Bias corrections of global models for regional climate simulations of high-impact weather
- 4.9k Downloads
All global circulation models (GCMs) suffer from some form of bias, which when used as boundary conditions for regional climate models may impact the simulations, perhaps severely. Here we present a bias correction method that corrects the mean error in the GCM, but retains the six-hourly weather, longer-period climate-variability and climate change from the GCM. We utilize six different bias correction experiments; each correcting different bias components. The impact of the full bias correction and the individual components are examined in relation to tropical cyclones, precipitation and temperature. We show that correcting of all boundary data provides the greatest improvement.
KeywordsCyclone Tropical Cyclone Bias Correction Community Climate System Model Version Bias Correction Method
Global circulation models (GCMs) provide the basis of our capacity to simulate, understand and predict climate variability and change. These models are based on established physical laws and have proven fidelity for assessing changes to global quantities (Randall et al. 2007; Anderson et al. 2004; Collins et al. 2004; Déqué et al. 1994; Flato et al. 2013; Pope et al. 2000; Roeckner et al. 2003). However, GCMs typically are of too a coarse resolution to directly infer climatology of high-impact weather at local scales and it is common to downscale over regions of interest using statistical techniques or nested regional climate models (RCMs). Unfortunately, biases that may be acceptable at global scales can be problematic for these downscaling applications to regional and extreme weather climate scales (e.g. Liang et al. 2008; Ehret et al. 2012; Xu and Yang 2012; Done et al. 2013).
One approach is to apply combined bias-correction and downscaling methods directly to the GCM data in the form of empirical relationships between the large scales and high impact weather (Camargo et al. 2007; Walsh et al. 2007; Bruyère et al. 2012). An obvious shortcoming of this method is that this bias correction is applied independently across time, space and variable, without taking into account feedback mechanisms between atmospheric processes. It is important to also remember that the GCM data were generated at a coarse resolution, where local processes and terrain heterogeneity were not taken into account. It also is possible that statistical downscaling methods developed on past climate might not hold true under climate change conditions.
An alternative, widely-used approach is to nest a RCM within GCM boundary conditions (Laprise et al. 2008; Bender et al. 2010; Knutson et al. 2007, 2008; Walsh et al. 2004; Done et al. 2013). Because of their smaller domain, RCMs can operate at higher resolution than GCMs to enable simulation of much finer scale features, which are required for assessment of many extreme weather phenomena. One shortcoming of this approach is the transmission of GCM biases through the RCM lateral and lower boundaries, which may have a severe impact on the interior climate (e.g. Warner et al. 1997; Done et al. 2013).
One approach to correcting these regional biases is to apply a correction to the RCM output (e.g. Dosio and Paruolo 2011). This approach suffers from the same limitations as the aforementioned statistical bias correction of GCMs and has the additional complication that GCM biases may irretrievably change—or even destroy—the high-impact weather signal of interest (Ehret et al. 2012; Done et al. 2013).
An alternative bias-correction approach is to construct boundary conditions from a current climate reanalysis plus a climate change perturbation, a technique known as pseudo-global-warming (Schär et al. 1996; Rasmussen et al. 2011). This approach is simple to apply and takes advantage of the improved ability of GCMs to simulate trends compared to absolute climates (Randall et al. 2007). However, there are substantial disadvantages arising from the inherent assumption of no change in synoptic and climate variability. Biases from current GCM simulations also may change into the future and alias into the imposed climate change perturbation.
A more recent approach takes advantage of the strengths in both the GCMs and RCMs by performing bias correction on the GCM boundary data. Using a common bias-correction method applied to all variables provides more balanced atmospheric conditions to drive the RCM. Variance is free to change into the future (within the resolution constraints of the driving GCM) and the RCM has the freedom to develop its own interior solution within the bias corrected boundary data. A number of variations on this theme have been attempted including; correcting bias in the mean and variance (Xu and Yang 2012), quantile–quantile mapping (Colette et al. 2012), and feature location correction (Levy et al. 2012). White and Toumi (2013) tested both the mean bias correction and quantile–quantile mapping methods, and found that the mean bias correction method is a more reliable and accurate method compared to the quantile–quantile mapping method.
In this study we investigate the applicability of bias correcting the boundaries in RCM simulations of high-impact weather. The environments for Atlantic tropical cyclones and North American summer precipitation and temperatures are used as examples, but the results are applicable to a wide range of weather extremes.
In Sect. 2.2 a bias correction method for GCM boundary conditions is developed that successfully reproduces the statistics of high-impact weather in the regional climate simulation. We then develop physical insight into the role of bias correction for the downscaled regional climate in Sect. 3.2 through analysis of the simulation sensitivity to bias correction of specific variables or sets of variables in the driving data. The results are presented in Sect. 3. Section 4 contains our conclusions.
2.1 Models and data
The GCM used here is the Community Climate System Model version 3 (CCSM3; Collins et al. 2006) run at T85 (~1.4° atmosphere and 1° ocean). CCSM3 is a coupled climate model with components representing the atmosphere, ocean, sea ice, and land surface as described in detail in Collins et al. (2006). The simulation was initialized in 1950 and run under twentieth century emissions.
The atmospheric reanalysis used to bias correct the CCSM3 data is the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCAR) Reanalysis Project (NNRP, Kalnay et al. 1996). Analysis SST data utilize the merged Hadley Centre and NOAA’s optimum interpolation (OI) SST data set (Hurrell et al. 2008).
2.2 Bias correction
Note that here we are using the term bias in the context of systematic errors in the model, as compared to some base ‘truth’ (specifically the NNRP). We also partially consider the ‘bias’ that may arise from sampling from relatively short time periods within a climate that varies on long and short time scales (e.g. Maraun 2012). This is accomplished through our use of a limited set of longer simulations. A related ‘bias’ arising from the essentially nonlinear nature of climate, which means that more than one internal solution may result from the same imposed boundary conditions is the subject of a separate study.
The mean climatological component is defined over a 20-year base period (to smooth out influence of short-period variations such as El Niño). Twenty years was chosen to avoid inclusion of any significant climate trends though we acknowledge that this may alias some decadal oscillations into the bias correction.
These bias-corrected climate data thus combine a seasonally-varying climate, as provided by NNRP and OI-SST, with the six-hourly weather from the GCM. This approach also retains the GCM longer-period climate variability and climate change.
Equation 3 is applied to all the variables required to generate surface and lateral boundary conditions for NRCM: zonal and meridional wind, geopotential height, temperature, relative humidity, sea surface temperature and mean sea level pressure.
3.1 CCSM bias corrections
We next examine the sensitivity of the revised climate to the choice of the base period arising from a possible non-stationarity of the bias. Choosing different base periods (1960–1979, 1965–1984, 1970–1989, and 1975–1994) result in nearly identical bias corrections over the entire simulation period (Fig. 4). This increases confidence that the bias will not change substantially in the future. The validity of this assumption is further addressed in the climate projection discussion.
The dashed red line in Fig. 4 shows the affect of including variance bias correction in addition to the mean correction (following the method of Xu and Yang 2012). Clearly, accounting for variance in addition to mean bias makes only a marginal difference. This is supported by the NRCM downscaling with mean bias-only correction. For current climate, the variance in 500 hPa temperature over the MDR is 0.88 for NNRP and 0.62 for the CCSM3 model. Yet, the NRCM with mean bias correction has a variance of 0.96, indicating that it is effectively spinning up realistic internal variance without the need for additional variance bias correction.
3.2 NRCM downscaling
Sensitivity to choice of variables used for bias correction is examined using a series of NRCM simulations with the following boundary conditions: raw CCSM3 data (NO_BC); bias corrected winds only (BC_UV); bias corrected SST only (BC_SST); bias correction of both the winds and SST (BC_SSTUV); all variables excluding SST corrected (BC_NoSST); and all boundary data corrected (BC). These simulations cover a 7 months period from May 1 to Dec 1, for an arbitrarily chosen year representative of current climate. Note that for the surface only the SST is prescribed, the land is free to evolve in NRCM.
Analysis of these sensitivity runs uses the ASO average large-scale flow, however, since the anomalies in a single year may not be representative of the anomaly over a longer period, we also compare the NO_BC and BC cases for a total of 11 years, using the first year as a spin-up year, and years 2–11 for the analysis period. These simulations are referred to as NO_BC10 and BC10.
3.2.1 Atlantic tropical cyclone environment
Winds (BC_UV, Fig. 5b) or SST (BC_SST, Fig. 5c) alone both reduce the shear bias substantially. This is expected: correcting the SST bias removes the anomalous Walker circulation that generates the strong vertical shear; applying wind corrections at the boundaries also suppresses this Walker circulation in the regional model. Notably, although both brought about a similar reduction in shear magnitude, leaving the cold SST in place (BC_UV) still suppresses all cyclone activity, whereas the warm oceans (BC_SST) combined with reduced vertical shear generates three cyclones (not shown).
Combining SST and wind corrections (BC_SSTUV, Fig. 5d) improves the shear values comparable to the sum of the shear improvement through correcting SST and winds independently (Fig. 5b, c) This improvement results in the genesis of five cyclones, some of which form in the MDR.
Applying a bias correction to all boundary variables excluding SST (BC_NoSST, Fig. 5f) indicates the importance of getting the surface correct; the shear increases substantially and only 2 cyclones develop.
Longer period simulations for NO_BC10 and BC10 produces similar results to those of the single season simulations (NO_BC and BC), with ASO mean shear values over the North Atlantic too high for NO_BC10, and realistic values being simulated for BC10 (Fig. 5g, h). These longer simulations also produce similar annual cyclone numbers to those for single seasons: ~1.5 for NO_BC10 that developed too far north (Fig. 3a), and ~10 for BC10 with much more realistic genesis locations and storm tracks (Fig. 3b).
3.2.2 North American summer precipitation and temperature
Biases in GCMs are transferred through lateral and lower boundary conditions to RCMs, impacting the downscaled results, sometimes severely. Here we examined application of a bias correction method that corrects the seasonally-adjusted mean error in the GCM but retains the weather variance, longer-period climate variability, and climate change from the GCM. The correction is nearly independent of the period over which it is developed, giving confidence that such corrections will be somewhat invariant in future projections. Corrections to both mean and variance were considered, but the variance correction made very little difference, as the NRCM was able to successfully reproduce the observed variance internally.
The impact of both the full bias correction and individual components were examined in relation to simulations of the North Atlantic tropical cyclone environment and North American precipitation and temperatures.
A consistent result was achieved for all three components. Using the uncorrected climate model boundary conditions resulted in substantial errors, including suppressing almost all tropical cyclones. Applying the full correction to all boundary variables substantially improved the simulations compared to observations: simulated tropical cyclones had realistic spatial distributions and annual frequency; North American precipitation distribution and magnitude was substantially improved; and the probability distribution of surface temperatures moved from a distinct cold bias to a better approximation of observations.
Correcting individual and groups of boundary variables in isolation indicates that the biggest single improvement came through correcting the SST. Correcting both SST and winds at the horizontal boundary provided the majority of the improvement. But in all cases correcting all boundary variables in a consistent manner was better than correcting any subset of variables.
These findings suggest that application of a relatively simple bias correction to the GCM boundary conditions for a RCM—in which only seasonal variability is included—may suit many regional climate applications. A particular strength of this approach is that it enables current-climate variability within the GCM (weather, decadal and climate change) to vary with future simulations while correcting for the major biases that can cause serious issues for regional climate downscaling.
NCAR is funded by the National Science Foundation and this work was partially supported by the Research Partnership to Secure Energy for America (RPSEA) and NSF EASM Grants AGS-1048841 and AGS-1048829.
- Collins W, Rasch PJ, Boville BA, McCaa J, Williamson DL, Kiehl JT, Briegleb BP, Bitz C, Lin S-J, Zhang M, Dai Y (2004) Description of the NCAR Community Atmosphere Model (CAM 3.0). NCAR Technical Note NCAR/TN-464 + STR. doi: 10.5065/D63N21CH
- Dosio A, Paruolo P (2011) Bias correction of the ENSEMBLES high-resolution climate change projections for use by impact models: evaluation on the present climate. J Geophys Res 116. doi: 10.1029/2011JD015934
- Flato G et al (2013) Evaluation of climate models. Contribution of Working Group I to the fifth assessment report of the intergovernmental panel on climate changeGoogle Scholar
- Holland GJ, Done JM, Bruyère CL, Cooper C, Suzuki A (2010) Model investigations of the effects of climate variability and change on future Gulf of Mexico Tropical Cyclone Activity. Paper OTC 20690 presented at the Offshore Technology Conference, Houston, Texas, 3–6 MayGoogle Scholar
- Hong S-Y, Lim J-OJ (2006) The WRF single-moment 6-class microphysics scheme (WSM6). J Korean Meteorol Soc 42:129–151Google Scholar
- Randall DA, Wood RA, Bony S, Colman R, Fichefet T, Fyfe J, Kattsov V, Pitman A, Shukla J, Srinivasan J, Stouffer RJ, Sumi A, Taylor KE (2007) Climate models and their evaluation. In: Climate change 2007: the physical science basis, contribution of working group I to the fourth assessment report of the intergovernmental panel on climate change. Cambridge University Press, CambridgeGoogle Scholar
- Roeckner E et al (2003) The atmospheric general circulation model ECHAM5. Part I: model description. MPI Report 349, Max Planck Institute for Meteorology, Hamburg, pp 127Google Scholar
- Skamarock W, Klemp JB, Dudhia J, Gill DO, Barker D, Duda MG, Huang X-Y, Wang W (2008) A description of the advanced research WRF Version 3. NCAR Technical Note NCAR/TN-475+STR. doi: 10.5065/D68S4MVH
Open AccessThis article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.