The groundwater dataset used in this paper was produced by GLOBGM v1.1, a two-layer transient MODFLOW-based groundwater model which represents confined, confining and unconfined aquifers . The model simulates monthly groundwater heads and water table depths at ∼1 km spatial resolution for the time period 1960–2019. The model distributes independent unstructured grids over three continental-scale groundwater models (Afro-Eurasia, America, Australia) and one representing the remaining islands. GLOBGM employs an offline coupling approach whereby local runoff, groundwater abstraction and groundwater recharge outputs at the 5 arcmin resolution (∼10 km at the equator) are derived from the PCR-GLOBWB2 global hydrological model . These inputs are then used as boundary conditions and forcing to simulate groundwater dynamics.

The groundwater abstractions in GLOBGM are based on data from PCR-GLOBWB2 . In PCR-GLOBWB2, the water withdrawal is equal to the gross water demand, which made up of both irrigation and non-irrigation sources. The demand from irrigation is based on the crop composition and the irrigated area and is derived from MIRCA2000 and the Global Crop Water Model . The non-irrigation demand covers three sectors (industry, households and livestock). For more information on the calculation of this demand we refer to . In PCR-GLOBWB2, water demand can be fulfilled from three sources: surface water, groundwater and desalinated water. The fraction of water which is abstracted from groundwater and surface water is calculated based on their relative abundance, where the abundance is calculated based on a 2-year running mean of recharge and discharge. Based on this, PCR-GLOBWB2 calculates how much water is abstracted from each source each month. The total groundwater abstraction is then given as an input to GLOBGM. Groundwater abstractions are highest in many of the recognized water scarcity hot-spots, with the largest average abstractions across the Indo-Gangetic Plain, northern China and western and central United States (see Fig. S8 in the Supplement). Declining water table depth trends in the GLOBGM simulations align well with known regions of excessive irrigation, groundwater abstraction and depletion e.g. the United States, the Arabian Peninsula and the Indo-Gangetic plain . For more information on the groundwater abstraction data we refer users to .

The validation of global groundwater models is notoriously difficult due to mismatches in spatial resolution and representation between observed groundwater data and model simulations of groundwater. Nonetheless, the outputs from GLOBGM have been validated in previous work . found that the GLOBGM non-parametric KGE skill score was more than 0 (suggesting that the model outperforms the mean flow benchmark) in 75 % of the simulated area. This proportion increased to 81 % after excluding locations where poor skill scores were attributed to metric artifacts associated with weak groundwater trends and very shallow water tables. The model performance is best where water table depths are shallow (0–20 m), but there is also a tendency for it to underestimate groundwater heads between 0–5 m. Contrastingly, groundwater heads were found to be overestimated at deeper water table depths, perhaps due to delays that occur in deeper unsaturated zones which the model does not accurately account for. Notably though, the ability of the model to reproduce groundwater variability does not appear to be linked to water table depth. GLOBGM has also been evaluated against other, similar, groundwater models. Compared to alternative 5 arcmin dynamic model simulations (e.g. de Graaf et al. 2017), GLOBGM displays a 19 % increase in the number of locations which have a non-parametric KGE skill score more than 0. For more information on the model evaluation we refer to .

Since this study focuses on drought, and the prior validation efforts have not considered this aspect of model performance, an extra validation was carried out to assess how effectively GLOBGM simulates patterns of groundwater drought against two sources of observed groundwater data which are introduced below.

Throughout this study we refer to two types of drought: Normal Droughts (ND) and Multi-Year Droughts (MYD). A MYD event is defined as any event that lasts for at least 12 consecutive months . Comparatively, a ND is defined as any event which lasts for less than 12 months. The onset and termination of a drought event is defined using run-theory, where a drought is defined as a continuous period below a threshold.

To understand the global vulnerability to groundwater droughts, we first aim to understand the role of the subsurface properties in the propagation of meteorological drought into groundwater drought. The following section introduces the metrics used in this manuscript to characterize the relationship between meteorological and groundwater drought. These metrics are also used to evaluate the model against observed groundwater data.

Figure  plots the anomaly correlation between the observed and simulated groundwater data. The mean anomaly correlation between observed GRACE and simulated groundwater data (Fig. a) is 0.23 (median 0.27), whilst the mean anomaly correlation between the point IGRAC observed data and simulated groundwater data (Fig. b, c, d) is 0.32 (median 0.35). The distribution of anomaly correlation coefficients is similar across the two observational datasets (Fig. e). Positive correlations dominate across most of the world, with consistently strong correlations observed across Europe and southern USA. The correlation is weakest south of the Sahara desert and strongest in the Arabian desert, both areas which are relatively dry. The model performs slightly better than average in the regions with abstractions in the 95th percentile. In these areas the mean anomaly correlation with the GRACE data is 0.36 (compared to the global average of 0.23) and the median is 0.39 (compared to the global average of 0.27).

In several of the areas where the anomaly correlations are lower, we find nuances in the GRACE data which can begin to explain these patterns. Since GRACE is a measurement of total water storage, in tropical regions such as the Amazon or central Africa, the GRACE signal can be dominated by changes in surface water and root zone soil moisture rather than changes in groundwater storage. In more arid regions, the lower anomaly correlations can be partially explained by the time it takes for surface water to reach the groundwater system, meaning that low correlations are not unexpected . Previous studies using the IGRAC groundwater point observations have also found challenges when using this data for model evaluation. identified that in many of the slow moving systems (e.g. the Sahel, Namibia), the observed and simulated data can show opposing trends, but note that in many of these cases the magnitude of change is very small. As a result in many locations the correlation can over exaggerate the differences in groundwater levels. Finally, we are also aware of flaws in the model input data, for example in sub-Saharan Africa where the forcing data is not always representative . Nonetheless, there are several regions where it is likely that the model does not have a strong anomaly correlation, and in these places the results should be interpreted with caution.

The groundwater model simulates drought events of a similar average duration to those recorded in the observational record (see Fig. S1 in the Supplement). The mean drought duration across the IGRAC observational data is 5.4 months and the median drought duration is 3.4 months. At the same locations, the simulated groundwater data has drought events with a mean duration of 6.4 months and a median duration of 2.9 months suggesting that the model records a greater number of long groundwater droughts at these locations which are not seen in the observational record. When comparing the global GRACE observational data to the simulated groundwater droughts, we see that GRACE records a mean drought duration of 5.5 months and a median of 4.3. Comparatively, in the regions which overlap with the GRACE coverage, the model simulates a mean drought duration of 9.9 months and a median of 5.4 months (note that this value differs from above because of only including data which overlaps with the observed GRACE data) suggesting that again the model simulates a greater number of long groundwater droughts. This is likely because of slower model response times and a higher degree of auto-correlation.

To the authors knowledge, this global analysis is the first to use hyper-resolution data to focus specifically on the drivers and patterns of multi-year groundwater drought. The lack of comprehensive global groundwater data has in the past meant that groundwater has been excluded from global drought propagation analysis , or has been represented by coarse and temporally limited data . Although drought propagation has received significant attention at local, regional and national scales, the variability in the global patterns and drivers of groundwater drought mean that these findings cannot be easily extrapolated for actionable use.

Our results suggest that subsurface geo-physical properties are the primary drivers that modulate meteorological droughts into multi-year groundwater droughts. Although drought propagation begins with the meteorology, we find that in much of the world groundwater resources are substantially disconnected from the meteorological input and that the sub-surface significantly amplifies and modifies the drought signal (meteo < GW). Less than 2 % of global groundwater grid cells have an average meteorological drought duration of more than one year, but by comparison, 35 % of global groundwater grid cells have an average groundwater drought duration which is multi-year.

Disentangling the relationship between meteorology and groundwater exposes areas which are vulnerable to multi-year drought in the present day, but also identifies those which will be more vulnerable in the future. At present, the regions where meteorology and groundwater are connected to one another do not frequently experience multi-year drought (only 6 % of global data has a multi-year average groundwater drought duration and also falls into the meteo = GW category). However, in the future, we expect an increase in the frequency and magnitude of multi-year meteorological droughts , and thus regions in this category will be increasingly at risk from multi-year groundwater drought. An increase in the frequency or duration of meteorological droughts might also push more points into the meteo < GW category. If meteorological droughts are lengthened, or more frequent, then there will be less time for recovery between drought events, which might encourage increased pooling of meteorological droughts into groundwater droughts . This process may also amplify the relationship between meteorology and groundwater for points already in the meteo < GW category (increasing the DDR), since increased pooling would extend the meteorological signal in a non-linear way. It is also likely that increases in groundwater abstractions or pumping might push data from the meteo = GW category to the meteo < GW category . Although we have not considered the impacts of abstractions on the typology we have presented in this paper, in many regions this is likely to have a significant impact on groundwater trends and should thus be considered in future work. However, strong trends in groundwater levels also have impacts on the proper characterisation of drought anomalies and thus require methodological advances in drought studies .

Understanding the dynamics of multi-year drought propagation can inform local drought mitigation and help to predict and prepare for drought impacts . This is particularly important in the context of multi-year groundwater drought, since groundwater is often used to supplement surface water and sustain vegetation , so a prolonged absence of this resource can have wide-reaching implications. The typology that we implement in this study provides a simple baseline for understanding global patterns of groundwater responsiveness, which in turn can inform drought mitigation. For example, in regions where the meteorology and the groundwater are connected (which in our analysis makes up 19 % of global groundwater grid cells), the magnitude and timing of a drought can be anticipated based on the observed anomalies in the meteorological conditions. This has significant advantages for early-warning systems, which can be particularly beneficial for managing groundwater abstractions and the conjunctive use of groundwater and surface water. If we know the systems to be closely connected, then the end of a meteorological drought (which is much easier to measure and observe) is likely to also signal the end of a groundwater drought. It is also beneficial to understand whether or not a location falls into the meteo < GW type, since in these cases even a short meteorological drought might signal the start of a multi-year groundwater drought for which society needs to prepare. Contrastingly, in places where the meteorological droughts tend to be longer than the groundwater droughts, a multi-year groundwater drought is extremely unlikely since multi-year meteorological droughts are rare.

Consistent with findings from other studies , our analysis confirms that even regions which are subject to similar meteorological forcing may exhibit a very different groundwater response. The modulation of meteorological drought is controlled by the presence (or absence) of several key groundwater processes which are tied to local catchment characteristics and aquifer properties.

Lag, attenuation, pooling and lengthening are all well-established mechanisms of drought propagation which we observe in this study . The existence of a multi-year groundwater drought is often facilitated by the lengthening of the meteorological signal, and we find that the most extreme examples of this are usually accompanied by pooling of several meteorological drought events. In these cases, the clustering of meteorological droughts could be made more likely by persistent weather patterns and decadal climate variability .

In general, the places where groundwater lags behind SPEI-12 also tend to be locations which substantially amplify the meteorological signal and are more likely to encounter a multi-year groundwater drought. In agreement with other studies, we find that longer lags tend to be associated with deeper groundwater and have longer average drought durations and lower anomaly correlations . Understanding the groundwater lag also has significant benefits for drought forecasting, where the characteristics of a meteorological drought can be used to anticipate the upcoming groundwater response.

Outside of the more well-established drought propagation processes, our analysis also identified a cluster of points which have a notably bimodal distribution of drought events. In these locations more time is spent in multi-year drought than normal drought, but the average drought duration is skewed down (to less than 12 months) by frequent short drought events. Groundwater here is typically shallow, and close to surface and potentially connected to surface water. We hypothesize that these regimes might have a mechanism whereby the groundwater system can buffer shorter, or less intense meteorological droughts, but when recharge is reduced beyond a certain threshold the groundwater becomes disconnected from the surface water, descending into an intense, multi-year drought event. Since small changes in groundwater management could prevent or trigger this mechanism, it is important that we can predict where and when this process might occur.

Throughout this analysis we encountered several difficulties associated with the definition of groundwater and meteorological drought. Firstly, variability in the response time of groundwater to meteorological anomalies meant that it was hard to characterise groundwater droughts with one uniform meteorological drought index. Although this was made more difficult by the diversity inherent in global data, other studies have also found this to be a challenge on smaller, more regional scales . For this study we chose to use SPEI with a 12 month accumulation which was consistent with multi-year drought literature , but also aligned with other global studies which found the average groundwater accumulation period to be close to 12 months . However, mismatches in SPEI-12 and groundwater accumulation still result in a portion of data falling into the meteo > GW type. In these locations the groundwater responds to the meteorology faster than the 12 month accumulation inherent in SPEI-12 and this results in shorter groundwater drought events. Since the main focus of this study was to identify the drivers of multi-year groundwater droughts, and they are not common in these fast-responding regions (there are no places with a multi-year average groundwater drought duration in the meteo > GW type), this has not hindered our analysis, but should be considered if this analysis was to focus on the drivers of shorter events. For a sensitivity analysis which demonstrates how the SPEI accumulation period impacts the distribution of data across our typology see Sect. S7. Furthermore, in many locations it is also not possible to distinguish the climatological controls on drought from the anthropogenic controls. Our study assumes that drought (and drought duration) are a reflection of the meteorological signal, when in some cases the drought and its persistence are actually a function of human activity (e.g. groundwater pumping). Therefore in regions with a high level of human interference, groundwater drought should not be considered as solely the manipulation of the meteorological signal.

Another challenge for groundwater drought definition was the presence of trends. Globally, many groundwater aquifers are experiencing significant decline both from unsustainable abstractions and from climate change . These trends remove the stationarity assumption that is often used in drought characterisation and can artificially inflate or deflate drought length. Solutions such as de-trending the data have not been successful in other global groundwater drought studies e.g., so instead, here we implement a simple screening to remove drought events likely to be a function of trends. By removing droughts which have already begun on the first timestep or not ended on the final timestep of our groundwater simulations, we see that the 54 % of global groundwater cells which have on average a multi-year groundwater drought duration is reduced to 35 %. To supplement this initial screening we also redistributed the global groundwater data into the response types after excluding places where a significant trend is observed. To calculate which locations are impacted by trends we fit a linear regression to the standardized monthly groundwater anomaly timeseries and calculated the total change in standardized anomaly over the time period. Where this is more than 3 standard deviations of change across the period and the p value is <0.01 we conclude there is a significant trend. Masking out this regions makes only a very small change to the global distribution of response types. The meteo > GW category goes from containing 43 % of data to 38 % when we mask out locations with trends, meteo = GW stays at 20 % and meteo < GW changes from 38 % to 43 %. This minimal change assures us that our initial screening procedure is removing most of the significant trend-influenced drought events, but it also highlights a potential artifact of the methodology where in a small number of locations, removing one particularly long event can instead skew the average groundwater drought duration down, leading to a slightly higher proportion of data in the meteo < GW category.

Finally, we also had difficulty defining droughts in regions with very little variability. In these locations, the monthly standardized anomaly was very small and thus sensitive to small changes in groundwater table depth.

The validation of global groundwater models is notoriously difficult due to mismatches in spatial resolution and representation between observed groundwater data and model simulations of groundwater. The relatively low anomaly correlations and mismatches in drought duration calculated between observed and simulated data should influence the confidence we have in the classification of groundwater response types and interpretation of the data in regions where the model does not perform well or can not be evaluated. In regions where the anomaly correlation is low, the model may not be well suited for regional use and the results must be interpreted with caution. While the ROC AUC scores show a skillful representation of the model, the magnitude of the anomalies in groundwater simulations could be better represented, as shown by the lower anomaly correlations. This indicates that while the groundwater anomalies are represented in the correct direction (either too wet or too dry) their magnitude is not always correctly estimated. There are several reasons why the anomaly correlations may be low which are unrelated to GLOBGM. Firstly, because of errors in the observational dataset. As discussed in Sect. 3.1, in many places the GRACE data is not suitable for direct comparison to the groundwater data, and the signal should not be compared to groundwater table depth. This has been observed by many other studies in the literature who have identified problems with using this data for drought-based evaluation . Although we tried to couple the GRACE evaluation with perhaps more reliable point observations, these are reasonably scarce. Similarly, comparing the length of droughts simulated by GRACE to those simulated by GLOBGM can be misleading. Using GRACE data alone only 9 % of global groundwater cells have a multi-year average drought duration, compared with 35 % using GLOBGM. This is likely to be influenced by the shorter length of this timeseries (18 years of GRACE data compared with 60 years simulated by GLOBGM), especially since it has already been shown that meteorological multi-year droughts can be clustered in time . As well as the limitations involved in comparing GRACE's total water storage to simulated groundwater table depth, the difference in resolution between the datasets (∼300 vs. 1 km) is also likely to influence the number of multi-year groundwater droughts in the record, where a coarser resolution can result in a more muted drought signal.

As a groundwater community, we need to push for more global groundwater data to aid model evaluation, but in its absence, we suggest that there is still value in running models in regions where they cannot be validated. Since GLOBGM has been proven to perform as well as, or better than many alternative global groundwater models , we believe there is still value in using the results for global analysis. Currently the model's performance is on par or better than other global dynamic groundwater model simulations. Choosing suitable metrics for model evaluation is also tricky, and in this case, in many areas the anomaly correlation could be over exaggerating a mis-match in simulated and observed data in regions with little water table variability where small trends oppose each other. It is also likely that the meteorological forcing input to GLOBGM could be decreasing the anomaly correlations in some regions such as sub-Saharan Africa and north-eastern Siberian . Uncertainties remain large for the sub-surface parameterization, however in the absence of better parameterization we have used the state-of-the-art available parameters to perform GLOBGM groundwater simulations. It is possible that the groundwater abstractions implemented in GLOBGM contribute to poor model performance in some regions. However, in general, GLOBGM simulates large scale groundwater depletion signals well, identifying many of the established regions of known depletion e.g. the Arabian Peninsula and California's Central Valley. It also recreates some of the regions where groundwater supplies are recovering e.g. the Guarani Aquifer System in South America .

This study has identified several new research avenues which could further enhance our understanding of multi-year groundwater drought. Firstly, although GLOBGM includes groundwater abstractions, this study has not attempted to disentangle their influence on groundwater drought. Future work could leverage the modelling framework introduced here to test how varying levels of groundwater abstractions impact multi-year groundwater drought development and evaluate whether changes to groundwater pumping could be an effective strategy for drought mitigation. Secondly, it would be beneficial to perform a similar analysis using model simulations which are forced with future climate data. Understanding how the drivers and patterns of multi-year drought might change in the future under climate change is important for preparedness and could inform long-term policy and planning. Finally, it would also be interesting to perform a similar analysis on the event-scale. Our work considers only the average relationship between groundwater and meteorological drought events, but this many vary on an event-to-event basis.