The importance of interannual-to-decadal sea surface temperature (SST) influences on drought in the United States is examined using a suite of simulations conducted with the T31×3 resolution version of the NCAR Community Earth System Model (CESM1.0.3). The model captures tropical Pacific teleconnections to North American precipitation reasonably well, although orographic features are somewhat enhanced at higher resolution. The contribution of SST anomalies is isolated by comparing two idealized, 1000-yr CESM1.0.3 experiments: a fully coupled control and an atmosphere-only (CAM4) run forced with the SST climatology from the control. Droughts are identified using the Palmer Drought Severity Index (PDSI), which is computed over four U.S. regions from the CESM1.0.3 experiments and compared with the North American Drought Atlas (NADA). The CESM1.0.3 reproduces the persistence of NADA droughts quite well, although the model underestimates drought severity. Within the CESM1.0.3 framework, SST forcing does not significantly affect drought intensity or frequency of occurrence, even for very persistent “megadroughts” of 15 yr or more in length. In both the CESM1.0.3 and NADA, with the exception of the Southeast United States, droughts in all regions have intensities, persistence lengths, and occurrence frequencies statistically consistent with a red noise null hypothesis. This implies that SST forcing is not the dominant factor in generating drought and therefore that many decadal megadroughts are caused by a combination of internal atmospheric variability and coupling with the land surface, with SST anomalies playing only a secondary role.
Droughts are among the costliest natural disasters in the United States; during recent events such as the 1930s Dust Bowl (Hoerling and Kumar 2002; Schubert et al. 2004a) and the droughts of the early 2000s (Fulp 2005), agricultural productivity was dramatically reduced. Potential climate change impacts on drought are therefore of critical importance, especially in the arid Southwest, where reductions in both mean precipitation and winter snowpack are projected for the twenty-first century (Seager et al. 2007; Sheffield and Wood 2008). Ensuring that the causes and frequency of observed North American droughts are well understood is a crucial step toward improving projections of future drought variability.
As the twentieth century is fairly short compared with the time scales of natural low-frequency climate variability, paleoclimate records [mainly from tree rings; e.g., Cook et al. (2004); Herweijer et al. (2007)] are often used to assess the intensity and recurrence rates of past drought periods. Proxy evidence suggests that past droughts may have been far more extreme than events seen during the modern era. These so-called “megadroughts” persisted for multiple decades (Woodhouse and Overpeck 1998), and the reduction in moisture availability was sometimes sufficient to dry out entire lake beds (Stine 1994). The physical processes driving these megadroughts are difficult to assess but may vary at very low frequencies; the spectral properties of tree ring records and climate model output differ, suggesting that models may have difficulties with simulating megadroughts (Ault et al. 2013). Given that quantitatively comparing climate model output with proxy data on time scales of decades to millennia is quite difficult, an intermediate step is to comprehensively evaluate the physical processes driving simulated megadroughts. This is the goal of the present study.
Because of the long time scales associated with oceanic variability, remote forcing from variations in sea surface temperature (SST) is commonly invoked as a driver for persistent drought conditions on decadal-to-multidecadal time scales (Seager et al. 2005a,b; Schubert et al. 2004b; Herweijer et al. 2007). In addition to forcing from the Atlantic (Enfield et al. 2001; McCabe et al. 2004; Oglesby et al. 2012), the tropical Pacific is known to cause a portion of the interannual precipitation variability over North America via teleconnections associated with the El Niño–Southern Oscillation (ENSO) phenomenon (e.g., McCabe et al. 2004; Cook et al. 2011a,b; Langford et al. 2014). In addition, many general circulation model (GCM) experiments have demonstrated the ability of persistent cold, La Niña–like tropical Pacific SST anomalies to generate prolonged droughts qualitatively similar to medieval megadroughts (Seager et al. 2007; Herweijer et al. 2006). The Pacific SST influence operates through transient eddy momentum fluxes, with a zonally symmetric component from enhanced subsidence over the southwestern United States (Seager et al. 2003, 2005a,b) and a zonally asymmetric component due to Rossby wave propagation (Hoskins and Karoly 1981). Atlantic SST variability may have further contributed to the generation of megadroughts, as demonstrated for the Medieval Climate Anomaly (Feng et al. 2008; Oglesby et al. 2012). However, SST forcing does not necessarily dominate; the amount of precipitation variance explained by ENSO is relatively small (10%–20%), and the precise amount of precipitation associated with ENSO may change from event to event [as noted by Kiladis and Diaz (1989); Cayan and Webb (1992); McCabe and Dettinger (1999); Gershunov and Barnett (1998); Cayan et al. (2010); and many others] as a result of atmospheric noise. This raises the question: how often do megadroughts occur in the absence of characteristic SST anomalies? It is the ability to characterize this background level of drought variability that will eventually determine the potential predictability of megadroughts, their onset, duration, and termination (Chikamoto et al. 2015).
Although recent work has demonstrated the importance of stochastic variability for generating megadroughts (Hunt 2011; Coats et al. 2013), the extent of this effect cannot be accurately determined in a coupled modeling framework. Toward that end, we have designed experiments that can isolate the component of drought variability unrelated to the ocean (although land–atmosphere coupling is still present; see section 6). These experiments are described in section 2 and validated against observational data products in section 3 to evaluate the fidelity of modeled ENSO and its midlatitude teleconnections. The statistics of drought properties are then discussed in section 4. The contribution of tropical influences to large-scale decadal atmospheric circulation anomalies is examined in section 5 and compared with the overall patterns of midlatitude variability in section 6. Conclusions and implications for future work are provided in sections 6 and 7.
2. Experimental setup
The main goal of this study is to determine whether megadroughts can occur in the absence of the SST forcing imposed by ENSO and other interannual-to-decadal modes of variability in the tropical Pacific. Then, if such megadroughts are found to occur, the secondary goal is to determine whether ocean coupling affects their intensity and/or persistence. These goals are accomplished by performing long simulations with the NCAR Community Earth System Model, version 1.0.3 (CESM1.0.3), and its atmospheric component, the Community Atmosphere Model, version 4 (CAM4; Neale et al. 2013). All simulations use preindustrial CO2 concentrations (280 ppm) with no external forcings applied. The control simulation (CTL) was performed with a low-resolution configuration of the CESM1.0.3; this is the T31×3, which has roughly 3.75° atmospheric resolution (Yeager et al. 2006). This simulation was first integrated for 1000 yr, of which the first 100 yr are discarded to minimize spinup effects. Analysis of the total ocean heat content and top-of-atmosphere radiative balance indicates that the climate has largely equilibrated by the start of the analysis period (not shown). For the analysis of model ENSO and teleconnections in section 3, output from two additional preindustrial coupled control simulations run at higher atmospheric resolution is used. These simulations were performed using the Community Climate System Model, version 4 (CCSM4; Gent et al. 2011), as part of phase 5 of the Coupled Model Intercomparison Project (CMIP5; Taylor et al. 2012). The CCSM4 preindustrial control simulations have resolutions of 1° and 2° and are therefore referred to as CTL1D and CTL2D. Model behavior in both control simulations is discussed in Deser et al. (2012), and the simulations have total lengths of 1300 yr (CTL1D) and 1000 yr (CTL2D). All model simulations used in this study are summarized in Table 1.
To test the influence of ocean dynamics on megadroughts, CAM4 simulations were performed forced with climatological SST and sea ice extent derived from the first three centuries of CTL. The boundary conditions used contain only seasonal variability, allowing the isolation of the effects of internal (uncoupled) atmospheric noise from the coupled atmosphere–ocean dynamics (although coupling with the land surface does remain).
Several different model resolutions are used for these atmosphere-only experiments: a T31 spectral truncation (matching the CTL resolution) (CLIM), and 1° and 2° finite-volume configuration (CLIM1D and CLIM2D). The modeling experiments are presented in Table 1; note that the higher-resolution simulations are shorter because of their larger computational cost and are here used primarily for model validation and for insight into the resolution dependence of circulation patterns associated with drought. The low-resolution CLIM simulation then provides more robust estimates of megadrought statistics.
Several observational datasets are used for evaluation of CCSM4–CESM1.0.3 performance. The Hadley Centre Sea Ice and Sea Surface Temperature dataset (HadISST; Rayner et al. 2006) is adopted for comparisons with modeled tropical Pacific SST variability; years 1900–2008 are analyzed. Sea level pressure (SLP) and geopotential heights in CCSM4–CESM1.0.3 are compared with the NCEP–NCAR reanalysis (Kalnay et al. 1996), covering 1948–2010. Two different precipitation products are used: the Global Precipitation Climatology Project, version 2 (GPCP; Adler et al. 2003), extending from 1979 to 2010, and the high-resolution University of Delaware dataset (UDel; Matsuura and Willmott 2009), which covers the 1950–2008 period. GPCP allows for a better assessment of large-scale precipitation anomalies associated with ENSO, as it includes precipitation data over the ocean; the UDel dataset only contains land-based measurements, but its higher spatial resolution (0.5°, as compared with 2.5° for GPCP) provides a useful comparison for evaluating the resolution dependence of North American precipitation in CCSM–CESM.
3. Model validation
a. ENSO representation
The low computational cost of the T31×3 CTL and CLIM runs makes them ideal for studies of long-term multidecadal climate variability. This is a fairly low-resolution version of the CCSM; however, the representation of ENSO in the T31×3 CCSM4 is known to be quite good compared with other models of similar resolution, and this configuration has therefore been previously used for multicentury ENSO studies (Jochum et al. 2009; Stevenson et al. 2012). To confirm that the T31×3 CESM1.0.3 has an ENSO of comparable fidelity to its predecessors, the T31×3 CCSM3.5 and 4, we investigate the spatial structure, spectral behavior, and seasonal phase locking of ENSO in the CESM1.0.3. These are commonly used metrics to characterize ENSO behavior and were part of the evaluation of the CTL1D and CTL2D simulations discussed in section 3 (Deser et al. 2012).
The spatial structure of tropical Pacific SST anomaly (SSTA) variability is shown in Figs. 1a and 1b for the CESM1.0.3 CTL and HadISST, respectively. CTL shows SSTA variance extending farther west along the equator than in HadISST, a known bias related to overly strong equatorial trade winds in the model (Deser et al. 2012). However, the amplitude of ENSO, as measured by the spectrum of the monthly Niño-3.4 (5°S–5°N, 190°–240°E) SST anomaly agrees remarkably well with HadISST (Fig. 1d). This is a marked contrast to the overly strong ENSO observed in both the CTL1D and CTL2D CCSM4 preindustrial control simulations (Deser et al. 2012). The spectrum of the simulated Niño-3.4 index is also realistic, exhibiting a broad peak at interannual periods; this peak is also captured by the 1° and 2° CCSM4 (Deser et al. 2012).
Finally, Fig. 1c shows the seasonally stratified standard deviation of Niño-3.4 SSTA, a commonly employed method of measuring the phase locking of ENSO to the seasonal cycle (Stein et al. 2014). The CESM1.0.3 shows an enhancement in variance relative to HadISST in all months, except for May and June, on the order of 10%–20%. Although nonnegligible, this error in magnitude is much smaller than for the 1° CCSM4 Deser et al. (2012). The offsets between the T31×3 CCSM4 and HadISST are well within the range of errors in current-generation climate models (Bellenger et al. 2014; Guilyardi et al. 2009) and could easily be explainable by undersampling natural variability in the observational record (Wittenberg 2009).
b. Midlatitude variability
Given the relatively realistic ENSO representation in the CESM1.0.3, we next confirm that that realism also holds for midlatitude variability, as well as tropical–extratropical teleconnections. If tropical SSTs do play a dominant role in generating megadroughts, this should be communicated to the midlatitudes via changes in the eddy-driven circulation (Seager et al. 2003; Seager 2007). Thus, it is important to determine whether different CCSM resolutions show differences in midlatitude circulation. This is done by analyzing years 800–899 of CTL1D and 800–999 of CTL2D.
To provide insight into the general structure of midlatitude variability, the dominant three modes of monthly mean sea level pressure anomaly are shown for the NCEP–NCAR reanalysis (Kalnay et al. 1996), for the T31×3 CTL simulation, and for the three CAM4-only simulations discussed in section 2 (Fig. 2). CTL is able to reproduce the overall structure of the first three EOFs (cf. Fig. 2, first two rows), although in EOF3 the midlatitude SLP anomaly extends somewhat too far across the basin. The SLPA EOFs are largely similar between CTL and CLIM (Fig. 2, second and third rows), indicating that these modes are generated by internal atmospheric dynamics (see section 5). The impact of model resolution may be seen by comparing the lower three rows of Fig. 2. Although the CLIM1D EOFs 2 and 3 are reversed in the order of variance the modes explain relative to CLIM1D and CLIM2D (i.e., cf. Figs. 2n,o to Figs. 2k,l and Figs. 2h,i), the spatial structure is nearly identical between the three climatologically forced simulations. Thus, resolution is not a dominant factor in shaping the modes of large-scale North Pacific atmospheric variability in CESM1.0.3.
The ENSO teleconnection structure is assessed in CTL by plotting the correlation between December–February (DJF) Niño-3.4 SSTA and other variables in Fig. 3. Here DJF has been chosen as the season of interest because the peak of the El Niño and La Niña SST anomalies tends to occur during boreal winter (Rasmusson and Carpenter 1982) and midlatitude teleconnections are strongest (Ropelewski and Halpert 1986; Trenberth et al. 1998). The correlation with tropical surface air temperature (Fig. 3c) shows the canonical horseshoe El Niño shape in surface temperatures, and comparison with the corresponding correlation generated using observational data (Fig. 3a) shows that the spatial structure is well represented in CTL despite its low resolution. Poleward of 30°, the sea level pressure and surface temperature teleconnections are weaker than the observations (contours in Figs. 3a,c). In spite of this caveat, the direct impact of ENSO in the tropics seems to be fairly robust. A comparison of Figs. 3b and 3d reveals that the CTL correlation of Niño-3.4 SSTA with North American precipitation in CTL is quite similar to the observed pattern, with the exception of discrepancies over higher-elevation regions, such as the Rocky Mountains and Pacific Northwest. This indicates that even at low (T31) resolution, where some large-scale circulation features are less sensitive to ENSO than observed, the precipitation sensitivity is nonetheless reasonably accurate.
To assess the impact of model resolution on ENSO teleconnections, Figs. 3e–h show correlations constructed from the CTL1D and CTL2D preindustrial control simulations conducted with the CCSM4. In all model versions, ENSO SSTA patterns extend too far west; this SST bias, common to the majority of GCMs (Bellenger et al. 2014), does not disappear at higher resolution. For both simulations, midlatitude teleconnections are stronger overall than in the T31×3 CTL run (Figs. 3c,d). This is especially apparent in the larger correlations between Niño-3.4 SSTA and precipitation–surface air temperature (TS). But interestingly, CTL1D, despite having the highest resolution, does not necessarily appear to be the most realistic. During DJF, the North American TS correlations are stronger and closer to the observed values in the 2° control versus the 1° control (Figs. 3e,g). Precipitation values are also qualitatively similar to observed values in CTL2D (Fig. 3f). The offsets remaining between CTL1D and the observed teleconnection patterns indicate that the problems with model performance cannot be solved simply by increasing resolution. Apparent model performance is also subject to compensating errors; the CESM may, in other words, be capturing the correct teleconnection structures for the wrong reasons. A complete assessment of these issues is outside the scope of the present study, but these possibilities are noted as caveats.
c. North American precipitation
The overall representation of precipitation over North America is next assessed. Here, the focus of the analysis is on the T31×3 simulations CTL and CLIM, since these simulations are longest and therefore have the most robust megadrought statistics. Figure 4 shows the overall North American precipitation variability in CTL and CLIM, as well as the annual and interannual precipitation variations for the University of Delaware dataset (Matsuura and Willmott 2009). Here, annual variability is defined as the amplitude of the mean seasonal cycle in precipitation (or climatology), and interannual variability as the variance of precipitation anomalies (deseasonalized). Figures 4b and 4e demonstrate that the general structure of precipitation variability is reproduced fairly well in CTL, although some biases remain. In particular, the Sierra Nevada shows weaker-than-observed variability in both annual and interannual bands; in the Great Plains, the annual cycle is too strong and interannual variability too weak; and interannual variability is too weak over the southeastern United States. The impact of atmosphere–ocean coupling may be seen by comparing Figs. 4c and 4f. This illustrates that the differences between CLIM and CTL are primarily in the interannual band (Fig. 4f), related to potential ENSO impacts. However, the intensification of interannual rainfall variability is only 10%–15%, suggesting that the influence of interannual SST variability is relatively minor (Fig. 4f).
Figure 5 shows the difference between the seasonal cycle and interannual precipitation variability in the CLIM modeling hierarchy, another measure of resolution dependence. As one would expect given the identical SST–sea ice climatologies used to force the three experiments, the annual cycle in precipitation is nearly identical in each, with the exception of regions with mountainous terrain, such as the Pacific Northwest and Colorado–Utah (Figs. 5a–c). Some differences do exist in interannual variability (Figs. 5c–f), and, in fact, those changes are statistically significant over much of the North American Continent. This likely relates to the passage of storm systems over continental topography; the relatively short temporal extent of CLIM2D and CLIM1D mean that natural variability may also contribute to offsets between the CLIM simulations.
The seasonal structure of precipitation is examined next by constructing composites from observations and the CESM1.0.3 during DJF, March–May (MAM), June–August (JJA), and September–November (SON) (Fig. 6). Figures 6a–d show climatological precipitation for CTL, and Figs. 6e–h for CLIM; as expected, the two rows have nearly identical precipitation structures. The magnitude of precipitation maxima (DJF) and minima (JJA) over the Pacific Northwest is comparable to observations, although the CESM1.0.3 shows precipitation anomalies extending further inland than observed. Agreement appears poorest during MAM and SON, although still fairly reasonable; these discrepancies may relate to known errors in timing of spring–fall onset in CESM1.0.3 (Peacock 2012).
The third and fourth rows of Fig. 6 show the precipitation climatologies for CLIM2D and CLIM1D. The structure is similar in all three cases, with more detailed orographic precipitation represented at higher resolution. In Southern California, the seasonal cycle becomes more intense in CLIM2D (Figs. 6i–l) and CLIM1D (Figs. 6m–p) relative to CLIM (Figs. 6e–h), with stronger anomalies present during both DJF and JJA. This is true for the southwestern United States as well; changes elsewhere are not as easily distinguishable. However, biases persist at all resolutions, indicating that higher resolution is insufficient to provide accurate representation of precipitation.
The results of the preceding two sections show that CESM1.0.3 exhibits a fairly realistic simulation of North American precipitation and its relationship with ENSO and is, therefore, a good tool for studying the tropical Pacific’s influence on megadroughts. Determining the properties of megadrought occurrence and associated dynamical mechanisms is the topic of the next few sections.
4. Changes to megadrought statistics
The properties of droughts are expected to vary spatially across North America, as many distinct climate zones exist throughout the continent. To provide a broad sense of the ocean’s role in North American drought, several averaging regions are therefore chosen. These regions are 1) the Great Basin of Meehl and Hu (2006) (32°–42°N, 106°–118°W); 2) the Southern California region (SoCal; 32.5°–36.5°N, 120°–115°W), covering Southern California and parts of Nevada and selected qualitatively for its climatologically low mean precipitation and high seasonal–interannual variance; 3) the Southeast United States (SE; 28°–36°N, 70°–90°W), selected to coincide with regions of high correlation between precipitation and the Niño-3.4 surface air temperature in CTL (5°S–5°N, 120°–170°W); and 4) the Great Plains (32°–42°N, 105°–90°W), covering the central portion of the United States and coinciding with the region containing the majority of tree ring data collected to date (Woodhouse and Overpeck 1998; Cook et al. 2007).
For all regions considered, time series of the Palmer Drought Severity Index (PDSI) are computed at each model grid point. Calculations are performed using the toolkit developed by Jacobi et al. (2013), and the method of Hamon (1961) is employed for the computation of potential evapotranspiration. PDSI values are computed using monthly temperature and precipitation, and annual JJA averages are calculated from the monthly PDSI time series. As one would expect, the resulting JJA PDSI time series has a correlation of 0.8 or better with modeled soil moisture in our study regions (not pictured), indicating a good correspondence between PDSI and the modeled hydrological balance.
Drought events are identified based on the 2S2E criterion of Coats et al. (2013), adapted from Herweijer et al. (2007). Here, a drought is considered to begin when PDSI is below 0 for two consecutive years, then to end when PDSI is above 0 for two consecutive years. This definition results in drought periods similar to the results of using a decadal-mean precipitation metric (e.g., Meehl and Hu 2006; not pictured). From this set of events, decadal megadroughts are defined as those droughts that persist for 15 or more years. These events have been highlighted in Fig. 7, which shows the PDSI time series for all regions in the CESM1.0.3 simulations, as well as the North American Drought Atlas (NADA; Cook et al. 2007). PDSI variability is stronger overall in the NADA than CESM1.0.3, but the number of megadrought events is on the same order of magnitude between the CTL run and the NADA in all regions.
The influence of ocean forcing on megadroughts can be qualitatively assessed by comparing event occurrence frequencies in CTL and CLIM; if SST influences dominate, megadroughts should be much more common in CTL. Yet there is no such detectable difference (cf. Fig. 7, left column vs middle column). In fact, both SoCal and the Great Plains show a larger number of megadroughts in CLIM than CTL (Fig. 7e vs Fig. 7d, Fig. 7k vs Fig. 7j). However, as these events remain fairly rare overall (roughly 0.3–0.8 per century), it is not possible to attribute significance using plots like Fig. 7. More rigorous statistics are presented in Fig. 8, which shows the properties of CESM1.0.3 and NADA droughts relative to a red noise [first-order autoregressive (AR-1)] null hypothesis model (see Delworth and Manabe 1988). Here the null is simulated by fitting an AR-1 model to the JJA PDSI time series for the CESM1.0.3–NADA in the appropriate region, then generating 1000 simulated time series; these time series have the same variance and lag−1 autocorrelation as the input data. From each simulated time series, mean drought properties are then calculated and used to form a distribution. The drought characteristics considered here include:
Frequency of megadroughts: the total number of occurrences of droughts persisting for 15 yr or more.
Intensity of megadroughts: the mean value of PDSI during megadrought periods.
Persistence of (all) droughts: the average length of all droughts detected using the 2S2E criterion described above (the full set of events is considered here, rather than strictly megadroughts, to avoid biasing the distribution through the application of a minimum persistence threshold).
The bars in Fig. 8 denote the 90% confidence interval on the mean frequency, intensity, and persistence for red noise–generated droughts–megadroughts in all four study regions; filled circles indicate the mean values for CTL, CLIM, and the NADA. We note that the NADA extends from AD 0 to 2006; to ensure consistency with the two 1000-yr CESM1.0.3 simulations, only the AD 1000–2000 portion of the NADA has been included in this analysis.
From Fig. 7, it was apparent that droughts in CESM1.0.3 are less intense than observed in the proxy record. However, the intensity plot in Fig. 8 indicates that the CESM–NADA offset is, in fact, consistent with the scatter one might expect from AR-1 noise. The same is true for drought persistences and megadrought occurrence frequencies; all values fall well within the AR-1 error bars, with the exception of NADA drought frequency in the Southeast region. Here, the AR-1 model indicates a higher drought frequency than observed in the NADA, consistent with a tendency for enhanced drought termination.
From this analysis, we conclude that, although SST forcing does indeed affect North American drought in CESM1.0.3, it does not lead to statistically significant changes in the number or character of such events.
5. Large-scale circulation and drought
The previous sections indicate that megadroughts occurring in the presence and absence of coupled ocean dynamics have similar frequencies and intensities. However, their driving mechanisms might nonetheless be quite different and altered by ocean coupling and SST anomalies. Such a situation could potentially arise if similar precipitation anomalies were created both by random migrations of the storm track and by changes in ENSO behavior. In that case, one might expect that the pattern of megadrought-related precipitation anomalies in a particular region might be dramatically different (Wang and Ting 2000) and that distinct precipitation signatures would be related to particular modes of large-scale variability (i.e., ENSO or the PDO, etc.). This possibility is investigated in this section.
First, to provide a view of precipitation variability in the frequency domain, spectra of deseasonalized monthly precipitation are shown in Fig. 9, calculated using the method of Bartlett (1948). Time series are broken into m subsections (here, length 40 yr), and periodograms are computed for each subsection individually, then averaged together to form an estimate of the overall spectrum. Confidence intervals are then computed according to a chi-squared distribution with 2m degrees of freedom (von Storch and Zwiers 2003). The spectra for all regions are characterized basically by white-noise behavior with no discernible differences in low-frequency variance in CLIM relative to CTL. This confirms that low-frequency precipitation variability is not predominantly a coupled atmosphere–ocean phenomenon but can be generated by either purely internal atmospheric processes or by land–atmosphere interactions. The only possible exception is the interannual band in the Southeast region spectrum (Fig. 9d), where there is a slight, but minimally statistically significant reduction in variance from CTL to CLIM (see also the reduction in interannual precipitation variability in the Southeast from CTL to CLIM shown in Fig. 4f). This is consistent with the PDSI results of section 4, although the CESM1.0.3 appears to have a somewhat weakened ENSO–Southeast connection relative to the NADA.
Next, to develop a more detailed understanding of the atmospheric circulation patterns that cause megadroughts in CTL and CLIM, we calculate the correlations of Great Basin monthly mean precipitation (Figs. 10a–d) and 11-yr running mean precipitation (Figs. 10e–h) with sea level pressure, surface temperature, 500-hPa geopotential height, and precipitation anomalies. In the absence of another source of long-term memory, the dominant mechanism for precipitation variability will be stochastic internal atmospheric modes, which will rely more heavily on high-frequency variations in precipitation. Thus, computing monthly and decadal droughts individually allows a direct visualization of these stochastically driven processes.
On seasonal time scales (monthly correlations in Figs. 10a–d), the SLP and 500-hPa patterns are strikingly similar between CTL and CLIM. These patterns bear a strong resemblance to the Pacific–North American pattern (Wallace and Gutzler 1981), which has previously been demonstrated to relate to changes in the position of the storm track [Lau (1998); see also review by Trenberth et al. (1998)]. Since CTL and CLIM share the same SST–sea ice climatology, the similarity of the high-frequency signal is to be expected; nonetheless, Figs. 10a–d confirm that precipitation anomalies driven by storm track variations do not change in the presence of atmosphere–ocean coupling.
On decadal time scales (Figs. 10e–h), we find that the precipitation excursions related to sustained Great Basin rainfall changes cover the majority of the western United States (Fig. 10e). In CTL, these events are accompanied by anomalous surface cyclonic circulation covering much of the North Pacific (Fig. 10e) and an upper-level trough over the western United States (Fig. 10f, contours); this feature is thus roughly barotropic, as for the monthly mean correlations in Figs. 10a–d. Figure 10e shows that the associated tropical Pacific SST anomaly is weakly positive and that a cold anomaly exists in the Kuroshio Extension region, with a stronger warm anomaly centered at 50°N. This analysis suggests that previously proposed effects of both tropical SSTs (Seager et al. 2003) and North Pacific SSTs (Timmermann et al. 1998) on the large-scale atmospheric circulation do lead to some rainfall changes over the Great Basin region. However, as demonstrated in Fig. 3, the associated explained variances for SST are relatively small in CESM1.0.3 (10%–20%). This is consistent with composites of surface temperatures during individual megadroughts in CTL; during some events, a cold tropical SST anomaly does appear, but during others the anomaly is absent (not shown).
Finally, Figs. 10g and,10h show correlations for 11-yr mean Great Basin precipitation in CLIM. Since SSTs in CLIM by construction contain no interannual variability, surface air temperature anomalies are negligible over the ocean, but substantial cooling over Texas–Arizona and Mexico is associated with enhanced Great Basin rainfall (Fig. 10g). The upper-air circulation pattern (Fig. 10h, contours) resembles a wave train, with alternating high and low geopotential height anomalies. In this experiment, the elongated circulation anomaly over the North Pacific, seen in the CTL run (Figs. 10e,f), is absent, which supports the notion that although the net effect is statistically insignificant, in CTL, tropical and North Pacific SSTs may play some role in steering hydroclimate variability over the continental United States, causing pluvials and droughts.
Our analysis of multicentury CESM1.0.3 simulations shows that North American megadroughts can be caused by internal atmospheric variability, such as changes to the Aleutian low–winter storm track; these changes can occur stochastically and therefore do not require forcing from tropical SSTs. Remote oceanic teleconnections are shown not to be the dominant control on precipitation changes, meaning that the forced response of the atmosphere to SST anomalies is small compared with internal variability. This is consistent with recent multimodel analyses showing substantial temporal variation in ENSO teleconnections and a lack of correspondence between tropical SSTs and megadrought events in coupled simulations of the last millennium (Hunt 2011; Coats et al. 2013).
More work remains to be done to quantify and understand the effects of tropical and North Pacific SST anomalies in driving decadal rainfall changes over the United States relative to the null hypothesis of internally generated atmospheric noise. Furthermore, the role of model biases remains unclear; the physical linkages between tropical SST anomalies and the midlatitude atmosphere are well understood (Hoskins and Karoly 1981) and operate in CESM1.0.3 (section 3). Additionally, the similarity of precipitation seasonality and large-scale variability between various CAM4 resolutions suggest that these results are robust for the CESM framework (Figs. 3, 4). However, an underestimation of decadal SST variability or of ENSO teleconnection strength could potentially lead to underestimating the oceanic influence on megadroughts (i.e., Ault et al. 2013), which effect is indeed observed in the PDSI time series of CESM1.0.3 (Fig. 7). To confirm the robustness of these results, replication with a variety of GCMs will be required to assess changes due to differing physical parameterizations.
Finally, our analysis highlights the need for dedicated suites of experiments to isolate various potential sources of decadal memory driving megadroughts. Clearly, there is a nonnegligible contribution from unforced atmospheric noise, which on its own would indicate a low potential predictability for megadrought-like events. However, the present simulations include an active land model, which could be a source of decadal memory for drought through vegetative or aerosol feedback processes, as well as subsurface aquifer storage (Chikamoto et al. 2015). Future work will be required to quantify the relative contributions of more predictable drought drivers.
The magnitude of the influence of internal atmospheric variability on prolonged droughts in North America is investigated using simulations conducted with the NCAR CESM1.0.3: a fully coupled 1000-yr control run (CTL), and a low-resolution CAM4-only simulation forced with climatological-mean SST (CLIM). Two shorter, higher-resolution CAM4 runs using the 2° and 1° finite-volume configurations (CLIM2D and CLIM1D) are used for evaluation of model performance, as are previously conducted preindustrial coupled control simulations at the same resolutions (CTL2D and CTL1D). All of the CAM4-only simulations use forcings derived from CTL to prevent biases because of mean-state differences. Model validation is performed relative to observations, which shows that CTL has a quite realistic ENSO for a model of its class. The structure of ENSO teleconnections into North America is also well captured compared with observations and is largely unchanged as model resolution increases, as is the seasonal cycle of North American precipitation. The T31×3 CLIM run is therefore used to generate multicentury time series of North American precipitation and soil moisture.
Droughts are identified based on time series of the Palmer Drought Severity Index calculated from CESM1.0.3 output, where events begin with JJA PDSI values persisting below zero for two consecutive years and terminate with this quantity persisting above zero for two years. In CESM1.0.3, megadrought occurrence frequencies and overall drought persistences are comparable to events observed in the North American Drought Atlas, although megadrought intensities are lower in the model than the NADA. Comparisons with AR-1 time series fit to NADA and CESM1.0.3 then indicate that drought persistences and megadrought occurrence frequencies–intensities are consistent with red noise in most cases; the exception is NADA droughts in the Southeast, which are less frequent than predicted by the AR-1 model. This lack of detectable signal is evidence for the importance of processes other than SST forcing in generating drought: the present study cannot completely distinguish between atmosphere-only and atmosphere–land coupled processes, but the tropical oceans are clearly not required for the generation of decadal megadroughts.
Some SST forcing can be detected in correlations of 11-yr running mean Great Basin precipitation with surface temperature in CTL, which is absent in CLIM; this confirms that SST does sometimes affect decadal droughts. However, correlations of monthly Great Basin precipitation with atmospheric conditions resemble midlatitude Rossby wave trains with minimal signal in the tropics in both CTL and CLIM. Thus, the stochastic component of precipitation variability does not necessarily rely on ocean dynamics. Given this result and the fact that the presence of SST forcing is insufficient to create a change in drought statistics between CTL and CLIM, we conclude that megadroughts in CESM are mainly driven by atmospheric noise, whereas SST influences play a much smaller role.
This work was funded by the NSF, under the project Investigation of Decadal Predictability and Hydroclimate Impacts on the Western U.S. (IDCPI; NSF Award 1049219). We thank the IDCPI group (N. Buenning, L. Kanner, D. Noone, L. Sloan, M. Snyder, and L. Stott) for numerous stimulating discussions. NCEP reanalysis-derived data, as well as the Global Precipitation Climatology Project land–ocean precipitation and the University of Delaware land precipitation datasets, were provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, on their website (http://www.esrl.noaa.gov/psd/). The HadISST dataset was obtained from the Met Office Hadley Centre website (http://www.metoffice.gov.uk/hadobs/hadisst/). CCSM4 preindustrial control data was obtained from the CMIP5 archive (http://pcmdi9.llnl.gov/esgf-web-fe/).