Influence of climate model biases and daily-scale temperature and precipitation events on hydrological impacts assessment: A case study of the United States

Authors

  • Moetasim Ashfaq,

    1. Department of Environmental Earth System Science, Stanford University, Stanford, California, USA
    2. Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA
    3. Department of Earth and Atmospheric Sciences, Purdue University, West Lafayette, Indiana, USA
    Search for more papers by this author
  • Laura C. Bowling,

    1. Department of Agronomy, Purdue University, West Lafayette, Indiana, USA
    Search for more papers by this author
  • Keith Cherkauer,

    1. Department of Agricultural and Biological Engineering, Purdue University, West Lafayette, Indiana, USA
    Search for more papers by this author
  • Jeremy S. Pal,

    1. Department of Civil Engineering and Environmental Science, Frank R. Seaver College of Science and Engineering, Loyola Marymount University, Los Angeles, California, USA
    Search for more papers by this author
  • Noah S. Diffenbaugh

    1. Department of Environmental Earth System Science, Stanford University, Stanford, California, USA
    2. Department of Earth and Atmospheric Sciences, Purdue University, West Lafayette, Indiana, USA
    3. Woods Institute for the Environment, Stanford University, Stanford, California, USA
    Search for more papers by this author

Abstract

[1] The Intergovernmental Panel on Climate Change's Fourth Assessment Report concludes that climate change is now unequivocal, and associated increases in evaporation and atmospheric water content could intensify the hydrological cycle. However, the biases and coarse spatial resolution of global climate models limit their usefulness in hydrological impact assessment. In order to reduce these limitations, we use a high-resolution regional climate model (RegCM3) to drive a hydrological model (variable infiltration capacity) for the full contiguous United States. The simulations cover 1961–1990 in the historic period and 2071–2100 in the future (A2) period. A quantile-based bias correction technique is applied to the times series of RegCM3-simulated precipitation and temperature. Our results show that biases in the RegCM3 fields not only affect the magnitude of hydrometeorological variables in the baseline hydrological simulation, but they also affect the response of hydrological variables to projected future anthropogenic increases in greenhouse forcing. Further, we find that changes in the intensity and occurrence of severe wet and hot events are critical in determining the sign of hydrologic change. These results have important implications for the assessment of potential future hydrologic changes, as well as for developing approaches for quantitative impacts assessment.

1. Introduction

[2] One of the expected consequences of increasing greenhouse forcing is intensification of the hydrological cycle [e.g., Intergovernmental Panel on Climate Change (IPCC), 2007; Santer et al., 2007; Trenberth, 1999]. Accelerated evaporation rates and earlier snowmelt, along with the increased occurrence of precipitation and temperature extremes, will likely exacerbate the risk of both seasonal floods and recurrent drought episodes [IPCC, 2007]. In the United States, decreasing diurnal temperature range [Easterling et al., 1997], declining snow-to-precipitation ratio [Feng and Hu, 2007; Knowles et al., 2006], increasing extreme precipitation events [Groisman et al., 2001; Karl and Knight, 1998], and changing streamflows [Lins and Slack, 1999; 2005; Mccabe and Wolock, 2002] provide discernable evidence that precursors of such hydroclimatic change are already unfolding and that many are most likely caused by anthropogenic increases in greenhouse gas concentrations [Barnett et al., 2008]).

[3] Hydrological extremes are historically the costliest natural disasters in the United States. For example, in terms of water, energy, ecosystem, and agriculture losses, the total cost of the 1988 Great American Drought alone was $40 billion [Kogan, 2001]. Similarly, based on the 1972–2006 records, average annual losses to flood-related catastrophes across the United States were $2.6 billion [Changnon, 2008]. Because of the economic and societal implications of potential future changes to the hydrologic cycle, the impacts of climate change on the different river basins in the United States has received considerable attention. These studies have been based on a variety of approaches, including regression-based analysis of general circulation model (GCM) simulations [Stewart et al., 2004], dynamic downscaling of GCMs through nested climate models [e.g., Diffenbaugh et al., 2005; Leung and Wigmosta, 1999; Leung et al., 2004; Pal and Eltahir, 2002], and the offline coupling of GCMs with hydrological models [e.g., Christensen et al., 2004; Miller et al., 2003; Sinha and Cherkauer, 2008]. Despite the differences in the methodology, most of these studies show that potential future changes in hydrologic variables such as surface runoff, streamflow and snow cover will be much larger than the observed changes over the past 50 years [e.g., Rauscher et al., 2008; Stewart et al., 2004].

[4] A key challenge to understanding the possible hydrologic response to elevated greenhouse forcing has been constraining and correcting errors in the climate model projections of future climate change [Fowler et al., 2007; Wilby and Wigley, 1997]. Earlier studies have shown that hydrologic simulations are sensitive to the biases in mean and spatial distribution of precipitation and temperature and to the resolution of forcing data [e.g., Hay et al., 2000; Shrestha et al., 2006; Wood et al., 2004]. Thus, substantial effort has been dedicated toward developing techniques for bias correction and spatial enhancement of climate model simulations for a variety of impacts applications [e.g., Andersson et al., 2006; Christensen et al., 2004; Payne et al., 2004; Steele-Dunne et al., 2008; Vanrheenen et al., 2004].

[5] Despite these important advances, most hydrologic studies to date have been limited in at least two respects. First, they rely almost exclusively on GCM-based climate change projections. Since GCMs can be challenged by fine-scale climate processes [Duffy et al., 2003], such GCM-based hydrological impacts assessments potentially miss important aspects of the physical response governing local changes in temperature and precipitation in spatially complex hydrological regimes [e.g., Arora, 2001; Diffenbaugh et al., 2005; Xu, 1999]. Second, they rely almost exclusively on monthly scale GCM variables. Since hydrological extremes are, in part, driven by the daily distribution of temperature and precipitation extremes [Trenberth, 1998], which have been found to be preferentially sensitive to enhanced greenhouse forcing [e.g., Diffenbaugh et al., 2005; Diffenbaugh et al., 2007b; Wehner, 2004], hydrological impacts assessments based on seasonal- or monthly scale information potentially misrepresent changes in the occurrence of high-frequency extreme hydrological events resulting from future changes in daily temperature and precipitation extremes.

[6] In this study we use daily outputs from a high-resolution nested climate model as inputs for a hydrologic model configured over the contiguous United States. Using a quantile-based bias correction technique, we quantify the effect of climate model biases on the simulated hydrological response to anthropogenic changes in greenhouse forcing, as well as the influence of changes in the intensity and occurrence of daily precipitation and temperature extremes in determining future hydrologic change.

2. Methods

2.1. Models

2.1.1. RegCM3

[7] The Abdus Salam Institute for Theoretical Physics Regional Climate Model (RegCM3) is a 3-D, hydrostatic, sigma coordinate, static vegetation, primitive equation nested climate model [Pal et al., 2007]. RegCM3 has been used extensively to study the potential response of regional climate to anthropogenic increases in greenhouse forcing [e.g., Ashfaq et al., 2009; Diffenbaugh et al., 2005; Gao et al., 2008; Gao and Giorgi, 2008; Seth et al., 2007; Trapp et al., 2007; White et al., 2006]. The RegCM3 performance over the contiguous United States has been detailed in a number of previous studies [e.g., Diffenbaugh et al., 2006; Diffenbaugh and Ashfaq, 2007; Rauscher et al., 2008; Walker and Diffenbaugh, 2009]. For this study, RegCM3 data are obtained from simulations with a domain covering the contiguous United States at 25 km horizontal grid spacing and 18 levels in the vertical, with initial and lateral boundary conditions provided by the NASA Finite Volume General Circulation Model (FVGCM) [Atlas et al., 2005]. The two nested climate model simulations cover a historic period (1961–1990) and a future period (2071–2100) under the SRES A2 emissions scenario [IPCC, 2000]. Further details of the RegCM3 configuration and simulations are described in the study of Diffenbaugh et al. [2005], while the FVGCM simulations are also described in the study of Coppola and Giorgi [2005].

2.1.2. VIC

[8] The variable infiltration capacity (VIC) model is a distributed, physically based, static vegetation, macroscale surface water and energy balance hydrologic model [Cherkauer et al., 2003; Liang et al., 1994; Liang et al., 1996]. The VIC model uses a mosaic approach to statistically represent the subgrid-scale spatial variability in infiltration and vegetation/land cover. The VIC model has been used successfully to study the hydrological impacts of climate change in a variety of configurations ranging from global- to single-basin scale [e.g., Christensen et al., 2004; Nijssen et al., 1997; Nijssen et al., 2001; Vanrheenen et al., 2004]. Several studies, including those conducted by the Land Data Assimilation System team, describe the calibration and evaluation of the VIC model over the contiguous United States [e.g., Cosgrove et al., 2003; Pan et al., 2003; Schaake et al., 2004]. For this study, the VIC model is implemented for the contiguous United States at a spatial resolution of one-eighth degree latitude by one-eighth degree longitude. The simulations are run using the VIC model version 4.0.5 at a daily time step, with no iteration to close the final surface energy balance (land surface temperature equals the air temperature). The snow model is resolved at a 3 hourly time step. For all simulations, the VIC model use three soil layers with depths of 0.1, 0.3, and 1.0 m (from top to bottom). Data for soil hydraulic and thermal parameters are based on the study of Miller and White [1998] and Maurer et al. [2002], as described in the study of Mao and Cherkauer [2009].

2.2. Climate Model Bias Correction

[9] We use a quantile-based bias correction technique to remove the biases in the RegCM3-simulated precipitation, minimum temperature, and maximum temperature fields. The basic concept of our bias correction technique follows the study of Wood et al. [2002, 2004]. They used a bias correction-spatial-disaggregation method to bias-correct monthly climate model fields on a coarse-resolution climate model grid and then spatially disaggregated the corrected climate model fields to the high-resolution hydrologic model grid. They then generated daily fields by tuning observed daily values to match the bias-corrected monthly climate model fields for both the historic and future periods, assuming no change in the daily precipitation and temperature distribution in the future period. This method of bias correction preserves the monthly autocorrelation of the climate model time series.

[10] For our experiments, we have made some modifications to the method outlined by Wood et al. [2002, 2004]. First, because of the availability of the nested climate model data at substantially higher spatial resolution than the GCMs used by Wood et al., we apply bias correction directly at the scale of the hydrologic model grid. Second, after applying the correction on monthly climate model fields (as in the study of Wood et al. [2002, 2004]), we disaggregate the corrected monthly fields to the daily climate model fields (see details below). The daily distributions of temperature and precipitation have been found to be highly sensitive to enhanced greenhouse forcing [e.g., Diffenbaugh et al., 2005; Wehner, 2004], including the crossing of critical thresholds [e.g., Diffenbaugh et al., 2007a; Diffenbaugh et al., 2008; White et al., 2006]. Our use of model-simulated daily fields attempts to capture that sensitivity, along with the associated sensitivity of the hydrological response to future changes in daily precipitation and temperature extremes.

2.2.1. Historic Period

[11] For the bias correction of the historic period (1961–1990), we use observed precipitation, minimum temperature, and maximum temperature fields from the corresponding period in the parameter-elevation regressions on independent slopes model (PRISM) data, available on a monthly timescale at 4 km horizontal grid spacing [Daly et al., 2000]. Using the simple inverse distance weighting technique, we first interpolate both the RegCM3 and PRISM fields to the VIC one-eighth degree geographical grid. For each field, we construct observed and modeled quantiles at each grid point, using the 30 monthly mean values from each of the 12 months of the year (30 months of January in 1961–1990, 30 months of February in 1961–1990, etc). We then adjust the magnitude of each RegCM3 field for each month by mapping each quantile to the corresponding PRISM quantile (i.e., we set the maximum January RegCM3 value at a given grid point to the maximum January PRISM value at that grid point, and so on for each quantile in each month at each grid point). Finally, we distribute the monthly correction from each bias-corrected month to the model-simulated daily time series so that the daily distribution is maintained, but the aggregate is equal to the bias-corrected monthly mean, as described here

equation image

where Tc(m,dm) and Pc(m,dm) are bias-corrected daily temperature (minimum or maximum) and precipitation, whereas To(m,dm) and Po(m,dm) are original daily temperatures (minimum or maximum) and precipitation for month m (m = 1–360) and day d of month m. Similarly, Tc(m,ave), Pc(m,ave), To(m,ave) , and Po(m,ave) are bias corrected and original average (minimum or maximum) temperatures and total precipitation for month m (m = 1–360), respectively.

2.2.2. Future Period

[12] For the bias correction of the future period (2071–2100), we first interpolate the original RegCM3 A2 data to the VIC one-eighth degree geographical grid. We then calculate the change in the magnitude of each variable between the simulated historic and future periods for each quantile (“quantile shift”) in each of the respective 12 months of the year (30 historic months of January and 30 future months of January, 30 historic months of February and 30 future months of February, etc). We calculate the quantile shifts as a difference (future minus historic) for average minimum and maximum temperatures and as a ratio (future/present) for total precipitation. We then generate the bias-corrected future period average minimum and maximum temperature quantiles by adding the quantile shifts to the bias-corrected historic period average minimum temperature and maximum temperature quantiles for each of the respective months (i.e., we add the difference between the maximum January RegCM3 value in the historic period and the maximum January RegCM3 value in the future period to the maximum January value in the bias-corrected RegCM3 historic period and so on for each exceedance probability in each month at each grid point). Similarly, we generate the bias-corrected future period total precipitation quantiles by multiplying the quantile shifts with the bias-corrected historic period total precipitation quantiles for each of the respective month. We then adjust the magnitude of each RegCM3 field for each month in the future period by mapping each simulated quantile to the corresponding bias-corrected future period quantile.

[13] Using the monthly corrected RegCM3 data for the future period, we generate two sets of future period bias-corrected daily data sets. Following the method described in section 2.2.1 for the historic period, the first data set (hereafter RegCM-BC) is generated by translating the monthly correction from each bias-corrected future period month to the corresponding future period RegCM3-simulated daily time series of that month (i.e., the correction from the bias-corrected January 2071 mean is translated to the RegCM3-simulated daily time series of January 2071, the correction from the bias-corrected January 2072 mean is translated to the RegCM3-simulated daily time series of January 2072, and so on). Similarly, the second data set (hereafter RegCM-BCD) is generated by translating the monthly correction from each bias-corrected future period month to the corresponding historic period RegCM3-simulated daily time series of that month (i.e., the correction from the bias-corrected January 2071 mean is translated to the RegCM3-simulated daily time series of January 1961, the correction from the bias-corrected January 2072 mean is translated to the RegCM3-simulated daily time series of January 1962, and so on). Since the RegCM-BCD data set is prepared by adjusting the historic period daily values to the future period bias-corrected monthly means, comparison between RegCM-BC and RegCM-BCD allows us to test the influence of changes in the daily temperature and precipitation distributions induced by elevated greenhouse forcing.

2.3. Hydrologic Model Experiments

[14] Our main experiments consist of two pairs of RegCM3-VIC simulations, described in Table 1. One pair is forced with un-bias-corrected RegCM3 (RegCM3-ORG) fields (“VIC-ORG”), and one pair is forced with bias-corrected RegCM3 (RegCM3-BC) fields (“VIC-BC”). Each pair consists of two 30 year VIC model integrations, one in the historic period and one in the future period. The un-bias-corrected simulations use the RegCM3-ORG fields, interpolated to the VIC model grid and the bias-corrected simulations use the RegCM3-BC fields.

Table 1. Summary of Different RegCM3 Driving Fields and VIC Model Simulations
 Historic Period (1961–1990)Future Period (2071–2100)
Driving Data
RegCM3-ORGOriginal daily precipitation, minimum temperature, maximum temperature, surface windsOriginal daily precipitation, minimum temperature, maximum temperature, surface winds
RegCM3-BCBias-corrected daily precipitation, minimum temperature, maximum temperature, original surface windsBias-corrected daily precipitation, minimum temperature, maximum temperature, original surface winds (bias correction applied to the future period daily values)
RegCM3-BCDSame as RegCM3-BCBias-corrected daily precipitation, minimum temperature, maximum temperature, original surface winds (bias correction applied to the historic period daily values)
 
Experiments
VIC-ORGRegCM3-ORGRegCM3-ORG
VIC-BCRegCM3-BCRegCM3-BC
VIC-BCDRegCM3-BCRegCM3-BCD

[15] In order to test the sensitivity of simulated hydrological changes to changes in the daily distribution of temperature and precipitation, we generate an additional future period bias-corrected simulation by forcing the VIC model with RegCM3-BCD fields (VIC-BCD). In addition, in the absence of multidecadal, high-resolution hydrological observations covering the contiguous United States, we generate a control VIC model simulation by forcing the VIC model over the historic period (implemented as described in section 2.1.2) with the Maurer et al. [2002] daily observed driving fields (“VIC-OBS”). This control simulation allows us to evaluate the RegCM3-VIC results and to test the usefulness of the daily scale bias correction.

3. Results

3.1. Bias Correction

[16] At the seasonal scale, RegCM3 exhibits a number of inconsistencies relative to the PRISM observations. For instance, the primary precipitation inconsistencies include a winter (December-January-February; DJF) and spring (March-April-May; MAM) wet bias over the northwest, a summer (June-July-August; JJA) and autumn (September-October-November; SON) wet bias over the southeast, and an all-season dry bias over the central United States (Figures 2a2d). By definition, all of these monthly scale precipitation biases are corrected after the adjustment of the monthly quantiles of the RegCM3 precipitation field to the monthly quantiles of the corresponding PRISM precipitation field through quantile mapping (Figures 2e2h).

[17] Bias correction also substantially improves the daily scale magnitudes of the RegCM3 fields. Following Walker and Diffenbaugh [2009], we evaluate the effectiveness of daily scale bias correction by comparing 95th and 5th percentile of daily temperatures (T95 and T05) and 95th percentile of daily precipitation (P95) from the original and corrected RegCM3 data with the North American Regional Reanalysis (NARR). Here the comparison is based on the 12 overlapping years from 1979 to 1990 (Figure 3). In each overlapping year, we define T95 (T05) as the mean of the 95th (5th) percentile daily temperature maxima (daily temperature minima) at each grid point. Similarly, we define P95 at each grid point as the mean of the 95th percentile precipitation value for the days above 0.5 mm/d. In general, RegCM3-simulated percentiles of precipitation and temperature capture the basic patterns seen in the NARR. The primary inconsistencies relative to the NARR include a cold T05 bias across the United States, a warm T95 bias over the central United States, and a wet P95 bias along the Atlantic coast. All of these biases are substantially corrected after the adjustment of RegCM3 daily fields through temporal disaggregation of quantile-based monthly corrections (Figure 3).

[18] Further, using the Taylor diagram [Taylor, 2001], we quantify how well the VIC model simulations forced with the un-bias-corrected (VIC-ORG) and bias-corrected (VIC-BC) RegCM3 fields compare with the VIC model simulation forced with observational data (VIC-OBS). The Taylor diagram quantitatively compares the pattern correlation, the ratio of variance (ROV), and the root-mean-squared difference (RMSD) (Figure 4). The radial coordinates represent the ROV and RMSD: ROV as a radial distance from the reference arc (labeled with a red arc in Figure 4) and RMSD as the radial distance from the point of reference (labeled VIC-OBS in Figure 4). Similarly, the angular coordinate represents the pattern correlation, which measures the extent to which maxima and minima in the reference data and the test data occur at the similar location. Because of the regional and seasonal heterogeneity across the United States, we quantify the results for four seasons (DJF, MAM, JJA, and SON) and four regions (northwest, northeast, southwest, and southeast, as shown in Figure 1). Comparison of VIC-ORG with VIC-OBS (Figure 4a) shows high pattern correlation of soil moisture and evapotranspiration between the two simulations, with the magnitude of soil moisture generally higher and the magnitude of evapotranspiration generally lower in VIC-ORG for most regions and seasons. However, with poor pattern correlation, large RMSD and large ROV, VIC-ORG-simulated surface runoff and base flow show a large deviation from those simulated by VIC-OBS. The differences are particularly large over the southeast and the southwest in summer and autumn, over the northeast in spring and summer, and over the northwest in summer (Figure 4a).

Figure 1.

Division of contiguous United States into 18 hydrological units as defined by the USGS. The dotted lines represent the boundaries of the four regions used in the analysis.

[19] Bias correction shows a substantial improvement in the VIC-simulated hydrologic fields (Figure 4b). The pattern correlation between VIC-BC and VIC-OBS variables is very high for most regions and seasons, with relatively small RMSD and ROV. However, VIC-BC shows minimal improvement in simulating base flow and runoff over the northwest in summer (Figure 4b).

3.2. Future Climate Changes

[20] RegCM3 simulates a heterogeneous precipitation response across the United States for the future period (Figures 5a5d). The changes in seasonal mean precipitation are mostly positive in winter and spring for most U.S. Geological Survey (USGS) hydrological units (Figure 1), except over the California unit in winter, and parts of the Pacific Northwest and California units in spring (Figures 5a5b). Similarly, positive changes are simulated in autumn across the United States, with the exception of parts of the South Atlantic-Gulf and Pacific Northwest units (Figure 5d). Summer precipitation changes are mostly positive except over parts of the Rio Grande, Lower Mississippi, Ohio, and Missouri units (Figure 5c).

[21] RegCM3 simulates an increase of 2°–7° in the mean seasonal maximum and minimum temperatures over the contiguous United States (Figure 5). This increase in the mean seasonal temperatures is most pronounced in winter over the Great Lakes, Ohio, New England, and Upper Mississippi units (Figures 5e and 5i), in summer and autumn over most of the United States (Figures 5g5h and 5k5l), and in spring over the Great Basin and Upper Colorado units (Figures 5f and 5j).

3.3. Future Hydrologic Changes

[22] By comparing the simulated changes in VIC-ORG, VIC-BC, and VIC-BCD, we can test the sensitivity of simulated hydrological changes to climate model biases and to changes in the daily temperature and precipitation distributions. The comparison is shown for soil saturation, evapotranspiration, runoff, and base flow (Figures 69). This comparison reveals that the hydrologic changes simulated by VIC-ORG and VIC-BC are substantially different. Using the USGS hydrological units (Figure 1), the dissimilarities are particularly large over the higher elevations of the Pacific Northwest and Missouri units in winter, spring, and summer, where VIC-ORG simulates changes in soil moisture, surface runoff, and base flow (Figures 6a6c, 8a8c, 9a9c) that are several times higher than those simulated in VIC-BC (Figures 6e6g, 8e8g, 9e9g). Substantial differences also occur over parts of the California, Great Basin, and Lower Colorado units, where VIC-ORG simulates relatively muted changes in surface runoff and base flow, and relatively smaller soil moisture changes in summer and autumn (Figures 6c6d, 8c8d, 9c9d). Similarly, VIC-ORG does not show most of the VIC-BC-simulated winter, spring, and autumn surface runoff changes over the Lower Mississippi unit (Figures 8a8d, 8e8h).

[23] In addition to the differences in magnitude of hydrologic change, VIC-ORG and VIC-BC simulate hydrologic changes that are opposite in sign over some areas of the United States. For instance, compared to a generally muted or positive response in VIC-BC over the South Atlantic-Gulf unit, VIC-ORG mostly shows a decrease in summer and autumn base flow (Figures 9c9d, 9g9h). Similarly, in summer in the VIC-ORG experiment, parts of the Missouri and Arkansas-White-Red units exhibit negative changes in evapotranspiration and positive changes in soil moisture, opposite in sign to the changes simulated in VIC-BC (Figures 6c, 6g, 7c, 7g).

[24] In general, the bias-corrected simulations with and without future changes in daily temperature and precipitation distribution show very similar patterns of change for all hydrologic variables (Figures 69). However, there are a few exceptions where VIC-BCD shows noticeable differences in simulating future hydrological change. For instance, the magnitude of simulated increase in summer evapotranspiration is much lower in VIC-BCD than that in VIC-BC over parts of the South Atlantic-Gulf and Lower Mississippi units, and over parts of the Great Basin and Lower Colorado units (Figures 7g and 7k). Similarly, VIC-BC-simulated increases in summer surface runoff over the South Atlantic-Gulf and Lower Mississippi units are absent in VIC-BCD (Figures 8g and 8k).

4. Discussion

4.1. Effect of Bias Correction

[25] It is well established that biases in climate model simulations affect hydrologic simulations and that bias correction is necessary for meaningful translation of climate signals to the hydrological scales [e.g., Fowler et al., 2007; Wood et al., 2004]. However, the effects of climate model biases on the simulated hydrologic response to future changes in precipitation and temperature are not well documented, particularly for high-resolution climate models. In a nested climate model, biases arise because of the combination of deficiencies in the nested model and the inheritance of biases from the driving GCM [e.g., Denis et al., 2002]. For instance, in our RegCM3 historic integration, wet and cold biases in winter and spring over the northwestern United States and the Rocky Mountains arise from excess snow in RegCM3 during winter and its persistence later in the spring [Diffenbaugh et al., 2006]. Likewise, the summer warm bias over the central United States is at least partially inherited from the driving FVGCM simulation, which exhibits a similar pattern of temperature bias (not shown). Such biases can propagate through the hydrologic coupling to create dissimilarities between the VIC-ORG and VIC-BC simulations. For instance, underestimation of evapotranspiration in most regions and seasons in VIC-ORG (Figure 4a) can be attributed to the fact that RegCM3 simulates below-normal daily minimum temperatures across the United States (Figure 3c). Similarly, overestimation of soil moisture, base flow, and surface runoff in the southeastern United States in VIC-ORG can be attributed to the fact that RegCM3 simulates above-normal summer and autumn precipitation in that region (Figures 2c2d, 4a). An indirect effect of these biases is also visible in the northeastern, northwestern, and southwestern United States, where wet and cold biases in RegCM3 (Figures 2a2b, 3c) affect the amount of snow cover in VIC-ORG (Figure 10), which subsequently leads to higher snowmelt-driven surface runoff and base flow (Figures 4a).

Figure 2.

RegCM3-simulated seasonal mean precipitation mean for the historic period (mm/d): (a, e) DJF, (b, f) MAM, (c, g) JJA, and (d, h) SON. Top row shows original RegCM3 fields. Bottom row shows bias-corrected RegCM3 fields.

Figure 3.

Magnitudes of 95th and 5th percentile of daily temperatures (T95 and T05) in Kelvin and 95th percentile of daily precipitation (P95) in millimeters per day. Top row shows results from North American Regional Reanalysis (NARR) fields. Middle row shows the results from RegCM3 bias-corrected fields. Bottom row shows the results from original RegCM3 fields.

Figure 4.

Quantitative comparison of VIC-ORG and VIC-BC-simulated soil saturation, evapotranspiration, runoff, and base flow with VIC-OBS. (a) VIC-OBS versus VIC-ORG and (b) VIC-OBS versus VIC-BC. The radial coordinates represent the ratio of variance (ROV) and root mean squared difference (RMSD): ROV as a radial distance from the reference arc (labeled with a red color) and RMSD as the radial distance from the point of reference (labeled VIC-OBS). The angular coordinate represents the pattern correlation. Four symbols represent four regions, four colors represent four seasons, and four numbers represent four variables. All values greater than 2.0 are set equal to 2.0.

[26] The differences between the simulated hydrological changes in the VIC-BC and VIC-ORG simulations demonstrate that systematic climate model biases can impact the net changes in hydrologic variables simulated in response to elevated greenhouse gas concentrations (Figures 69). These differences can be quantified in a Taylor diagram comparing changes in hydrologic variables simulated by VIC-ORG to the changes simulated by VIC-BC (Figure 11a). If the two simulations simulate statistically similar hydrologic change, their net changes in hydrologic variables should exhibit high pattern correlations and small ROV and RMSD. However, the Taylor diagram suggests that the two simulations simulate statistically different hydrological change, with the exception of a few variables in a few seasons (Figure 11a). It should also be noted that, by definition, the two data sets used to drive VIC-BC and VIC-ORG are similar in their monthly scale future changes but different in the magnitude of their daily scale quantiles. Because bias correction in the future period is applied with respect to the biases in the historic period, by definition, the bias correction does not affect the simulated monthly scale changes in RegCM3 fields. Given the (designed) agreement in monthly scale precipitation and temperature changes in the VIC-BC and VIC-ORG driving data sets, the disagreement between VIC-BC and VIC-ORG in simulating future hydrological change suggests that hydrological changes are driven not only by the monthly changes in meteorological fields but also by the absolute magnitude of their daily values.

[27] Differences in the spatial heterogeneity of the original and bias-corrected RegCM3 data also contribute to the differences between VIC-ORG and VIC-BC in simulating hydrological changes over the higher elevations. For example, VIC-ORG simulates excessive snow water equivalent over the higher elevations (Figures 10a and 10b). This bias affects both the magnitude of VIC-ORG-simulated hydrological variables (Figure 4a) and their response to the future-period changes in precipitation and temperature (Figures 6, 8, 9). While the VIC-ORG biases in simulating snow water equivalent primarily result from the errors in the original RegCM3 temperature and precipitation data, these RegCM3 errors arise partly because the high-resolution topography in RegCM3 is still poorly resolved relative to the actual topography. The spatial heterogeneity of VIC-simulated snow water equivalent is largely improved by correcting the RegCM3 temperature and precipitation fields to the more highly resolved PRISM data (Figure 10).

4.2. Effect of Changes in Daily Scale Extremes

[28] One of the strongest arguments in favor of using daily scale climate projections in process-based hydrological impact assessment is the sensitivity of hydrological processes to changes in climate extremes. Increases in the occurrence of precipitation in the form of high-intensity events can increase surface runoff production and reduce water infiltration, particularly in arid environments or other regions with soils with low infiltration capacity [Horton, 1945]. Similarly, increases in the occurrence of persistent daily temperature extremes can accelerate the rate of snow retreat and evaporative losses from the surface through changes in the daily temperature range, leading to early snowmelt in the cold season and strong surface drying in the warm season. Combined, these effects can support accelerated evaporation, higher surface runoff, and reduced soil moisture and subsurface recharge.

[29] Of particular interest in our simulations, heavy precipitation events appear to drive surface runoff changes over grid points where seasonal precipitation changes are negative but seasonal surface runoff changes are positive (Figures 5a5d, 8e8h). We examine these grid points by comparing the magnitude of surface runoff to the intensity of wet precipitation events (as represented by events exceeding the 75th percentile of precipitation (P75)) (Figure 12). The comparison is shown separately for each season and for each region as an average of all grid points where the surface runoff response is positive despite a net negative precipitation change (one symbol in each frame in Figure 12 represents the regional average of one season in a particular year). In general, surface runoff increases with the increase in the intensity of extreme precipitation events, except over the regions where runoff is heavily influenced by snowmelt such as in spring over the northwest (Figure 12c). The comparison of the historic and future periods indicates that in all regions and all seasons, both heavy precipitation events and high surface runoff events increase in occurrence and intensity in the future period. This comparison suggests that the increases in runoff over these regions are in part driven by increases in heavy precipitation events, which overcome decreases in seasonal precipitation.

Figure 5.

RegCM3-simulated seasonal mean changes (future minus historic) in daily precipitation (mm/d), daily maximum temperature (Kelvin), and daily minimum temperature (Kelvin): (a, e, i) DJF, (b, f, j) MAM, (c, g, k) JJA, and (d, h, l) SON. Top row shows changes in precipitation. Middle row shows changes in maximum temperature. Bottom row shows changes in minimum temperature.

[30] Similarly, many grid points in VIC-BC show a decrease in soil moisture where the seasonal precipitation increases. Considering only these grid points, we compare the changes in soil moisture (top two layers) with changes in the intensity of heavy precipitation events (Figure 13a). We again examine days above the 75th percentile of precipitation (P75). The increase in the magnitude of events above P75 suggests that increasing seasonal precipitation over these grid points is partly driven by the increase in the precipitation events above the 75th percentile (Figure 13a). Also, we find an inverse relationship between the changes in soil moisture and the changes in the intensity of these heavy precipitation events.

[31] We also find that the changes in soil moisture are sensitive to the occurrence of days above the 95th percentile of daily temperature maxima (T95) (Figure 13b). Considering the grid points where VIC-BC shows a decrease in soil moisture along with an increase in seasonal precipitation, we compare changes in soil moisture (top two layers) with changes in the number of T95 hot days. We find that substantial increases in extremely hot days in the future period are associated with large moisture deficits in soils (Figure 13b). Extremely hot conditions are often associated with subsidence of air masses and clear sky conditions. Under such conditions, the evapotranspiration rate can easily exceed the rainfall rate, leading to strong moisture depletion in soils.

4.3. Effect of Changes in Daily Distribution

[32] The importance of heavy precipitation and hot temperature events in the simulated hydrologic response raises the question of whether that response is driven by changes in the daily distribution of precipitation and temperature, or simply by changes in mean monthly values that create a uniform shift in the daily distribution. We do not find substantial differences between the VIC-simulated hydrological changes with and without changes in the daily distributions of temperature and precipitation (Figures 69, 10). In fact, the statistical comparison of changes in hydrologic variables simulated by VIC-BCD with those simulated by VIC-BC shows that the two simulations are very similar in simulating the hydrological change at regional scale (Figure 11b).

Figure 6.

VIC-simulated changes (future minus historic) in the seasonal mean soil saturation (%): (a, e, i) DJF, (b, f, j) MAM, (c, g, k) JJA, and (d, h, l) SON. Top row shows VIC-ORG-simulated changes. Middle row shows VIC-BC-simulated changes. Bottom row shows VIC-BCD-simulated changes.

Figure 7.

Same as in Figure 6 but for the seasonal evapotranspiration means (mm/d).

Figure 8.

Same as in Figure 6 but for the seasonal surface runoff means (mm/d).

Figure 9.

Same as in Figure 6 but for the seasonal base flow means (mm/d).

Figure 10.

VIC-simulated seasonal mean snow water equivalent (mm/d): (a, b) VIC-ORG and (e, f) VIC-BC. First column represents DJF. Second column represents MAM. VIC-simulated changes (future minus historic) in seasonal mean snow water equivalent (mm/d): (c, d), VIC-ORG and (g, h) VIC-BC. Third column represents DJF. Fourth column represents MAM.

Figure 11.

Quantitative comparison of VIC-ORG and VIC-BCD-simulated changes in soil saturation, evapotranspiration, runoff, and base flow with VIC-BC-simulated changes. (a) VIC-BC versus VIC-ORG and (b) VIC-BC versus VIC-BCD. The radial coordinates represent the ratio of variance (ROV) and root-mean-squared difference (RMSD): ROV as a radial distance from the reference arc (labeled with a red color) and RMSD as the radial distance from the point of reference (labeled VIC-BC). The angular coordinate represents the pattern correlation. Four symbols represent four regions, four colors represent four seasons, and four numbers represent four variables. All values greater than 2.0 are set equal to 2.0.

[33] This insensitivity to changes in daily distribution can be partly attributed to the fact that the data sets used to drive the VIC-BC and VIC-BCD future simulations (RegCM-BC and RegCM-BCD) show little difference in their daily precipitation quantiles at regional scale, particularly when compared with the difference in quantiles between the historic and future RegCM-BC periods (Figure 14). The major exception is summer in the southeastern United States, where VIC-BC and VIC-BCD simulate statistically different changes in surface runoff (Figure 11b), and the future period precipitation in RegCM-BC shows greater exceedance of the historic 75th percentile threshold than the future period precipitation in RegCM-BCD (Figure 15).

5. Conclusions

[34] We use a hydrological model (the VIC model) driven by a high-resolution nested climate model (RegCM3) to investigate the potential impacts of climate model biases and changes in the daily temperature and precipitation distributions on hydrology in the United States. We note a number of caveats to our study. There exists a large matrix of regional climate change uncertainty [Giorgi et al., 2008], but we considered only one climate model realization and one GCM-RCM combination. Moreover, our validation methodology did not involve the use of actual hydrological observations and did not account for uncertainties originating from the hydrologic model. Further, indirect bias correction of the daily climate model data could have introduced errors, which requires further exploration using daily temperature and precipitation observations. Likewise, the VIC model does not include direct CO2 effects on vegetation (such as changes in water-use efficiency), which can be significant for hydrology in some regions and seasons [Gedney et al., 2006]. Finally, we did not apply bias correction to the RegCM3-simulated winds, which could influence the hydrologic simulation.

[35] Despite these caveats, our results indicate that biases in the climate model fields not only affect the hydrologic model results in the historic period (Figure 4a), but they also lead to a statistically different simulated hydrological response to future increases in greenhouse forcing (Figure 11a). Moreover, we find that the future hydrological response is dictated not only by the monthly scale precipitation and temperature changes but also by increases in the occurrence of their daily extremes, indicating that fine temporal scales are potentially important for the hydrological response to elevated greenhouse gas concentrations (Figures 12 and 13, 14). Furthermore, in our results, future changes in the daily temperature and precipitation distributions contribute little to the regional-scale response of hydrological processes to elevated greenhouse forcing (Figure 11b). However, over some hydrological units, changes in the daily distributions seem to be important when they affect changes in the wet tail of the precipitation distribution (Figures 7g, 7k, 8g, 8k, 15c). Despite the limitations of our methodology, our results provide important insight into the sensitivity of hydrologic projections to climate model biases, as well as the potential role of changes in temperature and precipitation extremes in influencing the hydrologic response to anthropogenic greenhouse forcing.

Figure 12.

Seasonal surface runoff means (mm/d) versus magnitude of events (mm/d) greater than 75th percentile of the historic period daily precipitation (P75): (a) southeast, (b) southwest, (c) northwest, and (d) northeast. Four symbols represent four seasons in the historic period (blue) and the future period (red). Each symbol represents average of all those grid points where seasonal precipitation changes are negative but seasonal surface runoff changes are positive.

Figure 13.

(a) Seasonal changes (future minus historic) in mean soil moisture in the top two layers (mm) versus seasonal changes in the magnitude of events (mm/d) greater than 75th percentile of the historic period daily precipitation (P75). (b) Seasonal changes in the mean soil moisture in the top two layers (mm) versus seasonal changes in the number of events (days/season) greater than 95th percentile of the historic period daily maximum temperature (T95). Four colors represent four seasons. Each symbol represents average of all those grid points where seasonal precipitation changes are positive but seasonal soil moisture changes in the top two layers are negative.

Figure 14.

(a) Magnitude of daily precipitation quantiles (mm/d) in RegCM3-BC historic period (red) and RegCM3-BC future period (bias corrected future period RegCM3 with changes in daily distribution) (blue). (b) Magnitude of daily precipitation quantiles (mm/d) in RegCM3-BCD (bias-corrected future period RegCM3 without changes in daily distribution) (green) and RegCM3-BC (bias-corrected future period RegCM3 with changes in daily distribution) (blue).

Figure 15.

Difference (RegCM3-BCD minus RegCM3-BC) in number of events (days/season) above 75th percentile of the historic period daily precipitation (P75) with and without future changes in daily distribution: (a) DJF, (b) MAM, (c) JJA, and (d) SON.

Acknowledgments

[36] We thank two anonymous reviewers for their constructive and insightful comments. This work was supported in part by National Science Foundation awards 0315677 and 0450221 and Department of Energy awards DE-FG02-08ER64649 and DE-SC0001483. The model simulations and analyses were enabled by computational resources provided by Information Technology at Purdue (the Rosen Center for Advanced Computing, West Lafayette, IN). This is Purdue Climate Change Research Center paper number 0923.

Ancillary