SDHIs were computed using output from the global water availability and water use model WaterGAP 2.2d (Müller Schmied et al., 2021). WaterGAP 2.2d has a spatial resolution of 0.5∘ latitude by 0.5∘ longitude (55 km × 55 km at the Equator) and covers the whole global land area except Antarctica. WaterGAP consists of the WaterGAP Global Hydrology Model (WGHM) and five water use models for the sectors households, manufacturing, and cooling of thermal power plants (Flörke et al., 2013), as well as irrigation and livestock. WGHM computes daily time series of fast surface and subsurface runoff, groundwater recharge, and streamflow, as well as water storage variations in the canopy, snow, soil, groundwater, lakes, reservoirs, wetlands, and rivers. Model input includes time series of climate data between 1901 and 2016 and physio-geographic information, such as land cover, soil type, relief, and hydrogeology. For this study, WaterGAP 2.2d was forced by the WFDEI-GPCC climate data set (Weedon et al., 2014), which was developed by applying the forcing data methodology from the EU project WATCH on ERA-Interim reanalysis data. Daily model outputs of streamflow and surface water abstractions (WUs) were aggregated to monthly time series. WaterGAP total runoff is calibrated against long-term mean annual streamflow at 1319 gauging stations worldwide covering approximately 54 % of the Earth's land area (except Greenland and Antarctica). A detailed model description and evaluation can be found in Müller Schmied et al. (2021). Please note that while WaterGAP simulates the impact of reservoirs on streamflow, the accuracy is very low as it is unknown how all human-made reservoirs on Earth are managed such that a generic algorithm is used to simulate human reservoir management decisions.

In several model intercomparison and assessment studies, WaterGAP proved suitable for computing streamflow and SDHIs, although the discrepancies between simulated and observed low flows, seasonality, and interannual variability can be significant at the regional scale (see the literature review in the Supplement Sect. S1). A limited model validation of the WaterGAP version 2.2d applied in this study (Sect. S2) revealed that Q80 is overestimated by WaterGAP in 63 % of all months and stations with median percent deviations between 35 % in February and -7 % in July (Fig. S1). In another model validation exercise, SSI3 as modeled by WaterGAP was compared to observed SSI3 at 183 gauging stations (Sect. S2). With a median NSE of 0.5 and an interquartile range of 0.2–0.7, WaterGAP 2.2d model output showed moderate agreement with the observations. NSE exceeded 0.7 at 25 out of the 183 stations mainly located in central and eastern Europe and the United States.

Following the approach of Cammalleri et al. (2016a) to compute the low-flow index (LFI), the probability of drought events of a certain severity was computed for six cumulative indicators: CEP1(20 %), four CQDI1 variants (thresholds Q50, Q80, WUs, and WUs + EFR), and CRQDI1(-50 %). First, the partial duration series of drought events was derived based on the severities of all drought events of the reference period. Grid cells with fewer than six drought events were excluded. The exponential cumulative distribution function proposed in Cammalleri et al. (2016a) was used to estimate the probability of non-exceedance p of a certain cumulative streamflow deficit: 7pSi;λ=1-e-λSi(withSi>0), where the variable Si is the severity of drought event i, as quantified by a cumulative indicator, and the parameter λ is the inverse of the mean of the severities of all completed drought events. For instance, a value of p=0.7 in a certain month denotes that, if the drought event ended in this month, its severity would be larger than the severity of 70 % of the drought events in the reference period. Different from LFI, which is based on daily streamflow data, time series of monthly streamflow were used for all indicators and the 2mc (see Sect. 2.3.2) was applied. Since p was computed for each month of the reference period, it describes the non-exceedance probability (or rather frequency) of both completed drought events and continuing droughts. Following Sharma and Panu (2015) and Beguería (2005), the return period Tri of a drought event with severity Si is computed as 8Tri=1θ1-pSi, where θ is the average number of drought events per year during the reference period.

The selection of drought hazard indicators for a DEWS requires a clear definition of “the risk of what for whom”. Drought hazard indicators are risk-system-specific (Blauhut et al., 2022), and there is not one that fits all. Drought is usually conceptualized as an anomaly (“less water than normal”) and/or deficit (“less water than needed”). Consequently, the selection of an indicator requires a definition, often based on assumptions, about “what is normal or needed”, i.e., what the risk system is habituated to. In the case of streamflow, people and ecosystems are assumed to have adapted to certain characteristics of the flow regime. For example, if drought indicators are computed based on the calendar-month-specific distribution of streamflow values, it is implicitly assumed that the risk system has adapted to the seasonality of streamflow. But temporally constant thresholds, which have traditionally been used to define hydrological droughts (Stahl et al., 2020), are also suitable for certain systems, e.g., for computing drought hazard for electricity generation by thermal power plants, which require a certain minimum streamflow for operation.

At the global scale, it is unknown to which streamflow characteristics different risk systems such as drinking water supply, irrigation water supply, hydropower production, and the river ecosystem are accustomed. Therefore, the 11 global-scale drought hazard indicators analyzed in this study (Table 1) cover different types of habituation, including the habituation to a certain degree of interannual variability of streamflow, to streamflow seasonality, to a certain reduction from mean calendar month or mean annual streamflow, and to being able to fulfill the demand for surface water abstractions and environmental flow. It is up to the user of a large-scale DEWS, who understands the local risk-system-specific habituation to reduced water availability, to select the hazard indicator that is appropriate for the risk system of interest.

Percentile-based indicators including empirical streamflow percentiles, standardized indicators, and TLM indicators with a low streamflow percentile as a threshold are often applied in DEWSs (Bachmair et al., 2016; Cammalleri et al., 2016a). They are perceived as statistically consistent across different temporal and spatial scales, indicating the rarity of the event (Steinemann et al., 2015; WMO and GWP, 2016). Utilization of percentile-based indicators (e.g., SSI12, SSI1, and CQDI1(Q80) in Table 1) implies that people in different climate regions and social systems are equally habituated to a certain interannual variability, which is most likely not the case. The 20th streamflow percentile (or SSI1 = -0.84) would correspond to a low relative streamflow deviation (e.g., -20 %) in a humid region (low interannual variability) compared to a higher deviation (e.g., -50 %) in a semi-arid region (high interannual variability). Hence, percentile-based indicators might underestimate streamflow drought hazard in semi-arid areas where people (and ecosystems, although possibly to a lower degree) are often more vulnerable to reductions in water availability. Regions with high interannual variability are depicted in Fig. A1b. Here, drought hazard indicators that quantify relative deviations from the long-term mean or median (RQDI1, RQDI12 in Table 1) or TLM indicators with higher percentiles as a threshold (CQDI1(Q50) in Table 1) might be better suited to define drought conditions. Such indicators appear to be less preferred as periods with the same indicator value have different probabilities of occurrence in different regions and thus not the same rarity (Steinemann et al., 2015). Contrastingly, river ecosystems are, in the ideal case, perfectly adjusted to interannual variability of streamflow such that percentile-based drought hazard indicators are often suitable for drought hazard assessment for river ecosystems. In conclusion, percentile-based hazard indicators and relative deviations from the long-term mean or median should be used complementarily in large-scale DEWSs to cover different types of habituation.

The selected averaging period defines whether people are habituated to the annual or seasonal flow regime. One can assume that river ecosystems are generally accustomed to seasonality. Therefore, indicators with a short averaging period of, for example, 1 month (EP1, SSI1, RQDI1 and CQDI1 variants in Table 1) are appropriate for quantifying drought hazard for river ecosystems. Furthermore, short averaging periods are suitable in regions where farmers and other water users do not have access to large water storage such as reservoirs, lakes, or groundwater (either due to missing infrastructure or due to water use restrictions). As these users abstract water directly from the stream, they are very vulnerable to seasonal (monthly) streamflow deficits. Indicators with longer averaging periods (SSI12, RQDI12), on the other hand, are suitable in regions with large human-made reservoirs, which are usually replenished during the wet season such that streamflow deficits during the low-flow months are irrelevant. People in these regions are therefore only vulnerable to either interannual variability (SSI12) or mean annual conditions (RQDI12), but not to seasonality. Certainly, other averaging periods may be suitable depending on the region-specific storage capacity. Since volume-based indicators (TLM indicators) are also important components in water resources management (van Loon, 2015), the indicators CQDI1(Q80-HS) and CQDI6(Q80) are assessed as alternatives for SSI12 and RQDI12 (or rather the cumulated variants CSSI12 and CRQDI12) in regions with highly seasonal streamflow regimes (Fig. A1a) and large reservoirs.

For water managers, the status of the actual water deficit in terms of unsatisfied water demand might be as informative as the status of streamflow anomaly. Drought hazard is generally defined as a climate-induced anomaly, i.e., a period of below-normal water availability (McKee et al., 1993; van Lanen, 2006; van Loon, 2015). This concept can be broadened by assuming that a drought only occurs if the anomaly coincides with a water deficit for people or ecosystems (Cammalleri et al., 2016b; Popat and Döll, 2021; Wilhite and Glantz, 1985). Nevertheless, only a few studies exist wherein the combination of anomaly and deficit is translated into drought hazard indicators for soil moisture (Palmer, 1965; Cammalleri et al., 2016b; Popat and Döll, 2021) and streamflow (Popat and Döll, 2021). In the present study, the water deficit aspect of drought is represented by the indicators CQDI1(WUs) and CQDI1(WUs-EFR) (Table 1). Application of these indicators implies that the system at risk is habituated to the satisfaction of seasonal water demand. While CQDI1(WUs) neglects the water requirements of the ecosystem, CQDI1(WUs-EFR) assumes that the river ecosystem is habituated to the seasonality and magnitude of natural streamflow. As EFR might never be fulfilled in the case of strongly altered streamflow regimes, Qnat in the EFR computation can be replaced by Qant, implying that the river ecosystem has already adapted to the altered streamflow conditions (Table 1). Figure A1c shows regions where human water demand is high compared to available streamflow and where a drought hazard due to unsatisfied human surface water demand is likely.

Translating conceptual drought definitions into operational, quantitative drought hazard indicators is not straightforward due to the complexity of the underlying natural processes and the large number of indicators and methods that can be applied. In the literature, there is agreement about which drought characteristics are relevant for operational applications comprising the temporal component (onset, termination, duration) and the spatial extent as well as drought magnitude and severity, from which other metrics such as intensity, return period, and frequency or probability of occurrence can be derived (van Lanen et al., 2017). We understand drought magnitude as an anomaly or deficit occurring within a predefined period and severity as the accumulated deficit between the magnitude and a selected threshold since the onset of drought, which is defined by water availability dropping below the threshold (van Lanen et al., 2017). However, the terms drought magnitude and severity, which represent different levels of drought characterization, are not applied consistently in the literature. The terms are not made explicit and are sometimes interchanged (Steinemann et al., 2015; Vidal et al., 2010; López-Moreno et al., 2009). In particular, the commonly accepted classification of SDHIs into threshold-based and standardized indicators (van Loon, 2015) is somewhat misleading, since the former represents time series of severity and the latter time series of magnitude.

To facilitate a better understanding of the informative value of SDHIs, we suggest a new indicator classification that includes four types of indicators and distinguishes severity from magnitude indicators (Fig. 1). The indicator types (columns in Fig. 1) include the volume-based anomaly, the standardized or percentile-based anomaly, and the relative deviation (Sect. 2.3). Deficit anomaly indicators (last column in Fig. 1) combine an anomaly indicator with an indicator of the deficit with respect to optimal water availability (e.g., Popat and Döll, 2021). For each indicator type, two levels of drought characterization, drought magnitude (level 1) and severity (level 2), can be computed.

The dark grey boxes in Fig. 1 represent decisions regarding time step length and averaging period, as well as drought threshold and definition of drought events (minimum length of drought event, pooling of drought events). These decisions depend on the assumed habituation of people and ecosystems to certain streamflow conditions (Sect. 3.1 and Table 1). Beige and orange boxes contain indicators that are expressed in absolute or relative values and in frequency or probability of occurrence, respectively. Indicators applied in drought monitoring (CQDI1, low-flow index – LFI, percentiles, SSI, RDPI) or in the literature (pQ, cumulative SSI, streamflow deficit anomaly indicator QDAI) are written in bold. The units of the four indicator types differ at both level 1 and 2, but indicators can be directly compared when expressed in units of probability (or frequency) of non-exceedance.

Figure 1 shows that drought hazard indicators pertaining to one of the four indicator types can be transformed between level 1 (magnitude) and level 2 (severity) while still sharing the type-specific conceptual drought definition. Furthermore, the classification system clarifies that each indicator type requires a threshold setting either at level 1 or 2. Hence, the term “threshold-based” applies to any indicator of drought severity, and it is therefore not a suitable criterion for distinguishing types of indicators.

The differentiation of indicator types can be ambiguous. For instance, standardized and percentile-based anomaly indicators are subsumed in Fig. 1 (column 2), although there is a minor conceptual difference between them as highlighted by Tijdeman et al. (2020). While standardized indicators show the non-exceedance probability enabling extrapolation, empirical percentiles represent the historical non-exceedance frequency within the boundaries of observations. We account for this aspect by including the terms frequency and probability in Fig. 1.

On the other hand, volume-based and standardized or percentile-based anomaly indicators are presented as different indicator types, although they can be based on the same conceptual drought definition if equivalent thresholds are applied. If Q80 is used as a threshold for CQDI1 and -0.84 for cumulative SSI1 (corresponding to the 20th percentile for cumulative EP1 and a return period of 5 years), both indicators capture the same drought signal. Differences between the drought signals are then attributable to the computational methods for the standardization of streamflow. Analyzing the sensitivity of SSI1 to different parametric and nonparametric standardization methods in European river basins, Tijdeman et al. (2020) revealed considerable differences in computed SSI1 among seven probability distributions (and two fitting methods) and five nonparametric methods. A major difference between volume-based and standardized indicators is that the former detect absolute drought deficits and the latter relative drought deficits. This can result in different frequency values for the same drought event.

EP1 patterns (Fig. 2c for July 2003 and Fig. S3c for September 1993) are very similar to SSI1 (Fig. 2a and S3a) since both indicators are based on the same conceptual drought definition (Sect. 3.1). Both indicators generally identify the same drought regions. However, drought classes differ in many regions of the world, with EP1 indicating both higher and lower drought magnitude. For instance, in eastern France, EP1 indicates a higher drought magnitude class in July 2003 (return period RP > 20 years) than SSI1 (RP > 10 years) and vice versa in southern Germany. In September 1993, SSI1 indicates a higher drought hazard than EP1 for the Orange River along the Namibia–South Africa border, but a lower hazard in a few grid cells in central South Africa and Lesotho. These differences can be attributed to the fitting of the gamma distribution in the case of SSI1 and the assignment of the maximum rank among tied values within a streamflow sample in the case of EP1 (Sect. 2.3.3).

Comparing SSI1 with empirical percentiles, Tijdeman et al. (2020) identified several advantages and limitations for both indicators. SSI1 has the disadvantage that for different streamflow regimes, different parametric probability distributions would be required to achieve the best fit, which reduces consistency at the global scale. In this study, the gamma distribution showed the best fit among 23 parametric probability distributions for most grid cells and was applied in each month and grid cell. Of course, using only one distribution for the whole globe results in poorly fitting distributions for some cells and months (Tijdeman et al., 2020). Grid cells where gamma fitting was rejected in the calendar months July and September based on the KS test (Sect. 2.3.1) are shown in grey in Figs. 2a and S3a (18 % of all grid cells excluding Greenland). EP1 does not require fitting of a distribution and can therefore be computed in more grid cells than SSI1. Only if a sample includes more zero flows than the selected threshold is drought identification not possible (blue grid cells in Figs. 2c). On the other hand, if Q80 is zero and the current streamflow exceeds zero, it is possible to define the current month as not a drought month (shown in beige in Fig. 2a and c). EP1 has the disadvantage that it only allows the quantification of the historical non-exceedance frequency within the reference period, while probabilistic information, for example on extreme events such as a 100-year drought, cannot be derived (Tijdeman et al., 2020). Nonetheless, EP1 seems to be more suitable for a global-scale DEWS, as the indicator does not entail the possibly large uncertainties due to the fitting of a probability distribution and can be computed in more grid cells than SSI1.

With percentile-based indicators (e.g., EP1, SSI1), risk systems in different regions are assumed to be equally habituated to a certain interannual streamflow variability, which is most likely not the case as interannual variability varies strongly (Fig. A1b). Comparing two regions with high and low interannual variability, the same streamflow percentile or z score corresponds to a much higher relative deviation from mean calendar month streamflow (RQDI1) if interannual variability is high. For instance, at the Orange River and White River with high interannual variability (Fig. S2), SSI1 values below -0.84 (RP > 5 years) always correspond to RQDI1 values below -70 % and -60 %, respectively. At the Danube River and Angara River (Fig. S2) (low interannual variability), RQDI1 of -50 % is (almost) never reached, while maximum SSI1 values are higher than at the Orange River and White River. Hence, SSI1 might underestimate drought magnitude if interannual variability is high, especially for vulnerable systems.

At the global scale, RQDI1 (Figs. 2b and S3b) identifies most of the drought hotspots as indicated by EP1 (Figs. 2c and S3c), although the relative levels of magnitude differ. These differences correspond well to the interannual streamflow variability depicted in Fig. A1b. Drought hotspots according to EP1 in regions with low interannual variability (parts of North America, northern Europe, northern Russia) only show moderate relative streamflow deviations by global comparison. This is because RQDI1 values of -50 % or lower are never reached in these regions, as illustrated at the Danube and Angara stations (see above). Here, RQDI1 might underestimate drought magnitude. On the other hand, in regions with high interannual variability (e.g., large parts of Africa, central Asia, western US), both drought magnitude and the affected area are larger according to RQDI1. Here, RQDI1 can draw attention to potential drought impacts in regions with higher suspected vulnerability (e.g., southern Africa) that would otherwise be overlooked using EP1 or SSI1. In regions where people are probably well accustomed to the interannual variability of streamflow (e.g., western US), RQDI1 is less suited than EP1 to indicate drought magnitude. At the severity level, regions with low interannual variability are excluded using CRQDI1(-50 %)_f (Figs. 4d and S5d) due to the low threshold of -50 % (grid cells in light grey).

The water deficit indicators CQDI1(WUs) and CQDI1(WUs-EFR) (Figs. 3c, d and S4c, d) define drought as “less water than needed” as opposed to the anomaly indicator CQDI1(Q80) (Figs. 3a and S4a) indicating “less water than normal” (or rather less water in a certain month than in 80% of the years). Consequently, the spatial pattern of the former is very different from CQDI1(Q80) patterns. For instance, the drought event in 2003 in central and eastern Europe (Fig. 3) identified by CQDI1(Q80) is not indicated by CQDI1(WUs), while the latter shows an additional drought hazard in the northern part of South Africa (Fig. S4). This is because CQDI1(WUs) strongly depends on surface water stress, which is generally low in Europe and high in South Africa (Fig. A1c). The spatial patterns of CQDI1(WUs) correlate well with Fig. A1c, comparing human water demand for surface water as a fraction of mean streamflow. CQDI1(WUs-EFR) additionally considers the environmental flow requirement (EFR) computed as 80% of naturalized mean calendar month streamflow. Like RQDI1, the indicator thus depends on mean monthly streamflow, and the spatial pattern corresponds well to the map of interannual variability (Fig. A1b). A comparison between CQDI1(WUs) and CQDI1(WUs-EFR) shows that only in a few regions is human water demand the dominant component determining the water deficit. In most regions, EFR leads to high cumulative deficits even if seasonal human water demand is small (< 10 % of available streamflow, Fig. A1c). CQDI1(WUs-EFR) is the only indicator in this study that explicitly takes into account the health of the river ecosystem, an aspect that should be included in a global-scale DEWS. Alternatively, the cumulative anomaly deficit indicator (QDAI) (Popat and Döll, 2021), considering EFR based on a similar approach, can inform decision-makers and water users about the drought hazard for water supply. In strongly altered flow regimes, wherein simulated anthropogenic monthly streamflow (Qant) is always below 80 % of mean monthly naturalized streamflow (Qnat), time series of CQDI1(WUs-EFR) are continuously increasing, and it is not possible to distinguish drought events. In such cases, it is more meaningful to set EFR to 80 % of mean monthly Qant, implying that the altered flow regime is the “new normal” (see also Table 1).

In large-scale hydrological modeling, it is very difficult to accurately simulate how human-made reservoirs affect water availability, i.e., how they impact downstream streamflow and how reservoir storage varies in time. Therefore, it is more informative to use time series of reservoir inflow (streamflow data) instead of reservoir storage for assessing drought hazard for these risk systems. For water users that depend on large reservoirs, streamflow deficits during the low-flow months are not relevant, since reservoirs can store water from the high-flow season. Hence, drought magnitude should be assessed using SDHIs with longer averaging periods that either assume habituation to interannual variability (e.g., SSI12, EP12, Table 1) or mean annual conditions (RQDI12, Table 1), but not seasonality. At the four investigated gauging stations (Fig. S2), the relation between SSI12 and RQDI12 is the same as for SSI1 and RQDI1 (Sect. 4.2). If interannual variability is high (Orange River and White River), SSI12 values correspond to much higher RQDI12 values compared to the stations with low interannual variability (Danube River and Angara River). To obtain drought severity, these indicators can be cumulated using a suitable threshold. As described in Sect. 4.2, a threshold of -50 % for RQDI12 would exclude regions with low interannual variability, where this value is rarely reached, and where RQDI12 might underestimate drought magnitude.

In addition to these magnitude indicators, the volume-based severity indicators CQDI1(Q80-HS) and CQDI6(Q80) were assessed. With CQDI1(Q80-HS), an existing drought is allowed to continue in months in which the calendar month Q80 is zero, even if streamflow Q exceeds zero. In contrast, CQDI1(Q80) only allows a drought to continue if Q80 and Q are zero. A comparison of the two indicators (Fig. 4) reveals that the impact of the HS method is rather small at the global scale but can be relevant at the regional scale. Figure 4a depicts the fraction of drought months as a percentage of all 360 months during the reference period as indicated by CQDI1(Q80). Using Q80 as a threshold implies that the time series should be in drought 20 % of the time. The fact that this percentage is often reduced and sometimes increased can be attributed to the 2-month criterion (Sect. 2.3.2) (1-month droughts are ignored, and several droughts are pooled) and to drought prolongation if Q80 and Q are zero. The HS method leads to an increase in drought months by up to 3 percent points (corresponding to 11 out of 360 months) in 6 % of all grid cells, e.g., parts of India, Pakistan, Afghanistan, Iran, and the western US, all of which are regions with highly seasonal streamflow regimes (Fig. 4b). Larger increases of up to 12 percent points are only computed in 0.4 % of all grid cells. Hence, the additional information value of CQDI1(Q80-HS) in a large-scale DEWS would be small. Instead, CQDI variants with longer averaging periods like CQDI6(Q80) (Figs. 3b and S4b) are more suitable for assessing risk systems with reservoirs. The time series of CQDI6(Q80) at the four gauging stations (Fig. S2) illustrate how the maximum drought severity is shifted by 1 month or more compared to CQDI1(Q80), reflecting the fact that a reservoir storage requires several months of “normal” streamflow to be replenished.

A direct comparison between different severity indicators is possible when the time series of drought severity are transformed into frequency of non-exceedance. Figures 5 and S5 depict the probability (frequency) of non-exceedance p of drought severity in July 2003 and September 1993, respectively, between four CQDI1 variants, the cumulative relative deviation CRQDI1 with a threshold of -50 %, and the cumulative empirical percentile CEP1 with a threshold of 20 %. The indicators are denoted with the suffix “f” for frequency. A p value of 0.7, for example, indicates a high drought hazard, with the severity up to July 2003 being higher than the severity of 70 % of all completed drought events in the reference period. In both example months, the spatial extent of regions with p>0.7, i.e., severe droughts, is larger according to the indicators that do not assume habituation to interannual variability (CQDI1(Q50)_f, CQDI1(WUs)_EFR_f, and CRQDI1(-50 %)_f). Spatial patterns of CQDI1(Q50)_f and CQDI1(WUs-EFR)_f are rather similar. Correspondence between these two indicators is higher than between CQDI1(Q50)_f and CQDI1(Q80)_f. CRQDI1(-50 %)_f identifies fewer regions with severe drought status compared to CQDI1(Q50)_f but more regions compared to CQDI1(Q80)_f.

Spatial patterns of CQDI1(Q80)_f (Figs. 5a and S5a) and CEP1(20 %)_f (Figs. 5b and S5b) are very similar, since they are based on the same drought concept. Nonetheless, small differences occur in all identified drought hotspots, which can be explained by the fact that the former quantifies absolute and the latter relative streamflow anomalies per calendar month, leading to a different ranking of low-flow and high-flow droughts during the reference period. This relation is illustrated for the Danube gauging station in Fig. 6 and for the other three investigated stations in Fig. S6. Although CEP1(20 %) (in units of cumulative percent) and CQDI1(Q80) (in units of mean annual streamflow) capture the same drought signal at the four stations, the relative levels among the drought events differ. In Fig. 6, the three most severe droughts according to CQDI1(Q80) are the drought events in 1998, followed by 2014 and 2003. In contrast, the 2003 drought, which occurred mainly during the low-flow period (August to November), has the second-highest severity according to CEP1(20 %). The high-flow drought from March to May 2011, on the other hand, has a lower severity rank according to CEP1(20 %). The differences are more pronounced with higher seasonal variability (Orange River and White River, Fig. S6) but almost negligible if seasonality is very low (Angara River, Fig. S6). Consequently, in a large-scale DEWS, CEP1(20 %) appears to be more suitable in regions where the risk system is more vulnerable to low-flow droughts than to high-flow droughts. These differences would not occur if volume-based monthly streamflow deficits were normalized using mean monthly streamflow. They only occur if they are either not normalized (e.g., the low-flow index – LFI, Cammalleri et al., 2016a) or normalized against mean annual streamflow volume (e.g., van Loon et al., 2014, and all CQDI1 variants in this paper).