The SoilClim water balance model (Hlavinka et al., 2011; Trnka et al., 2020; Řehoř et al., 2021) is used to calculate the dynamics of plant-available soil moisture in four soil layers (0.0–0.1, 0.1–0.4, 0.4–1.0, and 1.0–2.0 m) by comparing the inflow and outflow water balance components. SoilClim accounts for soil water-holding capacity and vegetation cover type, seasonal phenological development, or leaf area index dynamics, and the model simulates root growth and snow cover (Trnka et al., 2010; Řehoř et al., 2021). SoilClim was applied to each grid with a daily input of meteorological variables that consisted of precipitation, temperature at 2 m above the ground, dew-point temperature at 2 m, wind speed at 10 m, and incoming shortwave radiation, which originates from ERA5-Land (Muñoz-Sabater et al., 2021), as well as input consisting of the leaf area index (LAI) and land use and terrain inputs, also taken from the ERA5-Land dataset. The plant-available water capacity of the soil was computed as the difference between the field capacity and wilting point based on the inputs from the SoilGrids database (Hengl et al., 2014, 2017). SoilClim reproduces changes in long-term soil moisture dynamics in topsoil well (Trnka et al., 2015; Řehoř et al., 2024b). The calculated volumetric soil moisture was converted into relative available water (AWR), where 100 % represents the full field capacity and 0 % represents the wilting point (for more details, see Řehoř et al., 2024b). This study used data for the 2.0 m soil depth obtained by aggregating all four modeled layers. The prepared SoilClim dataset covers global nonglaciated land with a 0.5° resolution, excluding latitudes above 72° N and all of Antarctica.

The mesoscale Hydrologic Model (mHM; Samaniego et al., 2010; Kumar et al., 2013) simulates hydrological processes at the mesoscale. This model considers the complex interplay of land surface and subsurface properties through multiscale parameter regionalization using land cover, terrain, and soil characteristics, and it simulates major water fluxes, such as evaporation, infiltration, river runoff, or groundwater flow. This study considers soil moisture (SM) simulations averaged over the entire 2 m soil column depth (aggregating values over six soil layers) to quantify shallow-water availability. The daily meteorological inputs (i.e., precipitation and air temperature) originate from the ERA5 reanalysis (Hersbach et al., 2020). The daily minimum, maximum, and mean temperatures are also used to obtain potential evapotranspiration estimates (Hargreaves and Samani, 1985). Our simulations are based on the existing model setup using the digital elevation model from the USGS (Danielson and Gesch, 2011), soil characteristics from SoilGrids (Hengl et al., 2017), land cover from ESA (Arino et al., 2012), and LAI climatology from NASA (Tucker et al., 2005). The mHM is one of the seven global hydrological models that the WMO has evaluated and used for the development of the annual State of Global Water Resources reports since 2011, focusing on >500 major hydrological basins worldwide (WMO, 2022). The mHM dataset prepared for this study covers global nonglaciated land with a 0.5° resolution, excluding latitudes above 72° N and all of Antarctica.

The daily AWR data obtained from the SoilClim model and the daily SM data from the mHM were aggregated into 10 d intervals. Then, 10th-percentile thresholds for AWR and SM were calculated for each grid and each 10 d interval (by applying a 30 d window to smooth the annual variation) for the entire 1980–2022 period to represent drought conditions. Drought occurrence has been identified using the 10th-percentile drought, which is in line with the use of this threshold in US Drought Monitoring (Svoboda et al., 2002) since 1995 for the “D2” category (“Severe Drought”) definition and in the Czech Drought Monitor System (Trnka et al., 2020; Intersucho, 2024) since 2012 as the “S2” (“Moderate Drought”) category. To assess the most severe drought, the 2nd-percentile drought (i.e., 50-year return period) was calculated using the same approach.

Both 10th-percentile drought datasets were further clustered by the “ordering points to identify the clustering structure” (OPTICS) method (Ankerst et al., 1999), which was applied to the whole three-dimensional dataset (spatiotemporal) covering 43 years. OPTICS is suitable for delimiting clusters of varying density and shape, without requiring the specification of the number of clusters beforehand. To eliminate regional cases with very small drought-affected areas, clusters that included fewer than 50 grids for one 10 d interval, that included fewer than 500 grids overall, or that appeared in fewer than three 10 d intervals were excluded from both datasets, after which 775 clusters (further drought events) remained in the SoilClim dataset and 630 remained in the mHM dataset. In the case of both models, selected global land drought events (GLDEs) comprise around 15 % of the original clusters but include over 80 % of all grids included in the clustering.

To describe the severity of the identified drought events, four spatiotemporal characteristics were calculated as follows: (a)

the maximum areal extent of a drought event during its duration,

For each identified drought event and each 10 d interval of its duration, centroids were calculated using the median of the longitudes and latitudes for individual grids. Subsequently, three dynamic characteristics of these centroids for individual drought events were calculated as follows: (a)

the total sum of geographic distances between centroid positions for all individual 10 d intervals,

With respect to the different numbers of detected GLDEs from SoilClim and mHM, the comparison relies on the relative proportions of seven severity (1s–7s) and dynamic (1d–7d) categories across individual continents (Fig. 8a, b). Additionally, the comparison extends to the proportions of dynamic categories 1d–7d in severity categories 1s–7s (Fig. 8c). In some cases, differences in relative proportions are very small; in others, they appear to be larger visually. For this reason, we used the two-proportion Z test (Sprinthall, 2011) to test their statistical significance. Notably, statistically significant differences were observed only for category 4s in Africa and 1d in North America (both p<0.10) and for proportions of dynamic droughts in severity categories for categories 1s and 6d (p<0.05) and 4s and 4d (p<0.10). These results indicate that outputs from different models generally agree regarding the continental distribution of delimited categories and connections between both classifications within a certain variability range partially connected to smaller sample sizes of GLDEs within individual continents/categories.

The primary sources of uncertainties within the drought classifications presented here originate from the AWR/SM datasets, which are connected to either the models' algorithms or the input data. Because both models used meteorological inputs from reanalyses (ERA5-Land and ERA5), inherent biases must be considered, particularly in terms of their precipitation and temperature data. As Cucchi et al. (2020) discovered, ERA5 (including the ERA5-Land dataset) generally indicates an underestimation of temperature and an overestimation of precipitation at high elevations above sea level. This underestimation could impact the delimitation of drought occurrence in these regions. Nevertheless, we argue that when using percentile transformation, drought characteristics are more robust against systematic biases in the absolute values of variables, such as temperature and precipitation, and the methodology enables us to capture the relative severity of drought characteristics.

Despite a possible lack of precision within the input data, some uncertainties are also intrinsic to the employed models. In the SoilClim model, there are uncertainties related to the schematization of individual processes, the soil profile vertical discretization, and other scale-related assumptions (e.g., each grid cell is represented by dominant land cover), described in detail in Hlavinka et al. (2011), Trnka et al. (2020), and Řehoř et al. (2021). These simplifications are inherent in any large-scale modeling scheme, and the uncertainty they create must be acknowledged. However, as we employed the AWR data in the form of percentile-based soil moisture anomalies, we were able to eliminate the effect of uncertainties that affect the long-term mean AWR within individual grids.

Furthermore, in both models, uncertainty originates from the underlying data, such as soil properties and land cover/type characteristics, which must be approximated due to limited observational data. We also acknowledge that the fixed/prescribed model functions representing hydrological processes might not be equally plausible across different parts of the world, and our approach also does not aim to study the impacts of sustained drought in one location. Long-term changes in land cover and irrigation can also cause larger uncertainties on a regional scale.

Some uncertainties are associated with the clustering technique used, i.e., density-based clustering OPTICS. As described by Ankerst et al. (1999), objects are connected to form clusters if there are a sufficient number of other objects nearby within a defined multidimensional space (in our case, three-dimensional). To illustrate the spatiotemporal behavior of the clusters, we included animation of their development during 2019 in Animation S1 in the Supplement. However, specific settings of the algorithm parameters must be partially derived empirically to fit the characteristics of a given dataset (particularly concerning its density) and prevent either the connection of all objects (in our case, grids in individual 10 d intervals) into one large cluster, including the whole dataset, or the failure of the algorithm to create clusters at all if the clustering parameters are too strict. Aside from these extreme cases, when the clustering failed, smaller changes in the parameters led only to minor changes in the event delimitation. Still, there is no objective method for defining “perfect” parameters; hence, clustering uncertainty is inherent and affects the length of the existence of individual clusters in our dataset. A good example of this is the massive drought that occurred in Eurasia, classified as an extremely long GLDE from 2004 to 2022 (and still ongoing) from SoilClim, whereas using mHM data, it was separated into two GLDEs that were disconnected in 2013/2014 (cf. Table 1). Potential alternatives for drought clustering have been proposed for example by Andreadis and Lettenmaier (2006), Vidal et al. (2010), or Samaniego et al. (2013), including stepwise selection of continuous drought areas, k-means cluster analysis, or density-based spatial clustering of applications with noise (DBSCAN), which is an alternative to OPTICS and uses similar approach.

Moreover, in cases of quickly varying densities of drought-affected grids and rapid spatial changes, some smaller, volatile clusters occasionally formed. These clusters may show questionably large values of their centroid's movement, which may be viewed as an artifact of the methods employed. Such an example could be the drought event from June 2014–March 2015 in North America (Table 2b), which appeared around the center of the continent between other, more consolidated, clusters. All centroid positions of this GLDE are shown in Fig. S2. A few problematic cases such as this represent some uncertainty in the classification integrity. However, the selection of multiple characteristics for both severity and dynamic classifications can partially mitigate the impact of such anomalous GLDEs. Another issue might be the arbitrary selection of percentile thresholds that were used to delimit seven categories in both classifications, with the particular aim to divide the extremes into smaller categories. However, slight changes in the thresholds would not alter the overall picture presented in this paper because the transition between categories is gradual, as shown in Figs. 1 and 4.

Numerous studies have explored different means of classifying global land droughts, employing a range of data, parameters, and methodologies. These studies typically present their findings through related global-scale maps. For example, Carrão et al. (2016) mapped global drought risk based on drought hazard, exposure, and vulnerability during 2000–2014. Spinoni et al. (2019) used the SPI and SPEI (at a scale ranging from 3 to 72 months) from 1951–2016 for 23 macroregions of the world to obtain approximately 4800 (SPEI-3) and 4500 (SPI-3) drought events. These events were divided into moderate, severe, and exceptional events based on the drought severity, intensity, area, top event, peak intensity, and area scores. He et al. (2020) developed a drought and flood catalog (GDFC) that spanned 1950 to 2016. To define drought, they used both the SPI and soil moisture percentiles based on the VIC land surface model output. In this dataset, they employed simple clustering that connected neighboring grids to define individual drought episodes. They subsequently studied the relationship between drought area and the intensity and return periods of severe droughts. Monjo et al. (2020) climatologically classified the duration of drought worldwide based on the dry–wet spell n index derived from the global gridded daily Multi-Source Weighted-Ensemble Precipitation (MSWEP) dataset for 1979–2016. By comparing the different relationships between the occurrence of dry and wet spells, they identified seven types and presented their worldwide distributions. Fuentes et al. (2022) mapped spatiotemporal drought propagation through different subsystems at the global scale over recent decades using different standardized drought indices. Drought propagation was established as the lag in the peak correlations between drought time series in different subsystems (for drought propagation, see, e.g., Li et al., 2023, for the Yellow River basin).

The classification approach for GLDEs presented in our study (Sect. 3.2 and 3.3) is not focused only on mapping any static drought state at the global scale, as in some of the above-reported studies, but rather, we present new views on the spatiotemporal variability and dynamics of drought events, combining several characteristics that describe their extent, duration, severity, and propagation. This research suggests a complex approach for monitoring droughts as individual dynamic events and offers many opportunities for future analyses concerning drought drivers and the propagation of droughts, including their possible self-propagation, as suggested by Schumacher et al. (2022). This approach could also help to link drought events to their impacts more directly (Meadow et al., 2013; Lackstrom et al., 2017), as it spatiotemporally delimits drought-affected areas in great detail. Compared with similar preceding studies, this study is based not on the use of standard drought indices but rather on selected soil moisture variables from two different physically based models (SoilClim and mHM), which express the compound effects of temperature, precipitation, evapotranspiration, and soil characteristics using land surface and hydrological modeling, respectively. The presented method is robust enough for the delimitation of GLDEs and is very flexible in the selection of basic drought parameters and their thresholds and the use of various drought datasets; i.e., the method is applicable to different datasets and is repeatable.

The analysis results presented in Tables 1–2 show that the most important GLDEs during 1980–2022 occurred on all continents and appeared mainly after 2000. Although Sheffield et al. (2012) reported only a slight change in global drought (particularly based on the PDSI) during 1950–2008 and He et al. (2020) did not discover a worldwide increase in drought during the last two decades, many important drought events were reported worldwide after 2000 (e.g., Shmakin et al., 2013; Van Dijk et al., 2013; Griffin and Anchukaitis, 2014; Erfanian et al., 2017; Ionita et al., 2017; Marengo et al., 2017; Spinoni et al., 2017, 2019; Deng et al., 2020; Chiang et al., 2021; Moravec et al., 2021; Rakovec et al., 2022; Liu et al., 2023; Arias et al., 2024; Garrido-Perez et al., 2024) and were well reflected in the selected GLDEs in our paper (Tables 1–2 and S1–S2). Moreover, Fuentes et al. (2022) reported the intensification of drought characteristics in recent decades in several regions of the world (particularly in southern South America, central Australia, southwestern Africa, and central and eastern Asia). To demonstrate this situation, Table 3 shows a comparison of the absolute and relative numbers of GLDEs during 1980–2000 and 2001–2022. As these periods are from a climatological point of view relatively short, the comparison should be taken with caution. Except for category 1d from mHM in the dynamic classification, in all categories for both classifications, the frequency of occurrence of the detected GLDEs was greater after 2000 than during the preceding period. According to the two-proportion Z test (Sprinthall, 2011), differences in the relative frequencies of given severity drought categories between the two periods were statistically significant only for category 2s from SoilClim (p<0.05), whereas for mHM they were significant for 4s (p<0.05) and 2s and 5s (p<0.10). For the dynamic drought categories, the related differences were statistically significant for category 5d from SoilClim (p<0.05) and for categories 7d (p<0.10) and 4d and 6d (p<0.05) from mHM.

Concerning the continental distribution of severe GLDEs, the lower relative proportions of categories 7s and 7d in Eurasia in both SoilClim and mHM are remarkable. This result could be influenced by the existence of one (SoilClim) or two (mHM) extremely large GLDEs during the last two decades in Eurasia (Tables 1–2), propagating particularly around Siberia and the Central Asian Plains, where the potential for large dynamic events is the highest due to the absence of prominent topographic features. On the other hand, most of the other droughts that occurred during this period and were identified in other parts of Eurasia (southeast Asia or Europe), which has a greater amount of fractured topography, were smaller and shorter. Both the smallest absolute frequency and relative proportion of 7s droughts occurred in Australia. This finding is due to the fact that it is by far the smallest of the analyzed continents; however, in terms of the 7d category, the relative proportion was much greater. With respect to the frequencies of 7s and 7d GLDEs, Eurasia and North America had the greatest absolute numbers of GLDEs but South America had their highest proportion.

In the context of recent global climate change, increases in the frequency and intensity of drought events are among the most impactful changes worldwide. Our effort to catalog individual GLDEs during the last four decades at a high spatiotemporal resolution could be effective and useful for benchmarking newly evolving droughts and contextualizing their potential impacts, while the developed methods can also be applied to new, emerging datasets with even greater accuracy. Moreover, as we consider droughts to be fully 3D (area and time) events, our approach is designed to provide tools for analyses with the aim of connecting drought occurrence with large-scale atmospheric circulation patterns and global climate variability modes; on the other hand, our approach can be employed to study processes leading to drought propagation (and possible self-propagation) at regional to continental scales, as our spatial delimitations are not restrained by predefined regional borders.