How do we identify flash droughts? A case study in Central European Croplands

Droughts are among the most extreme climatic weather events that threaten food security (FAO 2021). They have negative impacts on the global food–energy–water nexus and the sustainable development goals (D'Odorico et al. 2018). Droughts are generally characterised by unusually high levels of rainfall deficit, runoff deficit or soil moisture deficit (Palmer 1965; Mishra & Singh 2010) and are expected to increase in many regions of the world in terms of frequency, severity and duration under current and future climate change conditions (Lesk et al. 2016; Samaniego et al. 2018).

Flash droughts (FDs) are a special form of drought. In contrast to the descriptions of classical droughts (de Araújo & Bronstert 2015; Oikonomou et al. 2020), FDs are characterised by rapid onset and relatively short durations (Otkin et al. 2018; Lisonbee et al. 2021). They are associated with severe and immediate soil moisture depletions, resulting in plant water stress and mortality (Ford & Labosier 2017; Liang & Yuan 2021; Osman et al. 2021).

One of the first FD studies was by Peters et al. (2002), who were among the first to study single short drought events in late summer, which were characterised by the concurrence of low antecedent moisture and unusually high temperature. Interest in FDs has increased over the last few years, motivated by extreme FD occurrences in the USA, Russia and China, which caused extreme impacts on managed vegetation, disruption to the global food supply and increased wildfires (Otkin et al. 2013; Mo & Lettenmaier 2016; Christian et al. 2019a, 2019b; Liang & Yuan 2021). An FD in Australia during the spring of 2019 is thought to have played a central role in the massive forest fires that consumed over 1.6 million hectares (Nguyen et al. 2021).

Over the last two decades, multiple methods have been proposed to identify FD events (Lisonbee et al. 2021), yet there is no consensus on what an FD entails and how they may be defined in terms of onset, duration, the velocity of intensification and absolute or relative changes (Osman et al. 2021). In their recent literature review, Lisonbee et al. (2021) identified as many as 20 studies with different definitions using climate variables or indexes related to soil moisture, air temperature, precipitation, and actual and potential evapotranspiration (ET). Eleven of these definitions included an interval of intensification or rapid onset as part of flash flood delineation, whereas nine studies merely considered short-term drought events as FDs. A method comparison study by Osman et al. (2021) identified four core types where the definition of a heatwave FD (Mo & Lettenmaier 2015) was based on temperature anomalies, rapid soil drying (Ford & Labosier 2017; Yuan et al. 2019), actual and/or potential ET anomalies (Christian et al. 2019a, Pendergrass et al. 2020) and multi-criteria indexes (Chen et al. 2019). For the USA, they showed that the FD frequency, spatial extent and onset would vary significantly depending on which definition is used. They also suggested that a root-zone soil moisture-based method effectively captures the FD onset in both humid and semi-arid regions. Other than the study by Osman et al. (2021), which focused on the identification of spatial differences in delineated FDs, there have been no systematic studies on the temporal divergence and synchronicity of delineated FDs. At the same time, little is known about FD dynamics in Central Europe and it is not known which FD method would apply to this region, given that the present methods have so far been used for FD identification mainly outside Europe. Additionally, little efforts have been made towards an ensemble method for FD identification. In this study, we propose a method of this nature tested for areas in Central Europe.

Due to the lack of a general definition, we define an FD as the process of rapid, accelerated and unusually large depletion of root-zone soil moisture, in comparison with ‘average’ moisture conditions, due to the simultaneous or concurrent occurrence of two or more atmospheric and/or weather conditions over a short period of several weeks during the main growing season.

The objective of this study was to compare, in a local/point (climatological station) scale, the functioning of six recently developed FD identification methods with data from four well-monitored experimental cropland sites in Central Europe, by assessing co- and misidentification of FD time series using similarity and synchronicity metrics. We selected two soil moisture-based methods (Ford & Labosier 2017; Osman et al. 2021) and four indirect methods that used single or multiple climatic variables or indices for FD delineation (Christian et al. 2020; Noguera et al. 2020; Pendergrass et al. 2020, and a multi-criteria method proposed by the authors and validated for Central Europe). The methods were implemented in an R package and a Shiny app available to the public.

The following six FD identification methods were selected on the basis that they used station data as input and, following our definition, included a clear definition of the rapid onset of water limitation. The first two methods are based on soil moisture and the other four used indirect proxies of drought conditions, such as anomalies of rainfall, temperature and the ratio of actual and potential ET. The methods used are described as follows:

M2:Ford & Labosier (2017) identified FDs as periods when the pentad-average 0–40 cm volumetric water content declines, from at least the 40th percentile to below the 20th percentile, in four pentads or less (Figure 1(b)).

M3: The multi-criteria method is a new method that uses a set of 10 anomalies and indexes derived from weekly precipitation, temperature and potential ET data. It calculates a score for each week equivalent to the proportion of indicators that meet or surpass the respective pre-set thresholds. Weeks with a score higher than 0.65 and an absolute change of score (Δscore) higher than 0.25 over up to 3 weeks are classified as FDs (Figure 1(c)). The event duration is computed as the time from the beginning of intensification until the score is below 0.65. A full description of this method is provided in the Supplementary Material.

M4:Christian et al. (2020) used the standardised evaporative stress ratio (SESR), which is derived as the z-score of the quotient of actual to potential ET rate values for a specific pentad. They used four criteria which FD events were required to have: (1) a minimum length of five SESR changes, equivalent to a length of six pentads (30d minimum length); (2) a final SESR value below the 20th percentile of SESR values; (3) SESR changes must be at or below the 40th percentile between individual pentads and no more than one SESR change above the 40th percentile following the previous criterion and (4) an overall mean change in the SESR during the entire length of the FD must be below the 25th percentile in the SESR (Figure 1(d)).

M5:Noguera et al. (2020) used the standard precipitation evapotranspiration index (SPEI) on a short timescale (1 month) and performed calculations based on a temporal frequency of 1 week (four per month). To identify the rapid onset of a drought event, the change in the SPEI for each week, in periods of 4 weeks, was calculated and the onset of an FD was defined as involving a change in the SPEI equal to or less than −2 SPEI units (z-values) over an intensification period of 4 weeks. Further, final SPEI values had to be equal to or less than −1.28 SPEI units (Figure 1(e)).

M6:Pendergrass et al. (2020) utilised the evaporative demand drought index (EDDI) following Hobbins et al. (2016) and Lukas et al. (2017), which is calculated based only on the potential ET using the Penman–Monteith equation. The method identifies an FD when a 50% increase in the EDDI over 2 weeks is sustained for at least another 2 weeks (Figure 1(f)).

Workflow and key characteristics are summarised in Figure 1 and Table 1. Readers interested in a more detailed description of the percentiles and thresholds are referred to the original papers.

The six methods used one or more climate variables (rainfall, temperature, soil moisture and actual and potential ET). Further, they all share an underlying set of characteristics:

• 1.

  FDs evolve rapidly, with an intensification period lasting between 2 and 4 weeks.

• 2.

  The final conditions at the end of an FD lean towards extreme values, which are often characterised by the variable reaching values under the 20th percentile or, in some, a z-score value over ±1.

• 3.

  FDs are considered seasonal phenomena and are identified based on the expected values of climatic variables for each specific time of the year subdivided either in pentads or weeks.

• 4.

  FDs depend on crossing certain thresholds and are thought to be correctly identified if, and only if, environmental conditions meet a set of predefined rules.

The key differences between the methods are how the FD variables and durations are defined. Methods M1 – Osman et al. and M2 – Ford & Labosier take a direct approach to assess plant water availability using soil water data, whereas all other methods use proxy variables, which are likely to be less accurate, while simultaneously overcoming severe data limitations. Additionally, differences exist in the definition of the onset of the FD periods, the time resolution used (weeks and pentads), the minimum period over which it should be sustained and the maximum duration beyond which it might be considered a ‘normal’ drought (Table 1).

Additionally, the definition of event duration varies among methods. Most methods (M1 – Osman et al., M2 – Ford & Labosier, M3 – multi-criteria and M6 – Pendergrass et al.) assume that the duration of an FD event comprises the intensification phase (between 1 and 3 weeks), a period of persistence of the dry conditions (between 2 and 4 weeks) ending with the recuperation phase when the key variable (climatic or groundwater conditions) surpasses a threshold (Table 1). The methods of M4 – Christian et al. and M5 – Noguera et al. have different assumptions regarding duration. M4 – Christian et al. 2019a assume that the FD event is limited to the intensification period, with the FD event lasting only as long as there is a trend towards dry conditions and not having a threshold for the recuperation phase (Christian et al. 2019a). M5 – Noguera et al., although characterizing all three phases (intensification, persistence and recuperation), decided in their work not to consider the intensification phase as part of the FD duration. In this study, the application of the method of Noguera et al. (2020) includes the intensification period in the computation of the total duration of FD events.

All six methods were implemented in R programming language and are organised in an R package named ‘fdClassify’¹ and in a Shiny App named ‘FD-Viz’².

Two types of metrics were used to compare the six FD methods: synchronicity metrics and the confusion matrix. The method of Osman et al. (2021) was used as a reference method as it was considered the method that most closely followed our definition of an FD as given in Section 1 (using the rapid soil moisture decline as a key variable for FD identification). This method was evaluated against measured soil moisture data (see Section 2.3) in the root zone and, therefore, appeared to be particularly suited to reproduce the FD dynamics of croplands.

The concept of synchronicity based on the work of Kemter et al. (2020) was employed to compare the rate of identification of FD events and the intervals, which were correctly identified as intervals with no FD, for a weekly resolution. Synchronicity metrics were originally developed to analyse whether extreme floods occur concurrently with the same timing in larger basins (Kemter et al. 2020). It is based on the two synchronicity metrics sync₁ and sync₀, which are defined as the average proportion of the successful identification of FD events and no FD events, respectively. Here, the Osman reference method was compared separately with the other methods. The metrics can take values between zero and one; therefore, a perfect agreement between one method and the Osman et al. (2021) reference method occurs if both metrics are equal to one. If only sync₁ is close to one and sync₀ is close to zero, it would mean that all FD events were identified as in the reference method, but the interval times without FDs were all identified incorrectly (by identifying many more FD events, where the reference method did not find one). The two metrics sync₁ and sync₀ are further summarised by their harmonic mean (which is given by the reciprocal of the arithmetic mean of their reciprocals) to represent the trade-offs between the two synchronicities.

The four scores are then summarised into several confusion metrics that describe the degree of similarity (equations in Table 2):

• 1.

  The true positive rate (TPR) and positive prediction value (PPV) evaluate the performance of the tested method to correctly replicate an FD.

• 2.

  The true negative rate (TNR) and negative prediction value (NPV) evaluate the performance of the tested method to correctly replicate the intervals between FDs.

• 3.

  The Matthews correlation coefficient (MCC) summarises resemblance and considers imbalanced datasets for which one class of events, in this case, the FDs, is much smaller than the other class, that is, no FD periods (Matthews 1975; Delgado & Tibau 2019; Chicco et al. 2021).

Table 2

Definition of metrics from the confusion matrix

Metrics .	Equations .
True negative rate (TNR)
Negative predictive value (NPV)
True positive rate (TPR)
Positive predictive value (PPV)
Matthews correlation coefficient (MCC)

Metrics .	Equations .
True negative rate (TNR)
Negative predictive value (NPV)
True positive rate (TPR)
Positive predictive value (PPV)
Matthews correlation coefficient (MCC)

View Large

The metrics range from 0 to 1, with values close to 1 indicating a high degree of similarity and values close to 0 indicating little resemblance.

The climate series from cropland stations of the FLUXNET2015 dataset (Pastorello et al. 2020) was used for FD identification, as this dataset provided high-quality measured data for most of the required climate variables. Daily station data of the variables of precipitation, temperature, soil water content, latent and sensitive heat fluxes, wind speed and relative humidity, with durations ranging between 11 and 14 years, were available for four stations in Central Europe, where soil water data were the most limiting criteria for station selection. Table 3 contains information on soil type and location.

In the majority of the 11-year study period, all methods detected at least one FD per year, except in 2005 and 2008. The method M6 – Pendergrass et al. yielded the lowest number of events at all stations (4–7), whereas the other methods identified considerably more events varying between 8–10 (M2), 13–14 (M3), 5–15 (M4) and 5–15 (M5). M1 – Osman et al., with 11–18 events, identified the largest number of events.

Concurrent FD identification of five methods, marked with black frames (three methods, marked with blue frames), occurred 6 (2) times at Be-Lon, 10 (2) times at DE-Kli, 7 (1) times at DE-Geb and 6 (3) times at IT-Bci, with some small variations in the exact start time and duration. Station DE-Geb showed the overall best overlap for all methods, followed by Be-Lon. De-Geb and IT-Bci were characterised by longer periods, with no clear overlap between 2009 and 2012 for De-Geb. There was also no clear overlap between 2004, 2006 and 2010 for IT-Bci.

The two soil moisture-based methods, M1 and M2, showed the highest similarities overall. Every single event of M1 as the reference method was assessed for consistency with our FD definition. Other events that were not detected by the reference method were also evaluated to determine if they should have been delineated. It is noteworthy to mention that there were no concurrent FD identifications that did not include the reference method M1.

Concurrent identification among multiple methods would considerably increase if the just-missed events (yellow coding in Figure 3) were also considered, suggesting that some of the thresholds might have been set too rigidly. The red bars in Figure 3 mark obvious false identifications due to artefacts of the percentile approaches. This signifies some threshold crossing, but at the same time, the corresponding time series does not show any rapid decrease in absolute terms. Methods M3–M5 showed 24 false events altogether, with most of them in the early spring or autumn seasons. Instances of false identification may be reduced if additional rules regarding absolute changes in climate variables are introduced into the methods.

Methods M2–M5 presented intermediate values for the harmonic mean indicating similar behaviour regarding FD and no drought identification, in comparison to the reference method M1, with values ranging between 0.47 (M2: Ford & Labosier) and 0.39 (M5: Noguera et al.). The multi-criteria method M3 showed the highest sync₁ values overall, particularly for the station BE-Lon, directly followed by M2. For methods M2–M5, the sync₁ value at the station DE-Geb was significantly smaller than that of the other stations, indicating that the co-identification of FDs in comparison with the method M1 was the smallest at this station.

The method M6 (Pendergrass et al.) showed the largest differences from the reference method, with a value of only 0.2. However, M6 had large values for sync₀ (0.9), thus showing a great agreement in not identifying periods of no FD. Moreover, the small value of sync₁ (0.1) clearly shows that M6 rarely co-identified an FD simultaneously with M1.

All methods had high values for the TNR and the NPV, with M2 – Ford & Labosier and M6 – Pendergrass et al. having the highest TNR (above 0.9), and all NPV metrics varied between 0.7 (Pendergrass et al.) and 0.8 (Ford & Labosier). Hence, the methods did well in identifying intervals with no FD, relative to the reference method. However, there were significantly larger discrepancies for the TPR and PPV, M2 – Ford & Labosier had reasonably high values (0.75 and 0.43, respectively), directly followed by the multi-criteria method (0.55 and 0.51) and the method of Christian et al. (0.45 and 0.41), whereas Noguera et al. and Pendergrass et al. had very low values, particularly for the TPR metric, with values below 0.3. Thus, method identification deviated significantly in two ways: the identification of multiple events that were not identified by the reference method and the omission of multiple events (e.g. M3 – Noguera et al. and M6 – Pendergrass et al., both examples are also illustrated in Figure 3).

Misidentification and near-misses (as illustrated in Figure 6) might have been due to just one of the factors but were often due to a combination of them.

So far, we established that all six methods showed clear commonalities, but also disparities, in the detection and delineation of individual FD periods. The differences in the detected periods can be directly linked to the use of different critical variables and threshold values, and the different minimum durations and absolute or relative changes that are required for each method to detect an FD (Table 1).

We showed that the choice of the critical variable plays a major role in explaining the differences in identifications (Figure 6(a)). Different critical variables are already used in the analysis of ‘normal’ droughts, where the choice of variables such as rainfall, runoff and soil moisture deficits determines the type of drought form to be analysed. Normal droughts are commonly categorised into meteorological, hydrological and agricultural droughts (Mishra & Singh 2010). There are no comparable sub-definitions for FDs, although several studies have stated that some FDs are mainly influenced by heatwave dynamics and others state that they are more strongly influenced by rainfall deficits (Mo & Lettenmaier 2016; Pendergrass et al. 2020), but currently no clear definitions exist for such sub-types. FDs are known to be triggered or exacerbated by compound extreme climatic events (Otkin et al. 2018; Pendergrass et al. 2020; Lisonbee et al. 2021), but the relative magnitude and associated impact of each extreme have not yet been established (Mo & Lettenmaier 2016). At the same time, the methods differ in the extent to which they include compound dynamics in their detection procedure. Notably, the method M6 – Pendergrass et al. does not include any information on rainfall; therefore, it may be considered as an FD method that is solely driven by evaporative demand. Methods M3–M5 include a mixture of rainfall deficit, heatwave and ET information, whereas M1 and M2 use soil moisture as an integrative response variable for all vertical water exchange processes. We propose that by using multiple methods, one can identify different FD types, and future research efforts should concentrate on disentangling the extent to which different climatic components are responsible for the respective droughts.

The focus of this method comparison was placed on FDs in temperate croplands, where unusual rapid declining soil moisture during the growing season is considered the most direct indicator of severe plant water stress (Ford & Labosier 2017; Samaniego et al. 2018). However, methods using monitored, root-zone soil moisture data (M1 and M2) are severely limited by data restrictions, whereas indirect methods (M3–M6) using other proxy variables, such as rainfall or temperature, allow the use of much longer data series and more stations. Directly connected to the data limitation of the soil moisture-based methods are uncertainties in the derivation of the required percentiles. While they are impressive for soil moisture data, time series that are 11 or 14 years long, as the ones used here, are likely to result in skewed misrepresentation of the upper and lower percentiles required for the soil moisture-based methods. This fact highlights the dilemma of soil moisture-based approaches. While they should be preferred over other methods to correctly reproduce FD dynamics for croplands, the shortness of their series potentially leads to large uncertainties in their threshold values.

This study did not aim to determine whether some methods are better than others, but rather to verify whether the methods compare well, and thus evaluate whether any method could be used to assess FD dynamics for croplands in Central Europe. However, detailed synchronicity analysis does not confirm the latter. Similar to the comparison study of Osman et al. (2021) for the USA, we found different frequencies depending upon which definition was used. Nevertheless, three or more methods often detected the same FD (Figure 3). We propose that to balance the strengths and weaknesses of all methods, an ensemble approach for event identification be used as a way forward; thereby not just one, but several methods may be employed, which can be easily implemented with our R Shiny app. Multiple co-identifications would thus diminish the uncertainty of incorrect identifications and might give a more comprehensive picture when different types of FDs occur (Wang & Yuan 2018).

Several questions remain unanswered. Although a rapid depletion in soil moisture was detected multiple times during the investigated growing periods, the impact on vegetation health remains unclear and needs to be assessed with information other than climate data (plant mortality or remotely sensed imagery, as done by Peters et al. (2002) and Otkin et al. (2019)). The impact of different soil types and their water-holding and storage capacities, as well as the impact of different crop and vegetation species on FD identification, require future assessment. Finally, a thorough long-term analysis of FD dynamics using an ensemble approach, as suggested here, is pending. However, the preliminary analysis showed that FDs on croplands in Central Europe are no rare events but occurred on average once every 1–2 years during the investigated time period.

Defining and delineating FDs pose new challenges for hydrologists and climatologists. In this study, we compared six different delineation methods on a local scale (climatological station) and observed a large degree of synchronicity, but also some divergence, in the identified FD periods depending on which definition was used. The disparities of one method for detecting different drought intervals in comparison to others were narrowed down to four factors: (1) the opposing behaviour of proxy climate variables in comparison to prevailing soil moisture dynamics; (2) differences in the estimated velocity of drought intensification; (3) not exceeding pre-set threshold values for final depletion; and/or (4) differences in pre-set minimum drought lengths. Rather than seeing the detected divergence in identifying drought periods among the various methods as a weakness, we suggest using an ensemble approach for event identification to ensure that FDs of different sub-types are detected. These include different combinations of compound drivers, with differences in intensification dynamics, and were validated in this study for the temperate central European region.

Compound events, such as compound warm spells and rainfall deficit, have become more frequent in the past decades (Zscheischler et al. 2020; Vogel et al. 2021) and are expected to increase. However, it is not clear to what degree the different components that are relevant for FD development will develop. Thus, a single, commonly accepted definition for FD delineation may be the wrong goal, as it would take away the flexibility that an ensemble approach using multiple methods has.

This research was supported by CAPES (Finance Code 001) and DAAD (Award No. 91693642). We would like to thank the authors of the flash drought methods, which we used in this study, for their attention and support during our implementation in R. They include Dr Trent Ford, Dr Angeline Pendergrass, Dr Jordan Christian, Dr Mahmoud Osman, Dr Mike Hobbins and Ivan Noguera.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.