A range of hydro-meteorological data sets have been used to identify multi-annual droughts through the historical record. For rainfall, the key data set is a monthly 5 km × 5 km resolution gridded data set for the UK from 1910 to date, assembled using the methods of Perry and Hollis (2005a). This gridded data set is based on interpolated rain-gauge observations taking into account factors such as topography. It forms the basis of UK rainfall statistics produced by the UK Met Office National Climate Information Centre (NCIC). We term this data set “NCIC Rainfall”.

The station network comprises between 200 and 500 stations covering the UK from 1910 to 1960, a step increase to over 4000 for the 1960s and 1970s before a gradual decline to around 2500 stations by 2012. Despite the lower network density from 1910 to 1960, these data are still able to identify earlier historical droughts with considerable confidence. Long-term-average (LTA) values were obtained from a monthly 1 km × 1 km resolution LTA gridded data set for the period 1961–1990 (Perry and Hollis, 2005b).

River flow and groundwater level data were taken from the UK National River Flow Archive (NRFA) and National Groundwater Level Archive (NGLA). An NRFA regional river flow data set for the English Lowlands is available to characterise total outflows from the region from 1961 to 2012 (Marsh et al., 2015). The series is based on aggregated flows from large rivers and uses hydrological modelling to account for ungauged areas. The boundary shown in Fig. 1 was used to create the “English Lowlands” NCIC rainfall and NRFA regional river flow series used here. A regional groundwater level series was also created for the English Lowlands to directly compare with the English Lowlands rainfall and river flow series – further information on the derivation of the groundwater level series is provided in Sect. 2.3.

In addition to the regional English Lowlands outflow series, the flow record of the Thames at Kingston, the longest in the NRFA, from 1883 to the present, was used to provide a temporal coverage comparable with that of the NCIC rainfall. The river Thames has the largest catchment in the UK (9968 km2 at the Kingston gauging station) and constitutes 15 % of the English Lowlands study area. This series has been naturalised; i.e. the flows have been adjusted to take account of the major abstractions upstream of the gauging station. It should be noted that the homogeneity of the low flow record is compromised by changes in hydrometric performance over time (Hannaford and Marsh, 2006), although this is not likely to be unduly influential for the present study, which focuses on drought indicators rather than trends over time. The longest Chalk groundwater-level record (starting 1932) from the Thames catchment, the Rockley borehole series, is also used to provide a long-term picture.

Meteorological droughts are identified from monthly rainfall deficits, calculated as the monthly observed areal mean rainfall total minus the monthly 1961–1990 LTA. These deficits were accumulated over rolling multi-month time periods from 12 to 24 months long. All rainfall deficits over 170 mm (25 % of annual average rainfall) over 12 to 24 month timescales were selected to give 15 notable droughts from 1910 to 2012 lasting at least 1 year and encompassing at least one winter, i.e. likely to have a significant impact on groundwater resources. These droughts did not necessarily have below-average rainfall in all months from October to March; in some instances rainfall may also have been low during the summer half-year (April–September). Table 1 shows that two droughts just exceeded 24 months in length using this method. Figure 2 shows an example rainfall anomaly series, that for the 2010–2012 drought, which includes a few months before and after the chosen drought period to demonstrate a typical example of how drought beginning and end dates were chosen.

Meteorological droughts across the English Lowlands since 1910 identified here include 1920–1921, 1933–1934, 1975–1976, 1990–1992 and 1995–1997, consistent with earlier studies (Marsh et al., 2007), so their identification is not very sensitive to the criteria used. Of these, the 1975–1976 drought is generally regarded as a benchmark across much of England and Wales against which all other droughts are often compared (Rodda and Marsh, 2011). During only this and the 1920–1921 drought were rainfall totals below 65 % of LTA over the > 12 month timescale, including all or most of a winter half-year (Table 1). The most recent historical drought of 2010 to 2012 comfortably sits as one of the most significant prolonged droughts since 1910 (Kendon et al., 2013).

We also examined how spatially coherent on average these 20 major long droughts were over the UK. There is a well-known strong rainfall gradient between the English Lowlands and north-western Britain (an order of magnitude between the wettest parts of the Scottish Highlands and the driest parts of East Anglia). Because of the predominance of westerly airflows interacting with western uplands, eastern lowland areas are often in rain shadow. Accordingly, periods of very wet or very dry conditions in the lowlands often differ from those in the north-western UK. The atmospheric drivers of lowland UK droughts are therefore likely to be somewhat different to those in the north-west. To demonstrate this, Fig. 3 shows correlations between rainfall in the ten climatological rainfall districts covering the UK defined by the UK Met Office and gridded NCIC rainfall data elsewhere in the UK for both winter and summer half-years using the 15 long drought periods listed in Table 1. Although summer is not a focus of the paper, Fig. 3 shows considerable differences between winter and summer correlation patterns. Generally, there is a greater anticorrelation between south-eastern UK and north-western UK rainfall in the winter half-year than in the summer half-year. This implies that droughts have a greater tendency to affect the UK as a whole in the summer half-year than in the winter half-year. Indeed, Fig. 3 suggests that north-western Scotland is unlikely to be affected by drought at the same time as south-eastern England in the winter half-year. Rahiz and New (2012) have also recently confirmed a tendency for spatially coherent meteorological droughts in the south-east of England to be distinct in time from droughts in northern and western areas of UK.

In order to examine the impact of historical meteorological droughts on river flows and groundwater, consistent indicators are required to identify such drought events. A wide range of drought indicators are available (e.g. Mishra and Singh, 2010) and there is no current consensus on a single indicator appropriate for capturing the wide range of drought impacts. The Standardized Precipitation Index (SPI, McKee et al., 1993) benefits from being normalised to allow comparisons between diverse regions and through the annual cycle. The formulation of the SPI is described in detail elsewhere; in summary, it consists of a normalised index obtained by fitting a gamma or other appropriate distribution to the precipitation record, where fitting is done for each calendar month to account for seasonal differences. The monthly fitted distributions are transformed to a standard normal distribution and the estimated standardised values combined to produce the SPI time series. The index is fitted to precipitation data that are typically accumulated over 3, 6, 12 and 24 month periods. The SPI concept has been extended to river flows (e.g. Shukla and Wood, 2008), but numerous variants have been proposed and there is no consensus on the distributions that should be used for normalisation (e.g. Vicente-Serrano et al., 2012). More recently, the SPI concept has been extended to groundwater level records via a Standardized Groundwater level Index, SGI (Bloomfield and Marchant, 2013). This adopts a non-parametric normal score transformation rather than using a defined statistical distribution.

For the present study, the SPI has been applied to the English Lowlands rainfall series, and the SGI has been applied to 11 individual groundwater-level records from observation boreholes within the English Lowlands region. These are Ashton Farm, Chilgrove House, Dalton Holme, Little Bucket Farm, Lower Barn, New Red Lion, Rockley, Stonor House, Therfield Rectory, Well House Inn and West Dean (see Bloomfield and Marchant (2013) for more information on these groundwater records). The groundwater hydrographs have been averaged to create a regional SGI series of English Lowlands groundwater levels. Unlike the SPI, the SGI is not applied to time series that have to be accumulated over a range of durations, because groundwater levels exhibit autocorrelation or “memory”, which implies that a degree of accumulation is inherent in each monthly value. The same methodology was also applied to the English Lowlands regional river flow series (henceforth referred to as the Standardized Streamflow Index, SSI). Whilst the SGI was developed primarily for groundwater, its formulation is also highly appropriate for river flows – particularly in the English Lowlands, where a substantial proportion of the runoff comes directly from stored groundwater. As with groundwater levels, monthly river flows were not accumulated over a range of periods to produce the SSI for river flow.

Standardised indices were calculated for English Lowlands regional river flow and regional groundwater levels, and monthly SPI was calculated for all accumulation periods from months 1 to 24 (i.e. SPI1 to SPI24). Figure 4a shows a heatmap of the correlation between lagged English Lowlands river flow (as SSI) and English Lowlands precipitation (as SPI1 to SPI24). The maximum correlation of 0.79 occurs for lag zero between the two time series and for a precipitation accumulation period of 3 months. Figure 4b is a similar heatmap of lagged English Lowlands mean groundwater levels (as SGI) and English Lowlands precipitation (as SPI1 to SPI24). The maximum correlation is 0.82, also for lag zero, but only for a longer precipitation accumulation period of 12 months. In summary, the highest correlations between SSI and SPI and between SGI and SPI are associated with concurrent time series, although correlations > 0.75 between SGI and SPI are also seen at lags of a few months.

Figure 5 shows, for the English Lowlands, SPI rainfall series for several accumulation periods and the corresponding SSI and SGI river flow and groundwater series. Figure 6 shows the English Lowlands rainfall (SPI) series and equivalent series for the long Thames (SSI) record, and the Rockley borehole (SGI). Both figures demonstrate good agreement between the meteorological droughts and associated river flow and groundwater droughts – with some expected delays for the onset of given hydrological drought events, demonstrating the propagation between the meteorological and groundwater droughts in particular. Figure 6 also shows very good agreement between the severity of the major rainfall droughts identified independently in Sect. 2.2, suggesting that these long-duration events indeed had an identifiable and considerable impact on river flows and groundwater in the English Lowlands. However, a cluster of hydrological drought events in the mid-1950s, not identified in Sect. 2.2., is also apparent in Fig. 6. The magnitudes of the SPI/SGI/SSI anomalies in this period are not as great, but the duration is notable. Overall, these analyses demonstrate the strong link between meteorological droughts and their manifestation in hydro(geo)logical responses, but they also demonstrate some differences between the two, as expected. From this it is inferred that the major long meteorological droughts identified in Table 1, and the various hydrological drought metrics used to characterise them, provide a good basis for establishing links between potential climate drivers and the major historical droughts experienced in the English Lowlands. Nevertheless, links between the remote drivers of meteorological and groundwater hydrological droughts in particular are not expected to be identical, and the lag times identified above should be considered in interpreting these relationships.

Toniazzo and Scaife (2006) showed how El Niños (associated with significantly warmer than normal SST in the tropical eastern Pacific) affect winter, mainly January–March, extratropical Northern Hemisphere atmospheric circulation and temperature. The character and physical causes of the influences differ between moderate and strong El Niños (Ineson and Scaife, 2008). Moderate El Niños appear to influence winter extratropical Northern Hemisphere climate through a stratospheric mechanism, whereas very strong El Niños force a wave train through the troposphere from the tropics (Ineson and Scaife, 2008), giving very different patterns of winter atmospheric circulation response. Folland et al. (2012), in their Fig. 7b, show that the overall effect of El Niño on English Lowlands rainfall in December–February is towards modestly wetter than normal conditions, while La Niña (associated with significantly colder than normal SST in the tropical eastern Pacific) gives modestly drier conditions than normal conditions, consistent with the model results of Davies et al. (1997) and the observational results of Moron and Gouirand (2004). There is no evidence that strong La Niñas influence atmospheric circulation in different ways from moderate ones.

To investigate the influence of La Niña events, Fig. 7a first shows the mean global SST anomaly pattern associated with La Niña events where SST averaged over the Niño3.4 region (120–170∘ W, 5∘ N–5∘ S) has an anomaly <= -1.0 ∘C, compared to the 1961–1990 average. SST values averaging >= 1.0 ∘C above normal give a broadly opposite SST pattern. To provide dynamically consistent information about PMSL since the late nineteenth century, we use median results from the 20CR. This assimilates observed PMSL and surface temperature data into a physically consistent climate model framework every 6 h for most of the last 130 years using an ensemble of over 50 different, slightly differing analyses. Figure 7b, top panel, shows mean PMSL anomalies (from 1961 to 1990) for La Niñas where Niño3.4 region SST anomalies are < -0.92 ∘C for two independent epochs: 1876–1950 and 1951–2009. The value -0.92 ∘C is -1 standard deviation of Niño3.4 SSTs over 1951–2009. Both epochs show a finger of higher than normal PMSL stretching toward the southern UK, much stronger in the latter period, with lower than normal PMSL to the north. General similarities in the patterns tend to confirm the robustness of the PMSL pattern. PMSL anomalies project as expected onto the positive winter NAO in both epochs, but with higher PMSL over the south of the UK during La Niña than in the classical NAO pattern.

The central panel shows anomalies of atmospheric storminess from the NCEP reanalysis for 1951–2013 and western European rainfall anomalies for 1901–2011. These show significantly drier than average conditions and slightly reduced storminess over the English Lowlands during La Niña. The dry anomalies over the English Lowlands average around 5 mm month-1 (30 mm in the winter half-year), while north-western Scotland by contrast has significant slight to moderate wet anomalies exceeding 10 mm month-1. The average PMSL anomaly over the English Lowlands in 1951–2009 of 1.8 hPa in Fig. 7b corresponds to about a 38 mm rainfall deficit, 11 % of the 1961–1990 winter half-year average of 348 mm. The average effect is thus modest, as with all other individual climatic influences, though individual La Niña events can have a stronger influence. Details of the influence of La Niña on UK PMSL and rainfall vary through the winter half-year (e.g. Fereday et al., 2008), illustrated in Fig. S1 in the Supplement for each winter half-year month. Figure S1 shows no English Lowlands rainfall signal in January, though a dry signal appears to a greater or lesser extent in the remaining 5 months.

El Niño, by contrast, is associated with slightly wetter conditions than normal in the English Lowlands and slightly enhanced storminess (Fig. 7b, bottom right). Indeed, broadly opposite PMSL anomaly and rainfall anomaly patterns can be seen in the bottom panels of Fig. 7b in given locations over most of the UK and Europe during moderate El Niños (0.92 ∘C < Niño3.4 SST anomaly < 1.5 ∘C). For the relatively uncommon extreme El Niños, PMSL (Toniazzo and Scaife, 2006) and rainfall patterns change over the UK and English Lowlands (not shown).

Table 1 shows the mean winter half-year Niño3.4 SST anomaly during each drought. No moderate to strong El Niños occurred in these droughts, but there was one weak El Niño, four weak La Niñas (SST anomaly between 0.5 and 1 ∘C), seven “neutral” conditions (anomalies between ±0.5 ∘C, all here with weak negative SST anomalies) and three moderate to strong La Niñas. The mean winter half-year Niño3.4 SST anomaly in all 15 droughts is -0.45 ∘C. Table 2 looks at the problem in another way, showing the winter half-year rainfall anomaly associated with the strongest La Niñas and noting whether a Table 1 drought occurred. Many La Niñas are not associated with winter half-year components of Table 1 droughts. However, the probability of a Table 1 drought occurring during the top 20 winter half-year La Niñas is nominally 0.35, compared to a chance probability of 0.15, so the probability of a severe drought is approximately doubled compared to chance. The overall English Lowlands winter half-year rainfall anomaly during all top 20 Niño3.4 years is nevertheless weak at 25.2 mm or -0.39 standard deviations. So, a doubling of the chance probability is worth noting, but La Niña is inadequate for indicating a Table 1 drought with any confidence by itself. Moreover, La Niña winters can occasionally behave very far from expectation. The clearest example is 2000–2001, the wettest winter half-year in this record at 43 mm month-1, but accompanied by a weak La Niña with an SST anomaly of -0.70 ∘C. This very cyclonic winter may have been caused by the overriding influence of other strong forcings, especially in October–December (Blackburn and Hoskins, 2001).

Finally, Fig. 7c shows cumulative distributions of English Lowlands rainfall when Niño3.4 SST anomalies < -0.5 ∘C and Niño3.4 SST anomalies > 0.5 ∘C but < 1.5 ∘C were observed. The latter is an approximate lower Niño3.4 SST limit for extreme El Niños; these extreme years tend to be more anticyclonic over the English Lowlands, so on average drier than other El Niño years. Figure 7c shows drier conditions in La Niña compared to El Niño through almost all of the cumulative probability distribution of English Lowlands rainfall. A clear exception is the wettest winter half-year, 2000–2001. Including the three extreme El Niño years (not shown) slightly reduces the contrast between El Niño and La Niña influences.

In this section, we explore relationships between the various potential large-scale drivers identified in Sect. 3.2 and the hydrological drought indicators discussed in Sect. 2.

Figure 12 comprises box plots of the response variables, including winter half-year rainfall and river flow as well as the drought indicators (SPI, SSI and SGI) for low (< -0.5 SD) and high (> 0.5 SD) values of the predictors. This figure is intended to provide an overview of possible linkages between drought relevant hydro-climatic time series and the various climate drivers discussed in this study. The driving data include Niño3.4, the May SST tripole, the QBO, stratospheric volcanic aerosol loadings, TSI, and the AMO.

The data for the drivers and response variables in Fig. 12 are mostly averaged over October–March, so that the analysis is for concurrent data. However, the groundwater SGI is averaged with a lag of 2 months, and is thus shown for December–May, to reflect the temporal delay in groundwater formation. Because the SPI describes rainfall accumulated over a number of preceding months, these have also been lagged compared with the drivers so as to be centred on the target period October–March. Accordingly, the SPI3 is shifted forward by 1 month, and averaged for November–April; thus, the first 3 month accumulation starts in September and the last ends in April. Corresponding shifts for the SPI6 and SPI12 are 3 and 6 months respectively. The TSI precedes the hydrological response variable by 2 years to be consistent with the findings by Scaife et al. (2013) as discussed in Sect. 3.2.4. Significance levels are calculated using one-sided Welch two-sample t tests.

As perhaps expected, given the relationships discussed in Sect. 3.2, the majority of univariate relationships shown in Fig. 12 are very weak and non-significant, and the majority of individual drivers have little discernible impact on the means of the response variable. The only significant relationship for English Lowlands rainfall is with the Niño3.4 SST anomaly. Nevertheless, there is a clear tendency for El Niños (weak, moderate and strong) to be associated with wet conditions, and higher river flows and groundwater levels, and La Niña with dry conditions and lower flows and levels, consistent with Sect. 3.2 and Folland et al. (2012). As mentioned in Sect. 3.2, a strong note of caution, and a cause of the poor significance, is that the wettest winter half-year in Fig. 8c, 2000–2001, is associated with a weak La Niña and not with an El Niño. SPI3 shows a significant relationship with the SST tripole, which is only very weakly supported by the other variables. However, the spatial analysis shown in Fig. 8 (bottom panels) suggests that a stronger relationship exists for the upland north-west of the UK rather than the lowland south-east. Svensson and Prudhomme (2005) noted a positive concurrent winter (December–February) correlation between SSTs in the area corresponding to the centre of the SST tripole and river flows in north-western Britain (r = 0.36), consistent with Fig. 8b and d. For river flows in south-eastern Britain, encompassing the English Lowlands, they found a positive concurrent winter correlation with SSTs slightly further to the south (r = 0.43), partly overlapping the southernmost centre of the SST tripole.

For the majority of other potential climate drivers, the distributions of the drought indicators are typically not significantly different from one another for values > 0.5 or < -0.5 SD of the respective drivers. The key finding is that no single driver is close to compellingly explaining English Lowlands rainfall, river flows or groundwater levels. Combinations of drivers are of course difficult to test with the limited observational data available.

The predictability of winter droughts in the English Lowlands is a multiple forcing problem made more difficult by the relatively small scale of the English Lowlands compared to that of atmospheric anomalies. Temperature is a small additional factor in the winter half-year for drought but much more important in summer, when high rates of evapotranspiration can exacerbate hydrological drought. In winter, temperature could be influential in increasing the likelihood of snowfall as opposed to rainfall, which could confound links between the atmospheric drivers we have identified and precipitation, river flow and groundwater deficits. While water storage in snow/ice during the cold season can be a major influence on hydrological drought in parts of Europe (e.g. van Loon et al., 2015), generally, snowfall is limited in the English Lowlands. Some winter drought periods (e.g. 1962/1963, 2010/2011) were associated with major snowfall and persistent snow cover, but typically snow makes up a modest proportion of precipitation and is a minor runoff generation component (even in cold winters) on the monthly to seasonal scale.

Our work has focused on the winter half-year, but we acknowledge that a complete discussion of the multi-annual drought problem requires an investigation of the influences of remote drivers on summer half-year precipitation and temperature. Our current understanding of the drivers of atmospheric circulation in December–February over the UK and Europe has clearly improved, reflected in the new level of skill in dynamical forecasts of atmospheric circulation near the UK shown by Scaife et al. (2014a) mentioned in Sect. 3. Folland et al. (2012) point out that the magnitudes of the drivers we discuss in Sect. 3 can all be skilfully predicted in December–February winter or the winter half-year a season or more ahead. In other seasons, understanding is much less, and seasonal forecasting models commensurably much less skilful. However, the AMO is known to affect UK summer atmospheric circulation and rainfall (Folland et al., 2009; Sutton and Dong, 2012) as well as spring and autumn rainfall (Sutton and Dong, 2012), and is skilfully predictable a year or more ahead using persistence. Folland et al. (2009) also suggest an influence from strong La Niñas towards wetter than normal conditions in July and August. So, a major effort in studying drivers of predictability should be made for all seasons, particularly summer, when droughts can manifest themselves most severely. Whilst the winter season is most important for replenishment of water resources in the English Lowlands, intervening summers can be influential in dictating the outcomes of droughts – as was the case for the 2010–2012 drought, including its dramatic termination by the summer (Parry et al., 2013). In contrast, some of the most severe droughts have been associated with the combination of one or more dry winters with subsequent arid summers (e.g. in 1976, 1989). There is therefore a need to understand the drivers of both winter half-year and summer half-year deficiencies, and the likelihood of persistence between them in driving sequences of below-normal rainfall between seasons in long droughts. Folland et al. (2009) showed that, in summer, the summer NAO is the most prominent atmospheric circulation pattern and especially affects the English Lowlands. Its phase strongly modulates rainfall and temperature together such that both enhance drought or flood conditions. This is because high PMSL in summer, corresponding to the positive phase of the summer NAO, is associated with dry, sunny and warm conditions, while cyclonic conditions, associated with the negative phase, are associated with wet, dull and cooler conditions. Long droughts can also terminate at the end of summer dramatically, e.g. that of 1975–1976 (Folland, 1983).

Because many complex dynamical processes are involved, non-linear interactions may be important in creating the climatic outcome from a given combination of predictors. Only climate models can, in principle, represent these interactions, as observed data are too few for reliable non-linear statistical methods. Furthermore, the climate is in any case becoming increasingly non-stationary as global temperatures increase. It used to be thought that increasing greenhouse gases would most likely be associated with a slow tendency to an increasing positive, westerly phase of the winter NAO over the UK (e.g. Gillett et al., 2003). However, a recent tendency towards more negative winter Arctic and North Atlantic Oscillations casts doubt on this result (Hanna et al., 2014). Furthermore, ten dynamical models with high-resolution stratospheres suggest that increasing greenhouse gases may be associated with a tendency to more winter blocking over higher northern latitudes with perhaps some increased frequency of easterly winds over the northern UK in winter compared to the current climate (Scaife et al., 2012). The net effect on winter English Lowlands rainfall is by no means certain, though Scaife et al. (2012) find increased winter rainfall. In summer, there is more consensus that anticyclonic conditions may increase in the long term under increased greenhouse gases in the southern UK with decreased English Lowlands summer rainfall (e.g. Rowell and Jones, 2006; Folland et al., 2009). It is increasingly clear, though, that AMO fluctuations, which themselves may be influenced by anthropogenic forcing, may for decades reduce or hide this tendency or temporarily enhance it. However, Arctic sea ice reductions might affect long-term summer trends in hitherto unexpected ways (Belleflamme et al., 2013), and become an important influence in all seasons. Despite considerable uncertainty around changes in precipitation patterns, projections for future increases in temperature for the UK are more robust. The associated increases in evapotranspiration are likely to be a further factor increasing drought severity in future.

Recent developments in climate modelling (e.g. Hazeleger et al., 2010; Scaife et al., 2011; MacLachlan et al., 2014) provide the key way forward for investigating European climate mechanisms, supported by observational studies using improving and temporally expanded reanalyses. Dynamical climate models can be run in various complimentary ways. This includes running coupled ocean–atmosphere models, running their atmospheric component (AGCM) against observed lower boundary layer forcing, particularly SST and sea ice extents, and carrying out special experiments with specified forcings like observed SST patterns, including ENSO, or combinations of other forcings discussed above.

Recent research indicates that using AGCMs with specified SST and sea ice (e.g. HadISST1, Rayner et al., 2003) is a useful way forward for predictability studies, though there are limitations (e.g. Chen and Schneider, 2014). This may allow estimates of UK and perhaps English Lowlands rainfall predictability through the seasonal cycle, for example using the newly improved HadISST2 data set (Titchner and Rayner, 2014). An advantage of such runs is that SST variations are realistic, whereas they may not be in coupled models.

Coupled models have already shown great promise, as shown by the high skill of an ensemble of retrospective December–February European forecasts from a high-resolution version of the HadGEM3 coupled ocean–atmosphere climate model run for the last 20 winters (Scaife et al., 2014a). The SST predictions for this season also show considerable skill (MacLachlan et al., 2014). This work also shows that some aspects of the seasonal surface climate prediction can be further improved by basing them on forecasts of the governing atmospheric circulation pattern rather than the directly forecast surface conditions per se. For example, prediction of the NAO is more skilful than, say, the prediction of temperature across northern Europe, but because the NAO often governs regional climate fluctuations, European winter surface climate predictions may be improved if derived from the forecast NAO (Scaife et al., 2014a), at least in some regions. Thus, a good way to use dynamical seasonal climate predictions of regional UK rainfall in a hydrological context may be to combine dynamical atmospheric circulation predictions with statistical downscaling. A combination of atmospheric and coupled model approaches might be particularly valuable for studying the hitherto unknown causes of the large and persistent atmospheric circulation changes that resulted in the sudden ends of some major droughts like those of 1975–1976 and 2010–2012.

The 20CR stretching back to 1871, now in an enhanced version 2 form (http://www.esrl.noaa.gov/psd/data/gridded/data.20thC_ReanV2.html) and other existing and planned reanalyses will allow new observational studies of relationships between predictors, atmospheric circulation through the depth of the troposphere and rainfall for more than the last century. Thus, the late nineteenth century and very early twentieth century is an especially interesting period for study. It included several major English Lowland drought episodes, including a long drought from 1854 to 1860, a major drought from 1887 to 1888 and the “Long Drought” of 1890–1910 (Marsh et al., 2007; Todd et al., 2013). The latter was associated with several clusters of dry winters analogous to some recent multi-annual droughts. Such studies emphasise the importance of further digitising historical rainfall data. For example, digitised UK rainfall records from paper archives would enable key data sets such as NCIC rainfall to be pushed back well into the late nineteenth century. This, coupled with the longevity of the 20CR data, would open up new possibilities for examining the climatic drivers behind these multi-annual droughts of the nineteenth century. As indicated in Sect. 4.1, a key issue in long, multi-annual droughts is the sequencing between dry winter and summer half-years. The use of long hydrometric records opens up the possibility of exploring frequency–duration relationships to examine drought persistence in a probabilistic sense, e.g. using Markov Chain models to explore dry (wet) to dry (wet) season persistence (Wilby et al., 2015).

A key area for further study is improved understanding of the hydrological response to precipitation deficits during the onset, development of and recovery from drought episodes. This study has used consistent indicators of rainfall, flow and groundwater to shed new light on temporal correlations between meteorological drought anomalies (SPI) and their response in river flow (SSI) and groundwater levels (SGI). However, this has only been evaluated on a broad scale for the English Lowlands – the temporal relationships will vary widely across the study domain, depending on aquifer properties (Bloomfield and Marchant, 2013) and catchment properties (Fleig et al., 2011; Chiverton et al., 2015). The study highlights the need for more systematic studies of drought propagation using a combination of observational and catchment modelling approaches (e.g. as carried out for one English catchment by Peters et al., 2006, and for selected European catchments by Van Loon et al., 2012). Finally, it is important to emphasise that the manifestation of drought impacts in the English Lowlands will be heavily influenced by water management infrastructure and societal responses (e.g. the effects of surface and groundwater abstractions, reservoir operations, and the influence of societal demand during drought events). This study has examined the region on a coarse scale, but an examination of the finer catchment-/aquifer-scale links between climate drivers and flow/groundwater responses will require an appreciation of the moderating role these influences will have on the propagation of climate drivers through to streamflow and groundwater responses.