Denmark is located in northern Europe, covering approximately 43 000 km². The country has a temperate oceanic climate (Cfb, Köppen–Geiger classification) characterized by mild winters and warm summers, with a mean annual temperature of about 8–9° and a mean annual precipitation ranging between roughly 600 and 1000 mm, with a general east-west gradient and highest values in the west. Precipitation is relatively evenly distributed throughout the year (DMI, 2025).

The topography is low-lying, with elevations generally below 100 m a.s.l., and the landscape was shaped during the last glaciations, resulting in heterogeneous glacial deposits that in some parts cover shallow chalk and limestone aquifers locally affected by glaciotectonics (Schack Pedersen et al., 2018) and karstification (Nilsson et al., 2023). Quaternary glacial tills, meltwater sands, and clays dominate the near-surface geology in the eastern parts of Denmark, while western parts of Denmark and deeper layers comprise Neogene, Paleogene, and Cretaceous marine sediments. This results in a complex hydrogeological setting with interbedded aquifers and aquitards of varying permeability which are locally intersected by buried valleys primarily developed under ice sheets (Sandersen and Jørgensen, 2017). This gives rise to complex groundwater flow systems, varying groundwater age and travel time distributions both in shallow and deep aquifers (Hinsby et al., 2001; Troldborg et al., 2008), consequently affecting vulnerability of aquifers to droughts.

Denmark's hydrology is also strongly influenced by its land use. Almost two-thirds of the land area is agricultural, and artificial drainage systems are widespread, also in forests, to improve soil trafficability and vegetation growth (Olesen, 2009). Groundwater plays a central role in the Danish water cycle: Nearly all water supply (drinking water, industrial water use, irrigation in agriculture) is abstracted from groundwater resources (Henriksen et al., 2024). Moreover, there exist 65 000 km of water courses in Denmark, most of them only a few metres wide (Danish Agricultural Agency, 2025), with close interaction between surface and groundwater (Duque et al., 2023). Similarly, there exist 120 000 lakes with an area above 100 m². The uppermost groundwater table is found close to the surface, within a few metres below ground, for most of the country (Koch et al., 2021). Consequently, drought impacts in Denmark are not only reflected in meteorological deficits but also propagate through soil moisture and groundwater storage, potentially affecting streamflow and groundwater availability on seasonal to multi-annual timescales.

The DK-model is a distributed, integrated hydrological model covering all of Denmark, except for some smaller islands, at 500 or 100 m horizontal resolution (Henriksen et al., 2003; Højberg et al., 2013; Stisen et al., 2019; Henriksen et al., 2020). It has been under constant development over the last three decades (Henriksen, 2001), driven by projects for public authorities and research initiatives. It is being used as the basis for water resource assessments (Henriksen et al., 2024), climate change impact assessments (Schneider et al., 2022c; Seidenfaden et al., 2022), hydrological monitoring and early warning (Henriksen et al., 2018), nutrient transport studies (Andersen et al., 2025), or estimation of groundwater age and travel time (Musy et al., 2023).

The observational data for validating the simulated drought indices consists of three independent datasets: groundwater levels, streamflow, and soil moisture. Ideally, these data should cover the entire reference period, be continuous with at least monthly data, and have a wide spatial coverage. In addition, they should be minimally affected by direct anthropogenic factors such as groundwater abstraction, and therefore fluctuations in the observed time series should mainly be attributed to climatic variations. However, for some of the variables, the data temporal coverage requirements must be loosened due to data scarcity. Below, the selection process to identify data suitable for evaluating the drought indices simulated by the DK-model is described for each data source. Tables providing an overview over the final datasets used for the validation can be found in Appendix A.

A multitude of different drought indices exist and are described in the literature (see e.g., Zargar et al., 2011). They often cover different parts of the hydrological cycle, and thus, represent different variables, e.g., precipitation or streamflow. Often, they are either threshold-based or standardized (de Matos Brandão Raposo et al., 2023), meaning that they are either based on drought definitions characterized by crossing a certain threshold (e.g., a percentile of streamflow), or deviations of a time series from its normal (e.g., more than two standard deviations from the mean). Newer emerging indices are, for example, based on combining existing indices into composite ones (Raible et al., 2017) or indices modified for specific conditions, e.g., ephemeral streams (Aon and Biswas, 2024). Indices can be calculated from various observations of the hydrological cycle (Haas and Birk, 2017), remote sensing products or land surface models (Gaona et al., 2022), as well as hydrological models (Sutanto et al., 2024) or a combination of the above.

Generally, it is recommended to use standardised indices when comparing drought signals across different regions and compartments of the hydrological cycle (de Matos Brandão Raposo et al., 2023; Teutschbein et al., 2022; World Meteorological Organization (WMO), 2012). Due to the differences in precipitation regime from east to west in Denmark (Stisen et al., 2012), standardised indices were also chosen in this study. An overview of the indices used in this study is given in Table 1.

For the meteorological drought signal, we applied the commonly used SPI (Standardized Precipitation Index) (McKee et al., 1993) for precipitation, and the SPEI (Standardized Precipitation Evapotranspiration Index) for net precipitation (Vicente-Serrano et al., 2010), Eqs. (1) to (3), which we refer to as meteorological drought indices. The other indices, covering different compartments of the hydrological cycle, are referred to as hydrological indices. For soil moisture, we applied the ESSMI (Empirical Standardized Soil Moisture Index) (Carrão et al., 2016), Eqs. (4) to (6), as it allows for robust handling of bounds such as full saturation commonly occurring in simulated soil moisture values. For streamflow we use the SDI (Streamflow Drought Index) (Nalbantis and Tsakiris, 2009) with log-transformed streamflow values, Eqs. (7) and (8), as it is a common index in the literature (Gonçalves et al., 2023; Kim et al., 2024; Zhong et al., 2020), and its formulation originates from the SPI. This is also the case for the groundwater index used in the study, the SGDI (Standardized Groundwater Drought Index) (Bhuiyan et al., 2006; Bloomfield and Marchant, 2013), Eqs. (9) and (10). Examples of SGDI in the literature include Han et al. (2019), Ling et al. (2024), and Zhu et al. (2023).

Common for the standardized indices we used (Table 1) is that they indicate the deviation of the current status of, for example, groundwater levels, from the typical seasonal cycle, as defined by the mean monthly or weekly climatology over the reference period. For all indices, values below 0 correspond to below-average or dry conditions. Furthermore, the resulting index values are translated to categories of drought as established by McKee et al. (1993) and commonly used since: the different categories are “moderate drought” (-1 to -1.5), “severe drought” (-1.5 to -2), and “extreme drought” (below -2).

Figure 2 sums up the overall DK-model performance, showing cumulative distributions across the multiple conventional (i.e. not drought-related) calibration targets. Across 305 streamflow stations, a median KGE of 0.67 is reached, and the overall water balance error Fbal is 0.01, with a mean absolute error of 0.15. Fbal is calculated as (Qobs - Qsim)/Qobs. Across the 153 selected streamflow stations for the further drought analysis with long time series, performance is slightly better with a median KGE of 0.70, an overall water balance error of 0.02 and a mean absolute error of 0.14. In terms of groundwater performance, the mean absolute error across 39 514 wells with groundwater level observations is 3.67 m, with a mean error of 0.47 m. For the 53 selected wells for drought analysis, these errors are reduced to a RMSE of 2.37 m with a mean error of 0.46 m; which likely reflects their higher quality measurements. All mean errors are provided as obs - sim. Seasonal groundwater level amplitudes are reproduced with a mean absolute error of 0.65 m across 400 groundwater level time series with sufficient data to calculate average seasonal amplitudes. Across the 53 selected wells, the amplitude mean absolute error is 0.60 m. Observed amplitudes are 1.06 m on average, and mean absolute errors are skewed by outliers; the median absolute error is 0.41 m. Lastly, the drain fraction (average simulated drain flow per grid cell relative to precipitation) was included in the calibration: artificial drainage represents an important hydrological process in Denmark, with significant spatial variation, which is often overlooked. Hence, a Machine Learning generated map of drain fraction was used as a target (Schneider et al., 2025); panel (e) in Fig. 2 shows the residuals of the model against that map, indicating that the DK-model slightly underestimates the amount of artificial drainage. DK-model performance during validation periods 1990 to 1999 and 2011 to 2019 closely matches performance in the calibration period 2000 to 2010 reported here; see Fig. B1 in Appendix B.

Each of the individual observation time series of indices (obs) is compared to the simulated values at the respective locations (sim@obs). This is done based on monthly statistics for all drought indices, also those that originally were calculated on a weekly basis (SDI, ESSMI). To give the reader an impression of the time series, examples at observation points for the four drought indices are shown in Fig. 3, including the Pearson correlation coefficient r and the RMSE between the obs and sim@obs indices time series. The examples were chosen to be somewhat representative of overall performance (compare Fig. 4).

Figure 4 shows the correlation coefficients between obs and sim@obs across all observation points for each of the four indices. In panel (a), those values are shown on a map of Denmark. In panel (b), the cumulative distribution of the correlation coefficients is shown both for the entire time series, and separately for dry periods only (defined by a obs index < 0). Generally, performance is highest for the SDI, followed by SGDI and lastly ESSMI. Performance is slightly better in the western parts of the country. It can also be seen that the correlation coefficient r tends to be lower during dry anomaly periods. However, this does not necessarily reflect lower performance; rather it is related to the sensitivity of Pearson's r to the range of occurring values, which is restricted to roughly half if only looking at dry periods. This is also confirmed in Table 2, which shows the median values of correlation coefficients across all time series. Values are provided across the entire period, and separately for dry and wet periods, which are defined by negative and positive obs index values, respectively. Here, r decreases for both wet and dry periods compared to the full series, which is expected because truncating the distribution reduces variance and covariance, which mathematically leads to smaller Pearson correlation coefficients. Thus it is not necessarily indicative of a poorer performance of the model in dry or wet conditions.

Table 2 summarizes the correlation coefficients to median values across all time series for each of the four indices. Values are provided across the entire period, and separate for dry and wet periods, defined by negative and positive obs index values, respectively. In the “DK statistics” columns of Table 2, we also include correlation coefficients between aggregated drought index time series aggregated all observation locations (obs, sim@obs) or across all of the DK-model domain (sim). The overall performance at the individual time series level is at median Pearson's r values above 0.75 across all conditions, and values around 0.6 during dry periods only. The only exception to this is the ESSMI with lower r values mostly between 0.4 and 0.5. When looking at aggregated values across all of Denmark (DK statistics), the performance is better with r values close to or above 0.9 for SDI, SGDI_shallow and SGDI_deep. Again, the only exception is ESSMI which has lower correlations.

Table 3 summarizes the drought detection performance. This is done for three drought categories: Moderate drought, with drought index values below -1, severe drought with values below -1.5, and extreme drought for values below -2, all based on monthly values of sim and obs indices. The agreement in detection of those drought events then is provided as F1 scores, where performance is moderate to good. Exceptions are extreme droughts – which partly owes the fact that they are rare events badly defined by the observed and simulated data, occurring theoretically only during 0.75 times during a 30-year climatology. Also, as before, soil moisture drought performance is deteriorated compared to the rest. Drought detection performance was also evaluated for time series aggregated across all of Denmark (DK statistics), where agreement is better with F1 scores mostly around 0.7 or 0.8. The exception is soil moisture drought.

The SPI was calculated for different accumulation periods from 1 to 60 months, resulting in 60 time series from SPI_acc1 to SPI_acc60. For each of these time series, a correlation to the hydrological drought index was calculated. The accumulation period of the SPI that exhibits the highest correlation to the hydrological drought index indicates the dynamics of anomaly propagation from a precipitation anomaly to a hydrological impact. This is done separately for the obs and sim@obs index time series for ESSMI, SDI, and SGDI, and the resulting optimal SPI accumulation periods can be compared. If the model captures the development time and interconnectivity of the syst emsatisfactorily, the optimal SPI accumulation time for obs and sim@obs should be similar.

Figure 5 shows the results of this analysis. For each of the individual observation points, all 60 accumulation periods from 1 to 60 months for SPI_acc were tested. The resulting correlation coefficients are shown in the plots in the left column, separately for each of the obs and sim@obs time series. The SPI accumulation period yielding the highest correlation for obs and sim@obs, respectively (i.e. the peak of each curve in the left column plots), is then shown in scatter plots against each other in the right column, to evaluate whether obs and sim@obs indices reflect similar accumulation periods.

For streamflow (SDI), the DK-model tends to delay drought propagation more than seen in observations: the majority of obs indices (86 of 153 stations) show the highest correlation of SDI to SPI_acc1 or SPI_acc2, whereas sim@obs indices are correlated to accumulation periods of up to 10 months. However, it has to be noted that the optimal accumulation periods do not seem to be well defined; see especially the obs indices, which correlations only slowly decay up to roughly SPI_acc10. For groundwater (SGDI), the time series seem to be more grouped, with some having short accumulation time correlations (up to 6 months), others longer (7 to 12 months), and some very long (above 12 months). For the obs indices, roughly one third of the 53 wells fall in each of these categories: 19 with up to 6 months, 15 with 7 to 12 months, and 19 with above 12 months. The distribution for sim@obs indices is very similar, with 18, 15, and 20 wells in the respective groups. Also, the typically longer observed accumulation periods for SGDI_deep are reproduced in sim@obs indices. Again, ESSMI is the exception with the sim@obs indices having the highest correlation to SPI_acc2, whereas the obs indices exhibit high correlations for SPI_acc6 to SPI_acc8, but also around 30 months.

From the scatter plots, it can be confirmed that for SDI and SGDI, the optimal SPI accumulation periods broadly agree, with correlation coefficients between the optimal accumulation periods of obs and sim@obs of 0.49 and 0.66, respectively. This indicates that the DK-model can capture the major dynamics of drought propagation through the hydrological cycle, especially in the groundwater. The soil moisture performance is poorer, but the evaluation is also restricted by the limited amount of data.

The different accumulation periods across the observations suggest there are regional differences in the response time to precipitation. Figure 6 shows the accumulation period of SPI which yields the highest correlation to each of the sim indices, mapped for all of Denmark in the top row. The bottom row shows what the highest correlation is (between sim index and SPI_acc with the optimal accumulation period). Those correlations are generally high, with Pearson's r values mostly above 0.5, for SDI and SGDI often even above 0.75.

Note that the drought propagation from SPI to hydrological drought is very similar to the drought propagation from SPEI to hydrological drought. To maintain clarity, we focused on propagation from SPI here; corresponding versions of Figs. 5 and 6 for SPEI can be found in Appendix C, Figs. C1 to C4.

The skill of the DK-model to reproduce observed propagation expressed as accumulation periods of SPI was also evaluated by comparing it against correlations and simple linear regression models between geology-derived controlling variables and the optimal SPI accumulation periods. Table 4 summarizes the results, showing correlations between the controlling variables and the SGDI lag (SPI_acc to SGDI_shallow and SGDIdeep), separate for shallow and deep wells as well as combined across all 53 wells.

The correlation between the individual geological variables and SGDI lag is larger for shallow wells than for deep wells. Significant correlations, however, can only be found for filter depth and depth to observed groundwater table for the shallow wells and across all wells, and for overburden across all wells. No significant correlations exist for the deep wells. The DK-model simulated SGDI lag, conversely, shows significant correlation for all well groups, and demonstrates the highest predictive ability across deep wells and across all wells. The DK-model also outperforms a multi-variable linear regression model based on the five geological variables. Only for the shallow wells, single geological variables such as the depth to the observed groundwater table or the multi-variable linear regression model show better correlations to the SGDI lag than the DK-model.

In the “DK statistics” columns of Table 2, we include correlation coefficients between the combined drought index time series aggregated across all observation locations (obs), the same type of combined index for simulation time series (sim@obs), and one combined from the entire DK-model domain (sim). This sheds light on different aspects. First, the performance of the DK-model in simulating observed drought indices (obs vs. sim@obs), on an aggregated level. Here, the performance is even better than for individual time series, with r values close to or above 0.9 for SDI, SGDI_shallow, and SGDI_deep. Again, the only exception is ESSMI, which is not as well correlated.

Secondly, we can evaluate the performance of the observations concerning the entire domain (obs vs. sim), which proves very similar to the performance when comparing to simulations from the actual locations of the observed time series (sim@obs). Similarly, the index calculated for simulation data at observation points is very strongly correlated to the behaviour of the entire domain (sim@obs vs. sim).

Aggregated drought indices across all of Denmark, as monthly means for the years 1990 to 2023, are shown in Fig. 7. Generally, drought patterns between obs and sim@obs indices agree well, as already indicated by good correlation performance values reported in Table 2 above. Notably, there is also good agreement between the indices based on the relatively few observation points (obs and sim@obs) and the simulated Denmark-wide drought index dynamics (sim) across all grids or streamflow points in the DK-model. Thus, the observation points are thought to be representative of the behaviour of the entire domain and can therefore be used to evaluate the general DK-model drought performance.

Figure 8, top row, shows mean monthly drought indices for all of Denmark. Monthly SPI and SPEI values show little autocorrelation in time, and soil moisture anomalies (ESSMI) closely follow the anomalies in (net) precipitation. Effects of meteorological drought keep accumulating, though, when moving further through the hydrological cycle: SDI starts showing more continuous, more extended drought periods (or wet anomalies), and SGDI_shallow and SGDI_deep react with even more delay, exhibiting longer continuous drought periods in line with results shown in Figs. 5 and 6. The middle and bottom row then show the differences between two regions in Denmark, western Jutland and eastern Jutland (WJ and EJ in Fig. 2b), which are dominated by more sandy and clayey soils, respectively. In the more sandy western Jutland, drought signals propagate faster from meteorological to hydrological drought, especially visible in the deep groundwater (SGDI_deep). The more clayey eastern Jutland experiences slower drought propagation, particularly to the deep groundwater, as evidenced for example by a delay of few months in both the onset of and recovery from the SGDI_deep drought in 1996/97 compared to West Jutland.

Figure 9 shows maps of drought indices for all of Denmark for May 2020, and illustrates diverging response times of the different hydrological compartments. The first column shows SPI_acc2 and SPI_acc12, whereas the remaining maps show the four sim indices. The month of May 2020 is characterized by a soil moisture drought, with ESSMI values below -1 (moderate drought) in large parts of the country and partly below -1.5 (severe drought). Soil moisture is low as May and April 2020 had been experiencing unusually low precipitation amounts, reflected in SPI_acc2 values being mostly below normal values. The remainder of the hydrological cycle, however, remains in normal to wet conditions, as expressed by SDI, SGDI_shallow, and SGDI_deep. This is due to their slower response to precipitation anomalies than fast-reacting soil moisture (compare Figs. 5 and 6), and the wet preceding conditions in the entire 12-month period prior to May 2020, as expressed by high (i.e. wet) SPI_acc12 values.

For Danish conditions with winter being the main recharge season, meteorological droughts during the winter season have a comparably larger effect on groundwater drought development. This is illustrated in Fig. 10, where the 34 years 1990 to 2023 are separated into two seasons, the winter half-year October to March and the summer half-year April to September. Then, they are further split into their meteorological drought condition across each winter or summer, defined by the SPI_acc6 at the end of the respective 6-month period being below -0.5 (drought) or above 0.5 (non-drought), resulting in 11 drought winters and 10 non-drought winters, as well as 12 drought summers and 10 non-drought summers. Figure 10 then shows the developments of the SGDI_deep values throughout each of the drought or non-drought winters or summers, relative to the SGDI_deep value at the start of each season. The simulated sensitivity of SGDI_deep to a precipitation drought is higher during winter, with an average drop of -0.60 of SGDI_deep index value during a drought winter, panel (a), than during summer with an average drop of -0.28 during a drought summer, panel (b). This seasonal difference is similarly pronounced during non-drought, i.e. wet periods: the deep groundwater drought recovery during wet winters with an average increase of 0.63, panel (c), is larger than the drought recovery during wet summers with an average increase of only 0.29, panel (d).

The observational dataset for drought propagation compiled as part of this study was found to be robust for streamflow and groundwater. The quality assurance and selection criteria, such as excluding groundwater level observations significantly affected by abstractions, yielded a dataset suited for evaluating drought as a natural phenomenon, driven by climate variability instead of changing human interventions. Despite the inevitably incomplete spatial coverage, we could show that the 53 groundwater level and 153 streamflow time series are representative for Denmark-wide drought behaviour (compare Fig. 7). The only exception is soil moisture, where currently we are limited to observation time series of 10 years from only three stations, which potentially limits the validity of the evaluation. In the future, more long-term time series of relevant soil moisture observations will become available.

The subsequent evaluation of the DK-model's ability to simulate drought and its propagation through the hydrological cycle showed good results. Not only are the overall dynamics of drought indices captured well by the DK-model (Fig. 4 and Table 2), but importantly the lag times for propagation from meteorological drought to streamflow and groundwater drought are also captured (Fig. 5). Few outliers for accumulation periods between obs and sim@obs indices (Fig. 5f and i) can potentially also be explained by the peak correlation being weakly defined, see the relatively broad peaks in panels (d), (e), (g), and (h) Moreover, due to the limited time series length of soil moisture observations (10 years), long correlations of 30 months and above (Fig. 5a) should be interpreted with care.

The resulting patterns of (simulated) drought propagation lag from meteorological drought to soil moisture, streamflow, shallow and deep groundwater, respectively, show spatial variability (Fig. 6), which also can be represented to some degree by the DK-model (Fig. 5). A multitude of factors control those spatial patterns. The groundwater lag is a particularly complex variable, as not only does geology vary from well to well, but also the depths to the aquifer or groundwater table, local groundwater gradients, etc. Results (Table 4) of a correlation analysis between geology-derived controlling variables and groundwater lag at the 53 groundwater wells showed that drought propagation to deeper groundwater becomes increasingly complex and is controlled by a multitude of variables, going beyond simple information about aquifer depth or lithological information. The DK-model outperforms a multi-variable linear regression model built with the controlling variables, as it not only is informed by geological information but also adequately captures resulting regional patterns of recharge and groundwater flow, thus representing local differences in drought propagation lag. This spatial diversity is also apparent in the comparison of drought indices across the more sandy western Jutland and the more clay-dominated eastern Jutland in Fig. 8, where western Jutland generally shows quicker dynamics than eastern Jutland (see Fig. 1 for an outline of the regions and their surface geology).

Droughts are often perceived as more of a summer (or dry season) phenomenon. However, in particular groundwater droughts are controlled by groundwater recharge patterns instead of the meteorological variables directly. In humid temperate climates such as Denmark, groundwater recharge is concentrated during winter months (Hisdal and Tallaksen, 2003; Liu et al., 2025; Nygren et al., 2022). Hence, groundwater droughts show a lagged and seasonally dependent response to meteorological droughts. The phenomenon of groundwater sensitivity to winter drought (or, more general, to drought during a recharge season) is modelled well by the DK-model, meaning it reproduces observed larger sensitivity to winter droughts than to summer droughts (Fig. 10). Interestingly, a study by Wunsch et al. (2024) showed for mostly shallow groundwater wells across Germany that low-water periods at least during late summer are mostly driven by summer climate.

As outlined in the introduction, drought-specific evaluations of the ability of hydrological models to preproduce drought and its propagation across multiple hydrological compartments remain rare, underlining the research gap we are trying to fill with this work. Drought-specific evaluations often are limited to a single compartment, and multi-compartment evaluations cover overall hydrological signals, even though it is acknowledged that more holistic evaluations covering multiple parts of the water balance are crucial for, amongst others, drought monitoring (Rakovec et al., 2016). Examples for multi-compartment drought specific evaluation are: Hellwig et al. (2020) evaluated a distributed groundwater model across Germany, and report Temporal Agreement Indices (TAI) (Stahl et al., 2011) between observed and simulated groundwater heads and baseflow. Their drought definition is the lowest 20 % of each timeseries, i.e. more moderate and frequent events than the moderate drought definition used in our work, and they report median TAI of 0.25 on groundwater heads and of 0.28 on baseflow. We chose the F1 score as drought agreement metric, as it not only considers true positives. However, our TAI on moderate drought (i.e. 15.9 % of each timeseries) are higher, with a median of 0.33 across the streamflow stations and of 0.46 across the groundwater wells. Bruno et al. (2024) performed a diagnosis of the performance decrease of a distributed hydrological model over the Po River Basin during moderate and severe droughts. They found that model performance on streamflow, evapotranspiration and total water storage anomalies mainly deteriorated during severe droughts, and assumed the reason for that in their case are misrepresentation of irrigation during such severe droughts. This highlights the need for integrated assessments and models to capture drought dynamics across the hydrological cycle. Similar to our results, Wan et al. (2022) found also more difficulties in representing droughts in soil moisture than streamflow when evaluating the semi-distributed US National Water Model.

The DK-model is an operational model (Liu et al., 2026), running in real-time and forecast mode, and thereby offers potential for early warning and drought forecasting. Previous studies have noted that hydrological drought forecasts are generally more reliable than purely meteorological drought forecasts (Sutanto et al., 2020), particularly in systems with a strong groundwater component and long memory effects (Du et al., 2023; Pechlivanidis et al., 2020; Sutanto and Van Lanen, 2022). In Denmark, observed and simulated drought propagation lags (see Figs. 5 and 6) indicate that it often takes several months for meteorological droughts to translate into hydrological droughts, especially for groundwater. This implies that seasonal hydrological drought forecasts may achieve skill, as drought conditions several months in the future are partly affected by the current hydrological state. Such predictive capability is particularly relevant in a Danish context, where groundwater is the primary source for agricultural irrigation. Improved forecasts of groundwater drought could therefore provide an essential basis for early warning systems and proactive water management, supporting farmers and water authorities in preparing for increased irrigation demands during dry periods. The variability of groundwater extraction for drinking water and irrigation, both inter- and intra-annual, however, remains challenging to predict and incorporate in models.

Recent work has also shown that Machine Learning and deep learning models can predict hydrological drought indices (e.g. Liu et al., 2024b; Wang et al., 2023; Zellou et al., 2023). Also in the context of the DK-model it could be shown that LSTM (Long short-term memory) models, applied as hybrid models alongside DK-model output to predict streamflow, outperform the physically based hydrological model (Liu et al., 2024a). Besides that, drought indices based on a combination of remote sensing products and variable-driven indices have also shown great potential for drought monitoring (Choi et al., 2013). Such products can, for example, more accurately monitor vegetation response to drought stress.