3 projected increases in spring, with HRM3 projecting a smaller increase than HadCM3. This result agrees with those of Govindasamy et al. [2003] and Sushama et al. [2006] who mention the importance of regional-scale differences in snowpack changes between the high-resolution simulations relative to the low resolution of GCMs, because of better representation of topography at high resolutions. In summer, CCSM3 projected increases in runoff; however, WRF projected no change. CGCM3 projected increases; however, CRCM projected decreases. HadCM3 projected slight decreases; however, HRM3 projected larger decreases than HadCM3 (Table 6).

[40] Consistent with the spatial average over the entire basin, annual runoff change for all the RCMs and GCMs was controlled by the change signal in spring (Figure 13 and Table 6). Runoff change in spring was from changes in snow water equivalent (see SWE, Table 7), which was strongly related to temperature change (Table 8 and Figure 13). All RCMs and GCMs projected that, as temperatures rise, snow reductions were greater in spring than in winter. The spring trend was more influential than in winter because, while the change magnitudes were similar, the atmosphere's variability was smaller in the spring than in winter [Cayan et al., 2001]. Recent studies have documented that observed recent warming caused by earlier runoff during the spring snowmelt period was most pronounced in the mountains of the western states, where snowpack temperatures usually are not far below freezing [Cayan et al., 2001; Stewart et al., 2005; Mote, 2006; Clow, 2010]. Cayan et al. [2001] also noted that even the mountains of Colorado, with their high elevations and cold snowpacks, are experiencing substantial shifts in the timing of snowmelt and its runoff toward earlier in the year. They also state that increasing springtime air temperatures and declining SWE could account for a large portion of the variance in snowmelt timing.

Table 7. Annual and Seasonal Snow Water Equivalent (SWE) Change (2040–2069)-(1970–1999) for RCMs and GCMs Over the CRB (mm (%))
 ANNDJFMAMJJASON
WRF−1.9 (−32%)−4.5 (−27%)−2.3 (−41%)- -- -
CCSM3−11.3 (−57%)−15.4 (−51%)−18.9 (−73%)- -- -
CRCM−4.3 (−26%)−7.9 (−21%)−8.4 (−33%)- -- -
CGCM3−4.0 (−36%)−8.4 (−28%)−7.2 (−50%)- -- -
HRM3−3.8 (−28%)−7.7 (−24%)−6.7 (−32%)- -- -
HadCM3Not reported
Table 8. Annual and Seasonal Surface Air Temperature (T) (T Change, (2040–2069)-(1970–1999)) for Area Above 2250 m for RCMs and GCMs (°C)
 ANNDJFMAMJJASON
WRF1.6 (2.71)−9.71 (2.52)1.16 (2.29)13.49 (3.12)1.44 (2.89)
CCSM35.26 (3.3)−7.5 (2.85)3.13 (3.4)20.11 (3.73)5.32 (3.21)
CRCM0.11 (2.9)−12.02 (2.75)−1.23 (2.32)13.57 (3.52)0.11 (3.03)
CGCM34.92 (2.67)−7.81 (2.59)2.52 (2.46)19.15 (2.96)5.8 (2.69)
HRM34.3 (2.92)−7.4 (2.43)1.62 (2.37)17.65 (3.79)5.33 (3.08)
HadCM36.99 (3.09)−5.58 (2.96)5.45 (2.52)20.34 (3.72)7.74 (3.15)

[41] Annual surface air temperature predicted by the RCMs is 2–7°C lower than that predicted by GCMs above 2250 m (Table 8). The RCMs predicted smaller temperature increases in spring than GCMs (Table 8). As a result, RCMs projected smaller snow reduction percentages than their host GCMs. In spring, SWE reductions were −41%, −33%, and −32% for WRF, CRCM, and HRM3, respectively, compared to −73% and −50% for CCSM3 and CGCM3. Higher reduction percentages for the GCMs were likely because of the higher temperature projections for GCMs at high elevations where snow accumulation and melting occurs (Tables 6–8 and Figure 15). Generally, changes in snow distribution coincide with temperature change. Our results indicate that the RCMs and GCMs both projected that the headwater streams will be affected by a warming climate; however, they are much less susceptible in the RCMs than in the GCMs.

5. Value Added From RCM Simulations

[42] As mentioned above, the value of RCMs lies, not surprisingly, in their ability to add meaningful spatiotemporal detail to global-scale information with more highly resolved representations of surface forcing (e.g., topography and snow cover). Additionally, their superior formulations of model physics processes can contribute to improved skill in downscaled results beyond just changes in surface forcing [Duffy et al., 2006; Han and Roads, 2004]. In our results, RCM simulations at high elevations were more reliable compared to GCMs because of their finer resolution, but also because they had stronger upward lift triggered by their finer distribution in topography. In the CRB, stronger lift raises air parcels higher and results in higher snowlines in RCMs than in GCMs. As a result, runoff change in the CRB was less susceptible to climate change in RCMs than in GCMs. However, one can question the fairness of comparing at the finer resolution with elevation because GCMs inherently were constrained to have a smoother representation of topography than RCMs, especially at the highest elevations. To address this fairness, we aggregated the RCMs and GCMs to 2.5 degree spatial resolution (each GCM has a different resolution). This was close to the mean spatial resolution of the GCMs (Figure 16). When spatially aggregated, there generally was still more variation in grid cells between RCMs than GCMs, especially during the spring, which suggests that the differences we observe in the RCMs are not related only to their higher spatial resolution.

Figure 16.

Same as Figure 13 but at 2.5 degree resolution.

[43] Déqué et al. [2005] used the intermodel standard deviation as a measure of the dispersion of GCMs and RCMs about their centroid. We performed a similar calculation, results of which are shown in Table 9. The results show that the GCM projection spread is larger than for the RCMs not only for annual temperature, but also for annual precipitation and annual runoff. The seasonal spread is smaller for RCMs than GCMs in winter for temperature/precipitation and in spring for runoff. In this sense, there is less intermodel uncertainty in the RCM results relative to those from the GCMs.

Table 9. Intermodal Standard Deviation, as a Measure of Uncertainty, Over CRB for Temperature (°C), Precipitation (mm d−1), and Runoff (mm d−1) for the RCMs and GCMs Projection
 ANNDJFMAMJJASON
Temperature
RCM0.290.080.350.570.35
GCM0.320.330.440.460.34
 
Precipitation
RCM0.050.070.100.190.10
GCM0.060.230.110.190.26
 
Runoff
RCM0.050.070.170.020.01
GCM0.100.030.380.010.03

6. Conclusions

[44] Performance of the Regional Climate Model (RCM) historical simulations forced by four GCMs run for the North American Regional Climate Change Assessment Program (NARCCAP) Phase II was analyzed in comparison to the one-eighth degree historical North American Land Data Assimilation System (NLDAS) data set over the Colorado River basin (CRB). Future (2040–2069) simulations over the CRB using the A2 SRES emission scenario were then compared to historical runs (1970–1999) for the RCMs relative to their host GCMs. Through these comparisons we found that although the RCMs and their host GCMs often have similar patterns for surface temperature, precipitation, and evapotranspiration, in the CRB there are spatial and seasonal differences that result in runoff values in the RCMs being less susceptible to a warming climate than their host GCMs. Our key findings are:

[45] 1. The RCMs capture the primary features of the observed land–surface water budget and surface temperature better than their host GCMs in part because of their ability to better represent strong upward lift in topographically complex regions. Although the RCMs do not have significantly improved simulations of precipitation relative to the GCMs, their ability to better represent surface temperature in mountainous regions has important effects on simulations of evapotranspiration, snowpack, and runoff. This improved skill in simulating temperature is important for detecting climate change signals in the topographically diverse CRB.

[46] 2. Runoff generation and change in the CRB is highly elevation dependent. Headwater streams contribute most of the downstream flow, and annual runoff change signals are primarily controlled by changes in the snowmelt period. Therefore the response of cold season temperatures in headwater streams is key to determining the basin's susceptibility to a warming climate. In the CRB, runoff changes are generally impacted by the combination of three main factors: snow cover in winter, temperature change in spring, and precipitation change in summer.

[47] 3. The RCMs have similar annual climate change signals (although with slightly smaller magnitudes) to their host GCMs. All RCMs and GCMs had larger warming in the north than in the south and more warming in the warm season than in the cold season; however, the magnitude of these changes differ between RCMs and GCMs, and is notably different depending on the season. Compared to the host GCMs, the RCMs project larger decreases in precipitation and evapotranspration in summer, which translated to larger reductions in summer runoff in the RCMs. Alternatively, in the cold season, RCMs project less warming than GCMs. Therefore, although both RCMs and GCMs project decreases over the CRB, the RCMs projected less warming in the spring and thus have smaller decreases in runoff in the spring (arising from smaller temperature changes and better topographic resolution), which results in smaller decreases in annual runoff as compared with the GCMs. Thus, surface air temperature, runoff, and snow water equivalent changes at high elevation all indicate that headwater streams are less susceptible to a warming climate in the RCM climate change simulations than in GCM simulations.

[48] 4. Climate change susceptibility as simulated by the RCMs as contrasted with GCMs depends on the particulars of the region being investigated. Specifically, our results indicate smaller runoff differences in the future in RCMs than in GCMs, primarily because GCMs do not adequately simulate higher elevations where temperature changes have less affect on snow cover (where temperatures are still cold enough to retain snow). In regions where runoff originates in lower elevations, results may differ [e.g., Leung and Ghan, 1999b; Kim et al., 2002].

[49] It should also be noted that although the NARCCAP RCMs captured the primary features of observations and more closely reproduced the NLDAS estimates than do the host GCMs, the 50 km grid resolution may still not be fine enough to resolve orographic effects in parts of the domain. As shown by Ghan et al. [1997, 2006] subgrid variability in summertime precipitation can have substantial effects on surface hydrology, and subgrid variations in vegetation and soil properties also increase surface runoff and reduce evapotranspiration. Furthermore, there remain a number of other uncertainties associated with RCM climate change scenarios. First, the climate change signal for RCMs is dependent on the circulation from the driving GCM. Errors in the GCM response to global emissions forcing will be transferred to the RCMs. We have noted several situations where the RCMs clearly have inherited bias from the parent forcing. Second, the models differ in their ability to project climate change because of different physics, yet because only a limited number of GCM/RCM combinations have been completed and archived to date, understanding these physics is challenging. As Duffy et al. [2006] pointed out, RCM estimates of water content and projections of change in water content vary. Therefore, further synthesis investigations linking the atmosphere and terrestrial components of the water cycle that are based on more RCMs are necessary.

Acknowledgments

[50] We thank NARCCAP for providing the data used. NARCCAP is funded by the National Science Foundation (NSF), the U.S. Department of Energy (DoE), the National Oceanic and Atmospheric Administration (NOAA), and the U.S. Environmental Protection Agency Office of Research and Development (EPA). We acknowledge the contributing NARCCAP modeling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI), and the WCRP Working Group on Coupled Modeling (WGCM) for their roles in making available the GCM data. Support of this data set is provided by the Office of Science, U.S. Department of Energy. This work was supported by a DoE grant DE-FG02-08ER64589.

Advertisement