On the role of antecedent meteorological conditions on flash drought initialization in Europe

The fast depletion of soil moisture in the top soil layers characterizes flash drought events. Due to their rapid onset and intensification, flash droughts severely impact ecosystem productivity. Thus understanding their initialization mechanisms is essential for improving the skill of drought forecasting systems. Here, we examine the role of antecedent meteorological conditions that lead to flash droughts across Europe over the last 70 years (1950–2019) using ERA5 dataset. We find two major flash-drought types based on a sequence of development of antecedent hydro-meteorological conditions. The first type is characterized by a joint occurrence of two mechanisms, a decline of precipitation in conjunction with an increase of the evaporative demand, both occurring before the onset of a flash drought event. The second type, on the contrary, is characterized by high precipitation preceding the event’s start, followed by a sudden precipitation deficit combined with an increase in evaporative demand at the onset of the drought. Both drought types showed increased occurrence and higher spatial coverage over the last 70 years; the second drought type has increased at a much faster rate compared to the first one specifically, over Central Europe and the Mediterranean region. Overall our study highlights the differences between the two types of flash droughts, related to varying antecedent meteorological conditions, and their changes under recent climate warming.


Introduction
Flash droughts garnered significant attention in the last decade as they develop and intensify rapidly in contrast to traditional, slowly evolving droughts. During a flash agricultural drought event, the root zone soil moisture rapidly depletes, which can severely affect vegetation health, especially during critical plant growth stages (Otkin et al 2018, Koster et al 2019, Pendergrass et al 2020, Wang and Yuan 2022, Sungmin and Park 2023, Yuan et al 2023. Additionally, due to the pace at which it intensifies, drought adaptation activities will not occur on time, and they will further exacerbate the impacts. Flash droughts can lead to negative impacts across different sectors, such as agriculture (e.g. crop losses), ecological (e.g. forest fires), water management (e.g. less water availability), and loss of livestock and human health (e.g. Otkin et al 2016, He et al 2019. Flash drought occurrence has been reported from different regions worldwide, such as the USA, Europe (EU), Australia, India, and China, in the past (Otkin et al 2016, Nguyen et al 2019, Yuan et al 2019, Mahto and Mishra 2020, Shah et al 2022.
The overall losses due to drought are estimated at around $10 billion per event (www.drought.gov/ news/high-cost-drought) in the US and 9 billion per year in Europe and the UK (Naumann et al 2021), which makes it one of the costliest natural hazards. The aforementioned financial losses increase the relevance of skilful drought early warning and prediction from a risk management perspective. For drought prediction, it is essential to study and gain insight into what physical mechanism plays a role in the onset of drought in particular regions (Wood et al 2015). Droughts (of any type) usually start from a precipitation deficit (meteorological drought), and they can further lead to depletion of soil moisture (agriculture drought) and streamflow (hydrological drought) if they occur for a longer period (Van Loon 2015). Furthermore, when precipitation deficit is combined with high potential evapotranspiration (hereafter PET), then the propagation becomes faster (Trenberth et al 2014, Wang andYuan 2022). From the flash drought perspective, previous studies analyzed meteorological conditions (i.e. precipitation and temperature anomalies) during flash drought events which are necessary to maintain the drought event (Mahto and Mishra 2020, Shah et al 2022. Apart from the conditions occurring during the event, examinations of antecedent meteorological conditions leading to flash drought onset have received less attention (e.g. Labosier 2017, Osman et al 2022). This knowledge gap, combined with climate warming-which intensifies the flash drought occurrence (Wang et al 2016, Yuan et al 2019, Shah et al 2022increases the potential of future impacts. Therefore understanding the role of antecedent meteorological conditions leading to flash drought onset can help in the early detection of flash drought events and would eventually help minimising losses. Several large episodes of drought occurred in recent decades, e.g. 2003, 2010, 2015, and 2018-20 in Europe (Ionita et al 2017, Hari et al 2020, Moravec et al 2021. While the aforementioned role of meteorological conditions on initiating seasonal and multi-year droughts is relatively well studied (Fischer et al 2007, García-Herrera et al 2010, Hari et al 2020, Liu et al 2020, similar is not the case with short-term flash drought.
Thus, to fill these gaps here, we evaluate the role of antecedent meteorological conditions on the initialisation of flash drought across Europe. To that end, we first identify flash drought based on the rapid depletion of near-surface soil moisture. We then scrutinise the role of meteorological conditions such as precipitation deficit, high temperature or evaporative demand or both on initiating flash drought across Europe. Furthermore, we perform a datadriven exploratory analysis to identify the dominant flash drought initialisation mechanism that prevails across Europe. We also discuss their evolution over the last 70 years under contemporary warming conditions.

Data and methods
We used the top 30 cm soil moisture of the ERA5 global reanalysis product (Hersbach et al 2020, provided by the European Center for Medium-Range Weather Forecasts) to define flash drought conditions near surface root-zone. The original hourly data were aggregated to the pentad scale (i.e. 5 day, 73 pentads in a year) at the native resolution of 0.25 • for the period 1950-2019. To this end, we have used the ERA5 reanalysis data that combines the observation datasets provided by remote sensing with model data, to give the state of land and atmosphere seamlessly for the globe (Li et al 2020). The reason for selecting upper 30 cm is from the potential impact point of view: flash droughts that affect the agricultural areas with crops their dominant root biomass within the depth of 30 cm (crops: ≈ 70% of the roots in the top 30 cm; Temperate grassland: ≈ 83% of the roots in the top 30 cm) (Jackson et al 1996) prominently. Additionally, it was shown by Peichl et al (2021) across Germany that the top 25 cm of soil moisture is a better yield predictor than the total soil column.
Currently, there are no generally accepted conditions to define a flash drought. As a result, various definitions coexist to identify flash drought event (see, e.g. Pendergrass et al 2020). They are largely categorized into three groups. The first group defines a flash drought by identifying its short-term nature Lettenmaier 2015, 2016). The second group focus only on identifying how rapidly a drought event intensifies (Ford and Labosier 2017). And, the third one identifies a flash drought event by fulfilling two criteria: how rapid its intensification is and whether this intensity can be sustained for a fixed duration (Yuan et al 2019, Mahto and Mishra 2020, Pendergrass et al 2020. This third approach was also used in this study, according to earlier definitions given by Yuan et al (2019) and Shah et al (2022), as follows: • Convert the pentad soil moisture into soil moisture index (SMI) (quantile SMI(x) = F SM (x) = P(SM ⩽ x),) using an empirical non-parametric Gringorten plotting position (Gringorten 1963), to enable comparable analysis in space and time. This index ranges between 0 and 1; the lower values represent drier soil moisture conditions regarding the reference period, and the higher values represent wetter conditions in the soil. • A flash drought onset is defined when the SMI drops from above 0.4 (i.e. 40% of the SM pentad at a similar time of the year exceeds its 40th percentile) to below 0.2 within three pentads. • A drought event terminates when the SMI rises above 0.2 and stays there for at least two pentads.
• Once all the events have been identified, only flash droughts with a duration between 6 and 18 pentads are retained. A minimum of 6 pentads is necessary to remove all the events during which the SMI declines very quickly but then recovers. This kind of event is not likely to induce a severe impact on vegetation health. The upper boundary (a maximum of 18 pentads) is considered to focus only on events lasting up to one season. • Finally, select only those events occurring during the growing season (e.g. in Europe, from April to September) so that only those events that cause an impact on vegetation growth are selected.
Hydro-meteorological variables such as precipitation, temperature, potential evapotranspiration (PET) and actual evapotranspiration (ET) directly influence the state of soil moisture; they either act alone or together (Manning et al 2018). While precipitation deficit indicates a lack of water supply to soil moisture, ET represents moisture going away from the soil. Moreover, PET corresponds to ET, which would occur given an unlimited water supply. When soil moisture is ample, which is usually the case at the onset of flash drought, ET will respond to PET, and an increase in PET would lead to an increase in ET (Pendergrass et al 2020). Like soil moisture, we derived their pentadscale characteristics based on ERA5 from 1950-2019. Then we analysed both the antecedent period of a flash drought (up to three pentads) before and at the onset pentad by calculating total precipitation, temperature, PET, Precipitation-PET (hereafter P-PET) and ET anomalies. For regional analysis, we divided Europe into three regions: Northern Europe (NEU), Central Europe (CEU), and Mediterranean (MED) according to the IPCC climate reference regions (Iturbide et al 2020). Finally, to characterize different initialization mechanisms of flash droughts, we applied a simple and widely used unsupervised k-means clustering algorithm (Jain 2010). k-means clustering is a recursive partition-based clustering method. It starts by first randomly assigning k centers (centroids) representing k clusters, then sets each data point to the nearest cluster center. For every new data point assigned to a cluster, the algorithm recalculates the mean (centroids) of the clusters. This process is repeated till all the data points are assigned to a cluster. Different initialization allocates initial centroids at different locations; thus, final clustering will differ every time. The final selected clustering will be those for which the sum of squared or within-cluster variation over all the clusters will be minimum: where k is number of clusters, C k represent kth cluster, x n is set of variables in cluster C k and µ k is mean of clusters. For k-mean clustering, it is required to give number of clusters (k) in advance. One way to select optimal number clusters is Silhouette score method (Rousseeuw 1987): where a(o) is average distance between o and all the other data points in the cluster and b(o) is the minimum average distance o to all clusters to which o does not belong. Value of Silhouette score can vary from −1 to 1 with 1, 0 and −1 is clearly separated clusters, overlapping clusters and wrong assignment of clusters respectively. Therefore, we selected optimal number clusters for which the Silhouette score is maximum.

Role of antecedent meteorological conditions on flash droughts onset
Antecedent meteorological conditions responsible for flash drought initialization vary regionally (Ford and Labosier 2017). Figure 1 shows the antecedent meteorological conditions for soil moisture flash drought across Europe by quantifying the proportion of flash drought having precipitation deficit (i.e. ∆P < 0) and above normal evaporative demand (i.e. ∆PET > 0) for three pentads before onset (i.e. t − 3 to t − 1) and at onset pentad (i.e. t) pentad for all flash drought that occurred during 1950-2019. We observe a large variability with no clear signal for the wide-spread dominance of meteorological variables during t − 3 to t − 1 pentads (figures 1(a) and (b)). While nearly 50% of flash droughts have a precipitation deficit, 20-40% of above-average evaporative demand occurs just before the onset pentad (i.e. t − 1). Moreover, flash droughts which have above-average precipitation (i.e. ∆P > 0) also usually have below-average evaporative demand (i.e. ∆PET < 0) (see supporting information figure S1). Next we summarise the above results for three climatic regions across Europe (NEU, CEU, MED). Figure 1(c) shows that in all three regions, the proportion of flash droughts with precipitation deficit (i.e. ∆P < 0) and above average evaporative demand (i.e. ∆PET > 0) reduces consistently between t − 3 pentad to just before the onset of flash drought (i.e. t − 1) pentad. Moreover, at the onset of flash drought, precipitation anomalies are primarily negative, with nearly 50% of grid cells having 80% events with negative precipitation deficit. Additionally, a significant increase in the proportion of flash drought with ∆PET > 0 occurs between t − 1 and t pentads ( figure 1(c)). These findings are in line with previous studies, which also reported inconsistency in antecedent meteorological conditions signals leading to flash drought (Ford and Labosier 2017, Chen et al

Identification of different pathways of flash drought development
Considerable variability in all meteorological variables during all pentads before onset (figure 1) indicates that different pathways of flash drought initialization can exist. To objectively quantify unique mechanisms initiating flash droughts across Europe, we clustered SMI values from t − 3 to t pentad for each flash drought event (see Methods). The Silhouette score reached its maximum for k = 2; hence we selected it for k-means clustering. A similar Silhouette score and an optimal number of clusters when taking SMI for clustering from t − 2 to t and t − 1 to t imply that reducing the antecedent pentads does not change the optimal number of clusters or make them better separated (supporting information figure S2). shows two different pathways (clusters) for flash drought initialization based on the different behaviour of antecedent SMI values during four pentads (t − 3 to t) before the drought onset t. The SMI before the onset t of flash drought can prevail in any condition if the onset pentad (i.e. t) SMI is above 0.4. It is important to note that clustering here is performed only for exploration purposes rather than prediction. The first type represents wetter conditions (median SMI of 0.75) at three pentads before the onset, while the second type means drier conditions (median SMI of 0.25) at three pentads before the beginning. The SMI values of both types converge at the flash drought onset to the median SMI of 0.5. The first type is a more conventional way of drought initialization where SMI decline appears even before the onset pentad. In contrast, in the second type, the SMI values rise from nearly drought condition (0.25 percentile) in t − 3 to near normal SMI in t pentads. All four pentads of both types exhibit considerable variability in every individual pentad SMI. The Silhouette score of 0.4 at k = 2 (1 is the highest) indicates that the two identified clusters are not entirely separable. However, the clustering gives us a general idea about SMI movement before Figure 2. (a) Evolution of SMI for the two clusters during 3 pentads before onset (i.e. pentads t − 3 to t − 1) and onset pentads (i.e. t pentad). Boxplots represent anomalies of precipitation, and PET for three pentads before onset (i.e. pentads t − 3 to t − 1), onset (i.e. pentads t) and two pentads after onset (i.e. pentads t + 1 to t + 2) for all flash droughts. (b) Proportion (out of total) of each cluster for each grid cell. All data consider the period 1950-2019. the flash drought. Previous studies such as (Hunt et al 2009, Mozny et al 2012 identified the second type of flash droughts with in-situ soil moisture data. These studies show that when soil moisture is low, a high intensity of rainfall event can recover the soil moisture. However, when rainfall again breaks, soil moisture depletes very quickly, and this can onset a flash drought. This also indicates the effectiveness of the rainfall event. We also assess the robustness of antecedent wetness conditions on flash drought initialisation in Europe derived at deeper soils of ERA-5 and from independent hydrological datasets. The corresponding analysis of flash droughts using 1 m ERA-5 soil moisture is presented in supporting information figures S3(a) and (b)). Additionally, the same methodology for the soil moisture flash droughts is derived from a hydrological model ( (d)). supporting information figure S4 shows sensitivity of the method for different thresholds of drought definition on the silhouette width. Figure 2(a) (middle and right panel) further shows what meteorological conditions (precipitation and PET) lead to these two flash drought types using boxplots of each type. Note that other variables such as temperature, ET and P-PET are given in Supporting Information figure S5. Here we not only look at their antecedent conditions (i.e. pentads t − 3 to t − 1) and the onset conditions (pentad t) but also during intensification (i.e. pentads t + 1 to t + 2). Similar to SMI values, we observe two distinct pathways in meteorological conditions. Precipitation (PET) gradually decreases (increases) in every subsequent antecedent pentad (i.e. pentads t − 3 to t) during the first type of flash droughts. In contrast, the second type, where precipitation (PET) increases (decreases) in every subsequent pentad till just before the onset (i.e. pentad t − 1) of flash drought, followed by a sudden drop in precipitation and increase in evaporative demand at onset pentad. Consistent with previous studies Labosier 2017, Pendergrass et al 2020), both types of flash drought have a common behaviour of precipitation deficit with an increase in evaporative demand before the flash drought onset; they only differ in terms of the timescale it first appears. Furthermore, very comparable meteorological conditions during the intensification period of flash drought between the two types indicate that during second type flash droughts conditions change from wet/cold to dry/hot within a very short period. Overall, we notice that during the first type, meteorological signals such as a decline in precipitation, increasing temperature and evaporative demand prevail during the antecedent to flash drought. However, for the second type, we observe that because of precipitation deficit and high evaporative demand at onset pentad initializes the flash drought.
To analyze whether certain European regions favour a particular cluster type of flash drought, figure 2(b) shows the proportion of each type of event in a spatial manner. Although both flash drought types occur in almost all regions, the second type of flash drought is more common in the Mediterranean (MED), with 90% of flash droughts of this type in certain areas. On the other hand, the proportion (between 60% to 80%) of the first type is higher in NEU compared to the second type. A mixed signal is observed with a nearly equal balance of both types of event across CEU. It should be noted that the present study emphasizes on the antecedent meteorological conditions for the flash drought to occur.
We have also evaluated the possible role of large-scale atmospheric conditions prevailing flash drought events. To this end, we analyse the composite of changes in geopotential height at 500 hPa anomalies, corresponding to antecedent conditions of type 1 and type 2 events and segregated over three dominant climate regions (supporting information figure S6). Over all three regions, the type 1 eventsconventional flash drought events-are associated with atmospheric blocking conditions (i.e. above normal GPH anomalies) over the respective droughtaffected regions. The type 2 event does not have such a strong imprint of blocking (high) conditions over the drought-affected areas suggesting a role of other factors shaping drought events, such as land-surface atmospheric interactions. Further in-depth investigation of various physical mechanisms responsible for these conditions can be considered as a potential future study.

Spatio-temporal changes in occurrence of two flash drought types
Next, we evaluated spatio-temporal changes in the occurrence of both types of flash droughts in figures 3(a) and (b) between two periods (1950-1984 and 1985-2019). Both types of flash drought have seen an increased frequency in the recent period across Europe: 67%(73%) of grid cell frequency of the first type (second type) from 1985-2019 compared to . Furthermore, the increase in the second type of flash drought occurrence was more substantial for the first type during the period 1985-2019 when compared to . For instance, one-third of the domain had seen a more than two-fold increase in the second type of flash drought compared to only one-fifth of the domain for the first type.
We also analysed more recent trends with changes in decadal frequency and evolution of yearly spatial extent of both types of flash drought for three regions (figures 3(c) and (d)). For example, the decadal frequency and spatial extent in NEU increased until 2000 and remained stable afterwards. On the other hand, a significant increase in frequency and spatial extent occurred after the 1990s for both types of flash drought in CEU and MED. Moreover, in NEU, where the first type of flash drought was the dominant type during the entire study period, in CEU and MED in recent times second type's frequency and extent increased much faster than the first type. Regarding the Mediterranean region, there have been prevailing wetter and colder conditions until the 1970s, leading to the decline in the flash drought event Dell'Aquila 2012, Markonis et al 2021). After the 1970s, there has been an increased occurrence of drier and warmer conditions, leading to flash drought (Mariotti and Dell'Aquila 2012). Other large-scale atmospheric conditions and teleconnection patterns (e.g. NAO), deserve more detailed investigation, and we encourage future studies to do these investigations. Additionally, for example, in CEU, before the year 2000, only 30% of the time (i.e. 15 years out of 50 years), the areal drought extent of the second type was higher than the first type. Contrary, after the year 2000, already 60% of the time (i.e. 12 years out of 20 years), the second type of drought exhibited more extensive spatial coverage across CEU than the first one. In MED, we observe the equally balanced occurrence of the spatial extent of both drought types before the year 2000; however, after the year 2000, the second type dominates (80% of the time). In NEU, most of the time, the first type of droughts dominates over the second type.

Implications and discussions
The main implication of understanding and classifying the antecedent meteorological conditions of drought onset is its potential application to early warning systems. Since drought early warning systems rely on objective drought indicators (Funk and Shukla 2020), what we have classified as the first type can be directly used in flash drought prediction. Daily information for soil moisture state can now be easily assimilated daily from satellite-based remote sensing sources, such as Copernicus Global Land Service, available at a 1 km spatial resolution. Hence, the gradual depletion of soil moisture for two weeks over Central or Northern Europe could act as a red flag for an incoming flash drought emergency. This can be further analyzed by applying hydrological models to predict changes in other water cycle components, e.g. runoff and/or vegetation models, to assess the range of impacts for different scenarios of drought intensity.
An indirect, more alarming consequence of our results stems from the second flash drought dominating Southern Europe. The fact that the antecedent conditions do not appear to play a role in the onset and propagation of flash droughts highlights the low resistance of the hydroclimatic regime to abrupt shifts. This can be attributed to the water-limited conditions prevailing in most of the region during the growing season (Bouraoui et al 2010, Tague et al 2019. As the soil moisture levels are usually low at that time, our results show that they can be easily depleted in less than a week. Thus, the increase in SMI, a relative metric, before the drought onset could provide a misleading impression of resilience that does not hold. In the Mediterranean, flash droughts can occur with no warning.
Overall, we see that the level of drought predictability is related to the two flash drought types identified in our study. These are on good terms with an independent analysis and classification of the compound, warm-season droughts over Europe . The insights gained by the first flash drought type can be easily integrated in drought monitoring efforts such as the European Drought Observatory to improve our drought anticipation ability. The second drought type exposes our vulnerability in flash drought prediction. Still, since our analysis is limited to the meteorological perspective of flash drought onset, it is difficult to predict the actual impact on the vegetation and ecosystems in general. For example, the second drought type could result in lower vegetation stress and a milder ecosystem and socioeconomic disruption. Future research should investigate the relationship between the antecedent conditions of flash droughts and the vegetation response for each flash drought type. Since flash drought types appear to be increasing since 1950, improving our understanding and monitoring of these extreme phenomena is crucial.

Conclusions
Unprecedented drought-related economic losses in the past decades elevated the requirement for skilful drought monitoring and prediction systems (Hayes et al 2012, Thober et al 2015, Hao et al 2017. Flash droughts, often identified as the rapid depletion of soil moisture, can lead to significant impacts due to its main characteristic, namely, its short time from onset to maximum intensity. Consequently, water managers and stakeholders, need more time to implement strategies and solutions that mitigate the risk (Otkin et al 2018). Much of the efforts exists to understand the changes and associated physical mechanism of the long-duration traditional droughts over Europe (Vicente-Serrano et al 2014, Hanel et al 2018, Hari et al 2020. However, such efforts in understanding the flash droughts over Europe are lackingspecifically to understand the antecedent meteorological conditions during the flash drought. In this study we analyzed meteorological conditions associated with the onset of flash droughts defined using rapid decline in soil moisture across Europe. Our analysis showed that although the minimum conditions necessary for drought initialisation at the onset pentad are always significant precipitation deficits, high-temperatures, and high levels of evaporative demand (e.g. Otkin et al 2018, Pendergrass et al 2020, there is considerable variability of the antecedent state of these meteorological variables before the onset. We applied datadriven k-means clustering algorithm to identify the different mechanisms of flash drought initialization that dominate across Europe. We considered three pentads before onset and onset pentad SMI as a feature of each flash drought. Our analysis showed two dominant mechanisms: (1) The first type was where antecedent to flash drought onset soil moisture steadily declined because of a steady decline (increase) in precipitation (evaporative demand). (2) The second type was where soil moisture evolved from relatively dry to moderate conditions, followed by a rapid decline in soil moisture after the onset of flash drought. This flash drought type was usually initiated by changing meteorological conditions from wet/cold to dry/hot within a short period. Our results corroborate previous findings that not all flash droughts defined with soil moisture have prior warning (Osman et al 2022). The evolution of both types of flash drought in the last 70 years revealed that across Europe, both types of flash drought have been increasing in the recent period. Moreover, the occurrence of the second type was growing at a more rapid rate, especially in CEU and MED regions.
Our study highlights different mechanisms of flash drought development and provides essential insight, and can also help future studies when developing forecasting systems. Given the risk of an increased flash drought occurrence with contemporary climate warming (Yuan et al 2019, Shah et al 2022, the development of their skilful forecasting system should have immense importance. This study has focused on identifying flash drought based on a single variable (i.e. rapid decline in soil moisture). Future studies may explore whether multiple similar pathways exist when using different definitions of flash drought, such as using other or multiple variables to identify flash drought development (Osman et al 2021).

Data availability statement
The data that support the findings of this study are openly available at the following URL/DOI: https:// doi.org/10.24381/cds.adbb2d47.