Warm season dry spells over the central and eastern United States are classified into three canonical types via a hierarchical cluster analysis for the period 1950–2005. Four CMIP5 models exhibit diverging skill in representing the observed behavior, ranging from southern Great Plains dry spells that are reasonably simulated by all four models to southeastern U.S. dry spells that are only accurately captured by one model. A model’s skill in representing a particular dry spell cluster is positively correlated with the model’s ability to simulate the large-scale meteorological patterns (LMPs) accompanying the dry spell. The interannual variability and overall observed decreasing trend in dry spell days are represented with varying degrees of accuracy by the four models. The results 1) highlight existing shortcomings in the climate model representation of regional dry spells and 2) illustrate the importance of properly simulating the observed spectrum of LMPs in minimizing these shortcomings.
Droughts associated with large-scale atmospheric waves, jet streams, and subtropical high pressure systems constitute one of the most costly natural hazards in the United States (Rauber et al. 2005). Previous studies have examined the dynamical triggers for, and synoptic organization of, droughts over the central and eastern United States during the boreal warm season. For example, summer droughts over the Midwest tend to be driven by an anomalous high pressure over Canada, which transports dry air of the northern Canada southward and suppresses the moisture influx associated with low-level jets emanating from the Gulf of Mexico (Sud et al. 2003). Namias (1966) showed that negative sea surface temperature (SST) anomalies along the Atlantic coast force cyclones to track farther east than normal, leading to anomalous northerly winds and drought conditions over the northeastern United States during spring and summer. Low-frequency drought variability in the southeastern United States is associated with the strength and position of the North Atlantic subtropical high (NASH) during summer (Ortegren et al. 2011).
Global climate models [e.g., those from phase 5 of the Coupled Model Intercomparison Project (CMIP5)] have been widely used to assess water-cycle extremes such as U.S. drought events. Sheffield et al. (2013b) illustrated that during dry summers over the southern Great Plains, both precipitation deficits and an associated anticyclonic moisture flux anomaly pattern are well represented by the Community Climate System Model, version 4 (CCSM4; Gent et al. 2011) despite an overall bias in the simulated moisture source. Wuebbles et al. (2014) demonstrated that a subgroup of CMIP5 models simulates summer precipitation variability over the southeastern United States quite well; this is attributed to a good representation of the link between precipitation and NASH. In terms of the potential impacts of global climate change on regional hydrological cycles, Wuebbles et al. (2014) project a decreasing trend in the net surface water balance (precipitation minus evapotranspiration) over the south-central United States during spring by contrasting CMIP5 future projections (2073–99) to corresponding historical simulations (1979–2005). The fidelity of representing drought trends and variability in global climate models is crucial for assessing a wide range of detrimental drought impacts (e.g., decreased crop yields, increased wildfires, water supply deficiency, and rising food prices) (Witt 1997).
Meteorological drought is often measured in terms of negative precipitation anomalies (e.g., deviations from a long-term daily climatology) or indices derived from meteorological or hydrological data [e.g., the standardized precipitation index (SPI)] (Keyantash and Dracup 2002; Mo et al. 2012; Sheffield et al. 2013a,b). However, a negative precipitation anomaly is not necessarily equivalent to absolute “dry” conditions. Other studies have defined drought as times when the total daily precipitation rate falls below a threshold value [e.g., 0.1 in. (=2.54 mm) or 1 mm], which potentially better characterizes the existence of the true dry conditions (Groisman and Knight 2008; McCabe et al. 2010). Following a similar approach, the current study aims to systematically identify and document distinct warm season dry spells exhibiting daily precipitation rates of nearly zero lasting 10 days or more over the central and eastern United States. Specifically, we wish to evaluate the capability of state-of-the-art climate models in representing the basic character of such dry spells including their spatial distribution, occurrence frequency, interannual variability, long-term trends, and the associated large-scale meteorological patterns (LMPs).
2. Data and methods
The observed daily precipitation rate data used in our analysis are from the National Oceanic and Atmospheric Administration (NOAA)/Climate Prediction Center (CPC) (Chen et al. 2008) while the daily-mean sea level pressure (SLP) and horizontal winds are from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (Kalnay et al. 1996). Outputs from the historical climate simulations of four CMIP5 models, which are widely used in the United States and Europe for climate assessment and have relatively high resolutions, are compared with the observation for the same period (1950–2005) (see Table 1 for details). Because of the differences in spatial resolution between the observational datasets and the respective model output, the input data are transformed to a common grid prior to performing our analysis. Specifically, the analysis of daily precipitation rate and corresponding circulation anomaly (deviation from its long-term daily climatology) patterns are performed at the spatial resolution of the CCSM4 output and NCEP–NCAR reanalysis data, respectively. The analysis is focused on the boreal warm season (March–August) and considers a domain encompassing the central and eastern United States (20°–50°N, 55°–100°W).
b. Definition of dry spells
For the observation, dry days are defined at each grid point in the following manner: a day is considered to be dry if the daily precipitation at the grid point falls below 0.01 mm. To reduce the impact of potential systematic errors in model-simulated precipitation rates (e.g., Sheffield et al. 2013a), dry days in the model simulations are defined using a method similar to “quantile to quantile mapping” (Maraun 2013; Cannon et al. 2015). Specifically, in step 1, for each grid point over the central and eastern United States we assess the observed daily precipitation rate percentile (among all days in the warm season) corresponding to a rate of 0.01 mm day−1. In step 2, the same percentile measure is applied to identify a suitable threshold value at each grid point in a model. The resulting areal-average values of the precipitation thresholds used in our analysis are summarized in Table 2. Dry days are then identified at each grid point over the central and eastern United States during each warm season day for the period 1950–2005.
If the number of consecutive dry days at an individual grid point is greater than or equal to 10, we define each day falling within this period as a dry spell day for that grid point. To focus on dry spells having substantial regional impact, we further require that on any individual day the total number of dry spell grid points must exceed 5% of the total number of grid points within the domain considered (the central and eastern United States). The gridded daily precipitation rates for all days marked as dry spells then serve as the input into a hierarchical cluster analysis (e.g., Ward 1963; Park et al. 2014) that is used to categorize dry spell types in terms of distinct spatial patterns. The cluster analysis is terminated objectively at a step where a significant “jump” in the smallest Ward’s distance (which is used to determine which two clusters should be merged at each step) occurs. Following this approach, we identify a total of three dry spell clusters (types) for each dataset considered (the observation and each model simulation). To further check the robustness of the dry spell clusters identified from the observation, we perform several additional sensitivity experiments by perturbing 1) the duration of the warm season (varied from April–July to March–August), 2) the definition threshold for a grid-point-level dry day (ranging from 0 to 0.1 mm day−1), 3) the number of consecutive dry days of a dry spell for individual grid point (varying from 7 to 12 days), and 4) the percentage of the number of grid points over the central and eastern United States required for excluding very small-scale events (from 5% to 10%). The results obtained in these sensitivity experiments are all qualitatively similar to those in section 3a. Denoting a dry spell of 1-day duration as a dur-1 event and a 2-day duration as a dur-2 event and so on, we also assess the mean duration for each classified dry spell type as
where t denotes the duration of individual dry spells in days (T being the maximum duration), and ndur-t is the total number of dry spells with a duration of t days.
a. Dry spell clusters
The hierarchical clustering applied to the observed warm season dry spells results in three distinct structural types (clusters) for the central and eastern U.S. region. The cluster mean daily precipitation rate is shown in Fig. 1a (panels 1–3); only features statistically significant (according to Student’s t test) at the 0.01 level are plotted. The first dry spell pattern affects the southern Great Plains and has the highest occurrence frequency (691 days out of a total of 1659 days) among all three clusters (Table 2). Cluster 2 corresponds to dry spell events over the Midwest which have the shortest mean duration (7.6 days) compared to the other two clusters (Fig. 1a, panel 2). Cluster 3 is characterized by the longest mean duration (9.6 days) and lowest occurrence frequency (372 days) and represents an elongated dry zone extending northeastward from the Gulf States and covering most of the southeastern United States. (Fig. 1a, panel 3). To further understand the potential seasonality in the clusters, we examined the dry spell clusters identified separately for the spring (March–May) and summer (June–August). Results indicate that the three warm season dry spell clusters discussed above are more similar to the three clusters identified in the spring, suggesting a greater contribution of the spring season dry spells to the overall characteristics in the warm season (figures not shown).
Figures 1b–e show the corresponding dry spell cluster patterns derived from four parallel CMIP5 historical climate simulations. The total dry spell occurrence frequency (Table 2) is well simulated by the Geophysical Fluid Dynamics Laboratory Climate Model, version 3 (GFDL CM3; Donner et al. 2011) and the Max Planck Institute Earth System Model, medium resolution (MPI-ESM-MR; Zanchettin et al. 2013), while dry spell frequency is considerably underestimated in both CCSM4 and the EC-Earth Consortium model (EC-EARTH; Hazeleger et al. 2010). In general, these four models have reasonable skill in simulating the mean duration day of each dry spell type (Table 2). In terms of the spatial pattern, all four models can represent the observed southern Great Plains dry spell pattern (cluster 1). The Midwest dry spell (cluster 2) is well simulated by GFDL CM3, CCSM4, and MPI-ESM-MR despite the fact that for the latter two models the pattern includes a weak-amplitude extension into the southern and eastern United States. The cluster 2 spatial pattern in EC-EARTH is substantially different from the observation (Fig. 1e, panel 2). For southeastern dry spells (cluster 3), only GFDL CM3 captures this distinct pattern (Fig. 1b, panel 3). The results suggest that GFDL CM3 provides an overall better performance in capturing major patterns of warm season dry spells over the central and eastern United States.
b. Large-scale meteorological patterns associated with dry spell clusters
Figure 2a displays composite anomalies in 250-hPa streamfunction and SLP for the three observed clusters. Cluster 1 is characterized by an east–west series of three anticyclonic circulation anomalies located over the North Pacific, continental United States, and North Atlantic, respectively (Fig. 2a, panel 1). The high pressure anomaly feature over the central United States is responsible for a northward displacement of the jet stream from its climatological position, which typically enables rain-producing synoptic-scale disturbances to propagate into the central United States in summer (Rauber et al. 2005). The presence of an anticyclone in the central United States also effectively blocks the transport of moisture into the Great Plains from both the Pacific Ocean and Gulf of Mexico, helping to contribute to the formation of dry spells. The LMP associated with cluster 2 (Fig. 2a, panel 2) exhibits features of a stationary wave train originating from the central North Pacific and extending through North America to the North Atlantic, as discussed in Lyon and Dole (1995) and Chen and Newman (1998). One likely driver of this stationary wave train is the appearance of steady heating anomalies associated with the northward shift of the intertropical convergence zone (ITCZ) (Trenberth et al. 1988). Similar to cluster 1, cluster 3 is associated with a high pressure center over the central-eastern United States that reduces moisture transport from the Gulf of Mexico and Atlantic Ocean (cluster 3; Fig. 2a, panel 3). However, this feature appears more regionally localized than part of a zonal wave train.
Figures 2b–e show the corresponding dry-spell-related LMPs derived from the four models. For the southern Great Plains dry spells, which are well represented by all the models, the related LMPs are also generally well simulated by the models (Figs. 2b–e, panel 1). The LMP for Midwest dry spells is well reproduced by GFDL CM3, CCSM4, and MPI-ESM-MR, consistent with these three models’ good performance in representing the Midwest dry spell precipitation pattern. Compared to the observed LMP, the relevant wave train in EC-EARTH originates from higher latitudes with very weak amplitudes at lower latitudes, leading to the unsatisfactory representation of dry spells over the Midwest (Fig. 2e, panel 2). Unique among the four models, GFDL CM3 very well simulates the LMP corresponding to the southeastern U.S. dry spells albeit with considerably larger magnitude than found in the observation (Fig. 2b, panel 3).
c. Long-term trend and interannual variability of dry spell clusters
To quantify interannual variability and potential long-term changes in dry spells, we next assess linear trends of, and interannual standard deviation in (using detrended data), the annual number of dry spell days over the period 1950–2005 (Fig. 3). For the total (sum of all clusters) number of dry spell days, there is a significant downward trend (−0.65 days yr−1) in the observed events between 1950 and 2005 with cluster 1 dry spells providing the largest contribution to this trend. This indicates a long-term decrease in drought conditions over the southern Great Plains.
Parallel statistical analyses are performed for the model simulations (for each dry spell cluster, only the models providing a reasonable representation of a particular cluster are included in the calculation). It is interesting to note that, consistent with its overall better representation of dry spell types and associated LMPs, GFDL CM3 also qualitatively best captures the interannual variability and long-term trend behavior of observed dry spells, albeit with relatively weaker and less significant trend values. These discrepancies demonstrate the large deficiencies that exist in the representation of the variability and trends in regional drought conditions by modern climate models, presenting a major challenge for long-term planning efforts aimed at mitigating potential detrimental effects of climate variability and change on regional water supply. Note that the direct cause of model misrepresentations of “climatological” dry spell patterns is largely the misrepresentation of the relevant LMPs. However, the misrepresentation of dry spell variability and trends in a model more likely reflects the fact that regional circulation changes in association with global warming and/or phase changes of natural decadal-to-multidecadal variability [e.g., the Pacific decadal oscillation (PDO) and Atlantic multidecadal oscillation (AMO)] are either not well captured or temporally shifted by a model (Robinson et al. 2002; Wang et al. 2009). These circulation and phase changes manifest themselves in planetary-scale flow patterns that are fully interactive with LMPs producing dry spells, thus effectively project into regional dry spell behaviors.
4. Summary and concluding remarks
The canonical patterns of warm season dry spells occurring over the central and eastern United States are identified via a hierarchical clustering analysis for the period 1950–2005. The three identified types (clusters 1–3) correspond to dry spell patterns occurring over the southern Great Plains, the Midwest, and southeastern United States, respectively. Cluster-mean composites are constructed to isolate the LMPs (in terms of anomalies in 250-hPa streamfunction and SLP) associated with each dry spell type. These LMPs directly generate dry spell events via the (i) diversion of precipitation-producing weather disturbances and (ii) the regional suppression of moisture transport from nearby oceanic surfaces. The LMPs can be divided into two categories: 1) a series of three anomalous anticyclones located over the North Pacific, continental United States, and North Atlantic, respectively (in clusters 1 and 3), and 2) a quasi-stationary wave train originating from the central North Pacific and extending downstream through North America to the North Atlantic (cluster 2). The total of number of dry spell days over the central and eastern United States decreases during the period 1950–2005, which is mainly a result of the decreasing occurrence frequency of dry spells over the southern Great Plains.
The four CMIP5 models exhibit diverging skill in representing observed dry spells. While all four models reproduce the spatial pattern of southern Great Plains dry spells reasonably well, only GFDL CM3, CCSM4, and MPI-ESM-MR successfully capture the Midwest dry spell pattern, and only GFDL CM3 effectively simulates the southeastern U.S. dry spell pattern. An individual model’s skill in representing a particular dry spell cluster is positively correlated with its parallel skill in representing the accompanying dry spell LMP. This highlights the importance of correctly simulating the broad spectrum of large-scale atmospheric disturbances, either those internally generated or externally driven, in faithfully representing extreme hydrological events in global models. GFDL CM3 captures the observed decreasing trend of dry spell days over the central and eastern United States, particularly over the southern Great Plains. However, the magnitude of the trend is significantly underestimated by GFDL CM3 and all four models exhibit biases in representing the interannual variability in dry spell days over the southern Great Plains. These model biases are ultimately results of model biases and uncertainties in capturing regional circulation changes associated with global warming and also the right temporal evolutions of natural decadal-to-multidecadal modes of variability (e.g., PDO and AMO) that are fully interactive with dry-spell-producing LMPs, suggesting grand challenges in utilizing model climate projections in long-term planning efforts on minimizing the detrimental effects of climate change on regional water supply. The results presented here also provide some overarching general guidelines for improving modeling activities that target simulations of hydrological extremes. These include bias reduction via a structured examination of the simulated spectrum of large-scale atmospheric disturbances, which is an essential component of the authors’ ongoing research efforts. Finally, we want to stress that although LMPs are essential in initiating the dry spells over the central and eastern United States, local land feedback processes can be important in sustaining and amplifying the dry conditions. Additional research is required to fully delineate the dynamical (e.g., LMPs) and thermodynamical (e.g., soil moisture feedback) nature of dry spells over the United States.
This study is supported by the DOE Office of Science Regional and Global Climate Modeling (RGCM) program through a cooperative agreement between the DOE and Georgia Institute of Technology under Contract DE-SC0012554. Yi Deng is also partly supported by the National Science Foundation Climate and Large-Scale Dynamics (CLD) program through Grants AGS-1147601, AGS-1354402, and AGS-1445956.