Hydrological droughts propagate from a meteorological signal (precipitation deficit), through soils (soil moisture deficit), and, via reduced recharge, cause lowered groundwater levels. The spatio-temporal characteristics of the drought signal change as it passes through the terrestrial water cycle (Changnon, 1987; Peters et al., 2003; Tallaksen and Van Lanen, 2004; Van Loon, 2015). Due to the potentially large storage changes in aquifers, groundwater systems act to attenuate and add lags to meteorological drought signals, thus mitigating the effects of relatively short monthly to sub-seasonal droughts (Tallaksen and Van Lanen, 2004). However, groundwater resources are highly susceptible to prolonged (multi-seasonal to multi-annual) episodes of drought, where pooling, lagging and lengthening of multiple consecutive sub-seasonal or sub-annual drought periods can lead to extended drought events in groundwater systems (Van Loon and Van Lanen, 2012; Van Loon, 2015). The efficacy of these pooling, lagging and lengthening effects depends on catchment and aquifer characteristics (Van Lanen et al., 2013) and may be impacted by anthropogenic influences such as groundwater abstraction (Bloomfield and Marchant, 2013; Bloomfield et al., 2015; Tallaksen et al., 1997; Tallaksen and Van Lanen, 2004; Van Loon and Laaha, 2015; Van Loon et al., 2016).

The impacts of groundwater drought are wide-ranging and have important consequences for society and ecosystems. For example, public water supply, energy production and industry, agriculture and livestock farming, freshwater and terrestrial ecosystems, tourism and recreation may all be adversely impacted by major droughts (Lange et al., 2017; Stahl et al., 2016). Specifically, a reduction in deployable output from boreholes due to lower than normal groundwater levels during major droughts may also lead to costly restrictions to public supply, or to reduced abstractions for agriculture and industry, thus causing economic strain to society (Van Loon, 2015). Good examples of this are the last major groundwater drought in the UK in 2010–2012 when seven water companies in south and east England had to impose temporary use bans (colloquially known as “hosepipe bans”) on about 20 million people, including the farming sector in spring of 2012, and the 2015 drought in Europe with dried up wells and boreholes leading to issues providing drinking water for cattle (e.g. in Germany; Van Lanen et al., 2016).

Improved understanding of drought propagation in similar aquifers (such as the Chalk of the Anglo-Paris Basin) and in a wider range of hydrogeological settings across Europe would provide insights into generic groundwater drought processes, opening the possibility for a systematic analysis of drought propagation from meteorological, soil moisture and surface water to groundwater droughts at the continental scale (Van Lanen et al., 2016), and thus contribute to improved preparedness and more informed decisions associated with the use of drought triggers.

During the 2015 event, the EURO FRIEND-Water LFD-network collated and analysed data to characterise meteorological and streamflow drought conditions, and impacts across Europe (Ionita et al., 2017; Laaha et al., 2017; Van Lanen et al., 2016). However, an equivalent data synthesis and analysis of the 2015 groundwater drought situation was not possible due to the lack of co-ordination of groundwater level data and research across Europe (Laaha et al., 2017; Van Lanen et al., 2016; Van Loon et al., 2017). This highlights a major gap in European drought research capability. In addition, there have recently been initiatives to better inform policy makers, water managers, and the general public of drought status, for example in Europe through on-line monitoring services such as the European Drought Observatory (EDO, https://edo.jrc.ec.europa.eu/edov2/php/index.php?id=1000, last access: 28 July 2020) and the European Drought Centre (EDC, http://europeandroughtcentre.com/, last access: 28 July 2020). The EDO is mapping droughts and drought indicators but does not include groundwater droughts. EDC hosts the European Drought Reference (EDR) database, a collection of data on historic European meteorological and streamflow droughts, and a web-based viewer of meteorological drought status. EDC also hosts the European Drought Impact Report Inventory (EDII), an account of documented impacts of droughts. However, the EDII does not distinguish impacts of groundwater droughts and up to date, there is no attempt to establish a stand-alone database for the groundwater domain.

In summary, despite some early attempts to highlight the importance of research into groundwater droughts across Europe (Hisdal et al., 2004; Peters et al., 2001; Van Lanen and Peters, 2000) recent projects and initiatives mentioned above lack a systematic approach of clearly integrating groundwater droughts into their drought analysis. With few exceptions (Van Loon et al., 2013), they focus instead on droughts in the driving meteorology and consequences of droughts only in the soil and surface water compartments of the terrestrial water cycle. Therefore, the following research gaps can been identified: (i) although the skills are present across Europe, there is a major gap in European drought research capability related to groundwater droughts, primarily due to a lack of co-ordination between hydrogeologists and hydrologists with appropriate expertise (Staudinger et al., 2019); (ii) there is a gap in research capability to provide a timely, integrated and consistent overview of groundwater level data at the continental scale under conditions of drought, and hence to assess the status of groundwater drought in a manner comparable with other hydrometeorological drought signals; and, (iii) there has to date been no systematic attempt to undertake an assessment of the impacts of groundwater droughts at the continental scale.

Sixteen sites from four countries (Austria, France, UK, and Sweden) have been chosen for analysis and to illustrate the GDI workflow. The main selection criteria for these sites were the availability of relatively long, mostly continuous series of groundwater level data from unconfined aquifers and the representation of different hydrogeological settings. For a brief comparison between aquifer settings, we additionally chose two sites from confined aquifers (SW3 and SW4). The spatial distribution of the chosen sites, Fig. 1, covers different parts of France and Austria (though with one pair of neighbouring boreholes respectively, namely FR1 and FR2 in France, and AU3 and AU4 in Austria), and the southern parts of the UK and Sweden. The aquifer types differ across the sites (Table 1), and include fissured igneous and metamorphic rocks, fractured limestone aquifers such as the Chalk of the Anglo-Paris basin, consolidated sandstones and superficial glacio-fluvial and alluvium deposits. All groundwater level data were obtained from publicly available databases of groundwater levels for each country respectively (Austria: https://ehyd.gv.at/, last access: 28 July 2020, France: https://ades.eaufrance.fr/Recherche/Index/Piezometre?g=d5acb7 (last access: 28 July 2020), Sweden: https://www.sgu.se/produkter/geologiska-data/oppna-data/grundvatten-oppna-data/grundvattennivaer-tidsserier/ (last access: 28 July 2020), UK: https://www.bgs.ac.uk/products/hydrogeology/WellMaster.html, last access: 28 July 2020).

Figure 2 shows the overall workflow of the GDI. It is based on the analysis method developed by Marchant and Bloomfield (2018) that enables irregularly observed groundwater level data from a range of locations and variety of periods of observations to first be regularised into a common analysis frame, then standardised, and finally then to be used for characterisation of spatio-temporal variation in extracted episodes of groundwater drought. The raw time series of groundwater level data for each site is first pre-processed to provide a regularised monthly groundwater level time series using newly developed R-scripts run on a SHINY-interface. The regularisation step entails harmonisation of the observed groundwater levels to regular monthly time steps and stochastic modelling of groundwater levels (Marchant et al., 2016) to infill data gaps using the temporal correlation amongst the data and a modelled relationship with monthly precipitation data, obtained from the E-OBS gridded dataset (Haylock et al., 2008). This is achieved by estimating and predicting from a linear mixed model of the groundwater levels where the fixed effects are an impulse response function of rainfall plus a seasonal term. Random effects are temporally correlated according to a Matérn function – see Marchant and Bloomfield (2018) for a fuller description.

A variant of the Standardised Precipitation Index (SPI), the Standardised Groundwater level Index (SGI) approach (Bloomfield and Marchant, 2013; McKee et al., 1993; Svoboda and Fuchs, 2016) is then used to standardise the groundwater level data. The SGI is a non-parametric approach, which assigns a standardised value to the reconstructed groundwater levels for each month, thereby resulting in normalised monthly indices that allow for comparison across the sites (see Bloomfield and Marchant, 2013). Additionally, cluster analysis using the k-means approach is undertaken to obtain information about similarities between hydrographs across the study area. Due to the limited number of sites that have been modelled and analysed in the present study, only two clusters have been identified to illustrate the approach. As a more extensive set of hydrographs becomes available, the number of clusters used may be increased to help identify spatially coherent groundwater hydrographs and hence coherent groundwater drought behaviour. Simple statistics are applied to each site for comparison of drought frequency and duration across the sixteen sites. Drought is thereby defined as intervals with SGI < 0, in which there is at least one occurrence of SGI ≤ -1, following the convention of the World Meteorological Organization (Svoboda et al., 2012).

Data availability for the selected sites varied and ranged between near-complete time series from 1955–2018 (UK1) to datasets with observations starting only in the 1990s (AU3, AU4, UK3). An example of the driving precipitation data, the observed groundwater levels, and the regularised groundwater level time series for three of the sites is shown in Fig. 3, demonstrating different degrees of flashiness (response of the aquifer to precipitation events) across the sites.

Table 1 shows cross-correlation (CC) between the estimated SGI for each of the sixteen sites and the corresponding SPI for the period 1955 to 2018. There was a reasonably strong correlation (>0.5) between the SGI time series and the corresponding SPI series accumulated over an appropriate number of months for all sites except SW1. These accumulation periods (qmax), which maximise the correlation, were determined for each site and ranged 2–33 months. The CC can potentially be improved by selecting a shorter time period with a later onset, and by choosing sites with more homogenously distributed observations across time, thereby avoiding decadal gaps in the data. Table 1 also shows the maximum autocorrelation range (mmax) in the estimated SGI, a measure of the “memory” of groundwater levels at each site, ranging 5–28 months.

The SGI time series for the sixteen sites (Fig. 4a) demonstrate the different response patterns to driving meteorology. Some of the sites are characterised by fewer, but longer droughts (e.g. SW1, SW2, UK2) and others exhibit more frequent, but shorter drought periods (e.g. AU1, UK3). This is also reflected in differences in mmax and qmax, which are lower for the sites with frequent, short droughts, and higher for the sites with less frequent, longer droughts (Table 1).

The k-means clustering in two groups (Fig. 5) highlights the divide between slow- and fast-responding systems (Fig. 4b), Cluster 1 and Cluster 2 respectively. Even though only a limited number of sites are included in this analysis, there appears to be some spatial coherence of the hydrographs based on the two clusters. For example, there appears to be spatial coherence among the Austrian sites. However, the identification of spatial coherence in the clusters needs to be tested with the inclusion of more hydrographs in future analyses. In addition to the clustering capturing systematic differences in the responsiveness or flashiness of the hydrographs, the clusters also appear to reflect a systematic difference in the degree to which sites in each cluster were impacted by the drought in the early 1990s, with Cluster 1 being highly impacted and Cluster 2 showing no significant impact of drought. This can be seen both in the individual SGI hydrographs (Fig. 4a) and in the differences in the mean hydrographs of each cluster (Fig. 4b). In contrast, a major drought episode in the summer of 1976 is present in both clusters and most of the sites.

Drought statistics related to drought frequency, duration and intensity can be extracted from the prepared SGI series. For illustration, Table 2 summarises drought frequency and duration for each site. It shows a total number of droughts per site between as few as four 4 (SW1) to as many as 40 (UK3). It is noteworthy that the median drought duration differs substantially between aquifer types, ranging for example from only 6–7 months at limestone (FR3 and UK3) and gravel/sand sites (AU1 and AU2), to 11–22 at Chalk sites (FR1, FR2, UK1, and UK4), and 24–64 months in glaciofluvial sand aquifers (SW1, SW2); thereby confirming the importance of aquifer type in drought propagation as previously observed in other studies (Bloomfield and Marchant, 2013; Bloomfield et al., 2015; Tallaksen et al., 1997; Tallaksen and Van Lanen, 2004; Van Lanen et al. 2013; Van Loon and Laaha, 2015; Van Loon et al., 2016).