This study documents and investigates biases in simulating summer surface air temperature (SAT) variability over the continental United States in the Atmospheric Model Intercomparison Project (AMIP) experiment from phase 5 of the Coupled Model Intercomparison Project (CMIP5). Empirical orthogonal function (EOF) and multivariate regression analyses are used to assess the relative importance of circulation and the land surface feedback at setting summer SAT over a 30-yr period (1979–2008). Regions of high SAT variability are closely associated with midtropospheric highs, subsidence, and radiative heating accompanying clear-sky conditions. The land surface exerts a spatially variable influence on SAT through the sensible heat flux and is a second-order effect in the high-variability centers of action (COAs) in observational estimates. The majority of the AMIP models feature high SAT variability over the central United States, displaced south and/or west of observed COAs. SAT COAs in models tend to be concomitant and strongly coupled with regions of high sensible heat flux variability, suggesting that excessive land–atmosphere interaction in these models modulates U.S. summer SAT. In the central United States, models with climatological warm biases also feature less evapotranspiration than ERA-Interim but reasonably reproduce observed SAT variability in the region. Models that overestimate SAT variability tend to reproduce ERA-Interim SAT and evapotranspiration climatology. In light of potential model biases, this analysis calls for careful evaluation of the land–atmosphere interaction hot spot region identified in the central United States. Additionally, tropical sea surface temperatures play a role in forcing the leading EOF mode for summer SAT in models. This relationship is not apparent in observations.
There are many socioeconomic consequences of exceptionally warm summers in the continental United States. Extreme summer temperatures can cause hundreds of millions of dollars in crop damage (NOAA/NWS 2015) and strain water resources in drought-prone regions like the Southwest. In recent years, excessive heat has become the number-one weather-related killer, surpassing hurricanes, floods, tornadoes, and lightning strikes (Thacker et al. 2008). Around 500 heat-related deaths occurred in a one-week period during the 1995 Chicago heat wave (Whitman et al. 1997). Accurate forecasts of U.S. summer surface air temperature (SAT) will become more crucial as the climate system warms; it is anticipated that North American heat waves will become more intense, occur more frequently, and persist longer in duration in the second half of the twenty-first century (Meehl and Tebaldi 2004; Ganguly et al. 2009).
General circulation models (GCMs) must be shown to simulate the current mean state and variability of summer SAT to bolster confidence in future projections of summer climate, as processes governing SAT operate in response to increased CO2 (Xie et al. 2015). SAT climatology is determined by the distribution of solar insolation and orography. Interannual variations in SAT about the mean state are driven by atmospheric circulation and are much more challenging to simulate spatially than global mean SAT. Because the atmosphere loses memory of its initial state on subseasonal time scales, models must rely on the slowly varying states of the ocean and land surface for predictive skill (Koster et al. 2006). Even if the model atmosphere responds perfectly to forcings at the surface, near-term (from annual to decadal) predictive skill is diminished by internal variability, which is intrinsic to chaotic systems and inherently unpredictable because of its random temporal phase (Deser et al. 2012). Further uncertainty arises from computational limitations that require simplified numerical formulations and parameterizations of subgrid-scale physical processes.
Despite the complexity of the climate system and the practical limits of models, atmospheric GCMs forced with observed sea surface temperatures (SSTs) are generally skillful at capturing both the mean state and variability of winter SAT over the continental United States (Sheffield et al. 2013; Deser et al. 2014). In the winter [December–February (DJF)], tropical forcings are strong. SST anomalies associated with El Niño–Southern Oscillation (ENSO) peak and exert influence on the extratropical atmosphere through eddy-driven changes in mean meridional circulation and stationary planetary waves in the Pacific–North American (PNA) sector (Seager et al. 2003; Kushnir et al. 2010; Zhou et al. 2014). The distribution of solar radiation in winter sets up a strong meridional temperature gradient in the Northern Hemisphere. In the free troposphere, the meridional temperature gradient sustains tropical–extratropical teleconnections by maintaining the zonal jet stream. At the surface, atmospheric circulation anomalies dictate SAT patterns by advecting air parcels across these gradients. The predictability of U.S. winter SAT is enhanced because of the SST-forced component of atmospheric variability.
In contrast during the summer [June–August (JJA)], U.S. SAT variability is, to first order, the product of internal variability in the midlatitude atmosphere (Wallace et al. 1995, 2015). The strong temperature gradient that maintains the zonal structure of the westerly jet in the winter weakens in the summer. Meanders in the upper-tropospheric flow create high pressure systems that can persist for weeks in the free troposphere (Charney and DeVore 1979; Egger 1978). Subsiding air beneath these blocking anticyclones warms adiabatically and is accompanied by clear skies and light winds, which are conditions conducive to summer warming (Meehl and Tebaldi 2004; Lau and Nath 2012). At the surface, the semipermanent Pacific and Bermuda highs direct northwesterly flow into the Pacific Northwest and southeasterly flow over the eastern seaboard. The circulation reaching the continental interior also is anticyclonic in structure, with a southerly jet along the Sierra Madre that brings moisture from the Gulf of Mexico to the Great Plains (Nigam and Ruiz-Barradas 2006). Thermal anticyclones also tend to form in July over cold polar regions, like the snow and ice fields of Canada, and propagate southeast over the continental United States (Zishka and Smith 1980). The primary North American anticyclone track lies over the Great Lakes and north-central United States (Davis et al. 1997).
Although tropical SST is not the dominant driver of U.S. summer SAT variability (Barlow et al. 2001), many studies have noted a connection between tropical SST forcing regions on the midlatitude summer atmosphere (e.g., Shaw and Voigt 2015; Arblaster and Alexander 2012; Pegion and Kumar 2010; Schubert et al. 2009; Lau et al. 2006; Sutton and Hodson 2005; McCabe et al. 2004; Higgins et al. 2000). Protracted La Niña events have been associated with persistent droughts, notably the large-scale midlatitude drying between 1998 and 2002 (Hoerling and Kumar 2003), the Texas drought and heatwave of 2011 (Hoerling et al. 2013), and the Dust Bowl of the 1930s (Schubert et al. 2004; Seager et al. 2005). Some studies (Ding and Wang 2005; Ding et al. 2011) suggest that model skill derives from the predictable zonal mean component of the circumglobal teleconnection. Wang and Ting (1999) and Ting (1994) showed that nonlinear interactions among monsoonal heating-induced flows link U.S. summer climate to convective regions near Asia in the NCEP–NCAR reanalyses and GCMs. Kushnir et al. (2010) showed that warm SSTs in the tropical North Atlantic can exert an “upstream” control on SAT by weakening the subtropical North Atlantic anticyclone, driving northerly cold advection and anomalous subsidence over North America. SSTs associated with decadal modes of variability, such as the Atlantic multidecadal oscillation (AMO) and Pacific decadal oscillation (PDO), can influence U.S. summer SAT by modulating the Great Plains low-level jet, which brings warm, moist air from the Gulf of Mexico to the central United States (Weaver 2013).
Land surface conditions can also affect summer SAT variability (Seneviratne et al. 2006; Lorenz et al. 2010), in not too wet (energy limited) and not too dry (moisture limited) soil moisture (SM) “hot spot” regions like the central United States (Koster et al. 2004a,b, 2006; Dirmeyer 2011; Berg et al. 2014). To describe relationships between the atmosphere and the land surface, we use “interaction” to refer to a general association between two variables, “coupling” to refer to the degree one variable controls another, and “feedback” to refer to a two-way coupling, following Lorenz et al. (2015). For example, soil moisture–climate coupling refers to the soil moisture control on SAT variability through the sensible heat flux. Zhang et al. (2008) investigated the role of the land–atmosphere coupling on U.S. summer climate variability using regional climate models and found that the strong coupling between soil moisture and daily mean temperature contributed 30%–60% of the total interannual SAT variance in the southwestern, north-central, and southeastern United States.
Land–atmosphere interactions occur primarily through the surface energy balance. Terrestrial water acts to partition outgoing energy into latent heat flux at the expense of sensible heat flux QH, resulting in cooler SAT (Seneviratne et al. 2010). When atmospheric conditions are stable under anticyclones, the positive feedback between the land and atmosphere allows the sensible heat flux to enhance warm temperatures and increase SAT variability. Warm, dry conditions come with a high atmospheric demand for water, reducing soil moisture. Without water to evaporate, more outgoing energy is available to heat the atmosphere (Seneviratne et al. 2010; Miralles et al. 2012). Soil desiccation has been shown to contribute to mega-heat waves in Europe (Fischer et al. 2007a,b; Miralles et al. 2014), but fewer studies have assessed the impact of the land surface on severe temperature extremes in the United States (Diffenbaugh et al. 2005). Assessments of North American hydroclimate variability suggest that land–atmosphere interactions are overemphasized in models during the warm season and report that local evaporation in the central United States can be up to 4 times larger in models than in observationally constrained estimates (Ruiz-Barradas and Nigam 2005; Wu and Dickinson 2005; Ruiz-Barradas and Nigam 2006, 2013). Mueller and Seneviratne (2014) also document land hydrological and climate biases and highlight the overestimation (underestimation) of temperature (evaporation) in central U.S. climatology in CMIP5 simulations. Quantification of land surface coupling during the warm season is cited as a high research priority (Perkins 2015), and the land surface is thought to be key in improving model forecasts of SAT on seasonal time scales (Koster et al. 2011).
While the circulation and land surface controls on summer SAT variability are generally recognized, gaps exist in quantifying the relative importance of the involved physical mechanisms (Perkins 2015) and in systematically evaluating summer variability in climate models. Models are important for prediction, projection, and attribution, but careful investigation and documentation of their skills and errors are required before their output can be used with confidence. The present study evaluates the performance of atmospheric models in simulating summer SAT variability over the continental United States; this is one of few such CMIP5 assessments to our knowledge. Evaluations of model SAT benefit from the excellent accuracy and coverage of U.S. SAT measurements, which constrain gridded products from atmospheric reanalyses. In addition, summer SAT is strongly affected by land–atmosphere interactions, which in turn can be evaluated using SAT. To document model skill, we present maps of U.S. summer SAT variability and contrast them with their winter, higher-skill counterparts. In summer, the regions of high SAT variance differ spatially between observations and models, a discrepancy that we suggest is due to an excessively strong central U.S. land–atmosphere interaction in models. To evaluate this hypothesis, we assess the relative importance of circulation and the land surface at setting summer SAT and examine the effects of the SST-forced and internal components of variability in the midlatitude summer atmosphere. This study assesses the nature and sources of U.S. summer SAT variability in the AMIP experiment, with the aim of validating its predictive capability.
We organize the remaining sections as follows. The observational datasets, model experiments, and methods are described briefly in section 2. Observational and model estimates of U.S. summer SAT climatology and variability are documented in section 3, and model skill is contrasted between summer and winter. In section 4, the relative importance of circulation and the land surface on observed and simulated JJA SAT variability is evaluated. The land surface’s role in climatological SAT model biases is also assessed. Other potential causes for model biases are explored in section 5, including differences in SST forcing regions and the lead–lag response to ENSO variability. A summary of findings and discussion of the complexities of the land–atmosphere interaction are given in section 6.
2. Observational estimates, simulations, and methods
a. Observational estimates
We use the European Centre for Medium-Range Weather Forecasts (ECMWF) interim reanalysis (ERA-Interim) for monthly mean SAT, 500-hPa geopotential height Z500, and surface fluxes (Dee et al. 2011) on a 1.5° latitude × 1.5° longitude grid from 1979 to 2008. For verification, we use NCEP North American Regional Reanalysis (NARR; Mesinger et al. 2006). Spatial patterns of variability are similar for the reanalysis products, so for conciseness we present maps based on the ERA-Interim fields. (Relationships between SAT, Z500, and the land surface in Figs. 5 and 6, described in greater detail below, are presented for both ERA-Interim and NARR.)
ERA-Interim 30-yr JJA SAT, Z500, QH climatology, and regions of the United States used to describe the spatial structure of the considered fields are shown in Fig. 1. For clarity, we will refer to ERA-Interim SAT and Z500 as observed because the fields are well measured and thus well constrained in reanalysis products. However, it is important to acknowledge shortcomings in reanalysis surface flux estimates. By design, water and energy budgets are not closed in reanalysis products, and surface flux observations are spatially sparse and temporally limited. Because surface fluxes are not well constrained, studies recommend exercising great caution when using them (e.g., Trenberth et al. 2011; Ferguson et al. 2012). For this reason, we will consider interannual variations and spatial patterns of ERA-Interim surface flux estimates in this assessment but not base any conclusions on the magnitude of the fluxes themselves.
b. AMIP experiments
For comparison with reanalysis fields, we use monthly mean SAT, Z500, and surface fluxes from 14 models in the CMIP5 AMIP experiment (Taylor et al. 2012). In each model, AMIP runs are initialized months prior to January 1979 with different atmospheric states to form initial condition ensembles with between 2 and 10 ensemble members. All 14 initial condition ensembles are combined to form an all-model ensemble with 58 members. In addition to time-evolving external forcings, the AMIP runs are forced by observed SST and sea ice states. Land surface states interact with the atmosphere and vary among modeling groups. Salient information about each model used is given in Table 1, with further details given in Table 9.A.1 of the Intergovernmental Panel on Climate Change (IPCC) Fifth Assessment Report (AR5; Flato et al. 2013). The CMIP5 all-model ensemble climatologies are shown in Fig. 1 (right). Models vary in their ability to capture observed SAT climatology over the continental United States, and many feature warm biases in the central United States. We will consider model climatological biases as we assess model performance in depicting SAT variability.
SAT variability is characterized as the standard deviation in seasonally averaged, ocean-masked SAT fields within 24°–52°N and 45°–145°W. Winter averages are formed over DJF and summer averages over JJA. For the AMIP ensembles, the interannual variance of the 30-yr record (1979–2008) is computed for each member, then averaged over the ensemble and square rooted to avoid suppressing the internal component of variability. To emphasize spatial variability, the standard deviations depicted in Figs. 2 and 3 are scaled by their domain-average standard deviations, which are reported in the lower right corner of each panel. Model fields are plotted on their native grids in Figs. 2 and 3 and are regridded to the reanalysis grid (1.5° × 1.5°) for the remaining figures.
Empirical orthogonal function (EOF) analysis is used to describe dominant modes of JJA SAT variability. EOFs are computed after removing a mean and linear trend from the JJA SAT time series and applying a square root cosine latitude spatial weighting. Each EOF spatial mode has an associated temporal coefficient, or principal component (PC). For the AMIP ensembles, an EOF analysis is performed on both 1) the ensemble mean and 2) the concatenated n by 30-yr record of the JJA averaged fields, where n is the number of member runs for each model. Averaging across ensemble members suppresses the internal variability present in each individual realization, emphasizing the model’s response to external forcing (Deser et al. 2015). Concatenating realizations preserves both the forced and internal components that comprise the observed variability. Correlations of the leading JJA SAT principal component (PC1) with Z500 and QH are computed for each grid point. This method highlights regions of temporal covariability between SAT and potential dynamic (circulation) and thermodynamic (surface flux) controls. The 95% significance is determined by assuming a Student’s t distribution with N − 2 degrees of freedom, where N is the number of years in the record. When anomalies are spatially similar, temporally coincident, and physically consistent, we infer relationships between fields.
A local multivariate regression analysis and calculation of the soil moisture–climate coupling proposed by Dirmeyer (2011), Dirmeyer et al. (2013a,b), and Dirmeyer et al. (2014) are used to further examine relationships between JJA SAT, Z500, and QH. JJA SAT, Z500, and QH are averaged in boxed regions within the SAT centers of action (COAs), and the resulting time series are normalized by their standard deviations. We interpret the Z500 and QH regression coefficients bZ and bQ as indicators that SAT variability within the COA is associated with atmospheric circulation and local sensible heat flux, respectively.
The pathway for SM anomalies to influence the atmosphere in the COA is assessed through the Dirmeyer soil moisture–climate coupling metric described in Lorenz et al. (2015):
The metric accounts for the terrestrial segment (SM–QH) and atmospheric segment (QH–SAT) of the land surface feedback through correlations ρ and is suppressed in regions of low SAT variability by (Dirmeyer 2011; Dirmeyer et al. 2013a,b, 2014). We evaluate the relationship between mean SAT and mean evapotranspiration with a second soil moisture–temperature coupling estimate proposed by Seneviratne et al. (2006), ρ(evapotranspiration, SAT), which gives insight into climatological biases in the central United States. Finally, we consider the role of the global ocean through remote atmospheric pathways on U.S. summer SAT by regressing SSTs and 300-hPa heights Z300 on JJA SAT PC1.
3. SAT variability—DJF versus JJA
In boreal winter, the jet stream is strong and tropical SST anomalies peak, providing the foundation for teleconnections such as the PNA pattern to influence U.S. SAT. Standard deviation maps (Fig. 2) illustrate how the various models depict spatial patterns of winter SAT variability. The 30-yr DJF SAT climatology is shown in black contours, with values ranging from −15° to 15°C in 5°C increments. Both observed and modeled winter SAT climatologies have a zonal-banded structure, with the exception of lower temperatures at high altitude over the Rocky Mountains. From year to year, winter temperatures fluctuate most1 by up to 2.9°C in a region that extends down from western Canada over the north-central United States. Because advection of the mean SAT gradient by anomalous circulation dominates over other mechanisms (Thompson and Wallace 2000), this feature is captured in most AMIP models and the spatial structure of DJF SAT variability is well represented in CMIP5. The magnitude of DJF SAT variability in the domain tends to be higher in models than in observations, which is due to higher values of variability in the southeastern and south-central United States. Models with more members, such as CSIRO Mk3.6.0, have higher skill in reproducing observed variability, illustrating the power of larger ensembles to represent realistic internal variability in the presence of model error (Kay et al. 2015).
Model simulations of SAT variability over the continental United States are less similar to observational estimates in the summer than they are in the winter (Fig. 3). In the summer, observed climatological temperatures range from about 10°C over the Rocky Mountains to 35°C in the Mojave and Sonoran Desert regions of the Southwest. Aside from these regions, the observed JJA SAT climatology also has a zonal structure. A major discrepancy between observed and modeled JJA SAT climatology is a meridional warm bias spanning the central United States, which occurs in 9 of the 14 models and in the all-model ensemble [Fig. 1 (top right) and Fig. 3]. Causes of this warm bias are investigated in section 4.
Interannual variations in SAT are less localized and less pronounced in summer than in winter. ERA-Interim JJA temperatures vary by 1°–1.2°C over much of the United States, with lower variability (~0.5°C) along the southeastern seaboard due to the thermal inertia of the Gulf Stream. The largest JJA SAT variability (1.6°C) occurs in the southern northwest Great Basin cold desert and north-central Great Plains, on either side of the Rocky Mountain range. This pattern of summer SAT variability is poorly simulated in most of the AMIP models, which tend to feature regions of high variability in the central United States. Most models feature either a large swath of variability over the north-central United States [e.g., BCC_CSM1.1(m), CSIRO Mk3.6.0, and CCSM4] or a highly localized COA in the south-central United States (e.g., CMCC-CM, GISS-E2-R, and MRI-CGCM3). In MRI-CGCM3, SAT variability over Kansas and the Oklahoma Panhandle exceeds 2.1°C, doubling observed values in the region.
Models with JJA SAT COAs in the south-central United States tend to have SAT climatologies that are similar to observed. Models with JJA SAT COAs in the north-central United States have pronounced warm biases in the south-central United States. We conjecture that model skill in simulating U.S. summer SAT will improve if the causes of the warm bias and spurious variability in the central United States are identified and managed.
4. Relative importance of circulation versus land–atmosphere coupling
We examine summer SAT interannual variability in the AMIP models by considering the leading EOF mode (EOF1; Figs. 4a,d). The associated time series (PC1) is correlated with JJA Z500 (Figs. 4b,e) and QH (Figs. 4c,f) time series at each grid point. This method is used to demonstrate association between SAT and the main physical processes that are expected to contribute to its variability and to identify the relative importance of these contributions in the central U.S. model COAs. We focus on EOF1 because it features a dominant COA centered in the north-central United States and captures a sizable percentage (38%) of the domain-integrated, normalized variance in the 30-yr reanalysis record. Model EOF1s feature the regions of high SAT variability (Fig. 3) that we are interested in evaluating. The boxed region (9° × 15°) in each panel corresponds to the region of highest SAT variability in each EOF1.
In the observed COA, SAT interannual variability correlates highly with Z500, indicating the importance of midtropospheric highs in establishing warm SATs in this region (Meehl and Tebaldi 2004; Lau and Nath 2012). ERA-Interim QH and SAT are marginally correlated in the COA, positively correlated in the southeastern and south-central United States, and negatively correlated in the Southwest. ERA-Interim correlations that exceed ±0.36 are different than zero at the 95% significance level and are stippled in Fig. 5. Significant positive SAT–QH correlations in the eastern United States indicate a prominent land surface feedback, with warm SATs in the region amplified by soil desiccation (Findell and Eltahir 2003). Correlations in the southwestern portion of the United States fall below the 95% significance threshold likely because the region is moisture limited (Seneviratne et al. 2010) so surface fluxes are rarely large enough to influence SAT variability.
In the AMIP experiment, all but two models have JJA SAT EOF1s that resemble the observed north-central U.S. COA pattern. However, model COAs tend to extend either farther west (CanAM4, CSIRO Mk3.6.0, MIROC5, and IPSL-CM5A-LR) or south (CMCC-CM, GFDL CM3, MPI-ESM-MR, and NorESM1-M) than observed. The two model exceptions have COAs in the Southwest (GISS-E2-R) and south-central United States (MRI-CGCM3), similar to their respective variability patterns in Fig. 3. As in observations, models tend to capture the relationship between surface temperature and midtropospheric circulation, with high SAT–Z500 correlations in the COAs. However, the highest model SAT–Z500 correlations are not centered within the boxed region of highest SAT variability in all cases (e.g., CMCC-CM, HadGEM-A, and NorESM1-M).
A notable difference between the models and ERA-Interim is the high model SAT–QH correlations found within the COA regions (boxed regions in Figs. 4c,f). Moreover, the general shape and position of the COA closely resemble the QH projection. In contrast, ERA-Interim shows weak SAT–QH correlations in the north-central U.S. COA. We interpret the high SAT–QH model correlations as an indication that spurious SAT variability in the AMIP models may be due to QH fluctuations contributing an unrealistically high land–atmosphere coupling compared to observational estimates.
To examine further the relationships between JJA SAT and Z500 and QH, we apply a multivariate regression analysis within SAT COAs. Each variable is spatially averaged within the boxed regions indicated in Fig. 4 and normalized by the standard deviation. We regress SAT onto Z500 and QH to obtain regression coefficients bZ and bQ, respectively, with confidence intervals computed assuming serially independent time series. If bZ exceeds bQ, we categorize variability in the SAT COA as more closely associated with atmospheric circulation patterns than with the land surface (green region in Fig. 5a). In ERA-Interim, bZ (0.92) exceeds bQ (0.09), demonstrating a clear association between atmospheric circulation and SAT variability in the COA. A similar relationship is found in NARR, with bZ = 0.91 and bQ = 0.14. Only one model, FGOALS-s2, has similarly attributed COA SAT variability, with bZ = 0.93 and bQ = 0.07. The majority of models fall within the circulation-associated regime with less separation than ERA-Interim between the Z500 and QH regression coefficients; in the all-model ensemble, bZ = 0.62 and bQ = 0.46.
If bQ exceeds bZ, variability in the SAT COA is categorized as more closely associated with variations in the local sensible heat flux than with circulation (red region of Fig. 5a). AMIP models with shifts in the position of the COA relative to observed, such as GISS-E2-R, fall within the surface-flux-dominated regime. Models with striking regions of high SAT variability in the central United States (CMCC-CM, MPI-ESM-MR, and MRI-CGCM3) have a large discrepancy in SAT control attribution with their average bQ (0.72) doubling their bZ (0.36).
To accompany the regression analysis, the Dirmeyer soil moisture–climate coupling metric [Eq. (1)] and its constituents for the box-averaged regions within the COAs are shown in Fig. 5b. The metric comprises the correlations between soil moisture and sensible heat flux (the terrestrial segment, blue bar) and between sensible heat flux and SAT (the atmospheric segment, cyan bar), which are weighted by the standard deviation of SAT (red bar). For the CMIP5 models, we use total soil moisture content because it is an output provided by all but two modeling centers (GFDL and LASG). Root zone (~0–70 cm) soil moisture is likely more suitable for this analysis, but Dirmeyer (2011) showed that although total column soil moisture–based metric values are muted with respect to surface soil moisture–based values, their spatial patterns are consistent.
Most models and the all-model ensemble show values of that exceed both ERA-Interim (−0.35) and NARR (−0.44) values (average dashed in Fig. 5b), supporting the visual evidence of a strong relationship between the land and atmosphere in the COA. IPSL-CM5A-LR and MIROC5 have lower values of the soil moisture–climate coupling because of lower than average SAT variability combined with a weak terrestrial (MIROC5) or atmospheric (IPSL-CM5A-LR) coupling segment. Models with the largest are surface-flux-dominated models, MPI-ESM-MR (−1.18), MRI-CGCM3 (−1.07), and CMCC-CM (−1.00). In the same regions, ERA-Interim and NARR do not exceed −0.49 (not shown). These regions of strong land–atmosphere coupling in the surface-flux-dominated models roughly coincide with the land–atmosphere interaction hot spot region defined in other studies (e.g., Ruiz-Barradas and Nigam 2005; Koster et al. 2004a,b). However, land–atmosphere interactions in these hot spot regions appear to be both excessive with respect to observational estimates and a main contributor to spurious model SAT variability in the central United States.
Biases in mean summer SAT also have been related to biases in evapotranspiration (Mueller and Seneviratne 2014). We evaluate the relationship between mean SAT and evapotranspiration (Fig. 6) and focus on 33°–45°N, 92°–101°W, a region of climatological warm bias in the central United States, where SAT in the all-model ensemble exceeds observed SAT by more than 5°C (Fig. 6a). The region coincides with the meridional strip of evapotranspiration bias, where ERA-Interim values exceed all-model ensemble values by more than 1.5 mm day−1. Mean SAT in the region is plotted against mean evapotranspiration in Fig. 6b, and there is a statistically significant linear relationship between SAT and evapotranspiration. Higher (lower) evapotranspiration means less (more) energy is available to heat the atmosphere, so wetter (drier) models tend to be less warm (warmer) in the central United States.
Models with near observed mean SAT in the region are the same models that have excessive, land surface–linked SAT variability in the central United States (Fig. 6b, diamonds). In this case, the models have similar values of mean central U.S. evapotranspiration to ERA-Interim but larger interannual surface flux variations in the region, with standard deviations exceeding ERA-Interim by up to 0.45 mm day−1. Models with a climatological warm bias in the central United States can be up to 8°C warmer than observed but have circulation-associated SAT variability in their COAs, similar to observed (Fig. 6b, squares). Because these models have lower mean values of evapotranspiration, fluctuations in soil moisture have less effect on the atmosphere.
The land–atmosphere coupling strength is estimated through the correlations between seasonal mean SAT and evapotranspiration, a relationship that has shown pattern agreement with more rigorous measures of land–atmosphere coupling obtained through performing prescribed and freely varying soil moisture experiments (Koster et al. 2006; Seneviratne et al. 2006). Negative correlations suggest a soil moisture control on fluxes to the atmosphere and temperature, although correlations become less meaningful as the magnitude of evapotranspiration diminishes (Seneviratne et al. 2006). The central United States is considered a transition region, limited by the availability of neither moisture nor radiation (Seneviratne et al. 2010), so we feel that the evapotranspiration–temperature correlation gives insight into the land–atmosphere coupling in our region of interest. All models have high evapotranspiration–temperature correlations in the region, exceeding ERA-Interim (NARR) correlations by 0.55 (0.41) on average, which further indicates models have a more robust land–atmosphere coupling than has been observationally estimated.
In summary, local regression analysis supports the visual SAT–control relationships established in our EOF analysis. We find that the land surface plays a larger role in setting U.S. summer SAT variability in the AMIP experiment than in ERA-Interim. The land–atmosphere coupling is considerably stronger in models than in ERA-Interim, both in regions of high SAT variability and in regions of climatological warm biases. Models with high SAT variability in the south-central United States feature clear covariability between COA SAT and QH. Models with more realistic SAT variability in the central United States tend to have a climatological warm bias. We conclude that strong land–atmosphere interactions along with climatological surface flux biases are responsible for spurious U.S. summer SAT variability in the AMIP ensemble.
5. SST forcing
While we attribute spurious U.S. summer SAT variability to an enhanced land surface feedback in the AMIP ensemble, other forcing biases can also contribute. SST variability in the equatorial Pacific, tropical Atlantic, and Indian Oceans can influence the midlatitude atmosphere and thus SAT variability over land. Regressing global SSTs onto JJA SAT PC1 illustrates potential SST forcing regions that impact common patterns of U.S. summer SAT. SSTs are prescribed over the 30-yr AMIP period, so biases arise from how the modeled atmosphere responds to slowly varying ocean states.
To evaluate model bias in SST forcing, correlations among SAT model PCs, defined as ρ(Mi, Mj) for the correlation between model i and model j, are scattered against the average of the observed (obs)–model correlation pair in Fig. 7. Because observed SAT variability includes an internal component as well as an SST-forced component, both ensemble mean PC pairs [black circles in Fig. 7 (top), forced component] and the average of individual realization PC pairs [red triangles in Fig. 7 (top), both forced and internal components] are considered. Correlations between observed and model PCs are all positive, which suggests a robust SST forcing. The black circles fall to the right of the one-to-one line, with correlations among ensemble mean PCs exceeding those between ensemble mean and observed PCs. This indicates that the models have similar, robust responses to the prescribed ocean forcing, despite having differences in physical and numerical formulations. Individual realization PCs correlate less strongly with observed than ensemble mean PCs (Fig. 7, bottom), and the scatter of red triangles around the one-to-one line is consistent with the effect of random internal variability. In the presence of internal variability, models are no more correlated with each other than they are with observations. The bar graph in Fig. 7 (bottom) also shows which models share a forced component of summer SAT variability with ERA-Interim. Models in the AMIP experiment are forced with observed SSTs, so significant correlations between modeled and observed PCs arise from a common response to the shared boundary condition. For those models exhibiting significant correlations, we evaluate SST and circulation patterns associated with JJA SAT EOF1.
Potential regions of SST forcing (color) are highlighted in JJA SAT PC1–SST regression maps in Fig. 8. SSTs are regressed onto SAT PC1s, normalized by their standard deviation, to illustrate the variation (°C) in SST per standard deviation of SAT variability. Regions of the global ocean where SST correlates significantly (95%) with JJA (ensemble mean) SAT PC1 are stippled. SST influence can be felt over land through the atmosphere, so upper-tropospheric heights (i.e., Z300), regressed on normalized JJA SAT PC1, are contoured to show the associated circulation patterns.
In ERA-Interim, sizable patches in the North Pacific, the tropical and North Atlantic, and the Caribbean Sea covary with the leading mode of JJA SAT. Correlations are notably not significant in the equatorial Pacific ENSO forcing region, emphasizing that there is not a substantial pathway for ENSO to influence the extratropical atmosphere over the United States in boreal summer (Barlow et al. 2001). Significant correlations in the Caribbean Sea illustrate the pathway for surrounding oceans to influence summer climate in the continental interior through the Great Plains low-level jet (Ruiz-Barradas and Nigam 2006; Weaver 2013). The wavy structure of observed Z300 anomalies in the Northern Hemisphere midlatitudes is characteristic of spatially organized internal variability. Significant correlations between JJA SAT PC1 and midlatitude Atlantic SST appear to be due to the covariability between COAs within the internally generated midlatitude zonal wave train.
The AMIP models, however, appear to be highly sensitive to ENSO’s influence in the concurrent summer. Significant negative correlations in the eastern equatorial Pacific and positive correlations in the western tropical Pacific tie model summer warming to an ocean described as “perfect” for widespread midlatitude drying (Hoerling and Kumar 2003). Models shown do not feature significant correlations in the Caribbean Sea, suggesting an absence of Great Plains low-level jet–driven climate variability in AMIP simulations, consistent with the findings of Ruiz-Barradas and Nigam (2006). Model Z300 anomalies are generally positive in the midlatitudes in either hemisphere. The interhemispheric symmetry of these zonal bands of high pressure is indicative of tropical SST forcing.
Lead–lag correlations between U.S. summer SAT and the seasonally averaged Niño-3.4 index show how models are sensitive to ENSO conditions (Fig. 9). Observed lead–lag correlations (thick black line in Fig. 9) are compared to ensemble mean PC correlations (colored lines in Fig. 9a) and to individual model realizations (thin gray lines in Fig. 9b). The realization that most resembles observations in Fig. 9b is highlighted in blue. Correlations that fall within the shaded regions are significant at 95%. In observations, JJA SAT PC1 does not significantly correlate with the Niño-3.4 index until the subsequent winter–spring [from February–April (FMA) to May–July (MJJ)], indicating an association between JJA SAT EOF1 and a developing ENSO event. All but two ensemble mean PC1s, representing SST-forced SAT variability, correlate significantly with the Niño-3.4 index in the antecedent winter–spring, with correlations peaking at a two-month lead [April–June (AMJ)]. All ensemble mean PC1s correlate with the Niño-3.4 at lag 0, illustrating the role La Niña conditions may play in warm summer conditions over the United States in models. Individual realizations show a complex correlation picture owing to the presence of internal variability in each PC1. The average across all realization correlations is shown in red in Fig. 9, which crosses the 95% significance threshold in AMJ. Because the preseason ENSO-forced signal emerges even in the presence of internal variability, the response of the model atmosphere to SST variability must also be addressed to reduce summer SAT biases.
We show that while winter SAT variability over the continental United States is well represented in the AMIP experiment, the AMIP models do not reproduce observed patterns of U.S. SAT variability in boreal summer. The AMIP models evaluated here feature either a spurious region of high summer SAT variability displaced south and/or west of observed or a significant climatological warm bias in the central United States. To address these errors, we investigate potential sources of excessive U.S. summer SAT variability in the AMIP experiment. In observations, the leading mode of U.S. summer SAT variability between 1979 and 2008 is associated with a large-scale anticyclone over the north-central United States. Significant correlations between SAT EOF1 and QH suggest that the land surface contributes to variability on the margins of the observed COA but is a second-order influence in the region of highest SAT variability. The majority of models have JJA SAT EOF1s that resemble the observed COA, but all feature variability extending farther south and/or west than observed to coincide with regions of high sensible heat variability. Models with enhanced SAT variability over the central United States show high temporal correlations between SAT and QH in their COAs, which suggests a robust land–atmosphere coupling in that region.
To quantify the relative contributions of circulation (Z500) and the land surface (QH) on JJA SAT variability, a multivariate regression is performed in the regions of highest variability within each SAT EOF1 COA. Spurious central U.S. SAT variability associates more closely with variations in sensible heat than with variations in circulation and occurs in regions where the land and atmosphere are strongly coupled. Models with QH-dominated SAT variability have higher values of mean evapotranspiration in the central United States than models with circulation-associated SAT variability. These higher values aid in setting realistic SAT climatology in the central United States but contribute to additional SAT variability. Models with circulation-dominated SAT variability tend to have a climatological warm bias in the central United States, with mean summer temperatures up to 8°C warmer than observed. This warm bias is tied to lower mean evapotranspiration in the region than observed, which results in the land surface having a more realistic influence on SAT variability. All models evaluated appear to have land surface–associated SAT biases in the central United States.
Addressing issues with the bulk land–atmosphere interaction is challenging as the exchanges of energy, water, and momentum between the surface and atmosphere are complex and largely taking place at subgrid scales (e.g., Ek and Holtslag 2004; Sellers et al. 1997). Our understanding of large-scale land–atmosphere interactions is rapidly improving, but current observations are spatially and temporally insufficient to be used to “tune” relevant model parameters. Land surface model grid cells are forced by the model atmosphere, where biases in precipitation and cloud radiative processes affect water and energy cycles (e.g., Ruiz-Barradas and Nigam 2006; Vial et al. 2013). Grid cells are mosaicked with different land-cover types to account for urban and vegetated areas. Vegetated areas are further subdivided by plant functional type to account for variations in plant morphology, such as leaf and stem area and canopy height (Bonan et al. 2002). Many models account for land-use changes from harvests, fires, and urbanization on seasonal time scales, as green vegetation fraction impacts the surface energy budget (Ek et al. 2003). The type and vertical layer structure of soils must be considered to replicate the effects of surface albedo and soil moisture memory. The aggregate of physical parameterizations are ultimately integrated to pass information to the atmosphere on a gridbox scale. Process scale assessments and a detailed understanding of the model physics are necessary to identify specific processes that cause the land–atmosphere issues in the AMIP experiment.
Tropical SST forcing contributes to summer SAT variability over the United States. Models respond to SST patterns that are different than observed but similar to one another. High correlations between JJA SAT PC1 and SST in the equatorial Pacific, accompanied by interhemispherically symmetric bands of high pressure in the upper troposphere, indicate that the ENSO contribution to extratropical climate variability in the summer is too high in models. Most models are sensitive to ENSO conditions in the antecedent spring and all models significantly correlate with the Niño-3.4 index at lag 0. This forced signal emerges above internal variability.
We conclude that an excessive land surface feedback is responsible for spurious U.S. summer SAT variability in the AMIP experiment. Further examination of the AMIP land surface feedback is warranted, particularly in conjunction with regional observations and on time scales conducive to assessing the directionality of control–response relationships. The AMIP biases documented in this study affect regions identified from models as hot spots of land–atmosphere interaction (Koster et al. 2004a) and limit the skill of model-based attributions of extreme heat wave events (Stott et al. 2004). Evaluating physical parameterizations that contribute to the bulk land surface feedback, precipitation variability, and cloud biases is key to improving model skill in simulating summer climate variability.
This work is supported by the National Science Foundation Graduate Research Fellowship Program under Grant DGE-1144086. We are grateful to the World Climate Research Programme’s Working Group on Coupled Modelling for coordinating the CMIP5 experiment and to the participating climate modeling groups (listed in Table 1 of this paper) for providing the model output used in this study. NCEP reanalysis data are provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, from their website (http://www.esrl.noaa.gov/psd/). We thank Yu Kosaka for helping to initiate this project and Clara Deser, Karen McKinnon, Michael Ek, and Randal Koster for their valuable feedback during the writing process. We also thank the three anonymous reviewers, whose insights greatly improved this manuscript.