Potential predictability sources of the 2012 U.S. drought in observations and a regional model ensemble

Abstract

The 2012 drought was the most severe and extensive summertime U.S. drought in half a century with substantial economic loss and impacts on food security and commodity prices. A unique aspect of the 2012 drought was its rapid onset and intensification over the Southern Rockies, extending to the Great Plains during late spring and early summer, and the absence of known precursor large‐scale patterns. Drought prediction therefore remains a major challenge. This study evaluates relationships among snow, soil moisture, and precipitation to identify sources of potential predictability of the 2012 summer drought using observations and a Weather Research and Forecasting model multiphysics ensemble experiment. Although underestimated in intensity, the drought signal is robust to the way atmospheric physical processes are represented in the model. For the Southern Rockies, soil moisture exhibits stronger persistence than precipitation in observations and the ensemble experiment. Correlations between winter/spring snowmelt and concurrent and following season soil moisture, and between soil moisture and concurrent and following season precipitation, in both observations and the model ensemble, suggest potential predictability beyond 1 and 2 month lead‐time reside in the land surface conditions for apparent flash droughts such as the 2012 drought.

1 Introduction

Drought can be widespread and persistent, lasting from a few months to multiple years, causing significant impacts on agriculture, water supply, and the environment. The 2012 summertime drought over the Midwest was one of the most severe and extensive U.S. droughts since the 1930s Dust Bowl and brought the greatest summertime rainfall deficit to the Central Great Plains in 117 years, surpassing 1934, 1936, and 1988 [Hoerling et al., 2014]. Moderate to extreme drought conditions affected 80% of U.S. agricultural land in 2012 [U.S. Department of Agriculture, 2012] resulting in $30 billion in economic loss (https://www.ncdc.noaa.gov/billions/events). The 2012 drought has been referred to as a flash drought because of its rapid onset and intensification over the Southern Rockies, extending to the Great Plains during late spring/early summer, and the absence of precursor large‐scale patterns [Hoerling et al., 2013, 2014]. Wang et al. [2014] found only weak contribution from sea surface temperature forcing in the evolution of the 2012 drought. Hoerling et al. [2013] reported that the National Oceanic and Atmospheric Administration operational seasonal drought outlook issued in May 2012 for the subsequent June–August period failed to predict a trend toward increasing drought. They concluded that the 2012 drought was caused mostly by natural variability in synoptic weather (found by AghaKouchak [2014] to be driven by land‐atmosphere interactions), meaning that attempts to predict the 2012 drought would be extremely challenging.

Global climate models and regional climate models (RCMs) generally exhibit low skill in simulating seasonal precipitation [Fowler et al., 2007; Christensen et al., 2008; Lavers et al., 2009; Maraun et al., 2010; PaiMazumder et al., 2013; PaiMazumder and Done, 2014]. Prein et al. [2013] showed that a very high resolution RCM is able to simulate the statistics of summertime precipitation extremes, but regional climate simulations at these finer resolutions are computationally very expensive. Soil moisture, on the other hand, exhibits less variability than precipitation due to persistence of soil moisture memory and is therefore a source of drought forecast skill [Mo et al., 2012; Hao and AghaKouchak, 2014]. Hoerling et al. [2013] showed that a RCM driven by real‐time global seasonal forecast data was able to predict dry soil moisture anomalies better than dry precipitation anomalies over the Central Great Plains in 2012. AghaKouchak [2014] showed that a soil moisture‐based framework improves the skill of the 2012 summer drought forecasts for 1 and 2 month forecasts. In addition to soil moisture, winter/spring snow condition also plays a role in contributing to the skill of summer drought prediction [Mote et al., 2005; Staudinger et al., 2014]. Recently, Thomas et al. [2016] showed high potential predictability of seasonal temperature through improved initialization of snowpack and soil moisture. In many midlatitude and high‐latitude regions spring snow is a major component of water storage and the primary source of water supply [Mote et al., 2005]. Therefore, accurate prediction of soil moisture and snow condition is essential to agricultural and water resource management drought mitigation needs [Barnett et al., 2005; Bales et al., 2006; Kumar et al., 2013]. Staudinger et al. [2014] demonstrated the usefulness of snow accumulation and melt to predict hydrological droughts in snow‐influenced catchments, especially for those climatic regions where snowmelt and the rainy season coincide.

Given the apparent difficulties in seasonal prediction of the 2012 drought, it is important to further understand its predictability. Regional climate modeling is a powerful tool to explore regional physical mechanisms, and ensembles can provide a measure of the prediction uncertainty. Ensemble studies of seasonal prediction of extremes create ensemble members through the use of perturbed initial conditions [e.g., Toth and Kalnay, 1993; Buizza and Palmer, 1995; Houtekamer and Derome, 1995; Done et al., 2014]. However, these initial condition ensembles are generally unable to provide the spread obtained through varying model physics [e.g., Buizza et al., 1999; Stensrud et al., 2000]. Stegehuis et al. [2015] demonstrated that regional multiphysics ensembles could help to understand and constrain uncertainties.

The goals of this study are to explore potential predictability sources of the 2012 drought in observations and in a regional model ensemble. A multiphysics regional model ensemble simulation of the 2012 drought is used to assess the robustness of the 2012 drought signal to combinations of physical parameterizations. Sources of potential predictability arising from soil moisture memory and relationships to snowmelt are then explored in the observations. These sources are then evaluated in the ensemble to assess their robustness to the representation of model physics, thereby improving understanding of our capacity to predict drought, especially beyond 1 and 2 months.

The next section describes the data and experimental approach. Results are presented in section 3, and a discussion and conclusions follow in section 4.

2 Experimental Setup and Reference Data Set

2.1 Model Configuration

The Weather Research and Forecasting (WRF) model [Skamarock et al., 2008] is an atmospheric simulation system designed for operational forecasting, atmospheric research, and dynamical downscaling of large‐scale data at very high resolutions and, in offering several parameterization schemes for most processes, is suitable for a multiphysics ensemble experiment. Drought simulations are conducted with the WRF model, version 3.5.1 at 36 km grid spacing over an extended North American domain of approximately 25°S to 70°N and from the African Coast to the East Pacific (Figure 1). Two periods are simulated: January 1990 to January 2000 to provide an estimate of model climatology [Bruyère et al., 2015] and September 2011 to December 2012 to explore the 2012 drought and allowing the model to spin up the 2011/2012 winter snowpack.

Figure 1

WRF model domain. The region of the Southern Rockies is indicated, and the Midwest region domain (approximately 35°N–50°N and 110°W–85°W) is shown in the red box.

Initial and boundary conditions are provided by the European Centre for Medium‐Range Weather Forecasts Interim reanalysis (ERA‐Interim [Dee et al., 2011], http://rda.ucar.edu/datasets/ds627.0/). This use of reanalysis data, rather than forecast data, to drive the regional model represents a best case for drought prediction, and assessments of predictability are therefore referred to as potential predictability. The experimental physics ensemble consists of two planetary boundary layer (PBL) schemes (MYJ [Mellor and Yamada, 1982] and YSU [Hong and Pan, 1996]), three cumulus (CU) schemes (KF [Kain and Fritsch, 1990], NSAS [Han and Pan, 2011], and Tiedtke [Tiedtke, 1989]), two radiation (RA) schemes (CAM [Collins et al., 2006] and RRTMG [Mlawer et al., 1997]), and two microphysics (MP) schemes (WSM6 [Hong et al., 2004] and Thompson [Thompson et al., 2004]), giving a total of 24 ensemble members (Table 1). The parameterizations were chosen to maximize sampling of the uncertainty range of simulated temperature and precipitation in preliminary sensitivity experiments (not shown). Each physics parameterization combination is assumed to be equally likely and produces a physically credible solution. Only one land surface scheme (Noah) [Chen and Dudhia, 2001] is used to focus this study on variations of the atmospheric component of the land‐atmosphere interface and keep the ensemble size within practical limits. Noah is the most widely used scheme in WRF applications.

Table 1. Twenty‐Four‐Member WRF Physics Ensemblea^a Radiation: CAM and RRTMG. PBL: MYJ and YSU. Cumulus: KF, NSAS, and Tiedtke. Microphysics: WSM6 and Thompson.

	CAM		RRTMG
	MYJ	YSU	MYJ	YSU
KF
WSM6	ck6m	ck6y	rk6m	rk6y
Thomp	cktm	ckty	rktm	rkty
Tiedtk
WSM6	ct6m	ct6y	rt6m	rt6y
Thomp	cttm	ctty	rttm	rtty
NSAS
WSM6	cn6m	cn6y	rn6m	rn6y
Thomp	cntm	cnty	rntm	rnty

• ^a Radiation: CAM and RRTMG. PBL: MYJ and YSU. Cumulus: KF, NSAS, and Tiedtke. Microphysics: WSM6 and Thompson.

This study does not additionally explore the role of internal variability for the potential predictability of the 2012 drought. Internal variability is an intrinsic property of the climate system, emanating from regions of conditional or baroclinic instability [Nikiéma and Laprise, 2011a, 2011b] and for regional domain results in a range of solutions consistent with the boundary conditions [e.g., Done et al., 2014; Lucas‐Picher et al., 2008]. Perturbed physics, on the other hand, produce solutions to a slightly different model climate, and the resulting range is commonly found to be large compared to internal variability [Solman and Pessacg, 2012]. We therefore choose to focus computational resources on the robustness of potential predictability sources to representation of atmospheric physical processes.

In addition to ensemble averages, ensemble subsets are analyzed with common RA, PBL, CU, or MP physics schemes, hereafter referred to as ENS, ENS_CAM, ENS_RRTMG, ENS_MYJ, ENS_YSU, ENS_KF, ENS_NSAS, ENS_Tiedtke, ENS_WSM6, and ENS_Thompson (Table 1). For example, ENS_YSU contains the 12 ensemble members that use the YSU PBL scheme. To remove model bias from each ensemble member, anomalies of temperature, precipitation, 500 hPa geopotential height, soil moisture, and snowmelt for 2012 are calculated with respect to their 1990–2000 means. Initial focus is a region covering parts of the Central Great Plains and the Midwest (approximately 35°N–50°N and 110°W–85°W, Figure 1, and hereafter referred to as Midwest) shifting to a focus on the Southern Rockies (Figure 1) where the drought initiated in May 2012.

2.2 Observed Data

Gridded daily temperature, precipitation, soil moisture, and snowmelt data are used to evaluate the ensemble and physical processes at the soil‐atmosphere interface, and sourced from the NASA Modern‐Era Retrospective analysis for Research and Applications (MERRA‐Land) at 0.6° × 0.5° grid spacing, available from 1 January 1980 onward [Reichle et al., 2011; Rienecker et al., 2011]. MERRA‐Land data have been widely used in different climatic regions [Bosilovich et al., 2011; Golian et al., 2014; Wong et al., 2011; AghaKouchak, 2014] and are consistent with observations in the midlatitudes, while uncertainties in high latitudes are often large [Kennedy et al., 2011; Yi et al., 2011; Reichle et al., 2011]. Monthly 500 hPa geopotential height is derived from ERA‐Interim at 0.7° × 0.7° grid spacing.

3 Results

3.1 Evaluation of the Ensemble for 1990–2000

The 2012 drought in each ensemble member is analyzed later in this section using anomalies with respect to the ensemble member monthly means over the period 1990–2000. An evaluation of the ensemble for 1990–2000 over the Midwest is first presented here to establish the skill of the ensemble member climatologies.

Monthly temperature biases over the Midwest are computed against MERRA‐Land. In general, ensemble members have a cold bias (up to 5 K) over the Midwest, peaking in winter and spring, and largest in CAM and Tiedtke (not shown). RRTMG generally produces a slight warm bias (≈1 K) in summer and generally outperforms CAM for winter (not shown).

Monthly precipitation biases vary substantially among the ensemble members but generally show a wet winter/spring bias (up to 40%) and a dry summer bias (up to 50%) with the largest dry bias in combinations containing CAM and MYJ (not shown). Combinations containing RRTMG and KF minimize the summer dry bias and also enhance the winter wet bias. In general, RRTMG, YSU, and KF outperform CAM, MYJ, NSAS, and Tiedtke for the simulation of Midwest summer precipitation. Similar to precipitation, the ensemble has a wet winter bias in soil moisture and a dry summer bias. All ensemble members generally overestimate spring snowmelt over the Midwest while underestimating snowmelt elsewhere. Attributing characteristics of the solution to components of the physical parameterizations schemes is not the purpose of this study, but some insight may be gained from studies that have focused specifically on this topic [e.g., Li et al., 2014; Yang et al., 2012].

3.2 Performance of the Ensemble for the 2012 Drought

The performance of the ensemble to simulate characteristics of the 2012 drought is presented here using 500 hPa geopotential height, temperature, precipitation, soil moisture, and snowmelt anomalies. Figure 2 compares summer (May to August) 500 hPa height anomalies for ERA‐Interim and the ensemble experiments. In the summer of 2012, a region of anomalously high heights, peaking over the Midwest, limited the southward push of rain‐producing cold fronts from the north (Figure 2a). Given that this regional anomaly is far away from the lateral boundaries and the initial conditions, this provides a hard test of the ensemble. In spite of this limited help from the driving data, the ensemble mean reproduces anomalously high heights over the western U.S. that extends over the Northern Plains. But the anomaly is weaker than observed and peaks farther to the west (Figure 2b). The majority of members show similar patterns in 500 hPa geopotential height to one another (Figure 2b). CAM performs better than RRTMG to reproduce the anticyclone over the Northern Plains. KF and WSM6 also reproduce an anticyclone but still weaker than in ERA‐Interim. Hovmöller diagrams (Figures 2c and 2d) show the evolution of 500 hPa geopotential height anomalies over the Midwest, averaged between 35°N and 50°N, derived from ERA‐Interim and the ensemble from January to December 2012 with respect to the period 1990–2000. These Hovmöller diagrams also show that the majority of ensemble members reproduce anomalously high heights over the Midwest in spring and summer but weaker than observed. More specifically, the ensemble fails to capture the period of lower anomalously high heights during May and the magnitude of the second peak of high heights during June and July centered around 100°W.

Figure 2

The 500 hPa geopotential height anomaly (m) with respect to the 1990–2000 monthly mean for (a, b) the period May–August 2012 in (a) ERA‐Interim and (b) ensemble data and (c, d) shown in Hovmöller diagrams over the Midwest. Dotted stippling indicates that more than 75% of the ensemble members agree in sign, and the line stippling indicates that the spread of the members is greater than the ensemble mean.

Figure 3 shows Hovmöller diagrams of temperature, precipitation, soil moisture, and snowmelt anomalies over the Midwest derived from MERRA‐Land and the ensemble for 2012 with respect to the period 1990–2000. An observed warm anomaly, peaking at 7 K in early March and 5 K in early July, dominated the spring and summer (Figure 3a). ERA‐Interim shows a slightly weaker warm anomaly than MERRA‐Land (not shown), meaning that this evaluation of an ERA‐Interim‐driven regional simulation using MERRA‐Land data is a hard test of the model ensemble. MERRA‐Land depicts the onset of a dry anomaly in May over the Southern Rockies that gradually spread from west to east and continued for the rest of the year (Figure 3b). This analysis describes the meteorological drought over the Midwest and a strong inverse relationship between summer precipitation and summer surface air temperature [e.g., Madden and Williams, 1978; Hoerling et al., 2013]. The associated dry soil moisture anomaly throughout the year (Figure 3c) affirms the severe agricultural drought that occurred throughout the Midwest.

Figure 3

Hovmöller diagrams of anomalies for 2012 with respect to 1990–2000 over the Midwest for (a, e) 2 m temperature (b, f) precipitation, (d, g) soil moisture, and (d, h) snow melt, derived from MERRA‐Land (Figures 3a–3d) and ensemble mean (Figures 3e–3h). The dotted stippling indicates that more than 75% of the ensemble members agree in sign, and the line stippling indicates that the spread within the members is greater than the ensemble mean.

Errors in simulated height anomalies are reflected in surface variables. The weaker and westward shifted high height anomalies are consistent with the weaker surface warm anomalies although the spatial extent of winter and summer warm anomalies is well captured (Figures 3a and 3e). The weaker height anomalies are also consistent with a weaker dry precipitation anomaly (Figures 3b and 3f). In general, precipitation anomalies vary substantially among the members in terms of sign, magnitude, and spatial extent (Figure 3f) but most capture a summer dry anomaly but much weaker than observed. Similar to precipitation, the ensemble members fail to reproduce the magnitude of the observed dry soil moisture anomaly in summer and additionally produce a brief dry soil moisture anomaly in spring that was not observed (Figures 3c and 3g). It is likely that the ensemble missed much of the increase in surface temperature through the suppression of evaporative cooling by the large dry soil moisture anomalies. For snowmelt, all members fail to reproduce the dry anomaly seen in MERRA‐Land (Figure 3h) through the winter months. However, the ensemble produces large dry snowmelt anomalies in spring contributing to large dry spring soil moisture anomalies, which is absent in MERRA_Land and likely related to the ensemble spring warm temperature anomalies (Figure 3).

In summary, this analysis shows that the drought signal in the five variables considered here, although underestimated in intensity, is robust to the way atmospheric physical processes are represented. Relationships between these five variables provide insight into the physical mechanisms of drought onset and duration and are explored further in the next section.

3.3 Precipitation Versus Soil Moisture Memory

The 2012 drought initiated in early summer over the Southern Rockies, and this is our focus region for the analysis of physical mechanisms. Accumulated precipitation forecasts have been used for meteorological drought prediction [Lyon et al., 2012; Quan et al., 2012; Hao et al., 2014; Yoon et al., 2012; PaiMazumder et al., 2013; PaiMazumder and Done, 2014]. However, dynamical model precipitation forecasts are subject to high uncertainty and models generally have difficulties in predicting statistics of precipitation with a few months lead time [Goddard et al., 2003; Livezey and Timofeyeva, 2008; Lavers et al., 2009. In this section, monthly precipitation and soil moisture data from MERRA‐Land and the 24‐member ensemble are used to explore the persistence properties of accumulated soil moisture and accumulated precipitation in the region of the Southern Rockies as potential predictability sources. The autocorrelation of accumulated 6 month precipitation and soil moisture for 1 to 6 month time lags for the period 1990–2000 is shown for the initial month of July in Figure 4. For example, July corresponds to accumulated precipitation and soil moisture from February to July. The autocorrelations of the accumulated 6 month soil moisture decay at a slower rate than the accumulated 6 month precipitation (bottom row of Figure 4a). For example, the autocorrelation coefficients of 6 month soil moisture are generally higher than 0.75 even at a 6 month lag, while the autocorrelations of the 6 month precipitation drop below 0.7 after a 1 month lag for the initial month July (bottom row of Figure 4a). Taking initial months of June and August shows similar behavior (not shown). The higher persistence of the soil moisture relative to precipitation suggests that the use of soil moisture would lead to better prediction of drought compared to precipitation. These results are consistent with AghaKouchak [2014].

Figure 4

Autocorrelation coefficients of accumulated 6 month precipitation and soil moisture for (a) MERRA‐Land and individual members for the initial month of July 2012 and (b) differences between individual members and MERRA‐Land.

The ensemble also has higher persistence of soil moisture relative to precipitation in the Southern Rockies (Figure 4a). The ensemble reproduces observed soil moisture memory with slight underestimation, while precipitation memory varies substantially among the ensemble members (Figure 4b). Notably, KF produces far shorter precipitation memory than observed. The capacity of the ensemble to simulate soil moisture memory suggests the potential to use soil moisture rather than precipitation for drought prediction.

3.4 Relationship Between Precipitation and Soil Moisture

In this section, we explore the general relationships between summer precipitation and concurrent and preseason soil moisture anomalies over the Southern Rockies using MERRA‐Land for two periods: 1983–2012 and 1990–2000. The 11 year period may contain substantial decadal variability, and so two periods are used to explore how the relationships are sensitive to the data length, and the ensemble over the period 1990–2000 is evaluated in that context.

Figure 5a shows a strong relationship (significant at 99% confidence interval using a two‐sided test) between observed summer (May‐June‐July‐August, MJJA) precipitation anomalies and observed soil moisture anomalies in MERRA‐Land for MJJA (r = 0.83, 0.85), April‐May‐June‐July (AMJJ) (r = 0.74, 0.77), March‐April‐May‐June (MAMJ) (r = 0.65, 0.69), February‐March‐April‐May (FMAM) (r = 0.59, 0.62), and January‐February‐March‐April (JFMA) (r = 0.56, 0.58) for the periods 1983–2011 and 1983–2012, respectively. This strong correlation also persists for the period 1990–2000. Including 2012 improves the correlations, suggesting that strong correlation between summer precipitation anomaly and preseason soil moisture anomaly persists in 2012 over the Southern Rockies. The significant correlation between summer precipitation and concurrent and preseason soil moisture anomalies reveals that a moderate fraction of the 2012 May–August precipitation deficit could have been predicted based on preseason soil moisture over the Southern Rockies.

Figure 5

Correlation between summer (MJJA) precipitation anomaly and summer (MJJA) and preseason (AMJJ, MAMJ, FMAM, and JFMA) soil moisture anomaly for MERRA‐Land for 1983–2011 (brown line), 1983–2012 (red line), 1990–2000 (blue line), and the ensemble mean (black line), and the ensemble spread is shown with vertical gray lines. Scatterplots between 24‐member WRF ensemble simulated summer (MJJA) precipitation and soil moisture anomalies for (b) MJJA, (c) AMJJ, (d) MAMJ, (e) FMAM, and (f) JFMA for 2012 over the Southern Rockies. In Figures 5b–5e, bold and dotted lines represent the observed regression line for 1983–2012 and 95% confidence estimates and square makers represent observed values corresponding to the 2012 anomaly.

Ensemble skill in reproducing the observed relationship between summer precipitation and concurrent and preseason soil moisture anomalies for the period 1990–2000 over the Southern Rockies is also shown in Figure 5a. The ensemble members reasonably capture the observed relationship for the concurrent season (0.327 ≤ r ≤ 0.946) although correlations reduce at a faster rate with lag. However, for a few members, the correlation between summer precipitation and preseason soil moisture anomalies is as good as observation or even higher than observation. Only 12 ensemble members reproduce the relationship between summer precipitation and late spring soil moisture anomalies (with highest r = 0.798), and only four members reproduce the relationship between summer precipitation and winter soil moisture anomalies (with highest r = 0.614). The decreasing ability of the ensemble to capture the lead‐lag relationship as lag increases is likely associated with poor simulation of the persistence and magnitude of the summer dry anomalies (Figures 3b, 3c, 3f, and 3g). Members ck6y, rt6y, and rntm (Table 1) are able to reproduce the relationship for all seasons.

Ensemble skill in reproducing this observed relationship for the 2012 drought over the Southern Rockies is shown in Figures 5b–5f. All ensemble members underpredict the dry soil moisture anomaly, and 75% of the members fail to reproduce the magnitude of dry precipitation anomaly, but the positive relationship between summer precipitation and preseason soil moisture is well captured. It is possible that the simple linear relationship may not extend to extreme events such as 2012 meaning that models that capture the relationship may not necessarily also perform well at capturing the extreme anomalies. Overall, this analysis suggests potential predictability of the 2012 summer precipitation anomaly using preseason soil moisture anomalies. As expected, the number of ensemble members that reproduce the relationship (i.e., lie close to the observed regression line) reduces as the time difference between precipitation and soil moisture is increased.

3.5 Relationship Between Soil Moisture and Snowmelt

Similar to the general relationship between precipitation and soil moisture anomalies, there is significant (at 99% confidence interval using a two‐sided test) correlation between winter snowmelt and concurrent and following season soil moisture anomalies over the Southern Rockies in MERRA‐Land for 1983–2011, which persists in 2012 and for the period 1990–2000 (Figure 6a). Interestingly, the shorter period shows a much more rapid dropoff of correlation with lag than the longer period, suggesting a role for decadal variability in the strength of the relationship. These significant correlations suggest that winter/early spring snowmelt could have been a predictor of the 2012 May–August soil moisture deficit and summer drought over the Southern Rockies.

Figure 6

Same as Figure 5 but for winter (JFMA) snowmelt and soil moisture anomalies for JFMA, FMAM, MAMJ, AMJJ, and MJJA.

For the period 1990–2000, a majority of the ensemble members reproduce the observed relationship between winter snowmelt and soil moisture for winter (0.321 ≤ r ≤ 0.896). The rate of correlation dropoff with lag also agrees well with observations for the same period. Members ctty, cntm, rntm, and rnty (Table 1) are able to reproduce the relationship between winter snowmelt and soil moisture anomaly for concurrent and following seasons, sometimes with a stronger relationship than observations.

For 2012, similar to the relationship between precipitation and soil moisture anomalies, the ensemble underpredicts the 2012 dry condition (Figures 6b and 6f). For 2012, 65–50% of CAM simulations and 65–35% of RRTMG simulations reproduce (i.e., lie within the 95% confidence bounds) the observed relationship for MAMJ, AMJJ, and MJJA (Figures 6d–6f). In general, the snowmelt‐soil moisture relationship is stronger than the precipitation‐soil moisture relationship in the model.

4 Discussion and Conclusion

The 2012 drought was notable for its rapid onset and intensification over the Southern Rockies and the Midwest during late spring/early summer and in the absence of known precursor large‐scale patterns. Prediction of such rapid onset events remains a major challenge. Using observations and a regional model multiphysics ensemble experiment, this study evaluated relationships among snow, soil moisture, and precipitation to identify sources of potential predictability of the 2012 summer drought.

The model ensemble was first established as appropriate to study the regional climate of the Midwest using a reference period of 1990–2000. The performance of the ensemble in reproducing the 2012 drought characteristics was assessed using 500 hPa geopotential height, surface temperature, precipitation, soil moisture, and snowmelt anomalies over the Midwest. In general, a drought signal was captured but far weaker than observed, with strong variability among the ensemble members, particularly for precipitation. Notably, the ensemble reproduced anomalously high heights within the interior model domain indicating some ability to capture drought signals with little support from boundary driving data.

An investigation into sources of potential predictability for this apparent flash drought focused first on persistence properties of accumulated soil moisture and accumulated precipitation. Soil moisture exhibited stronger persistence relative to precipitation in observations and the ensemble and could improve lead times of drought prediction beyond 1 and 2 months. AghaKouchak [2014] also highlighted the value of soil moisture persistence in drought prediction. However, ensemble members that reproduced the summer precipitation anomaly did not correspond to an ability to simulate precipitation memory, indicating that precipitation memory alone may not be a useful source of potential predictability.

To further understand sources of potential predictability, relationships were explored among snow, soil moisture, and precipitation over the Southern Rockies. Empirical correlations were found between winter/spring snowmelt and concurrent and following season soil moisture, and between soil moisture and concurrent and following season precipitation, in both observations and the model ensemble. These results demonstrate sources of potential predictability beyond 1 and 2 month lead‐time reside in the land surface conditions for apparent flash droughts such as the 2012 drought. This extends the work of AghaKouchak [2014] by linking potential predictability back to winter/spring snowmelt. Appropriately accounting for these sources could potentially improve operational seasonal drought outlooks and also has the potential to support seasonal fire outlooks.

The extent to which these results transfer to other droughts and drought forecasting systems depends on the further exploration of a number of limitations. In particular, only one land surface scheme was used here to focus on the influences of representations of atmospheric physical processes on soil‐atmosphere interactions while limiting the number of simulations. However, there may lie sources of potential drought predictability in relationships among the land surface variables of snow, soil moisture, and rainwater. Therefore, this proof‐of‐concept study provides a foundation to explore the robustness of the drought signal and potential predictability sources to changed representations of land surface processes. Another limitation is the finite ensemble size. Kolczynski et al. [2011] suggested that large ensemble size is required to capture true error and variance at the weather or climate scales. However, only 24 ensemble members were used in this study to keep the ensemble size within practical limits. Finally, teasing out the relative importance of model physics versus internal variability would be important to guide the construction of operational drought forecasting systems.

Acknowledgments

This work is partially supported by NSF EaSM grants 1419563 and 1048829. NCAR is sponsored by the National Science Foundation. We thank two anonymous reviewers and Cindy Bruyère and Erin Towler for their useful discussions and comments. The WRF ensemble data used in this paper are available at the National Center for Atmospheric Research and can be obtained from lead author Debasish PaiMazumder ([e-mail]) upon request.