We inventory potential flash drought events using a range of definitions. As we are concerned primarily with drought impacts on agriculture and natural vegetation, we focus our analysis on spring (MAM), summer (JJA), and fall (SON) and do not consider winter months. We consider seven methods for identifying a flash drought. The first – the Soil Moisture Volatility Index (SMVI) – is a new definition proposed here. The next five are drawn from published literature on flash droughts, and the seventh is based on a remotely sensed product designed to be sensitive to rapid onset droughts. Where data coverage allows, we use the 1979–2018 period for index calculation and comparisons. For some products, there is a more limited data record, and in those cases, we use all available data. Differences in input dataset requirements and baseline period can affect comparisons across definition and are noted when relevant. Here we describe each definition and present the datasets used to calculate them.

The analyses presented here have been organized using Bukovsky regions. The Bukovsky regions are 29 ecoregions over United States, Canada, and northern Mexico designed to represent climatically homogeneous areas. They are similar to the National Ecological Observatory Network (NEON) (Kampe, 2010) ecological regions, with similar sensitivity to variations in regional climatology (Bukovsky, 2011). Analyses were conducted over the 17 unique regions within CONUS (Fig. 3) as well as the eight grouped regions as suggested by Bukovsky (2011). Here we present results for a subset of regions that capture a relevant diversity of results, while results for all regions are available at https://github.com/mosman01/Flash_Droughts/ (last access: 2 February 2021).

The flash drought inventories presented in this paper are based on flash drought occurrence: as soon as a flash drought is identified according to a given definition in a given grid cell, that grid cell is tallied as having experienced flash drought in that year. That is, we are concerned with spatial pattern and general seasonality of the occurrence of flash drought events as diagnosed by different definitions. Intensity and duration of drought are not evaluated. Also, since definitions differ in if and how they mark the end of a flash drought event, we count only the first flash drought identified for a grid cell in each year. The season of this flash drought (MAM, JJA, or SON) is assigned based on onset date. This approach risks missing cases in which two distinct flash drought events hit a single location in one growing season, but it allows for a consistent inventory across definitions on the basis of “years with flash droughts.” The problem of counting multiple events at the same location in a single year using different definitions is a point for further research, as differences and ambiguities in how different definitions define the end of a flash drought can lead to cases in which one definition diagnoses multiple flash droughts within a period that is classified as a single flash drought in another definition. We do note that this approach captures the first drought, so it undercounts late season droughts if they occur in the same location as an early season drought. When calculating frequency, we use all the available data for each definition from 1979 to 2018.

For results presented by Bukovsky region we calculate the percentage of area within each region hit by flash drought in each year. This metric is used for qualitative comparison of definitions for selected events and for quantitative comparison using Pearson correlations. Spearman and Kendall correlations were also calculated but yielded similar results and are not presented. Finally, an analysis of the trends in flash droughts annual footprint is carried out for each climatic region within the Bukovsky regions using the Mann–Kendall nonparametric trend test. Trend analysis is only performed for the definitions that can be calculated for the full 40-year period (1979–2018).

As flash droughts have become recognized as a significant climate hazard, one key question is whether certain regions have an elevated probability of experiencing flash drought. As shown in Fig. 4, the seven drought definitions considered in this paper offer different answers to this question. This figure depicts the frequency of flash drought onset at each grid point within the specified season over the period of data availability for each definition through 2018. As noted in Christian et al. (2019a), the SESR identifies the Great Plains and western Great Lakes regions as hot spots for flash droughts. This band of high flash drought frequency running down the middle of the country resembles the region of strong land–atmosphere coupling identified in Koster et al. (2004) and in subsequent studies of climate feedback zones. In this sense, the SESR, which depends directly on the ratio of actual to potential evapotranspiration, may be emphasizing flash droughts that emerge through land–atmosphere temperature and evaporation couplings, which are strongest in transitional climate zones. There is a tendency for this SESR hot spot to emerge in the southern Great Plains in the spring (MAM) and to move further north in the summer (JJA).

Interestingly, this SESR pattern is nearly inverse to the pattern seen for PDD. In PDD, we see the strongest hot spot in the southwest, with a secondary maximum in the more humid eastern United States. While PDD includes actual evapotranspiration and temperature rules in its definition, it is designed to capture short meteorological droughts triggered by precipitation deficit. This results in higher frequencies in semi-arid regions with high precipitation variability and, to some extent, in regions where average rainfall is high and a significant negative anomaly in precipitation generally occurs in concert with the warm conditions required by the PDD definition. In contrast to PDD, the HWD yields a relatively uniform pattern of flash drought frequency, with lower totals overall.

Looking at the two soil moisture definitions, SMVI and SMPD, we see differences in overall frequency and spatial and seasonal distribution – which may reflect choice of threshold values. SMVI shows a relatively muted spatial pattern, with a broad maximum extending across the middle of the country and the western northern tier in summer and a southwestern maximum in fall. SMPD has a springtime maximum in humid regions of the eastern United States and the Pacific Northwest, followed by a summertime pattern that includes significant frequency in the southwest. These differences trace to conceptual differences in the definition. Where SMPD focuses on soil moisture decline over several pentads and thus is likely to capture vegetation-enhanced soil moisture drawdown that occurs in warm or dry springs in highly vegetated areas, SMVI controls for steady decline in order to isolate very rapid soil moisture drops. This makes it relatively less sensitive to seasonal forcing (e.g., warm springs leading to steady drying) and more sensitive to subseasonal processes. SMPD shows a noticeably high frequency of flash drought onset due to the duration threshold of four pentads or fewer, which allows short meteorological droughts to be misclassified as flash drought events.

Considering the hybrid products, USDM and QuickDRI both show a summertime maximum in flash drought frequency but with distinctly different spatial patterns. In general, the QuickDRI areas of maximum frequency occur in drier regions in the western United States, while USDM shows a maximum in the middle of the country that resembles the summertime SESR and SMVI patterns, though with a stronger maximum in Texas and Oklahoma. While it is difficult to diagnose the source of these patterns in a precise way given the composite nature of both products and the subjective component to USDM, it is likely that USDM authors are particularly attuned to agricultural impacts and thus focus on rapid drying events that have severe impacts on crops and pastures, while the QuickDRI satellite-derived product may also be capturing variability in natural ecosystems and regions with less intensive agricultural activities. Different datasets and different algorithms involved within such complex model-based products could be a considerable source of uncertainty and variability.

The identification of geographic or seasonal flash drought hot spots, then, depends strongly on the definition. This choice of definition, in turn, will depend on the objective of the flash drought study. Investigating flash drought with an emphasis on vegetative impact, for example, might usefully apply a flux-informed definition like SESR and would consequently focus on flash droughts in regions with land cover types associated with denser vegetation (e.g., agriculture, grasslands, and forests). A study or forecast system primarily concerned with the rapid intensification of a flash drought over either a humid or semi-arid region might employ SMVI, which explicitly controls for more gradual drying in order to isolate the most rapidly intensifying portion of the events.

The definition-based differences in the geography and seasonality of flash drought frequency described above suggest that definitions might also differ with respect to interannual variability. This is a particularly relevant issue for forecasting, as differences in interannual variability imply differences in the prediction-relevant drivers of flash droughts. Indeed, if we examine interannual variability in flash drought extent – defined as the percent area that experiences at least one flash drought in a given year, within a specified region of interest – we see substantial differences between definitions. Figure 5 shows the Pearson's correlation coefficients between different definitions' area hit by flash droughts annually for four different climatic regions. At CONUS scale (Fig. 5a), the correlation between certain definitions, such as the two soil-moisture-based definitions (SMPD and SMVI) and the USDM, is relatively high (>0.7). This still leaves substantial unexplained variability between definitions, but the differences between definitions are larger when comparing definitions that include other variables. SESR and PDD, for example, have virtually no correlation in interannual variability at CONUS scale, which is consistent with the differences seen in Fig. 4 and with the fact that the two definitions are based on very different principles and variables.

These differences become even more pronounced at regional scale. Figure 5b–d show regions in which differences are particularly dramatic – the Southern Plains, Pacific Southwest, and North Atlantic Bukovsky regions – and Fig. S2 in the Supplement shows the remaining regions. We note that Fig. 5 is designed to highlight regions with substantial disagreement between definitions; the full suite of regions shown in Fig. S2 includes a number of regions for which definitions are in closer agreement with each other.

The Southern Plains is of particular interest, since it is a hot spot in the USDM-based definition and is an active agricultural region. Here we see that the PDD and HWD definitions have no positive correlation with the USDM definition, which is again consistent with differences in spatial patterns seen in Fig. 4 and with the fact that PDD and HWD are defined to capture short droughts rather than periods of rapid intensification. Across other definitions, the correlations for the Southern Plains also tend to be (though are not always) lower than the CONUS-scale correlations. In the North Atlantic region, the PDD shows very weak correlations with all definitions except the HWD since they share the common heat wave condition. Moving to the more arid Pacific Southwest and Desert regions, we begin to see extremely low correlations across definitions, which in part reflects low signal to noise ratio for drought indicators in dry climate zones and in part may point to implicit limitations in the useful climatic range of each definition. In the Pacific Southwest, SESR stands out as having no positive correlation with any other definition except with QD1.0, which is small, and the USDM also shows very weak association with other definitions. This is a complicated region that includes arid zones and irrigated agriculture, which would pose complications for an expert-informed composite indicator like USDM and which is not represented in NARR or NLDAS. Large expanses of arid areas with sparse vegetation coverage might also reduce the utility of a flash drought indicator based on the actual to potential evapotranspiration ratio, such as SESR. Nevertheless, it is still possible that rapid onset droughts matter in the region, particularly if they drive up irrigation demand or impact natural semi-arid ecosystems. Specifically, for the Pacific Southwest region, all definitions show relatively lower flash drought frequency (SMVI, SMPD, USDM, SESR, and QD1.0; local minimums in Fig. 4) except for PDD.

Though there is no single agreed-upon definition for flash droughts, a number of major events in the past decade are widely recognized as having flash drought characteristics, to the point that these events can be thought of as canonical flash drought events. In addition, several major droughts that occurred prior to the popularization of the term “flash drought” have since been recognized as being consistent with flash drought. To obtain a clearer picture of how different definitions capture flash droughts, we examine several of these canonical flash droughts in greater detail.

We begin with an event that pre-dates the term “flash drought” but has since been recognized as a member of the class (Basara et al., 2020; Jencso et al., 2019; Trenberth et al., 1988; Trenberth and Guillemot, 1996). The 1988 drought in the northwest, central, and midwest United States developed over a period of less than 5 weeks, resulting in severe to extreme dry conditions over more than 10 states that cost the nation at least USD 30 billion (National Oceanic and Atmospheric Administration, 1988). There was below-average precipitation prior to the onset of the event, which contributed to its evolution. However, the most dramatic meteorological forcings were the pronounced and extended series of heat waves that gripped the country in June, July, and August and which were in their own right responsible for thousands of deaths (Changnon et al., 1996; Ramlow and Kuller, 1990; Whitman et al., 1997). These heat waves occurred in combination with below-average precipitation in June and July (Lyon and Dole, 1995). As this event predates QuickDRI and the USDM, we present a simple comparison of the other five flash drought definitions (Fig. 6). All definitions capture widespread drought, but timing and patterns differ. For example, whereas HWD emphasizes acute drought associated with high temperatures in JJA in the northern tier, SESR is more sensitive to evapotranspiration deficits across the middle of the country, which appear as a MAM signal in these seasonal maps. Similarly, SMPD is sensitive to dryness that appears in MAM, particularly in the eastern United States (consistent with the general spatial pattern of this definition; Fig. 6), while SMVI has characteristics of both the dry signal in the MAM window and intensification in the JJA period. We note that our seasonal cutoff dates are arbitrary and could mask differences in timing within a season (e.g., March vs. May) while emphasizing relatively small timing differences that cross a seasonal break (e.g., May vs. June). Nevertheless, the analysis captures the general character of the seasonal timing of events.

Jumping forward, in 2011, the Southern Plains experienced a rapid onset, geographically focused flash drought that led into an extended drought during the remainder of the year, making this one of the driest years in Texas since 1917 (Ejeta, 2012; Nielsen-Gammon, 2012). The different flash drought definitions show signs of an early onset in spring in Texas and the southeast (Fig. 6), which was the actual scenario according to the Office of the State Climatologist in Texas (Nielsen-Gammon, 2012), that then spread to other regions during the summer. SESR shows a more eastern pattern (where it is more humid), while the QD1.0 has a broad drought signal across the southern tier of the county, but overall agreement across definitions is quite good. This suggests that the 2011 flash drought has a consistent signature in multiple meteorological and hydrological variables, which can be explained due to the strong relationship between surface fluxes in the Southern Great Plains region (Mo and Lettenmaier, 2016).

The following year, 2012, produced one of the largest and most well documented flash droughts to date (Basara et al., 2019; Fuchs et al., 2015; Hoerling et al., 2013, 2014; Mallya et al., 2013; Otkin et al., 2016). According to post-event analysis, large-scale teleconnections may have set the stage for the flash drought onset in spring and early summer (Basara et al., 2019; Fuchs et al., 2015), with rapid intensification coming in summer as vegetation stress and heat set in. Results from the definitions (Fig. 6) show different patterns for the spread of the drought. While an extensive drought in the middle of the country was in some form by all definitions, the geographic pattern differed. Both HWD and SMVI, for example, capture a rapid drying in spring in Missouri and surrounding regions, as abnormally warm conditions led to rapid soil moisture drawdown. The USDM-based definition, in contrast, shows only limited drought in the MAM window, with widespread flash drought emerging in JJA. This likely reflects the fact that the USDM did not make extensive use of vegetation indices in 2012, such that it is not optimized to capture rapid droughts (Senay et al., 2008), and the warm spring conditions that set the stage for the catastrophic drought of summer are not identified as flash drought when using the USDM as the input variable.

In 2016, the southeast was hit by an “exceptional drought” (Svoboda et al., 2002), which sparked unusual wildfires that covered area more than had ever occurred since 1984, leading to the destruction of thousands of structures (Park Williams et al., 2017) and severe ecological and socioeconomic impacts (Konrad II and Knox, 2018). The southeast region has generally experienced an exceptional precipitation deficit since 1939 beside a rapid substantial increase in maximum air temperature and solar radiation (Konrad II and Knox, 2018; Park Williams et al., 2017) which amplified the event and resulted in the observed severe flash drought event over the months of the fall (Otkin et al., 2018). The 2016 flash drought was expected to extend eastward towards the Carolinas, but heavy precipitation from the tropical storms and hurricanes (Hermine and Matthew) that hit the region ended the catastrophic event (Konrad II and Knox, 2018). Results from SMVI and USDM-based definitions (Fig. 6) show similar spatial patterns; however, the USDM one shows an early timing for the onset in MAM and JJA, which is similar to what is captured by the QuickDRI-based definition. The SESR definition underestimated the spread of the drought event, capturing only very few spots of onset in spring and summer months. Despite the high temperatures and precipitation deficit, HWD and PDD did not show a clear pattern for the onset, which may be due to the lack of the rapid intensification criteria in both definitions (Otkin et al., 2018).

Finally, we examine the 2017 Northern High Plains flash drought. This was a geographically focused drought event that primarily affected Montana, North Dakota, and South Dakota (Jencso et al., 2019). In contrast to the geographically focused flash drought event of 2011, which was captured in a relatively similar way by most definitions, there is little consensus in the representation of the 2017 event (Fig. 6). Both USDM and SMVI show spotty areas of drought in the northern high plains in MAM that expanded during JJA, which is similar to the observed onset (Gerken et al., 2018; He et al., 2019; Jencso et al., 2019). This pattern is almost entirely absent in HWD (despite the likelihood of being driven by reduction in snowpack due to an early spring heat wave; Kimball et al., 2019) and is evident only in spots in Montana for PDD and North Dakota for SESR. SMPD identifies flash drought in this region in MAM and in some areas in JJA, but the region does not stand out relative to the rest of the country. Similarly, QuickDRI shows widespread drought conditions that are not focused on the northern high plains. These results show that the 2017 event qualified as a flash drought for some but not for all methods.

Building on the event analysis presented in the preceding section, we now examine meteorological fields in the region of maximum drought intensity for the 2011 and 2017 events – i.e., two regionally focused events, one of which presents relatively similar results across all of the definitions (2011) and one which does not (2017). To simplify the problem, we examine only the main climate variables used in creating the flash drought definitions (precipitation, RZSM, temperature, and actual and potential evapotranspiration).

During the 2011 flash drought event, temperatures rapidly went extremely high and stayed that way for most of the spring and the whole summer, as did potential evapotranspiration. While precipitation anomalies remained negative with very few exceptions, actual evapotranspiration decreased just after the rapid increase in potential evapotranspiration. The RZSM shows a relatively rapid decline in early summer, which occurs on top of a negative RZSM anomaly inherited from spring (Fig. 7a). In short, all of the key variables applied in the flash drought definitions show a clear signal of rapid change to dry and hot conditions that were sustained throughout the event, while precipitation stayed consistently low. For this type of event, choice of definition may not be critical when attempting to characterize, monitor, or predict the drought.

In contrast, during the 2017 Northern High Plains drought (Fig. 7b) temperature was highly variable, and SM and ET did not fulfill the HWD conditions for drought onset, so the HWD does not capture the observed drought onset. Precipitation was also less consistent, explaining why PDD is spotty and may have missed the onset in multiple locations. Potential evapotranspiration, interestingly, is fairly consistent even though temperature was noisy, so SESR captures the onset in some areas (though mostly misses Montana), and RZSM gives the clearest signal, which is why SMVI and, to some extent, SMPD do well. In essence, the 2017 event is a flash drought primarily from the perspective of rapid soil drying, likely reinforced by high evaporative demand. It is not a cleanly defined heat wave flash drought, and the rainfall signal is noisy. This suggests that efforts to understand and forecast an event like 2017 will be concerned with different variables and different biophysical intensification processes than were active in events like 2011.

Over the past century there has been an increase in precipitation over much of the United States (IPCC, 2018). Studies over the CONUS (Andreadis and Lettenmaier, 2006) also show positive trends in soil moisture and runoff, which lead to fewer hydrological drought events. At the same time, temperature has increased for much of CONUS in recent decades, and Mo and Lettenmaier (2016) show that there was a dramatic increase in HWD events in the 90s due to this rapid warming. An increasing trend in flash drought frequency according to this definition may be attributed to anthropogenic climate change as the rising temperature increases evapotranspiration in humid and densely vegetated regions, which consequently causes a decrease in soil moisture (Wang et al., 2016; Yuan et al., 2019).

In our analysis of flash droughts trends from 1979 to 2018 (USDM and QuickDRI definitions are not included due to the short period of data availability), we see an increase in areas hit by HWD and PDD over most of the CONUS region in the past decade (2009–2018) compared to 1979–1988 and almost no difference in SM-based and evaporative-demand-based flash drought definitions (Figs. 8 and 9). Insomuch as HWD and PDD indices capture acute drought anomalies rather than the rapid intensification targeted by other definitions, these results suggest that there is consensus across definitions that the frequency of rapidly intensifying flash droughts did not increase between 1979–1988 and 2009–2018.

Considering each Bukovsky region (Fig. 8), however, we do observe different patterns of change in the percentage of area experiencing flash droughts over time. For example, the western coast (Pacific regions and Southwest) shows an increase in areas experiencing flash droughts, while the Northern Plains and Rockies are characterized by a decrease in flash-drought-impacted areas. PDD shows positive trends in almost all regions, and about half of the regions show a statistically significant trend. HWD is also positive in almost all regions, with the majority of these trends showing statistical significance (Mann–Kendall test at p<0.05; Fig. 8). Trends in PDD and HWD are also positive and significant for CONUS on the whole. Trends for SMVI, SMPD, and SESR are mixed in sign and generally not significant.

The presence of significant trends in PDD and HWD can be attributed to the fact that both directly depend, in part, on air temperature. The other definitions are indirectly influenced by air temperature through its impact on evaporative demand and soil moisture, but trends in those mediating variables are not as clear as the trend in temperature over the period of study. Insomuch as there are systematic trends in flash drought across CONUS, then, it appears that those trends are only prevalent in definitions that include the meteorological drivers of flash drought in the definition of the event. In this study, those definitions are limited to PDD and HWD, which are definitions that target acute drought anomalies rather than rapidly intensifying flash droughts. The trends are not evident when a definition depends only on a drought outcome of interest, such as soil moisture or evaporative stress. We do note that there are very few cases of direct disagreement in sign between statistically significant trends across definitions. This only occurs in the Central Plains, where SMPD differs in sign from HWD and PDD, and in the arid Great Basin region, where SMVI shows a significant positive trend, while SESR is significantly negative.