The analysis has been undertaken for a 30 year period (1984–2014) using precipitation, evapotranspiration, and groundwater level time series. This time period includes at least four major droughts with national spatial extent, namely 1988–1994, 1995–1997, 2003–2006, and 2010–2012 .

Precipitation and potential evapotranspiration data were obtained from the GEAR data set and the CHESS data set . The gridded (1 km2) GEAR data set contains interpolated monthly precipitation estimates derived from the UK rain gauge network. The CHESS data set is also gridded (1 km2) and contains climate data from which potential evapotranspiration estimates are computed using the Penman–Monteith equation. We aggregated daily potential evapotranspiration estimates to monthly sums. For both gridded data sets (GEAR and CHESS), grid cells were extracted corresponding to groundwater well locations. The 1 km2 gridded precipitation and potential evapotranspiration sums were compared to monthly groundwater observations of the same location. This point-scale comparison relies on the assumption that the influence of precipitation is largest surrounding a groundwater monitoring site .

Precipitation estimates were converted into standardised precipitation indices (SPIs) following the method of . A gamma distribution was fitted to precipitation estimates, but alternative distributions were also tested (normal, Pearson III, and logistic) . Considering the use of SPIs to account for delayed recharge, a large range of accumulation periods of precipitation (1 to 100 months) was calculated in order to find the optimal correlations between precipitation and groundwater time series. For this particular use of the SPI, the “best” fitting distribution varies . Alternative distributions showed minimal differences from the gamma distribution in the computed correlations between standardised precipitation and groundwater time series; hence, we decided to use the gamma distribution.

Groundwater level time series were obtained from the national groundwater database in the UK, which contains time series for both reference wells and regular monitoring wells. A total of 209 wells (or sites) have been included in the analysis, of which 39 are reference sites and 170 regular monitoring sites. Reference sites were taken to represent near-natural conditions in the 30 year time period. These sites were selected from the index and observation wells listed in the UK Hydrometric Register and have previously been assessed by the British Geological Survey. Well descriptions indicate near-natural conditions or possible (intermittent) influence of groundwater abstraction. Wells selected for this study are categorised as near natural, reflecting regional variation in groundwater levels with minimal abstraction impacts. This selection of reference wells includes 30 wells in the chalk and nine wells in the Permo–Triassic sandstone. Regular monitoring sites are part of the monitoring network in the four water management units. Initially, 660 monitoring sites were considered for the regional groundwater drought analysis that was truncated to the 30 year analysis period and quality checked. Unrealistic observations were cross-validated with available meta-data and, if unexplained, removed from the data set. Missing data were linearly interpolated from the last observation to the next observation in case of short sequences of missing data (less than 6 months) . Sites with records containing longer sequences of missing data were removed from the data set prior to the analysis, leaving a total of 170 (out of the original 660) groundwater level time series that were deemed of good quality, of which 38 were located in Lincolnshire, 45 in Chilterns, 36 in Midlands, and 51 in Shropshire.

All groundwater level time series were standardised into the Standardised Groundwater level Index (SGI) , which is briefly explained here. Monthly groundwater observations were grouped for each calendar month, and within each group observations were ranked and assigned a SGI value based on an inverse normal cumulative distribution of the data. No distribution was fitted, but SGI values were assigned to monthly observations, accounting for seasonal variation within the calendar year. The resulting SGI time series represent extremely low to below-normal (-3 < SGI < 0) and above normal to extremely high (0 > SGI > 3) monthly groundwater levels in the groundwater time series. Groundwater level observations are physically constrained by the length of the screened interval of the borehole. Therefore, the lowest SGI value might indicate that groundwater levels fell below the borehole screen, and the highest SGI value can indicate that groundwater levels reached the surface.

Qualitative information about groundwater use was provided for each water management unit by the national regulator (Environment Agency (EA) in England). Detailed maps were made available regarding the purpose and recent (dated at 2015) licensed abstraction volumes (see Fig. S1 in the Supplement). In addition, reports describing the EA's regional groundwater resource models and location-specific groundwater studies were used as reference material to indicate changes in groundwater use (Table ).

The developed methodological framework consists of two approaches for investigating the impact of groundwater use on groundwater droughts. The first approach uses a regional near-natural groundwater drought reference based on reference sites. SGI time series of reference sites are clustered to identify common spatial and temporal patterns in near-natural groundwater levels in the two aquifers. Reference sites are thereby taken to represent regional groundwater variation that is primarily driven by climate and hydrogeology. Then, monitoring wells in each of the four water management units were paired to these regionally coincident clusters of reference wells (Fig. ), and human-influenced sites are identified using the correlation between SPI and SGI. Drought occurrence, duration, and magnitude in monitoring wells were compared with those in paired reference clusters to assess the potential effects of abstraction on groundwater droughts. The second approach consisted of a groundwater trend test that quantified long-term trends as a consequence of the continuous impact of groundwater use. The spatial distribution of identified trends was evaluated according to the location of annual abstraction licenses, changes in water use, and hydrogeological features in the water management units.

The near-natural groundwater reference clusters, based on SGI clusters of the reference wells and the clustering criteria, were defined by Ward's minimum clustering technique, showing the least overlap between clusters of the three tested clustering techniques (Fig. S3). A total of eight clusters are identified, of which five clusters are located in the chalk (C1–5) and three in the Permo–Triassic sandstone (S1–3; Fig. ). The spatial distribution of chalk clusters (C1, C3, and C4) is consistent with clusters identified by . Two additional clusters are identified, of which one is located in East Anglia (five reference wells in C2) and one in southeast England (two wells in C5). The cluster dendrogram shows a small difference in the similarity between C4 and C5, which is located close to the coastline (cluster dendrogram result not shown; difference between C4 and C5 is shown in Fig. S3). C1 and C3 are coincident with water management unit 1 and 2, respectively, and are used as near-natural reference for monitoring sites in those units. In the Permo–Triassic sandstone aquifer, only one spatially coherent cluster (S2) is found when all nine SGI time series are clustered (Fig. ). The cluster composition of the other two smaller clusters (S1 and S3) is not spatially coherent, and there is no evidence of previous clustering studies available that can confirm these two clusters. Hence, only S2 is used as near-natural reference for monitoring sites in water management units 3 and 4.

The optimal SPIQ–SGI correlations of near-natural wells are high on average (0.79), with a range of 0.66 to 0.89. These correlations are found using the optimal accumulation period, which accounts for a delay in recharge that is different for each reference cluster. High SPIQ–SGI correlations are found for both short and long accumulation periods, and there was no systematic relationship between the SPIQ–SGI correlation and the SPI accumulation period Q or SGI autocorrelation in the near-natural wells. C1 represents a relatively fast-responding section of the chalk and has a short Q of 12.6 ± 5.4 months. The Q of C2 and C3 is higher, respectively 18.2 ± 4.3 and 24 ± 8.6 months. This corresponds to the delay in groundwater recharge due to the Quaternary deposits present in these regions . In the southeast, the chalk is highly fractured, which is reflected by a short Q of 8 ± 2.2 months for C4 and C5. In the Permo–Triassic sandstone, the Q of S2 is 35 ± 4.5 months, which confirms a slow-responding groundwater system .

In the monitoring sites, the majority of the SPIQ–SGI correlations are as high as or higher than the minimum correlation of paired reference clusters. Hence, these monitoring sites are considered, on average, uninfluenced by abstraction. The percentage of uninfluenced sites varies between the water management units. The largest percentage is found in the Chilterns (71 %), followed by the Midlands (63 %), Shropshire (53 %), and Lincolnshire (31 %). Monitoring sites with a SPIQ–SGI correlation below the minimum correlation of the paired reference cluster are treated as, on average, influenced by abstraction.

The found optimal precipitation accumulation periods within the management units is variable and appears to be in part a function of aquifer depth and the local nature of aquifer confinement (Fig. S4). For example, shorter accumulation periods are found in shallow sections of the aquifer (east Shropshire and west Chilterns), and in outcrops (east Lincolnshire). Longer accumulation periods are found in deep sections of the Permo–Triassic aquifer (west Shropshire) and semi-confined sections of the Permo–Triassic (Midlands) and chalk aquifer (east Chilterns and southeast Lincolnshire).

Groundwater droughts observed in the reference clusters reflect both spatial and temporal variation due driving precipitation and hydrogeology setting. In general, the four UK-wide droughts (1988–1993, 1995–1998, 2003–2006, and 2010–2012) are reflected in near-natural groundwater time series. Spatial patterns in driving precipitation, however, result in variable groundwater drought occurrence (Fig. ). For example, in C1, groundwater levels are low in 2003–2006 but not below the drought threshold. In C2, groundwater levels are slightly lower, and a short drought event is observed in the SGI cluster mean. In C3–C5 and S2, however, the 2003–2006 drought event was a major drought event. Spatial variation in the hydrogeology also results in varying drought duration for the chalk clusters. In central England, longer drought durations are found in clusters C2 and C3. This region is partly covered by Quaternary deposits that delays recharge. Shorter (and more frequent) events are observed in C4 and C5, which are located in highly fractured chalk.

On a smaller scale in the water management units, average drought characteristics (duration in months, magnitude in accumulated SGI over the drought period, and frequency) for monitoring sites show differences due to abstraction influence, which we have classified in, on average, uninfluenced and influenced sites (Table ). Shorter and less intense, but more frequent, drought events are observed in the influenced sites in Lincolnshire, Chilterns, and Shropshire. In these water management units, the difference in average drought duration and frequency between uninfluenced and influenced sites is significant. Droughts are observed twice as often in influenced compared to uninfluenced sites in Lincolnshire and Chilterns, although a smaller difference is found in Shropshire. The distribution of recorded drought frequency (Fig. S5) shows that the difference between, on average, uninfluenced and influenced sites is actually less pronounced in Lincolnshire and Shropshire. In the Midlands, the average drought duration of influenced sites exceeds the duration in uninfluenced sites. Longer and more intense groundwater droughts occurred less often in influenced sites, which is in contrast with the other water management units. The distribution of recorded drought frequency (Fig. S5) shows a majority of sites recording fewer droughts and some sites that record a higher frequency. On average, this results in a small difference between the influenced and uninfluenced sites.

Presented drought characteristics in Table  suggest that drought events vary significantly within and between water management units. These different drought events are shown in a combined time series plot (Fig. ) capturing reference droughts and droughts recorded in monitoring sites showing drought occurrence, duration, and magnitude. Monitoring sites are sorted based on their SPIQ–SGI correlation (high to low). The cluster minimum SPIQ–SGI correlation is indicated with a dashed line, i.e. 0.75 for Lincolnshire, 0.71 in the Chilterns, and 0.69 in the Midlands and Shropshire. Below this minimum correlation, drought occurrence in uninfluenced sites aligns mostly with that of droughts in the reference clusters. Observed droughts in influenced sites (i.e. those with SPIQ–SGI correlations lower than the cluster minimum) are typically shorter, but drought events are of a lower magnitude in Lincolnshire, Chilterns, and Shropshire. The distribution of drought duration in Fig. S6 shows that the majority of these additional droughts is recorded in influenced sites compared to uninfluenced sites in Lincolnshire, Chilterns, and Shropshire (drought deficit distribution is shown in Fig. S7). Contrastingly, longer and more intense droughts are observed in all Midland sites in 1990–1995. Droughts observed in influenced sites are also longer in 1984–1986, 1997–2001, and 2005–2006 compared to the reference cluster and fewer droughts are observed in 2010–2012.

The additional events in influenced sites coincide with low SGI values in the reference wells that sometimes occur prior to a long drought event. For example, additional droughts are observed in 1984, 1995–1996, 2005–2006, and 2014 in Lincolnshire and in 1984–1986, 2004, and 2009–2010 in the Chilterns. In those periods, the reference cluster mean was below 0 but not below the drought threshold. In the case of 1995–1996, 2004, and 2009–2010, these additional drought events occurred prior to a long drought event. However, there was no consistent evidence found among the study areas in relation to the timing of these shorter drought events. In Lincolnshire, minor droughts occur more often during reference droughts compared to Chilterns and Shropshire, where more droughts are detected prior to reference droughts (Table S8). All minor droughts are shorter than the groundwater auto-correlation, suggesting that these minor droughts are less likely to be related to propagated precipitation deficits and more likely to be related to groundwater abstraction.

Drought descriptions in the literature show an increase in water demand during the 1995–1997, 2003–2006, and 2010–2012 drought . For example, found that during the 1988–1993 drought event evapotranspiration was exceptionally high, and groundwater use increased. Impacts were mostly felt in the chalk, particularly in regions where groundwater is the principal source of water supply. An extreme rise in water use was also found during the 1995–1997 drought event, putting strain on the drinking water supply systems in northeast England . Sections of the Permo–Triassic sandstone were amongst the worst affected, prolonging drought conditions until 1998 . During the 2003–2006 and 2010–2012 droughts, a sudden increase in groundwater use was found that was attributed to dry weather and hot summers . reported low SPI3 values in the summer months for 1995, 1996, 2003–2006, and 2010/2011, highlighting exceptionally dry weather that led to surface water use restrictions prior to droughts to maintain low flows. Consequently, reduced surface water abstractions were replaced by groundwater for which use was rarely restricted , resulting in lowered groundwater levels that could also potentially aggravate groundwater droughts.

Over the whole investigation period, drought magnitude seems to be decreasing since the 1995–1997 drought event. Droughts observed in 2003–2006 and 2010–2012 are shorter and of lower magnitude than the 1995–1997 drought in most sites. This is seen most convincingly in Lincolnshire, Chilterns, and the Midlands, where the magnitude of droughts decreases dramatically over the 30 year time period. In Shropshire, this tendency is less strong, as the 2010–2012 drought was of a similar magnitude as the 1995–1997 drought.

Significant trends in groundwater level have been detected in 38 % of all monitoring sites in the water management units. Of these 38 %, half of the trends are upward (positive) and the other half is downward (negative) trends (Fig. ). Overall, upward trends are dominating (61 % of sites including significant and non-significant trends), indicating a sustained rise in the 30 year groundwater level time series. Fewer (39 % including significant and non-significant) downward trends are detected, indicating sustained lowering of groundwater levels. The presence of these significant trends in groundwater is notable given the weak, non-significant, range of trend Z values in the 30 year precipitation and potential evapotranspiration data (P: Z=-0.75–1.53; PET: Z=0–0.65).

The direction and spatial coherence of trends in groundwater show different patterns within the water management units (Fig. ). In the chalk water management units, positive trends dominate. In Lincolnshire, five out of the total of 25 positive trends are significant, compared to three out of 32 in Chilterns. There are fewer negative trends detected in both water management units, but more of these are significant with, respectively, seven out of 13 in Lincolnshire and four out of 12 in Chilterns. In Lincolnshire, sites with a negative trend are, for all but one, located in the semi-confined chalk. This is in sharp contrast with the semi-confined chalk in Chilterns, where mainly (significant) positive trends are found. In the Permo–Triassic sandstone water management units, more significant trends are detected compared to the chalk (63 % in the Midlands and 43 % in Shropshire). In the Midlands, more positive than negative trends are detected. In total, 17 out of 25 positive trends are significant, compared to six out of 11 significant negative trends. Negative trends are mainly found in the centre of the water management unit. Positive trends are found north and south of that. In Shropshire, more negative than positive trends are detected. A total of 31 sites have a negative trend, of which 15 are significant. These trends are mainly detected in the west of the water management unit. Positive trends are mainly located in the east in between two fault lines (Ollerton and Childs Ercall fault; ). A total of seven of these positive trends (from 20 in total) are significant. In Fig. , the maximum licensed abstraction volumes are also shown. These licenses show in which aquifer sections groundwater is primarily abstracted. However, without a record of the actual use of these licenses or the change of licensed abstractions over time, it is impossible to directly relate detected trends to these abstraction locations.