Constrained dynamical downscaling for assessment of climate impacts

Authors

  • M. Harkey,

    Corresponding author
    1. Nelson Institute Center for Sustainability and the Global Environment (SAGE), University of Wisconsin-Madison, Madison, Wisconsin, USA
    • Corresponding author: M. Harkey, Nelson Institute Center for Sustainability and the Global Environment (SAGE), University of Wisconsin-Madison, Madison, WI, USA. (mkharkey@wisc.edu)

    Search for more papers by this author
  • T. Holloway

    1. Nelson Institute Center for Sustainability and the Global Environment (SAGE), University of Wisconsin-Madison, Madison, Wisconsin, USA
    Search for more papers by this author

Abstract

[1] To assess climate change impacts on hydrology, conservation biology, and air quality, impact studies typically require future climate data with spatial resolution high enough to resolve urban-rural gradients, complex topography, and sub-synoptic atmospheric phenomena. We present here an approach to dynamical downscaling using analysis nudging, where the entire domain is constrained to coarser-resolution parent data. Here meteorology from the North American Regional Reanalysis and the North American Regional Climate Change Assessment Program data archive are used as parent data and downscaled with the Advanced Research version of the Weather Research and Forecasting model to a 12 km × 12 km horizontal resolution over the Eastern U.S. Our results show when analysis nudging is applied to all variables at all levels, mean fractional errors relative to parent data are less than 2% for maximum 2 m temperatures, less than 15% for minimum 2 m temperatures, and less than 18% for10 m wind speeds. However, the skill of representing fields that are not nudged, such as boundary layer height and precipitation, is less clear. Our results indicate that though nudging can be a useful tool for consistent, comparable studies of downscaling climate for regional and local impacts, which variables are nudged and at what levels should be carefully considered based on the climate impact(s) of study.

1 Introduction

[2] To assess climate change impacts on air quality, food production, water availability, and other sectors, future climate information is often needed at finer scales than provided by global general circulation models (GCMs). Though the resolution of GCMs participating in the upcoming fifth report of the Intergovernmental Panel on Climate Change (IPCC) ranges from 2.8° × 2.8° to 0.5° × 0.5° [e.g. Vavrus et al., 2011; Zhao et al., 2009], even this increased resolution cannot resolve sub-synoptic scale phenomena such as thunderstorms and the effects of urban areas, especially in areas of complex topography [Giorgi, 2006a]. Regional scale climate processes (those with a horizontal resolution of 10–50 km) have been shown to have an important impact on air quality [e.g., Lin et al., 2010], conservation biology [e.g., Wiens and Bachelet, 2010], hydrology [e.g., Teutschbein et al., 2011], and precipitation [e.g., Li et al., 2011]. Thus, regional scale climate data are an important input into many analyses of future climate impacts.

[3] Over the last two decades, regional and local climate impacts have been explored using statistical downscaling from GCMs [e.g., Holloway et al., 2008; Yoshimura and Kanamitsu, 2008; Norton et al., 2011; Notaro et al., 2011; Seidou et al., 2012], dynamical downscaling with regional climate models (RCMs) [e.g., Hogrefe et al., 2004; Tagaris et al., 2007; Nolte et al., 2008; Chen et al., 2009; Liao et al., 2009], and the delta-change approach [e.g., Hay et al., 2000; Diaz-Nieto and Wilby, 2005; Tryhorn and De Gaetano, 2011; Litschert et al., 2012; Olsson et al., 2012]. The delta-change method, widely applied in hydrologic studies, assumes that GCMs credibly model relative changes to climate, and those changes are imposed on local time series of climate variables. Statistical downscaling isolates empirical relationships between large-scale features and localized variables—GCM predictors and the finer-scale predictands. Once these spatial and temporal relationships are determined using past climate data, they are applied to future GCM conditions to create local and regional future conditions. The technique has the advantages of being computationally inexpensive and easily tailored to a specific region [Wilby et al., 2004]. The value added by statistical downscaling depends on the choice of training period, domain, and predictor variables [Leung et al., 2003]. Dynamical downscaling, by comparison, is physically based: An RCM is driven by a parent GCM, with the GCM providing boundary conditions and RCMs resolving “regionalized” climate [Giorgi, 2006a]. Dynamical downscaling affords higher resolution of topography and coastlines, which results in a more accurate representation of the hydrologic cycle compared to GCMs [Leung et al., 2003; Giorgi, 2006a]. As with statistical downscaling, skill with dynamical downscaling is sensitive to the parent GCM [Racherla et al., 2012].

[4] Here we examine how varying means of constraining climate simulations to the lower-resolution “parent” data perform in preserving the benefits of traditional dynamical downscaling while maintaining consistency with a chosen climatic starting point (e.g., parent GCM, or other—lower resolution—RCM simulation). This approach employs nudging, or a relaxing of variables within the entire RCM domain toward the parent data set rather than only at the lateral boundaries. We focus on downscaling with nudging as a means of reducing additional uncertainty caused by the choice of downscaling model and the parameterizations within it, and of improving the temporal and spatial resolution of the parent data, rather than a means of simulating a more accurate climate, which may not be possible [e.g., Chase et al, 2003; Lynn et al., 2009b; Pielke and Wilby, 2012]. A nudging approach has been widely used for simulations of historic climate [e.g., Tang et al., 2010; Heikkilä et al., 2011; Song et al., 2011]. There are different means of nudging: analysis (or grid) nudging, with nudging towards gridded analyses [e.g., Stauffer et al., 1991]; observational nudging, nudging towards observations; and spectral nudging, in which nudging is applied to large-scale features within the spectral domain [e.g., von Storch et al., 2000; Miguez-Macho et al., 2004].

[5] Miguez-Macho et al. [2004] highlight the value and skill of constraining an RCM to global re-analyses when spectral nudging is applied to long waves over a large domain, reducing the potential error sources caused by varying domain positions. Using the Weather Research and Forecasting (WRF) model, Lo et al. [2008] note that continuous analysis nudging shows the most skill compared to year-long simulations with varying initialization frequencies and maintains the consistency of downscaled parameters with large-scale features. In their study of value added by nudging in modeling tropical cyclones in the Pacific Basin, Cha et al. [2011] found intermittent spectral nudging 20% more accurate than continual nudging for convective systems, as continual nudging has a suppressive effect on meso-scale phenomenon. Still, they note the value added by nudging that results from a reduction of model uncertainty, and that intermittent spectral nudging is only one out of many factors that may potentially improve the technique.

[6] While both nudging techniques show skill, they have only recently been compared to one another. Using the WRF model, Liu et al. [2012] found spectral nudging of horizontal winds at all levels, and temperature and geopotential height only above the boundary layer retains “large-scale” features and allows for finer-scale variability, with improved representation of temperature, kinetic energy, and precipitation compared to analysis nudging of horizontal winds at all levels, and temperature and water vapor mixing ratio above the boundary layer. Similarly, Bowden et al. [2012] noted that the differences in performance of analysis and spectral nudging in WRF are small for temperature, horizontal wind, and 500 hPa geopotential height and that analysis nudging suppresses small-scale variability compared to spectral nudging. However, they found analysis nudging better represents precipitation intensity and frequency. The performance of each technique may depend on other factors, such as which parameters are nudged and where they are nudged.

[7] Whether nudging is appropriate for downscaling future climate is an area of ongoing study. Castro et al. [2005] found application of interior nudging in the Regional Atmospheric Modeling System (RAMS) reduced errors in 500 hPa geopotential height but describe how nudging reduces small-scale variability, which weakens meso-scale features and increases errors in precipitation amounts. In a multi-decadal study using the WRF model, Otte et al. [2012] found that reasonable application of either nudging method not only decreased errors in 2 m average and extreme temperatures but also more accurately simulated extreme precipitation. However, dynamical downscaling with nudging may not reduce errors in the parent data set, as the ability of the downscaling RCM to freely simulate local and regional processes will be limited [Rockel et al., 2008], and the downscaled simulations are highly dependant on the parent data set and downscaling model configuration [e.g., Giorgi, 2006b; Lynn et al., 2004; Lynn et al., 2009a; Lynn et al., 2009b; Pielke et al., 2012; Racherla et al., 2012].

[8] However, for climate impact studies, the limitations of dynamical downscaling with nudging may be key to study design: whereas historic simulations may be compared to observations, future climate simulations must be compared to other model estimates, and inter-study comparisons will be hindered by potentially divergent downscaled simulations. Though nudging has been applied to and shows skill in historic simulations of air quality [e.g., Kim et al., 2010; Maurizi et al., 2012; Otte, 2008], to date no published simulations of impacts of future climate change have used nudging. We hereafter refer to constrained dynamical downscaling (CDD) to address the general approach of nudging a future climate simulation toward a parent data set. This study analyzes historic and future climate simulations using a CDD approach to evaluate which analysis nudging options produce a downscaled climate most similar to the parent fields. These high-resolution results will be used as input to summertime air quality simulations over the Eastern United States (Harkey et al., in preparation) using the U.S. Environmental Protection Agency (EPA) Community Multiscale Air Quality (CMAQ) model, which is commonly used with WRF meteorology as input [e.g., Holloway et al., 2012; Im et al., 2011; Lin et al., 2010; Yu et al., 2012].

[9] We conducted our CDD analysis of future climate using the Advanced Research version of the Weather Research and Forecasting (WRF-ARW) model, hereafter referred to as WRF. Our goal is to select a parent climate projection appropriate to our air quality study goals and employ WRF as a translational tool to generate climate data with both the temporal resolution needed by CMAQ (which requires hourly values of more than 50 meteorological parameters) and the horizontal resolution of 12 km × 12 km needed to capture urban-rural gradients. The paper is organized as follows: Section 2 details our selection of parent data and WRF setup. In section 3, we compare the CDD results to a traditional dynamical downscaling driven only by input from the lateral boundaries. The effects of varying CDD parameters are explored in section 4, and a discussion is in section 5.

2 Choice of Constraint Data

2.1 Historic Summer Data

[10] For historic summertime simulations, we have employed the North American Regional Reanalysis (NARR) data set [Mesinger et al., 2006]. The NARR data are on a 32 km × 32 km grid, output every 3 h, available from January 1979 to the present [Mesinger et al., 2006]. NARR data provide a dynamically consistent meteorology based on observations, with no discontinuities or gaps in data as might exist in an observational data set. Comparing soundings averaged over January 1999–December 2001, Kennedy et al. [2011] found that from the surface to about 300 hPa, NARR winds are within 0.5 m/s of sounding data at the Atmospheric Radiation Measurement Program (ARM) Southern Great Plains (SGP) site, temperatures are generally within 0.5 K, and relative humidity within 5%. They note a slight dry bias in the boundary layer (<3%) and a humid bias near the tropopause (up to 15%). NARR data have been shown to have particular skill with precipitation, representing organized convection and extreme events well [Bukovsky and Karoly, 2007].

[11] Here we focus on summers of 2006 and 2007, which correspond to years of interest in our air quality modeling study (Harkey et al., in preparation). The months of June, July, and August of 2006 were slightly warmer than average, with temperatures above normal across much of the Eastern U.S. and temperatures much above normal in the mid-Atlantic region, along the coast of Virginia northward to Massachusetts. Precipitation over this summer was above average in the Northeast and mid-Atlantic states and below average in the Southeast. The summer of 2007 was also warmer than average but drier than the summer of 2006: temperatures were near normal in the Northeast, and above normal in the Southeast and Midwestern states. Precipitation was also near normal in the Northeast, and below normal in the Midwest and Southeast, and the south-central U.S. was abnormally cool and wet [http://www.ncdc.noaa.gov/temp-and-precip/maps.php].

2.2 Future Summer Data

[12] For future summers, we leverage the breadth of model simulations archived through the North American Climate Change Assessment Program [NARCCAP, Mearns et al., 2012]. The NARCCAP archive includes results from six RCMs, used to dynamically downscale projections from five GCMs. The NARCCAP domain covers most of North America at a 50 km × 50 km horizontal resolution and 3 h temporal resolution for years 1968–2000 and 2038–2070, including model spin-up for both time periods. All model simulations in the NARCCAP archive assume the A2 emissions scenario from the IPCC Special Report on Emissions Scenarios, characterized by a growing population (10 billion by 2050), an emphasis on regional economies, and slow but steady technological development [Nakicenovic et al., 2000]. Although data archived as part of NARCCAP are higher resolution RCM data, they are not available at a fine enough spatial or temporal resolution to directly support an air quality analysis, for which we require gridded 12 km × 12 km fields with meteorological data provided every hour. Here then, NARCCAP data are taken as the parent data for simulations of the future, although other global data sets could also be used.

[13] The NARCCAP archive allows us to compare multiple GCM-RCM pairings, as well as 30 years representing mid-century climate, 2041–2071. Although any GCM-RCM pairing from NARCCAP could be used as input to WRF, because air quality modeling is computationally expensive, we cannot simulate all future years in the ensemble. However, knowing the distribution of climatic characteristics—historic and future means and extremes, spatial patterns, and a wide range of meteorological variables—allows us to select a limited number of model years for further analysis. Based on analysis of historic NARCCAP RCM-GCM performance against NARR (supporting information Text1, Table S1, and Figure S1), we have selected the climate simulated by the National Center for Atmospheric Research (NCAR) Community Climate System Model (CCSM) [Collins et al., 2006] as downscaled by the Weather Research and Forecasting model (WRF) [Knievel et al., 2007] for further study. Here we focus on the summer of 2069, the warmest future summer simulated by the WRF-CCSM, as it provides a potential upper bound of risk to air quality.

3 Experiment Design and Methodology

[14] We have chosen the WRF model, version 3.2 [Skamarock et al., 2008], to downscale both the NARR and NARCCAP/WRF-CCSM data. The WRF model is a non-hydrostatic, meso-scale model with a terrain-following vertical coordinate system. For historic simulations with NARR, we have employed two horizontal grids on a Lambert conformal conic projection: the outer with 36 km × 36 km resolution, 137 × 173 grid points; the inner grid with 12 km × 12 km resolution, 190 × 229 grid points. Future simulations with NARCCAP differ only with the size of the outer domain, which is 105 × 145 grid points. Domains in all experiments have 27 vertical coordinate levels from the surface up to 150 hPa, as the WRF requires a model top lower than the parent data (100 hPa from NARCCAP). For consistency, we have employed the same model top in our downscaling experiments with NARR data. The model time step is 3 min on the 36 km domain, and meteorological variables are output every hour. The main physics options used in our downscaling experiments, as well as those used by the models generating the NARR and WRF-CCSM data sets, are given in Table 1.

Table 1. Physics Settings Used to in Parent Data Set Generationa and by This Study
 NARRWRF Driven by CCSMWRF (This Study)
  1. a

    NARR data generated using the operational NCEP Eta model [Mesinger et al., 2006].

CumulusBetts-Miller-Janjic [Janjic, 1994]Grell-Devenyi [Grell and Devenyi, 2002]Kain-Fritsch [Kain and Fritsch, 1990; Kain, 2004]
Land SurfaceNOAH [Chen and Dudhia, 2001]NOAHNOAH
MicrophysicsZhao et al. [1997]Mixed phase with water, ice, snow, and rain (WSM5) [Hong et al., 2004]Lin et al. [1983]
Planetary Boundary Layer (PBL)Mellor-Yamada [Chen et al., 1997]Yonsei University (YSU) [Hong et al., 2006]YSU
Surface LayerPaulson [Chen et al., 1997]Monin-Obukhov [Monin and Obukhov, 1954]Monin-Obukhov

[15] Results below show July of 2006, 2007, and 2069 as a case study month to test the sensitivity to whether, where, and which variables are nudged. For each month, WRF was allowed to spin-up for 4 days, starting on June 27 and running through July 31. The sensitivity to nudging over July is similar to nudging over the entire JJA period for each year (supporting information Text2, Table S2, and Figures S2 and S3). In the first experiment, all variables available to be nudged (potential temperature, water vapor mixing ratio, and the horizontal wind components) are nudged at all levels and at all time steps, all with a nudging coefficient of 3 × 10−4 s−1 on both domains, which describes the relative magnitude of nudging compared to the other model forcing terms, such as the Coriolis effect and advection [Stauffer and Seaman, 1990]. This corresponds to the default analysis nudging coefficient settings in the WRF documentation [http://www.mmm.ucar.edu/wrf/users/docs/user_guide_V3/contents.html], as well as employed by Bowden et al. [2012], Lo et al. [2008], Racherla et al. [2012], Stauffer and Seaman [1990], and Stauffer et al. [1991]. In section 4.1, we investigate departures from parent data in the experiment with nudging all variables, as Choi et al., [2009] have done, and we are able to gauge the bias introduced by our choice of model and model parameterizations, providing context for results of CDD of other scenarios.

[16] A priori knowledge of model strengths and biases may warrant differing CDD configurations: parent data may be too coarse to represent complicated near-surface conditions, or the downscaling model may show high skill in the boundary layer. In such cases, it is preferable to allow the downscaling model to resolve the land-atmosphere coupling independently and in a manner consistent with finer-scale topography, nudging only above the boundary layer [Bowden et al., 2012, Liu et al., 2012; Lo et al., 2008; Miguez-Macho et al., 2004]. Miguez-Macho et al. [2004] recommend spectral nudging to be applied to temperature and horizontal winds above the boundary layer in order to accurately represent precipitation when the model domain is at least a few thousand kilometers across. However, they did not compare these results with the skill of analysis nudging. Lo et al. [2008] found little differences between simulations with analysis nudging in WRF at all vertical levels and only above the PBL, except for precipitation amount and areal extent, where nudging above the PBL performed with higher skill. This result was hypothesized as caused by differences in topographic resolution between the parent data set and downscaling model. Though Bowden et al. [2012] found analysis nudging above the PBL effective in reducing errors in 2 m temperature, 850 hPa horizontal winds, 500 hPa geopotential height, and precipitation amounts, they note a need for sensitivity studies for nudging within the PBL.

[17] Additionally, it may be preferable to nudge fewer variables: work by Alexandru et al. [2009], Cha et al. [2011], de Elía and Côté [2010], and von Storch et al. [2000] use spectral nudging of only the horizontal components of wind. Alexandru et al. [2009] note that spectral nudging of only the horizontal components of wind may have little variation from nudging additional variables. Cha et al. [2011], de Elía and Côté [2010], and von Storch et al. [2000] only nudge winds above the boundary layer, while Alexandru et al. [2009] investigated the effects of varying the levels at which the horizontal winds were nudged. We discuss the results of these varying CDD configurations in section 4.2.

[18] To quantify the differences among varying CDD configurations and to using WRF-CCSM output solely as lateral boundary conditions, we examine daily averages of 2 m temperature, maximum temperature, minimum temperature, accumulated precipitation, boundary layer height, and 10 m wind speed, calculated using the same physics options as described above for the following experiments: “unconstrained,” where no variables are nudged, “CDD,” where all available variables are nudged at all heights, “CDD above PBL,” where temperature, water vapor mixing ratio and the horizontal wind components are only nudged above the boundary layer, and “CDD only wind,” where only the horizontal wind components are nudged at all levels. Our analysis is limited to points over land, as we did not employ a lake model with WRF. Where the WRF was unconstrained, the lateral boundary conditions were updated with WRF-CCSM data every 3 h. Since our 12 km × 12 km WRF output has a finer resolution than the parent data, for comparison the parent NARR and NARCCAP have been linearly interpolated to the WRF grid. The interpolation will contribute to the differences among parent data and downscaled simulations, which has not been quantified here. Our metrics are comparable to those used by Bowden et al. [2012] and Otte et al. [2012], who analyzed 2 m temperatures and precipitation averages, and Tai et al. [2010], who included wind speed in their analysis, and are also of interest for air quality studies: temperature affects reaction rates, precipitation affects wet deposition, and winds and boundary layer height affect pollutant dispersion.

4 Results

4.1 Experiments With and Without CDD

[19] When all available variables are constrained at all levels, the resulting temperatures, winds, and boundary layer heights have the smallest departures from the parent data. The differences in average daily maximum temperatures between CDD and unconstrained experiments compared to the parent data sets are shown in Figure 1, differences between parent and CDD experiments' average temperatures are shown in the supporting information Figure S4. When all variables are constrained, daily average maximum temperatures have departures from the parent data that are less than 0.25°C (Table 2), and when no variables are constrained, the WRF overestimates daily average maximum temperatures. Differences in temperatures that result from increased resolution of the Appalachian mountains are visible in Figure 1. Maximum temperatures along the Atlantic coastline are overestimated by the CDD experiments; this and the effect of the mountains are most pronounced in July 2069 when CDD is applied to downscaling the coarser-resolution WRF-CCSM.

Figure 1.

July average daily maximum 2 m temperatures from 2007 (a) and differences between the NARR parent data with experiments with CDD (b), without (c), CDD above PBL (d), and CDD only wind (e); from 2069 (f) and differences between parent WRF-CCSM data and experiments with constrained dynamical downscaling (g), without (h), CDD above PBL (i), and CDD only wind (j). All temperatures are in degree Celsius. White dots indicate where grid points were excluded from analysis for having less land than water (lakes).

Table 2. Mean Values and Error Metrics for July 2006, 2007, and 2069 Average Daily 2 m Maximum Temperaturesa, and Between the NARR (2006 and 2007) and WRF-CCSM (2069) and Downscaling Experiments With WRF
Average Daily Maximum 2 m Temperatures
Parent Mean (°C)WRF ExperimentMean (°C)RMSE (°C)Mean Bias (°C)Mean Error (°C)Mean Fractional Bias (%)Mean Fractional Error (%)Correlation Coefficient
  1. a

    Mean, mean bias, and mean error are in degree Celsius, averaged over the Eastern U.S.

(2006)
35.95CDD35.851.17−0.080.68−0.041.900.98
Unconstrained37.532.561.231.603.574.490.93
CDD above PBL36.171.600.170.970.662.710.96
CDD only wind37.762.691.401.674.034.730.93
(2007)
33.67CDD33.700.930.020.520.121.560.96
unconstrained35.622.781.511.774.385.110.83
CDD above PBL34.091.340.320.780.992.340.93
CDD only wind35.902.721.731.814.995.250.90
(2069)
44.70CDD44.501.20−0.160.65−0.341.480.93
Unconstrained40.385.09−3.353.40−7.827.940.56
CDD above PBL42.933.06−1.371.88−3.054.250.62
CDD only wind41.184.59−2.732.99−6.266.920.45

[20] As shown in Figure 2, minimum temperatures are often underestimated in the CDD experiment compared to the unconstrained, with a negative mean bias ranging from −0.81°C (2069) to −1.19°C (2006) (Table 3). The mean biases are smaller, and mean errors larger in the unconstrained case, as departures from the parent data vary more widely in sign (Figure 2). The effects of higher resolution on minimum temperature are seen in all WRF downscaling experiments, with urban areas warmer than the parent data—the WRF experiments are clearly warmer than the parent data minimum temperatures in cities such as Chicago, Detroit, Atlanta, and Pittsburgh. Similar to differences in average and average maximum temperatures, differences in minimum temperatures exist in areas with varying topography (Figure 2); however, these are both small and expected as the WRF resolves conditions related to the topography on a finer scale than the parent data.

Figure 2.

July average daily minimum 2 m temperatures from 2007 (a) and differences between the NARR parent data with experiments with CDD (b), without (c), CDD above PBL (d), and CDD only wind (e); from 2069 (f) and differences between parent WRF-CCSM data and experiments with constrained dynamical downscaling (g), without (h), CDD above PBL (i), and CDD only wind (j). All temperatures are in degree Celsius. White dots indicate where grid points were excluded from analysis for having less land than water (lakes).

Table 3. Mean Values and Error Metrics for July 2006, 2007, and 2069 Average Daily 2 m Minimum Temperaturesa, and Between the NARR (2006 and 2007) and WRF-CCSM (2069) and Downscaling Experiments With WRF
Average Daily Minimum 2 m Temperatures
Parent Mean (°C)WRF ExperimentMean (°C)RMSE (°C)Mean Bias (°C)Mean Error (°C)Mean Fractional Bias (%)Mean Fractional Error (%)Correlation Coefficient
  1. a

    Mean, mean bias, and mean error are in degree Celsius, averaged over the Eastern U.S.

(2006)
13.28CDD11.741.96−1.191.33−11.6712.790.96
Unconstrained12.562.28−0.561.39−7.1013.760.90
CDD above PBL12.631.97−0.511.26−6.3211.490.93
CDD only wind12.232.36−0.821.51−11.2415.550.92
(2007)
12.17CDD10.651.98−1.181.30−13.1714.450.97
Unconstrained11.302.13−0.681.36−9.2514.400.93
CDD above PBL11.442.03−0.571.26−8.1313.810.94
CDD only wind11.262.28−0.711.46−11.2516.490.94
(2069)
13.98CDD12.931.84−0.811.15−8.1110.930.95
Unconstrained15.062.200.841.337.2711.290.92
CDD above PBL15.582.531.241.629.7212.660.92
CDD only wind14.662.080.531.275.0210.660.91

[21] Wind speeds at 10 m are also consistently underestimated in the CDD experiments, with the smallest slow bias of −0.43 m/s in July 2069 (Table 4), the reason for which unclear. However, the differences are within the level of uncertainty of radiosonde observations within the lowest kilometer of the atmosphere [Gilliam et al., 2012]. The spatial distribution of differences of 10 m winds between the parent data sets and CDD experiments is shown in the supporting information Figure S5. In the historic CDD and unconstrained experiments, the slower winds and lower minimum temperatures result in less vertical mixing, lowering the height of the boundary layer relative to the parent data by over 100 m (Figure 3). Boundary layer heights are low relative to NARR throughout the domain, especially in the Dakotas and along the Applachian mountains, with small areas of elevated PBL heights relative to NARR in southern New Jersey, Delaware, southern Mississippi, and urban areas such as Chicago (Figure 3). In the CDD and unconstrained experiments for July 2069, however, the boundary layer height averages over 40 m higher than the WRF-CCSM PBL height (Figure 3; also summarized in the supporting information Table S3).

Table 4. Mean Values and Error Metrics for July 2006, 2007, and 2069 Average 10 m Wind Speeda, and Between the NARR (2006 and 2007) and WRF-CCSM (2069) and Downscaling Experiments With WRF
Average 10 m Wind Speed
Parent Mean (m/s)WRF ExperimentMean (m/s)RMSE (m/s)Mean Bias (m/s)Mean Error (m/s)Mean Fractional Bias (%)Mean Fractional Error (%)Correlation Coefficient
  1. a

    Mean, mean bias, and mean error are in meters per second, averaged over the Eastern U.S.

(2006)
3.80CDD3.070.82−0.570.57−17.0517.110.79
Unconstrained3.630.49−0.130.30−3.748.320.70
CDD above PBL3.530.54−0.210.34−6.219.490.72
CDD only wind3.190.74−0.470.49−13.9014.320.74
(2007)
3.57CDD2.840.81−0.560.56−17.8817.940.77
Unconstrained3.660.460.070.291.877.990.66
CDD above PBL3.260.52−0.240.33−7.179.710.69
CDD only wind2.960.72−0.470.48−14.5914.810.71
(2069)
4.15CDD3.590.73−0.430.46−11.9612.750.93
Unconstrained3.590.81−0.430.51−10.8313.230.88
CDD above PBL4.240.520.070.331.998.060.91
CDD only wind3.690.67−0.360.42−9.6011.320.92
Figure 3.

July average boundary layer heights from 2007 (a) and differences between the NARR parent data with experiments with CDD (b), without (c), CDD above PBL (d), and CDD only wind (e); from 2069 (f) and differences between parent WRF-CCSM data and experiments with constrained dynamical downscaling (g), without (h), CDD above PBL (i), and CDD only wind (j). Boundary layer heights and departures from parent data are shown in meters. White dots indicate where grid points were excluded from analysis for having less land than water (lakes).

[22] As shown in Table 5, the greatest variance among the experiments and their parent data sets appears in the daily averaged precipitation. Spatial correlations are higher, and mean errors in precipitation are lower when the WRF is unconstrained. The mean daily average precipitation is always higher than the parent data when WRF is unconstrained regardless of year of study or parent data set used. Differences in daily average precipitation between the parent data sets and constrained and unconstrained experiments are mapped in Figure 4. In the historical simulations with NARR, both the CDD and unconstrained experiments overestimate precipitation in the southeastern states (Figure 4), but the mean daily precipitation in CDD experiments using NARR is closer to the NARR mean than the unconstrained experiments. The CDD experiment of July 2007 has a lower RSME for precipitation (2.28 mm/day) and lower mean bias than the unconstrained case (−0.17 mm/day) (Table 5). In the 2006 simulation, the advantage of CDD is not as clear: the errors are lower in the unconstrained experiment (RMSE of 2.19 mm/day, mean error of 1.18 mm/day), while the CDD experiment has a lower bias (0.48 mm/day) and higher spatial correlation (0.42) (Table 5). In the future simulation with WRF-CCSM, the only location where the CDD overestimates precipitation is in northern Georgia; everywhere else, the precipitation average is underestimated (Figure 4). When the WRF is unconstrained, precipitation is overestimated over the majority of the domain (Figure 4)—the spatially averaged mean precipitation in the unconstrained experiment is 0.43 mm/day higher than the WRF-CCSM, and the correlation coefficient a low 0.28 (the CDD experiment's correlation coefficient is 0.35). In all experiments, where the parent data indicate heavier daily rainfall averages—such as northeastern Texas, northern Louisiana, and southern Mississippi (2007), and Alabama, Georgia, and eastern North Carolina (2069)—both CDD and unconstrained simulations have less precipitation on average, a dry departure from their parent data sets as much as 8 mm/day in 2007 and 2069 (Figure 4).

Table 5. Mean Values and Error Metrics for July 2006, 2007, and 2069 Average Daily Precipitationa, and Between the NARR (2006 and 2007) and WRF-CCSM (2069) and Downscaling Experiments With WRF
Average Daily Precipitation
Parent Mean (mm)WRF ExperimentMean (mm)RMSE (mm)Mean Bias (mm)Mean Error (mm)Mean Fractional Bias (%)Mean Fractional Error (%)Correlation Coefficient
  1. a

    Mean, mean bias, and mean error are in millimeters per day, averaged over the Eastern U.S.

(2006)
2.72CDD3.352.580.481.33−2.3844.060.42
Unconstrained3.412.190.531.189.0837.110.41
CDD above PBL4.272.751.201.4824.0339.460.54
CDD only wind4.212.681.561.4521.7537.980.61
(2007)
3.11CDD2.902.28−0.171.27−18.4745.160.40
Unconstrained4.372.970.981.5320.1137.910.37
CDD above PBL4.342.580.951.3716.0735.400.66
CDD only wind4.672.841.201.5725.2037.720.50
(2069)
1.02CDD0.051.35−0.740.74−140.79140.780.35
Unconstrained1.451.380.330.7029.1457.540.28
CDD above PBL0.181.18−0.650.65−121.68122.030.61
CDD only wind1.621.430.470.7435.3758.870.49
Figure 4.

July average daily accumulated precipitation from 2007 (a) and differences between the NARR parent data with experiments with CDD (b), without (c), CDD above PBL (d), and CDD only wind (e); from 2069 (f) and differences between parent WRF-CCSM data and experiments with constrained dynamical downscaling (g), without (h), CDD above PBL (i), and CDD only wind (j). Daily average precipitation values and departures from parent data are in millimeters per day. White dots indicate where grid points were excluded from analysis for having less land than water (lakes).

4.2 Experiments Varying Nudging Parameters

[23] Here we evaluate the results of changing the configuration of analysis nudging, from all variables at all levels to all variables only above the PBL, and to only the horizontal wind components at all levels. When CDD is limited to only the horizontal components of wind (at all levels), the departures from the parent data sets' average daily maximum temperatures are similar to those when the WRF is unconstrained; when nudging is limited to above the boundary layer, departures from the parent data are not as great as when unconstrained but not as small as when variables are constrained at all levels (Figure 1). The unconstrained, CDD only wind, and CDD above the PBL experiments overestimate historic maximum temperatures, with warm mean fractional biases ranging from 0.66% to 4.99%, while in the future simulation, the average daily maximum temperature is underestimated, though less so than in the unconstrained case (Table 2). The spatial patterns of departure from the parent data sets' maximum and minimum daily temperatures are also similar among the unconstrained, CDD above PBL, and CDD only wind experiments (Figures 1 and 2). Unlike the statistics for experiments' departures from the parent data sets' average maximum temperatures, however, the errors and biases in minimum temperatures are not lowest for the CDD experiment—in the historic simulations, they are lowest overall in the CDD above PBL experiment, and there is no best nudging configuration for minimizing the differences in minimum temperatures in the simulation of July 2069 (Table 3).

[24] The height of the boundary layer is, on average, consistently underestimated relative to NARR and overestimated relative to WRF-CCSM regardless of nudging configuration. One feature of interest appears in the unconstrained, CDD above the PBL, and CDD only wind simulations of July 2069: lower-than-WRF-CCSM PBL heights north and west of the Ohio River through Iowa, Minnesota, and North Dakota (Figure 3), corresponding to lower average temperatures in the same region (supporting information Figure S4). This pattern is analogous to the “warming hole,” which, as Bukovsky [2012] summarizes, may be caused by changing dynamical conditions associated with climate change, or other variability, such as changes in the land cover. The spatial pattern of cooler temperatures and lower PBL heights in our future simulations (Figure 3) corresponds to locations where Diffenbaugh [2009] found JJA cooling as a result of cropland replacing historic distributions of short grass and interrupted forest. Since our WRF simulations use a present-day land-use classification [Loveland et al., 2000], and the A2 emissions scenario used by the WRF-CCSM includes a decrease in crop and increase in pasture land in the Midwest and plains states [Haim et al., 2011] (hence a decrease in the cooling effect of cropland), it is possible that differing land cover relative to the parent data has affected our simulations where temperature was not constrained in the boundary layer.

[25] Daily average precipitation is overestimated in every experiment except when CDD is applied to all available variables at all levels and in the case where CDD is applied only above the PBL in our future simulation. There is no single nudging configuration that consistently has the lowest departures from precipitation in the parent data sets, though nudging only the horizontal components of wind tends to be the wettest simulation of each summer month (Table 5). Increases in precipitation relative to NARR may be tied to our use of the Kain-Fritsch cumulus parameterization, which, though more realistic than the Grell and Betts-Miller parameterizations, has been found to overestimate the frequency of rainy days in the summertime over the southeastern U.S. [Lynn et al., 2007].

[26] Constraining of variables only above the PBL in July 2069 results in drier conditions than the WRF-CCSM, with a mean bias of −0.65 mm/day, which is only slightly less than the mean bias of the CDD experiment (−0.74 mm/day) (Table 5). Since the CDD and CDD above the PBL experiments of July 2069 both underestimate daily precipitation relative to WRF-CCSM, and the unconstrained and CDD of only horizontal winds both overestimate daily precipitation relative to WRF-CCSM, it appears the nudging of temperature and/or moisture—within or above the boundary layer—may act to inhibit precipitation. For example, if humidity is not nudged within the boundary layer, there may be too much evaporation (this would explain maximum temperatures being lower than WRF-CCSM in all but the CDD experiment), but that humidity may not translate to precipitation if it is then nudged above the boundary layer, the “extra” moisture creating more widespread clouds. This is likely the reason for the elevated minimum temperatures (Figure 2i) and relative dryness of the CDD above the PBL simulation (Figure 4i), as there are more widespread low clouds in the CDD above the PBL simulation of July 2069 (not shown).

5 Conclusions

[27] A need for future climate conditions on finer spatial and temporal scales than are currently available from global and regional climate models drives the need to regionally downscale climate data. In using CDD, one may avoid creating climate conditions unique to the downscaling model, which would complicate comparison of models and results. Here we have evaluated multiple CDD configurations and quantified consistency of each with the parent data set. As with any downscaling method [e.g., Chase et al., 2003; Pielke et al., 2012; Ray et al., 2010], the skill of CDD is limited primarily by the choice of the parent data set. In addition, we have found that CDD skill varies with nudging configuration: which variables are nudged in the downscaling model and where they are nudged.

[28] Our simulations of July 2006 and 2007 underestimated minimum temperatures in all experiments, with the underestimate of minimum temperatures relative to the NARR data set most pronounced when temperature and moisture were constrained within the boundary layer. Whereas historic simulations had cooler minimum temperatures than the parent data set, our July 2069 simulations overestimated minimum temperatures for all experiments except the case where all variables were nudged at all levels. These differences in minimum temperatures may be related to boundary layer heights, since there is similar past-to-future reversal of the sign of differences in boundary layer height relative to the parent data sets. All downscaling experiments underestimate PBL height relative to the NARR data set (July 2006 and 2007) and, on average, overestimate PBL height relative to WRF-CCSM data (July 2069). Embedded within the overall elevation of PBL height relative to WRF-CCSM data in July 2069, PBL heights are underestimated in those simulations in the Midwest, which may be a result of differing land cover between our downscaling model and the WRF-CCSM data set. The reason for the lowered minimum temperatures and PBL heights in the historic simulations relative to the NARR data set is unclear, but it may be related to cloud cover and precipitation.

[29] Most downscaling experiments overestimate precipitation relative to NARR, which may have multiple causes. Historic simulations have shown that the Kain-Fritsch cumulus parameterization scheme overestimates precipitation [Lynn et al., 2007], as does analysis nudging of all variables above the PBL [Otte et al., 2012], either of which may have affected precipitation within our simulations of July 2006 and July 2007. Which is the stronger effect, and whether there is a relationship between overestimated precipitation and departures of temperature relative to the NARR data set, needs further study. The WRF does not consistently overestimate summertime precipitation across all years studied and nudging configurations, however. In the simulations of July 2069, the CDD and CDD above the PBL experiments underestimate precipitation relative to WRF-CCSM data. This may be the result of the parent WRF-CCSM simulation July of 2069 being exceptionally warm and dry—with an average maximum temperature 8.75 ° C higher and precipitation average of 1.7 mm/day lower than July of 2006, which was a record-setting warm and dry July at the time. It is possible that in these extreme conditions, WRF overestimated the latent heat flux from the surface, but when moisture was nudged, more widespread, non-precipitating clouds resulted, lowering maximum temperatures and raising minimum temperatures relative to the parent WRF-CCSM. Though Bowden et al. [2012] found that analysis nudging of water vapor mixing ratio above the boundary layer results in a better representation of parent precipitation compared to spectral and no nudging, limiting nudging of moisture to only above the boundary layer may not be appropriate in extreme future conditions.

[30] We have shown that regardless of time period and parent data set, daily average and maximum temperatures are closest to the parent data set temperatures when temperature is constrained within the PBL; though expected, reasonable differences in temperature where the downscaling model provides additional information related to high topographical variation and the distribution of urban areas. The potential value added in these geographical areas is unconfirmed, and we recommend future work comparing observations to CDD results in areas of urban-rural gradients and high topographical variation. Nudging temperatures in the boundary layer also insures against the possible effects of varying land cover classifications between parent and downscaling model. Though nudging temperature at all levels clearly results in smaller departures with temperatures from the parent data set, nudging of the horizontal components of wind at all levels does not result in the smallest differences in the wind field relative to the parent data set. Differences in 10 m wind speeds among experiments are within 18% of the parent data, with the smallest departures from NARR and WRF-CCSM when the horizontal wind components were either not nudged or only nudged above the boundary layer.

[31] Although our motivation is the impact of climate change on air quality, and the metrics we used to assess CDD were parameters that affect atmospheric chemistry and chemical transport in the boundary layer, the potential benefits of CDD are not limited to the study design detailed here. Our findings describing the benefits of CDD align with those described for numerical weather prediction, including cloudiness [Meinke et al., 2006] and precipitation [Falkovich et al., 2000; Otte et al., 2012; Tang et al., 2010], as well as 2 m temperatures [Heikkilä et al., 2011; Otte et al., 2012] and 10 m winds [Heikkilä et al., 2011]. The CDD technique shows skill with multiple resolutions, both of the parent data and downscaled conditions. Here we have downscaled 50 km × 50 km parent data to a 12 km × 12 km grid. Falkovich et al. [2000] nudged tropical precipitation towards 1° × 1° NCEP global analysis and towards 0.5° × 0.5° satellite-derived precipitation with low errors compared to simulated precipitation when no nudging was applied. Heikkilä et al. [2011] nudged 1° × 1° ERA-40 parent data to 30 km × 30 km and 10 km × 10 km grids, finding the results reduced biases seen in the parent data caused by coarser topographical resolution. Given the successes of CDD when used for numerical weather prediction and in studies of historic climate, CDD may be a particularly useful tool for future climate impact studies. However, further testing is needed to maximize the skill of representing fields that are not nudged, such as boundary layer height and precipitation, and the choice of which variables are nudged and at what levels should be carefully considered based on the climate impact(s) being studied.

Acknowledgments

[32] M. H. and T. H. were funded by NIH Grant 1R21ES020232-01, and T. H. was additionally supported by the NASA Applied Sciences Program through the NASA Air Quality Applied Sciences Team (AQAST). The authors wish to thank William Lewis and Steve Vavrus for valuable comments. They gratefully acknowledge the anonymous reviewers for their thoughtful comments for improving this manuscript and Melissa Bukovsky, who provided WRF-CCSM soil data necessary to run the WRF with the NOAH LSM. They also acknowledge the North American Regional Climate Change Assessment Program (NARCCAP) and the National Climatic Data Center (NCDC) for providing the data used in this paper. 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 (EPA) Office of Research and Development. The NCDC provided North American Regional Reanalysis (NARR) data from the National Centers for Environmental Prediction (NCEP), a division of NOAA/National Weather Service.