An Improved Dynamical Downscaling Method with GCM Bias Corrections and Its Validation with 30 Years of Climate Simulations

15 SEPTEMBER 2012 X U A N D Y A N G 6271

ZHONGFENG XU Department of Geological Sciences, The Jackson School of Geosciences, The University of Texas at Austin, Austin, Texas, and RCE-TEA, Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China

ZONG-LIANG YANG Department of Geological Sciences, The Jackson School of Geosciences, The University of Texas at Austin, Austin, Texas

(Manuscript received 22 December 2011, in ﬁnal form 15 March 2012)

ABSTRACT

An improved dynamical downscaling method (IDD) with general circulation model (GCM) bias corrections is developed and assessed over North America. A set of regional climate simulations is performed with the Weather Research and Forecasting Model (WRF) version 3.3 embedded in the National Center for Atmospheric Research’s (NCAR’s) Community Atmosphere Model (CAM). The GCM climatological means and the amplitudes of interannual variations are adjusted based on the National Centers for Envi- ronmental Prediction (NCEP)–NCAR global reanalysis products (NNRP) before using them to drive WRF. In this study, the WRF downscaling experiments are identical except the initial and lateral boundary conditions derived from the NNRP, original GCM output, and bias-corrected GCM output, respectively. The analysis ﬁnds that the IDD greatly improves the downscaled climate in both climatological means and extreme events relative to the traditional dynamical downscaling approach (TDD). The errors of downscaled climatological mean air temperature, geopotential height, wind vector, moisture, and precipitation are greatly reduced when the GCM bias corrections are applied. In the meantime, IDD also improves the downscaled extreme events characterized by the reduced errors in 2-yr return levels of surface air temperature and precipitation. In comparison with TDD, IDD is also able to produce a more realistic probability distribution in summer daily maximum temperature over the central U.S.–Canada region as well as in summer and winter daily precipitation over the middle and eastern United States.

1. Introduction of physical principles and therefore have the potential for capturing fine spatial-scale nonlinear effect, which in- A relatively accurate regional projection of future cli- creases confidence in their abilities to downscale future mate and its impacts on society and environment has be- climate. Additionally, RCMs can better resolve orographic come crucial for public policy and decision making. effects than the coarse-resolution GCMs. RCMs have been Regional climate projection has been a major topic in the widely used to provide a better projection of future climate Intergovernmental Panel on Climate Change (IPCC) as- at the regional scale and have proven to be a powerful tool sessment report (Solomon et al. 2007). High-resolution in the dynamical downscaling of regional climate since the regional climate model (RCM) simulation (‘‘dynamical 1980s (e.g., Dickinson et al. 1989; Giorgi et al. 1994; Wang downscaling’’) or statistical methods (‘‘statistical down- et al. 2004; Lo et al. 2008; Heikkila et al. 2010). scaling’’) are the common approaches to obtain fine spa- The common approach to future climate dynamical tial-scale information from long-term global circulation downscaling employs a continuous integration of RCMs model (GCM) simulations. RCMs are formulated in terms where GCM data are used to provide initial conditions (ICs), lateral boundary conditions (LBCs), sea surface temperatures (SSTs), and initial land surface conditions. Corresponding author address: Zong-Liang Yang, The Jackson School of Geosciences, 1 University Station #C1100, The Univer- These GCM-driven simulations are herein called the sity of Texas at Austin, Austin, TX 78712-0254. traditional dynamical downscaling (TDD) approach in E-mail: [email protected] which GCM outputs are directly used as ICs and LBCs for

DOI: 10.1175/JCLI-D-12-00005.1

Ó 2012 American Meteorological Society Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 6272 JOURNAL OF CLIMATE VOLUME 25

RCMs (e.g., Giorgi and Hewitson 2001; Seth et al. 2007; wind shear over the tropical North Atlantic. Jin et al. Bukovsky and Karoly 2011). This approach has been (2011) developed a statistical regression model between employed by numerous regional climate simulating and GCM and reanalysis data to eliminate the GCM clima- assessment projects such as the Regional Climate Model tological bias, and then the bias-corrected GCM data Intercomparison Project for Asia (RMIP; Fu et al. 2005), were used to force an RCM to simulate winter pre- Ensembles-Based Predictions of Climate Changes and cipitation over the western United States. This approach Their Impacts (ENSEMBLES; van der Linden and also produced a precipitation simulation closer to obser- Mitchell 2009), and the North American Regional Climate vations than the TDD approach. Change Assessment Program (NARCCAP) (Mearns et al. The previous downscaling studies have not performed 2009). It is known that GCMs are not perfect and all in-depth comparisons between the TDD and the GCM simulations suffer from systematic biases to a certain ex- bias-corrected dynamical downscaling methods in which tent. The downscaled simulation is strongly influenced by the GCM bias correction has been employed. It is still not the skill of the GCM. The TDD certainly brings GCM bias quite clear to what extent the GCM bias correction is able into RCMs through the LBC of the RCMs and degrades to improve the downscaled regional climate projection. the downscaled simulation (Kim and Miller 2000; Rojas Additionally, the previous bias corrections only corrected and Seth 2003; Wu et al. 2005; Seth et al. 2007; Cook and the GCM climatological mean bias. No effort has been Vizy 2008; Caldwell et al. 2009). made to correct GCM variance bias yet. The GCM vari- Some bias correction methods have been used to ance may impact extreme events simulated by an RCM improve the regional climate downscaling simulation. Wu through the LBC of the RCM. In this study we develop an and Lynch (2000) investigated the response of seasonal improved dynamical downscaling method (IDD) for carbon cycle in Alaska to future climate change through regional climate projection, which incorporates both the a dynamical downscaling approach in which they con- GCM climatological mean and variance bias corrections. structed the LBC of an RCM by adding projected changes In-depth analyses are performed to assess the performance of temperature and specific humidity in a GCM simulation of IDD relative to TDD. Many previous validations of to reanalysis climate. A similar method was employed by RCM performance were based on reanalysis-driven Sato et al. (2007) to investigate the influence of global RCM simulations (e.g., Leung et al. 2003; Seth and warming on regional precipitation over Mongolia. The Rojas 2003; Lo et al. 2008; Kanamitsu et al. 2010, Ainslie bias correction was applied to wind speed, temperature, and Jackson 2010; Heikkila et al. 2010). The difference geopotential height, specific humidity, and sea surface between reanalysis-driven RCM simulations and obser- temperature in Sato et al. (2007). Their results suggested vations can be regarded as RCM errors if we take the that the rainfall intensity simulated with the new method reanalysis data as the ‘‘perfect’’ IC and LBC. In compar- was closer to observations than the traditional method. ison with the reanalysis-driven simulations, the GCM- Using a method similar to that of Sato et al. (2007), Cook driven simulations usually show larger bias due to the and Vizy (2008) investigated the effect of climate change GCM bias, which is an additional error source in RCM on the Amazon rain forest and Patricola and Cook (2010) simulations. Therefore, differences between GCM-driven projected the Northern African climate at the end of the simulations and observations result from both the GCM twenty-first century. The climatological LBCs in the and RCM biases. Two biases could appear with the same aforementioned studies maintain variations on the sea- or opposite sign. The opposite sign biases would offset sonal time scale but eliminate the diurnal and synoptic each other. In this case, the GCM-driven simulation may effects. The climatological mean boundary conditions are be closer to observation than the reanalysis-driven simu- not suitable to investigate transient climate variabilities lation, which does not mean that the GCM-driven simu- (e.g., interannual variability) or climate extremes. lation is actually better than the reanalysis-driven Holland et al. (2010) developed a more complex bias simulation. In this study, to examine the influence of GCM correction method for hurricane simulation. Their bias bias correction on the performance of an RCM down- correction retained the diurnal and synoptic effects and scaling simulation, the GCM-driven simulations will be the interannual variations in the LBC by correcting GCM compared with reanalysis-driven simulations rather than climatological mean bias with 6-hourly National Centers with observations, which is different from previous for Environmental Prediction (NCEP)–National Center validations (e.g., Sato et al. 2007; Jin et al. 2011). In for Atmospheric Research (NCAR) reanalysis data and addition, long-term simulations are necessary to compare GCM outputs. Their results suggested that the dynamical the climatological means between different dynamical downscaling simulation with GCM bias correction can downscaling experiments. produce reasonable tropical cyclone frequency because We describe the GCM bias correction method in the bias correction reduced the unrealistic high vertical section 2. Section 3 introduces the numerical models and

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 15 SEPTEMBER 2012 X U A N D Y A N G 6273 experiment design. Section 4 presents the dynamical mean climate change between the future (1980–2010) downscaling simulations by comparing the differences and the past (1950–79), and future weather and climate between the IDD and TDD simulations. Conclusions variabilities simulated by the CAM. However, the CAM and discussion are given in section 5. simulations can have also biases in temporal variations. We therefore improved the bias correction method further by correcting both the climatological mean and 2. Methodology for dynamical downscaling with variance biases in the CAM outputs over the future GCM bias corrections period: The global model used in this study is the Community Atmosphere Model (CAM), which is the atmospheric CAM** 5 NNRP 1 (CAM 2 CAM ) component of the Community Earth System Model F P F P S (CESM) developed at NCAR. CAM biases were corrected NNRPjP 1 CAMF9 , (5) by using NCEP–NCAR global reanalysis products (NNRP) SCAMjP data over the period 1950–79. To facilitate the validation of RCM simulations on future climate, we broke down where SNNRP and SCAM represent the standard deviation 61 years (1950–2010) into two periods: the past (1950–79) of NNRP data and CAM simulations, respectively. and the future (1980–2010). No future NNRP data were Clearly Eq. (4) only adjusted the climatological mean used when we produced the bias-corrected CAM data over states of CAM simulations, which is the same as the bias the future period. The future CAM-driven RCM simula- correction method used by Holland et al. (2010). In con- tions were then compared with the future reanalysis-driven trast, Eq. (5) adjusted both the climatological mean states RCM simulation to assess the performance of CAM-driven and the variances of CAM simulations. In the meantime, RCM in simulating the future climate. Six-hourly NNRP thebiascorrectionmethodallowedretainingtheCAM data and CAM outputs were broken down into a climato- simulated climatic change in the mean seasonal state, di- logical mean plus a perturbation term: urnal cycle, and variance of interannual variation. The bias correction to CAM variance described herein has not CAM 5 CAM 1 CAM9, (1) been attempted in previous GCM-driven dynamical downscaling simulations to the authors’ knowledge. NNRP 5 NNRP 1 NNRP9. (2) The bias corrections were applied to the air temperature, zonal wind, meridional wind, geopotential height, The future 6-hourly mean CAM data, CAMF, were and relative humidity at each model grid vertical level written as follows: for every 6-hourly CAM output ﬁle. The surface variables are not bias corrected in the experiments reported CAMF 5 CAMF 1 CAMF9 in this paper. As an example, Fig. 1 shows the original 5 NNRP 1 (CAM 2 NNRP ) and the bias-corrected 200-hPa July temperature in P P P a single CAM model grid (558N, 1408W) versus the 1 (CAMF 2 CAMP) 1 CAMF9: (3) NNRP data during 1981–2010. The 30-yr averaged value and standard deviation in each data are showed on the The subscript P and F represent the past (1950–79) and top of the ﬁgure. In comparison with the NNRP, the the future (1980–2010), respectively. Four terms on the original CAM simulation underestimates 200-hPa tem- right-hand side of Eq. (3) are the NNRP climatological perature by 9.7 K and its standard deviation by 1.1 K. In mean over the past period, the CAM mean value bias Fig. 1, CAMbc_ave and CAMbc_std represent the relative to NNRP over the past period, the CAM sim- CAMF* and CAM**F in Eqs. (4) and (5), respectively. ulated climate change between the future and the past The mean temperature bias is remarkably reduced from periods, and the CAM anomaly in the future. The cli- 9.7 K in CAM to 0.3 K in CAMbc_ave, while the stan- mate mean bias-corrected CAM data were constructed dard deviation remains the same as that in the original by removing the CAM bias CAMP 2 NNRPP in Eq. (3): CAM output. In contrast, both the CAM mean value and variance biases are largely reduced in the CAMF* 5 NNRPP 1 (CAMF 2 CAMP) 1 CAMF9 CAMbc_std experiment. To be brief, the bias correction 5 NNRPP 1 CAMF 2 CAMP: (4) method removes CAM biases of mean value and variance through shifting and scaling the original CAM

The bias-corrected 6-hourly CAM data, CAMF*, over simulation based on the NNRP data. However the bias the future period (1980–2010) have a base climate pro- correction methods do not change the climatic trend and vided by the NNRP data from the period of 1950–79, phase of interannual variability simulated by the CAM.

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 6274 JOURNAL OF CLIMATE VOLUME 25

FIG. 1. Interannual variation of 200-hPa temperature at the northwest boundary of the WRF domain (558N, 1408W) in July. NNRP indicates NNRP data; CAM, original CAM output; CAMbc_ave, CAM output with mean value bias correction; and CAMbc_std, CAM output with both mean value and variance bias corrections. The 30-yr averaged value and standard deviation for each individual data are also shown on the top of the ﬁgure.

3. Model description and experimental setup SST were given by the CAM 6-hourly outputs or NNRP 6-hourly data. The Community Atmosphere Model (Neale et al. There is a 6-h mismatch of the diurnal cycles between 2010) was coupled to an active land model (CLM) the NNRP data and CAM simulations because each (Oleson et al. 2010), a thermodynamic only sea ice 6-hourly CAM output saved the past 6-hourly mean data model (CICE), and a data ocean model (DOCN). In this and each NNRP saved the future 6-hourly mean data. study, a 63-yr simulation was performed by using the A 6-h shift to the CAM output was performed to have CAM at a resolution of T42 (approximately 2.8832.88) the diurnal cycle in the CAM output match with that in and 30 vertical layers with observed monthly SSTs from the NNRP. As summarized in Table 1, four dynamical 1948 to 2010. The first two years were discarded for downscaling simulations were carried out to assess the spinup. Model outputs were saved in 6-h interval. performance of the IDD relative to the TDD in down- The regional model used in this study is the Weather scaling the future climate. The WRF configurations were Research and Forecasting Model (WRF) with the Ad- kept the same in all experiments except that the ICs and vanced Research WRF (ARW) dynamic core version LBCs were derived from the NNRP data (WRF_NNRP), 3.3 (Skamarock et al. 2008). This model has been de- the original CAM output (WRF_CAM), CAM output veloped and maintained by NCAR. It is a non- with mean value bias correction (WRF_CAMbc_ave), hydrostatic model with 28 vertical levels designed to and CAM output with both mean value and variance bias serve both atmospheric research and operational fore- corrections (WRF_CAMbc_std). All experiments em- casting needs. The WRF domain is centered at 408N, ployed the same SST, sea ice, vegetation distributions, 978W with dimensions of 106 3 76 horizontal grid points and background albedo. Each WRF simulation was in- (Fig. 2). Horizontal resolution of 60 km was used with 28 tegrated over 31 years from 1980 to 2010. The first year vertical levels, and the time step was 360 s. The main was discarded for spinup. Model outputs were saved in 3-h physical options we used included the new Kain–Fritsch intervals. convective parameterization (Kain 2004), CAM short- wave and longwave radiation schemes (Collins et al. 4. Dynamical downscaling simulations 2004), the Noah land surface model (Chen and Dudhia 2001), and the Yonsei University planetary boundary Since the differences between RCM experiments and layer scheme (Hong et al. 2006). The ICs, LBCs, and observations result from both the GCM and RCM

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 15 SEPTEMBER 2012 X U A N D Y A N G 6275

FIG. 2. WRF domain (shaded area) and topography. The thick outline is the extent of the verification region. Contours represent the terrain height (m). biases, we therefore compare the WRF_CAM, a. Upper air variables WRF_CAMbc_ave, and WRF_CAMbc_std experi- Figure 3 shows the annual mean profile of RMSEs of air ments with the WRF_NNRP experiment to isolate the temperature, geopotential height, wind vector, and water GCM bias correction influence. Given the identical vapor mixing ratio. The annual mean RMSEs are able to model setup, the differences between various WRF represent the overall performance of each simulation in simulations result from the differences in IC and LBCs. one year since RMSE is always equal to or larger than zero. We take NNRP as the ‘‘perfect’’ IC and LBC and all bias The RMSEs of air temperature and geopotential height corrections are based on the NNRP data. We therefore are greatly reduced, especially over the upper troposphere, expect that the WRF_CAMbc_std simulation will be when the GCM bias corrections are applied (Figs. 4a,b). closer to the WRF_NNRP simulation than the We also computed the difference of area- and seasonal- WRF_CAM simulation is. averaged air temperature and geopotential height between The WRF dynamical downscaling experiments are the WRF simulations with and without GCM bias correc- assessed by some statistical verification techniques. tions (not shown). Generally the WRF_CAM experiment Root-mean-square error (RMSE) is one of the com- shows a cold bias in the troposphere relative to the monly used error statistics, and it is defined as WRF_NNRP experiment throughout the year. The max- sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi N imum cold bias appears in winter. As a result the simulated 1 2 geopotential height is systematically lower over the mid to RMSE 5 å (Mi 2 Oi) , N i51 upper troposphere in the WRF_CAM experiment than where N is the total number of grid points and M and O TABLE 1. Brief summary of downscaling simulations. represent the CAM-driven WRF simulations and NNRP-driven simulations, respectively. Following Experiment identifier Description Shukla and Saha (1974) with a slight modification, the WRF_NNRP WRF experiment with the NNRP data wind vector RMSE is defined by as the initial and lateral boundary conditions qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 N WRF_CAM Traditional dynamical downscaling 2 2 approach. WRF experiment with the RMSEV 5 å (Mu,i 2 Ou,i) 1 (My,i 2 Oy,i) , N i51 original CAM output as the initial and lateral boundary conditions where u and y subscripts indicate the zonal and meridio- WRF_CAMbc_ave Same as WRF_CAM except the nal wind, respectively. All statistics presented in this pa- climatological mean biases in CAM output are corrected. per are based on the 30-yr WRF simulations from 1981 to WRF_CAMbc_std Improved dynamical downscaling 2010. The thick outline in Fig. 2 indicates the verification approach. Same as WRF_CAM except region where the performance of dynamical downscaling that both the climatological mean biases simulations will be examined. The verification region and the variance biases in excludes all buffer zones of the WRF. CAM output are corrected.

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 6276 JOURNAL OF CLIMATE VOLUME 25

FIG. 3. Annual mean RMSEs of (a) air temperature, (b) geopotential height, (c) wind vector, and (d) water vapor mixing ratio. The RMSEs are computed over the verification region by monthly mean data. that in the WRF_NNRP experiment. The cold bias in low-level air temperature is largely dominated by the land WRF can be largely removed in spring, summer, and au- surface process or by the sea surface condition where no tumn when the GCM bias corrections are applied. In SST bias correction is applied yet. Lower tropospheric contrast, only part of the cold bias is removed in lower moisture is also improved, which indicates that the mois- troposphere in winter. Similar to the air temperature and ture transfer simulated by the IDD experiment is more geopotential height, the wind vector simulation is also reasonable than that in the TDD experiment. improved in the IDD simulation relative to that in the Generally the WRF_CAMbc_ave and WRF_CAMbc_ TDD. The improvement appears to be more remarkable at std experiments show very similar performance in the upper troposphere than the lower troposphere for all simulating climatological mean variables, which means variables except the moisture. One possible reason is that that the bias correction to variance only plays a minor the CAM shows larger bias in the upper troposphere. This role in modulating the climatological mean states of the feature has also been found in an earlier version of CAM downscaled climate (Fig. 3). The errors of 500-hPa ge- (e.g., Khairoutdinov et al. 2005). As a result, the bias cor- opotential height for the simulations with and without rection influence is larger in the upper troposphere and its GCM bias corrections in winter [December–February influence can be more easily transported into the RCM (DJF)] and summer [June–August (JJA)] give a spatial domain through advection effects than in the lower tro- overview of the model performance. A very significant posphere, and the lower troposphere is strongly impacted improvement to 500-hPa geopotential height simulation by the land surface condition (Gold 1967). The change of has been found in winter when the GCM bias corrections

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 15 SEPTEMBER 2012 X U A N D Y A N G 6277

FIG. 4. Errors of 500-hPa geopotential height (m) of the (a),(c) WRF_CAM and (b),(d) WRF_CAMbc_std relative to the WRF_NNRP in (a),(b) winter and (c),(d) summer. are applied (Figs. 4a,b) and very significant improvements simulation. Given the importance of surface air tem- can also be found in spring and autumn (not shown), perature in climate simulation, Fig. 6 shows the monthly whereas the improvement in summer is not as remarkable RMSEs of climatological mean 2-m air temperature as in other seasons although the simulation is also been (T2m) for three CAM-driven WRF downscaling ex- improved (Figs. 4c,d). The reason is that the CAM bias is periments. All experiments show relatively large bias in generally smaller in summer than in other seasons. There winter and small bias in summer. The IDD simulation is no room for large improvement in summer season in shows smaller RMSEs than the TDD simulation the IDD experiment. Similar results are found in 850-hPa throughout the year except in June. Generally the new winds. The winter wind vector bias is much smaller method performs better in spring [March–May (MAM)] over the North American continental region in the and autumn [September–November (SON)] than in WRF_CAMbc_std than in the WRF_CAM experiment winter (DJF) and summer (JJA). The WRF_CAMbc_ (Figs. 5a,b), while the improvement in summer is not as ave and WRF_CAMbc_std experiments show similar significant as that in winter (Figs. 5c,d). Given the iden- performance in the simulation of T2m, which means that tical model setup, these differences between the WRF_ the GCM variance bias correction only improves the CAMbc_std and WRF_CAM experiments are clearly downscaled mean climate slightly. due to the differences in their ICs and LBCs. The GCM Figure 7 illustrates the spatial pattern of errors of bias corrections significantly improve the air temperature, climatological mean T2m in winter and summer. In wind, geopotential height, and relative humidity in the winter, the WRF_CAM shows a pronounced cold bias CAM output, which in turn leads to the better WRF dy- within the model domain. Similar cold biases were also namical downscaling simulations. found in previous studies (Caldwell et al. 2009; Jin et al. 2011). The cold bias is slightly reduced in the b. Surface air temperature WRF_CAMbc_std experiment as compared to the Surface air temperature is one of the principal climatic WRF_CAM experiment (Figs. 7a,b). The cold bias may elements influencing human life and is also an important be partially attributed to the fact that the CAM un- indicator in measuring the performance of climate derestimates the observed air temperature. In this study

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 6278 JOURNAL OF CLIMATE VOLUME 25

21 FIG. 5. As in Fig. 4, but for 850-hPa winds (m s ). Shaded areas indicate the overestimated or underestimated wind speed reach 99% conﬁdence.

the CAM model was forced by the observed monthly remove CAM mean value bias CAMP 2 NNRPP in SST. The greenhouse gases were ﬁxed at year 2000 level Eq. (3) while the climate change term CAMF 2 CAMP (Solomon et al. 2007). However, the lower-tropospheric all depends on the GCM simulation. There is still a 0.38C (below 500 hPa) air temperature simulated by the CAM cold bias in the bias-corrected CAM data relative to the still shows a 1.18C cold bias relative to the NNRP during NNRP data, which may partly account for the winter cold the period of 1980–2010, which in turn leads to a lower bias in the downscaled climate. In addition to the cold bias temperature in the WRF_CAM experiment than in the in the LBC of WRF, the cold temperature–snow–albedo WRF_NNRP experiment. The bias corrections can positive feedback may also account for the cold bias in

FIG. 6. Root-mean-square errors (RMSEs) of 2-m air temperature over the veriﬁcation region. RMSEs are computed based on climatological mean monthly 2-m air temperature.

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 15 SEPTEMBER 2012 X U A N D Y A N G 6279

FIG. 7. Errors of climatological mean 2-m air temperature (8C) of (a),(c) WRF_CAM and (b),(d) WRF_CAMbc_std relative to WRF_NNRP in (a),(b) winter and (c),(d) summer. The shaded areas indicate the errors reach 99% significance level. the WRF_CAM and WRF_CAMbc_std experiments. Fig. 9 illustrates the differences of precipitation between We compared the albedo and snow cover simulated the CAM-driven WRF simulation and the NNRP-driven by the WRF_CAM, WRF_CAMbc_std, and WRF_ WRF simulation in winter and summer. The shaded area NNRP experiments. Generally the WRF_CAM and indicates the difference reaches the 99% confidence level. WRF_CAMbc_std experiments overestimate snow In winter the WRF_CAM experiment overestimates pre- cover, which leads to a higher surface albedo. The cold cipitation over Mexico and the western Atlantic Ocean bias pattern in Fig. 7a well corresponds to the region and underestimates the precipitation over the middle with higher land surface albedo. The WRF_CAM also United States (Fig. 9a). These errors are greatly removed underestimates the summer temperature by 28C over when the GCM bias corrections are applied (Fig. 9b). The the central U.S.–Canada region (Fig. 7c). The cold bias WRF_CAM experiment significantly overestimates is reduced when the GCM bias corrections are applied in summer precipitation by 0.5–1.5 mm day21 over the the WRF_CAMbc_std experiment (Fig. 7d). The sum- central U.S.–Canada region and 2–12 mm day21 over the mer cold bias is closely related to the wet bias in the western Atlantic Ocean and Gulf of Mexico (Fig. 9c). WRF_CAM experiment, which will be elucidated in the The WRF_CAM overestimates low troposphere mois- following section. ture by 3%–6% over the central U.S.–Canada region, which likely plays a substantial role in the higher-than- c. Precipitation average precipitation simulated by the WRF_CAM Figure 8 illustrates the RMSEs of precipitation from (not shown). The overestimated precipitation over the the WRF_CAM, WRF_CAMbc_ave, and WRF_ central U.S.–Canada region in the WRF_CAM ex- CAMbc_std experiments. The comparison of the periment leads to a higher soil moisture content and WRF_CAM with the WRF_CAMbc_ave and WRF_ enhanced evaporation, which in turn leads to a lower CAMbc_std experiments indicates that the errors of surfaceairtemperature(Figs.7cand9c).Thesesig- downscaled precipitation are smaller throughout the nificant errors of precipitation in the WRF_CAM ex- year when the GCM bias corrections are applied. The periment are largely removed in the WRF_CAMbc_std most significant improvement appears in August, experiment because of the GCM bias corrections (Fig. September, and October. To show spatial difference 9d). The new dynamical downscaling method shows sig- between different dynamical downscaling experiments, nificant improvement in precipitation simulation over the

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 6280 JOURNAL OF CLIMATE VOLUME 25

FIG. 8. As in Fig. 6, but for precipitation.

North American domain although the new method may event occurring on average once every N years (i.e., N is also degrade the dynamical downscaling performance the return period). For example, the 2-yr return level of somewhere, such as the winter precipitation over the daily maximum temperature in a region is 408C, which boundary of the United States and Canada and the indicates that the extreme event with daily maximum summer precipitation over western Canada. temperature higher than 408C occurs on average once every 2 years. Return level has been widely used in cli- d. Extreme events mate extreme studies (e.g., Wigley 2009; Cooley 2009; To examine the inﬂuence of GCM bias corrections Pausader et al. 2011). The 2-yr return level shown in Fig. on RCM simulation of extreme events, the difference of 10 is equivalent to the 99.5th percentile value of maxi- 2-yr return levels of daily minimum and maximum mum or minimum temperature approximately. We also temperature in winter and summer are respectively pre- computed the 1-yr and 5-yr return levels. It turns out that sented in Fig. 10. An N-yr return level is deﬁned as the the conclusions presented in this study are the same no

21 21 FIG. 9. As in Fig. 7, but for precipitation. Contour interval is 0.5 mm day for the contours between 22 and 2 mm day , and 2 mm day21 for the remaining contours.

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 15 SEPTEMBER 2012 X U A N D Y A N G 6281

FIG. 10. Errors of 2-yr return levels of (a)–(c) daily minimum temperature in winter and (d)–(f) maximum temperature in summer. The shaded areas denote the difference of daily minimum temperature is smaller than 228C in winter and the difference of daily maximum temperature is larger than 18C or smaller than 218C in summer. matter which return level is selected. The daily minimum WRF_CAMbc_std experiment further improves the per- temperature in the WRF_CAM suffers from problems formance of dynamical downscaling simulation by re- similar to climatological mean temperature in the moving almost all biases to the south of 508N by means of WRF_CAM simulations: both the climatological mean the CAM bias corrections in both mean value and variance. and extreme daily minimum temperature are under- Similarly, the WRF_CAMbc_std also improves the estimated, and the CAM bias corrections only slightly re- extreme precipitation simulation as illustrated in Fig. 11 move the cold bias over the northern part of the for the 2-yr return levels of daily precipitation in winter veriﬁcation domain (Figs. 7a,b and 10a–c). In contrast, the and summer. All CAM-driven simulations underestimate CAM bias corrections do signiﬁcantly improve daily the extreme precipitation by 3–15 mm day21 over the maximum T2m simulation in summer. The TDD simu- central to eastern United States and overestimate ex- lation, WRF_CAM, underestimates the extreme daily treme precipitation by 3–18 mm day21 over Mexico in maximum temperature by 28–68C to the east of Great winter. The comparison of Fig. 11a with Figs. 11b and 11c Lakes in summer (Fig. 10d). The WRF_CAMbc_ave indicates that the GCM bias corrections are able to im- experiment greatly reduced the cold bias through prove winter extreme precipitation simulation charac- the CAM mean value bias correction (Fig. 10e). The terized by the smaller errors of 2-yr return level of

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 6282 JOURNAL OF CLIMATE VOLUME 25

FIG. 11. As in Fig. 10, but for the 2-yr return levels of daily precipitation in (a)–(c) winter and (d)–(f) summer. precipitation. For the downscaled summer extreme pre- bias in daily maximum temperature. The maximum cipitation, the GCM bias corrections moderately improve probability appears in 268–288C in the WRF_CAM sim- the simulation in general (Figs. 11d–f). The WRF_ ulation and 288–308C in the WRF_CAMbc_ave, CAMbc_ave produces a more realistic extreme pre- WRF_CAMbc_std, and WRF_NNRP simulations. The cipitation over most parts of the United States and WRF_CAM underestimates the frequency of extreme Canada than the WRF_CAM does except for western high temperature events. Clearly there are errors in both Canada. The WRF_CAMbc_std produces an overall the mean and variance of the WRF_CAM relative to the better simulation of extreme precipitation over the veri- WRF_NNRP experiment. The WRF_CAMbc_ave ex- ﬁcation domain relative to the WRF_CAM experiment periment is able to remove the error of mean and partly except over the southwestern United States, where the remove the error of variance. The WRF_CAMbc_std WRF_CAMbc_std slightly overestimates the extreme experiment is able to produce a distribution of daily precipitation. maximum temperature very close to that in the IDD improves not only the extreme events simulation WRF_NNRP experiment. Both errors in mean and but also the possibility distributions of temperature and variance are removed in the WRF_CAMbc_std experi- precipitation. Figure 12a shows the daily maximum T2m ment. The improvement is likely related to the better distribution computed over the central United States and simulation on summer precipitation in the WRF_ Canada (408–508N, 1008–858W) where TDD shows a large CAMbc_std experiment characterized by reduced wet

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 15 SEPTEMBER 2012 X U A N D Y A N G 6283

FIG. 12. Probability density distribution for (a) daily maximum temperature in summer, (b) daily minimum temperature in winter, and daily precipitation in (c) summer and (d) winter. The probability is computed over the central U.S.–Canada region (408–508N, 1008–858W) for the air temperature and over the central and eastern United States (308–458N, 1058–758W) for the precipitation. The temperature over the Great Lakes region is not included in the statistics.

bias in precipitation and soil moisture over the central CAM-driven simulations is lower than the reanalysis- U.S.–Canada region, which in turn increases extreme driven simulation (Figs. 4a,b). IDD failed to improve high temperature events due to the reduced evaporation the distribution of winter minimum temperature, which in association with the decreased soil moisture. How- mightbeforthesamereasonthattheIDDfailedto ever, the GCM bias corrections do not help improve the improve climatological mean temperature in winter. distribution of winter minimum temperature. The winter Similarly, the possibility distributions for summer and minimum temperature at a given probability is generally winter daily precipitation are shown in Figs. 12c and lower in the CAM-driven simulations than in the 12d, respectively. TDD overestimates the heavy pre- reanalysis-driven simulation (Fig. 12b). The daily aver- cipitation in summer and underestimates heavy pre- aged temperature also shows a distribution similar to the cipitation in winter. IDD improves the distribution of daily minimum temperature in winter (not shown). As precipitation characterized by a more realistic tail of a result, the climatological mean temperature in the precipitation (Figs. 12c,d). The comparison of various

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 6284 JOURNAL OF CLIMATE VOLUME 25 dynamical downscaling experiments shows that the climatological means and extreme events through re- WRF_CAMbc_std also improve the rainfall days simu- moving GCM systemic bias, the performance of the dy- lation in summer. The error of summer rainfall days is namical downscaling simulation for regional climate is still less than 65% over the central U.S.–Canada region in sensitive to the GCM, especially to its projected climate the WRF_CAMbc_std as opposed to 10%–15% in the change. The GCM bias corrections reported in this paper WRF_CAM experiment (not shown). do not correct GCM bias in climate change simulation. Additionally, the performance of dynamical downscaling is also strongly impacted by RCM bias, which is not consid- 5. Conclusions and discussion ered in this study. We selected the CAM radiation scheme A new dynamical downscaling method with GCM bias in the WRF configuration, which is especially suited for corrections for the regional projection of further climate regional climate simulation (Skamarock et al. 2008). We was developed and validated by comparing the GCM- did not find a significant climate drift during the WRF driven WRF simulations to the NNRP-driven WRF simulation although both the reanalysis-driven and CAM- simulation. The GCM bias corrections involved shifting driven long-term (31 yr) WRF simulations show a similar and scaling to adjust the GCM mean and variance so warming trend of 0.18–0.28C decade21 in the surface air they matched with those in the NNRP data. The com- temperature. The warming trend should result from the parison of various dynamical downscaling simulations CAM radiation scheme (Collins et al. 2004) employed in shows that the improved dynamical downscaling WRF that prescribes the CO2 mixing ratio from the IPCC method with GCM bias corrections is superior to the Special Report on Emissions Scenarios A2 scenario. traditional dynamical downscaling method character- This study has demonstrated that the IDD is superior to ized by the significant improvement in downscaled air the TDD when it is applied to North America, and it is temperature, geopotential height, wind vector, and wa- expected that the method can be implemented for regional ter vapor. The most significant improvement appears in climate projection elsewhere. In addition to the future the middle-upper troposphere for the air temperature, climate projection, IDD can also be applied to downscale geopotential height, and wind vectors, and in the lower GCM sensitivity simulations. In doing so, the GCM con- troposphere for the water vapor. As a result, the trol (sensitivity) simulation is corresponding to the past downscaled surface air temperature and precipitation (future) GCM simulation in the paper. The difference are also improved. The most significant improvement of between the GCM sensitivity and control simulations is summer precipitation appears in the central U.S.–Canada corresponding to that between the GCM future and past region where the TDD overestimates precipitation by simulations as described in the present paper. Thus the 0.5–1.5 mm day21. The overestimated precipitation difference between the GCM sensitivity and control sim- over the central U.S.–Canada region in TDD leads to ulations can be retained and passed onto the RCM, a higher moisture content and enhanced evaporation, thereby further impacting the downscaled simulations. It is which in turn leads to a cold bias of surface air tem- known the LBC becomes a stronger constraint to the re- perature. These significant errors in precipitation and gional climate simulation when the regional model do- surface temperature are largely removed in the IDD main size is smaller. We can therefore expect that the IDD because of the GCM bias corrections. will perform better when the regional model domain is Additionally, IDD is also able to improve the down- smaller. Alternatively, spectral nudging may also help scaled extreme events and the possibility distribution of improve IDD simulation further, which will be presented temperature and precipitation. The 2-yr return level of in another paper. Although the influence of RCM sys- summer daily maximum temperature simulated by the tematic biases on the performance dynamical downscaling TDD is underestimated by 28–68C over the central U.S.– was not investigated in this study, the IDD would be Canada region. In contrast, the bias is generally less than expected to continue to produce more accurate down- 618C in the IDD experiment. IDD simulates a more re- scaled climate when the RCM biases are reduced with the alistic probability density distribution of daily maximum improvement of RCM in the future. We only examined temperature over the central U.S.–Canada region. Simi- the influences of the LBC bias corrections, while the sur- lar conclusions are found in the downscaled precipitation. face condition influences (e.g., SSTs) were not considered. TDD overestimates summer extreme precipitation and It is known that the SSTs simulated by a fully coupled underestimates winter extreme precipitation. In contrast, climate system model also show significant biases. The bias IDD is able to produce better extreme precipitation in corrections of temperature, zonal wind, meridional wind, both seasons. geopotential height, and relative humidity only impact the Although the new dynamical downscaling method is ICs and LBCs of RCM. In contrast, the SST bias correc- able to improve regional climate simulation of both tion would continuously impact low boundary conditions

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 15 SEPTEMBER 2012 X U A N D Y A N G 6285 inside the RCM domain during the model integration, Heikkila, U., A. Sandvik, and A. Sorteberg, 2010: Dynamical which also plays a very important role in regulating the downscaling of ERA-40 in complex terrain using the WRF RCM simulations. The relative importance of bias cor- regional climate model. Climate Dyn., 37, 1551–1564, doi:10.1007/s00382-010-0928-6. rections applied to LBCs versus SSTs is worth exploring in Holland, G. J., J. Done, C. Bruyere, C. Cooper, and A. Suzuki, 2010: the future study. Model investigations of the effects of climate variability and change on future Gulf of Mexico tropical cyclone activity. Proc. Offshore Technology Conf., Houston, TX, ASCE, OTC 20690. Acknowledgments. The authors thank the Texas Ad- [Available online at http://www.netl.doe.gov/kmd/RPSEA_ vanced Computing Center for providing computer re- Project_Outreach/07121-DW1801_OTC-20690-MS.pdf.] sources and technical support. NCEP–NCAR reanalysis Hong, S.-Y., Y. Noh, and J. Dudhia, 2006: A new vertical diffusion package with an explicit treatment of entrainment processes. data were obtained from the Research Data Archive Mon. Wea. Rev., 134, 2318–2341. (http://dss.ucar.edu/datasets/ds090.0/), which is main- Jin, J. M., S. Y. Wang, and R. R. Gillies, 2011: An improved tained by the Computational and Information Systems dynamical downscaling for the western United States. Cli- Laboratory (CISL) at NCAR. Financial support was mate Change: Research and Technology for Adaptation and provided by NASA (Grant NNX11AE42G), ‘‘Na- Mitigation, J. Blanco and H. Kheradmand, Eds., InTech, 23–38. tional Basic Research Program of China’’ Project Kain, J. S., 2004: The Kain–Fritsch convective parameterization: 2011CB952004, National Natural Science Foundation of An update. J. Appl. Meteor., 43, 170–181. China General Program (Grant 40905042), and KAUST. Kanamitsu, M., K. Yoshimura, Y.-B. Yhang, and S.-Y. Hong, 2010: Errors of interannual variability and trend in dynamical downscaling of reanalysis. J. Geophys. Res., 115, D17115, doi:10.1029/2009JD013511. REFERENCES Khairoutdinov, M., D. Randall, and C. DeMott, 2005: Simulations Ainslie, B., and P. L. Jackson, 2010: Downscaling and bias correcting of the atmospheric general circulation using a cloud-resolving a cold season precipitation climatology over coastal southern model as a superparameterization of physical processes. British Columbia using the regional atmospheric modeling J. Atmos. Sci., 62, 2136–2154. system (RAMS). J. Appl. Meteor. Climatol., 49, 937–953. Kim, J., and N. L. Miller, 2000: A seasonal precipitation and stream Bukovsky, M. S., and D. J. Karoly, 2011: A regional modeling study flow hindcast and prediction study in the western United of climate change impacts on warm-season precipitation in the States during the 1997/98 winter season using a dynamic central United States. J. Climate, 24, 1985–2002. downscaling system. J. Hydrometeor., 1, 311–329. Caldwell, P., H.-N. S. Chin, D. C. Bader, and G. Bala, 2009: Leung, L. R., Y. Qian, and X. Bian, 2003: Hydroclimate of the Evaluation of a WRF dynamical downscaling simulation over western United States based on observations and regional California. Climatic Change, 95, 499–521. climate simulations of 1981–2000. Part I: Seasonal statistics. Chen, F., and J. Dudhia, 2001: Coupling an advanced land surface– J. Climate, 16, 1892–1911. hydrology model with the Penn State–NCAR MM5 modeling Lo, J. C., Z.-L. Yang, and R. A. Pielke Sr., 2008: Assessment of system. Part I: Model implementation and sensitivity. Mon. three dynamical climate downscaling methods using the Wea. Rev., 129, 569–585. Weather Research and Forecasting (WRF) model. J. Geo- Collins, W. D., and Coauthors, 2004: Description of the NCAR phys. Res., 113, D09112, doi:10.1029/2007JD009216. Community Atmosphere Model (CAM3.0). NCAR Tech. Mearns, L. O., W. J. Gutowski, R. Jones, L.-Y. Leung, S. McGinnis, Note NCAR/TN-4641STR, 214 pp. A. M. B. Nunes, and Y. Qian, 2009: A regional climate change Cook, K. H., and E. K. Vizy, 2008: Effects of twenty-first-century assessment program for North America. Eos, Trans. Amer. climate change on the Amazon rain forest. J. Climate, 21, 542– Geophys. Union, 90, 311–312. 560. Neale, R. B., and Coauthors, 2010: Description of the NCAR Cooley, D., 2009: Extreme value analysis and the study of climate Community Atmosphere Model (CAM 5.0). NCAR Tech. change: A commentary on Wigley 1988. Climatic Change, 97, Note NCAR/TN-4861STR, 268 pp. [Available online at http:// 77–83. www.cesm.ucar.edu/models/cesm1.0/cam/docs/description/ Dickinson, R. E., R. M. Errico, F. Giorgi, and G. T. Bates, 1989: cam5_desc.pdf.] A regional climate model for the western United States. Climatic Oleson, K. W., and Coauthors, 2010: Technical description of Change, 15, 383–422. version 4.0 of the Community Land Model (CLM4). NCAR Fu, C., and Coauthors, 2005: Regional Climate Model Inter- Tech. Note NCAR/TN-4781STR, 257 pp. comparison Project for Asia. Bull. Amer. Meteor. Soc., 86, Patricola, C. M., and K. H. Cook, 2010: Northern African cli- 257–266. mate at the end of the twenty-first century: An integrated Giorgi, F., and B. Hewitson, 2001: Regional climate information— application of regional and global climate models. Climate Evaluation and projections. Climate Change 2001: The Scientific Dyn., 35, 193–212. Basis, J. T. Houghton et al., Eds., Cambridge University Press, Pausader, M., D. Bernie, S. Parey, and M. Nogaj, 2011: Computing 583–638. the distribution of return levels of extreme warm temperatures ——, C. S. Brodeur, and G. T. Bates, 1994: Regional climate for future climate projections. Climate Dyn., 38, 1003–1015, change scenarios over the United States produced with a nes- doi:10.1007/s00382-011-1032-2. ted regional climate model. J. Climate, 7, 357–399. Rojas, M., and A. Seth, 2003: Simulation and sensitivity in a nested Gold, L. W., 1967: Influence of surface conditions on ground modeling study for South America. Part II: GCM boundary temperature. Can. J. Earth Sci., 4, 199–208. forcing. J. Climate, 16, 2454–2471.

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC 6286 JOURNAL OF CLIMATE VOLUME 25

Sato, T., F. Kimura, and A. Kitoh, 2007: Projection of global Change 2007: The Physical Science Basis. Cambridge Uni- warming onto regional precipitation over Mongolia using versity Press, 996 pp. a regional climate model. J. Hydrol., 333, 144–154. van der Linden, P., and J. F. B. Mitchell, Eds., 2009: ENSEMBLES: Seth, A., and M. Rojas, 2003: Simulation and sensitivity in a Climate change and its impacts: Summary of research and nested modeling study for South America. Part I: Reanalysis results from the ENSEMBLES project. Met Office Hadley boundary forcing. J. Climate, 16, 2437–2453. Centre Rep., 160 pp. [Available online at http://ensembles- ——, S. A. Rauscher, S. J. Camargo, J.-H. Qian, and J. S. Pal, 2007: eu.metoffice.com/docs/Ensembles_final_report_Nov09.pdf.] RegCM3 regional climatologies for South America using re- Wang, Y., L. R. Leung, J. L. Mcgregor, D.-K. Lee, W.-C. Wang, analysis and ECHAM global model driving fields. Climate Y. Ding, and F. Kimura, 2004: Regional climate modeling: Dyn., 28, 461–480, doi:10.1007/s00382-006-0191-z. Progress, challenges and prospects. J. Meteor. Soc. Japan, 82, Shukla, J., and K. R. Saha, 1974: Computation of non-divergent 1599–1628. streamfunction and irrotational velocity potential from the Wigley, T. M. L., 2009: The effect of changing climate on the fre- observed winds. Mon. Wea. Rev., 102, 419–425. quency of absolute extreme events. Climatic Change, 97, 67–76. Skamarock, W. C., and Coauthors, 2008: A description of the Wu, W., and A. H. Lynch, 2000: Response of the seasonal carbon Advanced Research WRF version 3. NCAR Tech. Note cycle in high latitudes to climate anomalies. J. Geophys. Res., NCAR/TN-4751STR, 113 pp. [Available online at http:// 105 (D18), 22 897–22 908. www.mmm.ucar.edu/wrf/users/docs/arw_v3.pdf.] ——, ——, and A. Rivers, 2005: Estimating the uncertainty in Solomon, S., D. Qin, M. Manning, M. Marquis, K. Averyt, M. M. B. a regional climate model related to initial and lateral boundary Tignor, H. L. Miller Jr., and Z. Chen, Eds., 2007: Climate conditions. J. Climate, 18, 917–933.

Unauthenticated | Downloaded 09/28/21 02:39 PM UTC