Hydrological drought risk recurrence under climate change in the karst area of Northwestern Algeria

Human-induced climate change has significantly altered local climates (Zhang et al. 2007). This has also led to accelerated changes in severity and/or frequency of extreme values of climatic variables (Dai 2013), and consequently affects precipitation patterns and hydrological cycles. Severe, persistent and extended droughts, as well as their risks, are the principal consequences of this global warming. It can trigger famine, conflicts and displacement of populations, affect economic development and ecosystem health. In the semi-arid regions of the Mediterranean Basin, including northern Algeria, many studies have highlighted increases of drought events during the last decades, although the frequency and severity of this drought changes from one region to another (Raymond et al. 2018a, 2018b; Tramblay & Hertig 2018; Zkhiri et al. 2019; Achour et al. 2020). This region has been referenced as one of the most vulnerable to climate change and has been defined as a primary ‘Hot-spot’ by Giorgi (2006), based on the results from global climate change projection scenarios. The situation is confirmed by a consensus between models regarding the future increase in both severity and frequency of drought and exacerbation of water scarcity (Forootan et al. 2019).

Among the many drought types, meteorological drought is recognized by the Intergovernmental Panel on Climate Change (Sheffield & Wood 2008) as the principal factor that triggers other drought types (e.g. hydrological, agricultural, etc.). They occur either simultaneously or with different time lags (Wilhite & Glantz 1985). After a few months, a prolonged lack of rainfall can lead to a decrease of storage reservoir levels below normal. This, together with overexploitation of groundwater, may put at risk crop yields and food safety. Hydrological drought is an intermediary transitional link between meteorological drought and agricultural drought; therefore, special importance must be placed on understanding its propagation mechanisms (Dinar et al. 2019; Zhou et al. 2019). Additionally, the propagation time from meteorological to hydrological drought remains a challenge to understand the interactions between different components of the hydrological cycle, particularly in drainage areas with diverse reservoir lithology, physical characteristics (Tijdeman et al. 2018), and water management practices (Bachmair et al. 2016; Wu et al. 2017). For example, in karst regions, meteorological drought can directly affect the karst aquifers' recharge and accelerate hydrological drought; i.e. spring discharge feeds the streams considerably during the humid period and, therefore, helps to reduce or delay hydrological drought. However, during meteorological drought, this spring supply is reduced and, therefore, the hydrological drought is triggered earlier. Thus, a complete and thorough assessment of drought by exploring the relationship between hydrological and meteorological droughts with respect to the basin's characteristics is considered to be the first essential step for early warning of hydrological drought risk (Bachmair et al. 2016; Barker et al. 2016).

To study the effects of climate change on drought risk, several drought prediction approaches such as artificial intelligence (Jalalkamali et al. 2015; Ghorbani et al. 2018), hybrid artificial intelligence (Zhang et al. 2017c; Komasi et al. 2018; Soh et al. 2018; Qasem et al. 2019), support vector regression (Moazenzadeh et al. 2018), and data mining-based methods (Bahrami et al. 2019) have been developed based on a variety of drought indices. Nevertheless, because drought is characterized by highly correlated random characteristics, others introduced univariate and multivariate frequency analysis for assessing the future risk of extreme drought (Shiau 2006; Hao & AghaKouchak 2013; AghaKouchak et al. 2014; Salvadori et al. 2016).

González & Valdés (2003) emphasized that the risk of extreme drought estimated using frequency analysis in both a univariate and multivariate framework has enormous significance, particularly to define the expected consequences of future climatic changes and to correctly establish new strategies. The univariate frequency analysis enables the characterization of drought separately for the duration and severity (Chen et al. 2013; Mesbahzadeh et al. 2019). Accordingly, this analysis remains limited to understanding thoroughly the multivariate behavior of drought and may bias the associated drought events (AghaKouchak et al. 2014; Kim & Chung 2019; Mesbahzadeh et al. 2019). Kim & Valdés (2003) recommended that the frequency analysis of the drought variables need to be strengthened by their joint conditional distributions and their marginal distributions. Among the available approaches for multivariate analysis, copulas are still one of the most used functions to construct the joint distribution function of drought characteristics (Shiau 2006; Huang et al. 2014; Zhang et al. 2017a; Dodangeh et al. 2019).

Although the majority of studies on dryness have been performed using observed data, more studies about hydrologic risk are necessary in order to investigate the impact of climate change on the availability of water resources and the future of extreme drought (Salvi & Ghosh 2016; Apurv et al. 2017; Dai & Zhao 2017). Most studies that have highlighted the response of hydrological systems to climate change by the 21st century are, however, based on the conceptual rainfall-runoff hydrological models (Giuntoli et al. 2018). Through their robustness and very low input data requirement (potential evapotranspiration and precipitation), calibration/validation data (streamflow), the conceptual hydrological models are extremely successful not only in terms of results but also for proving accuracy in many regions over the world (Zeroual et al. 2013; Okkan & Fistikoglu 2014; Todorovic & Plavsic 2016; Allani et al. 2019; Al-Safi et al. 2019).

In Algeria, despite the fact that a number of studies have been devoted to identification of regional and local drought episodes (Djerbouai & Souag-Gamane 2016; Habibi et al. 2018; Zerouali et al. 2018; Achour et al. 2020), few studies have been dedicated to identifying the future hydrologic risk of extreme droughts and their probability of recurrence. To our knowledge, all past studies have focused on the study of persistence of drought based on observed data. Some of them reported that the persistence of these events in the western and the central part of the region was mainly recorded in 1973, 1981 and especially since 1991 while no significant event has been noticed in the east (Meddi et al. 2014; Zeroual et al. 2017). Specifically, the northwestern region of Algeria has been affected by this phenomenon during the 1980s, 1991, 2001 and 2002, causing a significant decrease in dams' storage levels and agricultural production, drawdown of groundwater levels, and drying-up of some karst springs (Meddi & Hubert 2013; Bensaoula et al. 2019; Achour et al. 2020). Based on the above insights, the overall goal of this study is to estimate the hydrologic risk for extreme drought in three basins located in the karst area of northwestern Algeria and assess the possible increase of the drought risk in the future period under two climate change scenarios. The main specific goals include: (1) inspect the relationship between meteorological and hydrological droughts at three watershed scales in the karst area of northwestern Algeria through comparison between Standardized Precipitation Index SPI and Streamflow drought index SDI and their characteristics; and (2) evaluate the joint return periods of 2, 10, 20, and 50 years based on the reference period (1941–2011) and future hydrological drought events projected under two Representative Concentration Pathway scenarios (RCP4.5 and RCP8.5).

A description of the study area, observed and climate model's data acquisition and the methodology are presented in the next section. Results concerning calculation of drought indices, the construction of joint distribution for return periods estimation and future drought projections are presented in the Results section. Discussion about the relation between meteorological and hydrological droughts and on hydrological drought risk and its possible increase under climate change are discussed in the following section. Finally, conclusions and future directions are summed up.

The study has been founded on three basins in the extreme northwest of Algeria, located at the southern Mediterranean shore near to the Moroccan boundaries (Figure 1). Beni Bahdel and Chouly basins are situated in the Tafna basin while El-Hcaiba basin belongs to the wide Macta watershed. The study zone extends to a total area of 21,634 km². The Beni Bahdel basin is located in the Tlemcen Mountains, and Chouly basin is also located in the northern foothills of the Tlemcen Mountains made up mainly of upper Jurassic and lower Cretaceous formations (Bensaoula et al. 2019). El-Hcaiba basin is located in a mostly karstic dominant system under the high plains. The Beni Bahdel station is located upstream of the Beni Bahdel dam (Figure 1).

The zone of study is characterized by a high temporal variation of annual rainfall. The climate is typified by a wet season with maximal precipitation in December, January and February, and a long dry period, nearly without rain, from June to September. The average annual rainfall ranges from 300 to 350 mm and the mean monthly temperature ranges from 11° in January to 26 °C in July and August. Eleven dams with a total capacity of 611.1 Mm³ have been built mainly for the irrigation of western plains (around 4,693.49 km² of irrigated perimeters).

The region contains large fertile plains and an important industrial area and it plays an important role in the regional economic development. However, the discharge in rivers crossing the study area has decreased significantly during recent decades (Achite et al. 2017). This is related to droughts that have affected northwestern Algeria since the mid-1970s (Meddi et al. 2014).

The choice of these catchments was guided by the availability of long-time series of flow measurements covering the period 1941–2011 (71 years) and the lithological and hydrogeological nature of these catchments which are located in the most extensive karst system in northern Algeria and that present the widest natural groundwater reservoir in the west (Collignon 1986; Bensaoula 2007; Bensaoula et al. 2019). Karst aquifers of this region are considered as the largest natural reservoirs of rainfall in north Algeria with a total renewable groundwater reserve estimated to be about 200 Mm³/year (Collignon 1987). Therefore, karst springs, which have an annual mean outflow that reaches up to a few cubic meters of water per second, feed the streams considerably during the humid period and help to reduce or delay hydrological drought. The existence of karst springs in Chouly and El-Hcaiba basins helped to feed the streams and to supply the drinking water demand. However, during rainfall deficit (meteorological drought), this spring supply is reduced or even suspended and therefore the hydrological drought is triggered earlier. During the last two decades, a drop in the spring discharge has been seen which is caused by a decrease in annual and winter rainfall in this region, i.e. a 20% decrease in annual rainfall according to Meddi et al. (2010), Taibi et al. (2013) and Zeroual et al. (2017). For example, the last events of drought in the Chouly basin (e.g. 2001, 2002 and 2008) have caused total drying up of some of their springs. We note that spring discharge data are not available and thus they have not been used in this study.

The data are provided from the National Hydraulic Resources Agency (ANRH). The rain gauge stations are selected from seven gauge stations over the three basins, two stations' data records are from 1941 to 2011 (Meddi & Meddi 2009; Meddi et al. 2010; Meddi & Hubert 2013) and others from 1960 to 2011. The three hydrological gauged stations selected for this study are also shown in Figure 1. Beni Bahdel and the Chouly stations cover the period 1941–2011, while El-Hcaiba station covers the period 1960–2011. All rainfall data series over northwestern and northcentral Algeria have registered a break-point located in the 1970s and that is related to a decrease of rainfall (20% decrease in annual rainfall since 1970) (Meddi et al. 2010; Zeroual et al. 2017). This break-point is caused by climate change and has also been detected in several regions of the Mediterranean basin (Brunetti et al. 2006; Ramadan et al. 2013; El Kenawy et al. 2019; Scorzini & Leopardi 2019; Achour et al. 2020; Royé et al. 2020; Tsiros et al. 2020). The missing data in the rainfall times series (gaps do not exceed 5%) have been computed at monthly scale by normal ratio method. The missing value is estimated from nearby stations by weighing the rainfall by the ratios of normal as suggested by Subramanya (1994). It is important to note that all time series were found to be outlier free using the Grubbs' test (Grubbs 1950). The homogeneity of data series has been tested according to Buishand (1982), Alexandersson (1986) and González-Rouco et al. (2001).

The gauge stations are selected according to long-term availability of the data in order to better present the temporal comparison between meteorological and hydrological data. Moreover, the determination of the drought return period is sensitive to the size of the data used for bivariate distributions (Zhang et al. 2015; Zhao et al. 2017). Similarly, the longer the data, the more the validation of the hydrological model and bias correction of the climate model are accurate. We note that climatic and hydrological data in Algeria are roughly available. The climatic and hydrologic gauge-instruments were installed mostly after independence (1962) and there were only a few implemented during French colonization.

We used climate simulations (monthly precipitations and temperatures) from nine Atmosphere–Ocean Global Climate Models (GCMS) of the fifth phase of the Coupled Model Intercomparison Project (CMIP5) (Jones et al. 2011), dynamically downscaled by the latest version of the Rossby Centre Regional Climate Model (RCA4) (Table 1). The nine GCMS were used as the boundary conditions of RCA4 by the Swedish Meteorological and Hydrological Institute (SMHI) to drive an ensemble of Regional Climate Model simulations over Africa with a grid spacing of about 0.44° resolution over Africa (Jones et al. 2011). The projection period was forced by the representative concentration pathway scenarios RCP4.5 and RCP8.5 with boundary conditions from the GCMs. These data are available within the CORDEX-Africa framework via the Earth System Grid Federation (ESGF) data portals. Therefore, for a further discussion on the changes in the future of the hydrological drought, we have considered two existing scenarios for future projections: RCP8.5 which is the most pessimistic scenario; and RCP4.5 which is projected to be in the near future or in the case of moderation of climate change (Taylor et al. 2012). To extract the corresponding data (monthly precipitations and temperatures) from model simulations over the 1951–2100 period, the nearest neighbour method (Ly et al. 2011; Luhunga et al. 2016) has been used for each climatic station.

The data extracted from the nine RCA4-GCMs models (monthly rainfall and temperature) differ from the observed data due to the imperfect conceptualization, spatial averaging within grid cells and non-inclusion of local features that impede their direct use in the assessment of climate change effects (Ghimire et al. 2019). Thus, the correction of biases between the climate models and observed values is required in order to improve the accuracy of hydrological studies. Among the various approaches that have been presented in the literature, we have used the empirical quantile mapping (EQM) technique (Themeßl et al. 2012; Xu 2018) to correct RCA4-GCMs monthly rainfall and temperature during the 2006–2100 period. This method has been considered as one of the preferred techniques in the context of improvement of hydrological simulation abilities of monthly GCM-simulated rainfall and temperature (Ghimire et al. 2019).

The Standardized Runoff Index (SRI) (Shukla & Wood 2008) and Streamflow Drought Index (SDI) (Nalbantis & Tsakiris 2009) are the most widely known hydrological drought indices. The principal advantage of these indices is that they result in standardized hydrological drought indices that we can simply classify and compare to a meteorological drought index. In this study, we use the SRI and the modified SDI introduced by Madadgar & Moradkhani (2013). It is based on a monthly streamflow time series where the streamflow volume is fitted to the best distribution function and then transformed into a standard normal function with zero mean and unit standard variation.

and are the mean and standard deviation of y_i,j,k while V_i,j,k is the observed cumulative streamflow. When we replace the observed streamflow by the simulated runoff volume from the hydrological model in SDI calculation, the index becomes the SRI (Shukla & Wood 2008). The SDI and SRI monthly time series are estimated for each hydrometric station at six-month time-scales aggregation.

The spatiotemporal connection between meteorological and hydrological droughts is based on the analysis and the comparison of the evolution of their characteristics in the study zone. Drought characteristics are mainly duration, severity, intensity and inter-arrival time. The duration is defined as the number of months where SDI or SPI values are continuously below the threshold –0.8 (Agnew 2000; Svoboda et al. 2002). The severity is the absolute aggregated negative values along the drought event duration (Dracup et al. 1980), while the time lapse separating the onset of a drought event and the onset of the next one is defined as the inter-arrival time (Shiau 2006).

Copulas are divided into distinct families: Archimedean, extreme value, elliptical, and others. The Archimedean copulas are the most applied in the hydrology field (Hao et al. 2017; Hangshing & Dabral 2018; Wang et al. 2019). The Gumbel–Hougaard, Clayton and Frank copulas from the Archimedean copula family were used to construct the joint probability distributions of hydrological drought characteristics. The Clayton and Gumbel–Hougaard copulas have the ability to model the dependence structure of two random variables that have a positive correlation (Fang et al. 2018). The structure and parameters of the copulas used are briefly given below:

The method gave satisfying results, especially in hydrology studies (Huang et al. 2015, 2017a, 2017b; She & Xia 2018; Yang et al. 2018).

In order to simulate hydrological drought events in the future, the parametric GR2M conceptual hydrological model (rural engineering model with two parameters on a monthly scale) (Mouelhi et al. 2006) was used to simulate the future runoff for the three basins. The GR2M is a rainfall-runoff model that requires the monthly precipitation and potential evapotranspiration. The model has two parameters, X1 and X2. Parameter X1 is for the soil moisture estimation, and X2 is used for computing the groundwater exchange with neighbouring catchments (Mouelhi et al. 2006). The two parameters are well optimized during this step and are employed for the simulation of the future runoff. We applied the Nash–Sutcliffe Efficiency (NSE; Nash & Sutcliffe 1970), mean absolute error (MAE) and root-mean-square-error (RMSE) (Willmott & Matsuura 2005; Chai & Draxler 2014) as objective functions for the model calibration for each basin. Next, the nine RCA4-GCMs future climatic outputs (precipitation and temperature) under the two scenarios (RCP4.5 and RCP 8.5) were used as inputs for the GR2M model in order to simulate the future runoff. The simulated runoff data were then employed to calculate SRI at six months' time scale.

The results of this section are presented in two levels showing: (1) the calculation of drought indices SDI and SPI over the three basins at different time scales during the studied period and (2) a comparison between drought characteristics issued from hydrological and meteorological droughts.

First, for the calculation of the SDI, the best-fit distribution of accumulated streamflow volumes was selected from five distributions, Gamma, Lognormal, Normal, Weibul, and Exponential according to the Kolmogorov–Smirnov test (K-S) where the null hypothesis supposed that the tested distribution is fitted to the data.

Results for the selection of the best fitted distribution to the aggregated streamflow volume using the K-S test and their parameters estimated by the maximum likelihood method (MLM) are exhibited in Table 2. The MLM is considered as the most efficient parameter estimation method (Rao & Hamed 1997).

Second, for the calculation of SPI, the SPI series was fitted to the gamma distribution with consideration of the zero-precipitation probability. Overall, it has to be noted that Gamma distribution is not always the best distribution for the calculation of SPI and can be different across regions or aggregation scale (Hong et al. 2013; Blain & Meschiatti 2015; Vergni et al. 2017). For the Mediterranean basin, including our study region, several studies were carried out to test different distributions to fit SPI at various time scales (SPI-3; -6; -9; -12) and demonstrated that the gamma distribution dominates for long-accumulation distributions (Stagge et al. 2015). These results have been confirmed by Achour et al. (2020) for the western plains of Algeria.

After computing the two indices, we chose the threshold of –0.8 to define a drought event, according to Svoboda et al. (2002), in order to calculate drought severity, duration and inter-arrival time. The primary motive to consider −0.8 as a threshold is that according to drought severity classification, a drought index below −0.8 starts triggering damage to crops and pasture and water shortages (Svoboda et al. 2002). This threshold was agreed by the U.S. drought monitor (https://droughtmonitor.unl.Edu/ About/AbouttheData/DroughtClassification.aspx) and was chosen in many semi-arid regions studies (Agutu et al. 2017; Zhang et al. 2017b; Das et al. 2020).

Figure 2 shows a comparison of the severity of drought events computed from SPI-12 and SDI-6 scale for the three basins within the study period 1941–2011 for Chouly and Beni Bahdel watersheds and 1962–2011 for El-Hcaiba basin, where the horizontal line lengths represent the duration in months. For the three studied basins, we notice some simultaneous dry periods, whether in hydrological or meteorological records, since 1970.

The Beni Bahdel and Chouly basins have experienced more severe and longer drought periods since 1976. Episodes of hydrological drought have also increased significantly. However, the Beni Bahdel Basin experienced more severe and longer-lasting sequences compared to the Chouly basin, except for the hydrological drought of 1996–2008 where the Chouly basin was affected by severe events and reached a total severity of 237 during 147 months (Table 3). The episodes of meteorological drought reached their maximum severity between 1980–1985 and 1996–2003. Hydrological droughts were also marked by this dry period when they reached their maximum severity between 1996 and 2008 for the two basins. For El-Hcaiba watershed, we distinguish a high persistence of droughts during the studied period where the basin was affected from multiple drought episodes.

Figure 3 confirms the previous findings and further reveals the influence of rainfall deficit on the response of the basins. On the left side the monthly spatial SPI-12 are presented while the SDI-6 are shown on the right side to easily compare the two types of drought. The variations of SDI-6 and SPI-12 are in agreement. This confirms that seasonal flow, presented by SDI-6, is well described by annual variation of SPI.

The Beni Bahdel basin is more sensitive to meteorological drought (Figure 3), i.e. the two types of drought occur simultaneously. However, the response of the basin to rainfall deficit is longer for the Chouly watershed and more for El-Hcaiba watersheds, where the hydrological drought is observed only when the meteorological drought has reached its maximum severity (less than –1.5). This corresponds mainly to the existence of karst springs that feed the streams sub-continuously and influence the basin's memory. However, for the Chouly basin, an exceptional hydrological drought was observed from February 1996 to April 2008. This prolonged event is the result of consecutive severe meteorological drought events (March 1993–March 1994, June 1996–December 1996, January 1997–November 1997, May 1998–August 2001, November 2003–April 2006 and May 2006–January 2008).

From another viewpoint, the analysis can also determine that drought events are a very frequent hazard in the studied area (Table 3). The meteorological drought events are more frequent than the hydrological drought events at all studied basins. However, all mean and maximum, duration and severity are greater in hydrological drought. This can be particularly related to the propagation from meteorological to hydrological drought where the hydrological drought is related to the antecedent and/or concurrent meteorological drought. The occurrence of both antecedent and concurrent meteorological drought increases the risk of hydrological drought occurrence and makes the hydrological drought more severe and longer lasting. This situation is inversed at El-Hcaiba basin where the response of the basin to meteorological drought is slower due to the springs' discharge that feed the streams and help to impede or break drought propagation.

It is worth highlighting that the Beni Bahdel basin is the most affected area where drought events are more frequent. These findings also highlight that the potential drought risks over the studied basins are better defined by the assessment of drought characteristics.

In order to acquire the best marginal distribution to fit the hydrological drought characteristics, the goodness-of-fit values of the probability distributions are calculated with the K-S test. The results of the best fitted distributions are illustrated in Table 2. The Gumbel–Hougaard copula has shown the highest weights for all stations and the values of the parametric estimate are very close to the empirical estimate indicating that the Gumbel–Hougaard copula can describe well the tail dependence (Table 4).

The joint return periods of the duration and severity series were computed by Gumbel–Hougaard copula and contours of the different return periods are presented in Figure 4. The historical drought events are also included in the graphs. The most damaging drought event for the Beni Bahdel basin occurred from May 2002 to January 2008, with a severity of 82.47 and duration of 69 months. The joint return period of this event is more than 65 years while the is less than 45 years (Figure 4(a)).

For the Chouly basin (Figure 4(b)), most drought events have return periods of less than 50 years, except the extreme event that occurred in February 1996 until April 2008. This event has a duration of 147 months and corresponding severity of 237 with joint return periods, and ⁠, that are greater than 200 years. For the El-Hcaiba station (Figure 4(c)), all of the dry events have return periods less than 25 years, except the exceptional drought that occurred from November 1991 to March 1995 with a severity of 45. The joint return periods and of this event are greater than 200 years.

The impact of climate change on future droughts has been evaluated in this section. For this purpose, first the simulated precipitations and temperatures are used, after EQM correction, as inputs to the previously calibrated and validated GR2M model to simulate the future runoff. The SRI series were computed for all studied basins for the near (2021–2060) and far (2058–2100) future periods.

The SRI simulations derived for the three basins from all RCA4-GCM are illustrated in Figure 5. The figure indicates a similarity in the variation of mean SRI series across model projections. According to the two scenarios, the far future is more pessimistic in the projection of drought events compared to the near future period. However, all nine models exceptionally project negative means with a larger range of SRI variations, especially for Chouly and El-Hcaiba basins and more extreme dry events for the far future period under RCP8.5 scenario.

Figure 6 presents a comparison of monthly temporal variation using the mean projected SRI from the nine models under the two scenarios. The comparison of the severity, duration and projected frequency of drought events shows that dry episodes are expected to be more severe and persistent in Beni Bahdel and Chouly watersheds. However, El-Hcaiba basin will witness more frequent episodes. All three basins will experience more frequent drought events compared to the historical events, with accentuated severities and during longer periods. The temporal analysis also reveals that drought will be the dominant situation at the three basins, but it seems to intensify particularly under the RCP8.5 emission scenario.

Table 5 summarizes drought events for three basins within the two future periods in terms of frequency, mean duration and severity and maximum severity and duration for each of the nine models included in this study under the two RCP scenarios.

For drought frequency (Table 5), all models project an increase in frequency of drought event in the far future (2058–2100) under the two scenarios compared to the reference period (1963–2005) at the three basins, except RCA4-CanESM2 which projects a balance in drought frequency under the RCP4.5 scenario. The RCA4-MIROC5 under the RCP8.5 scenario is the most pessimistic model to project drought frequency at both Beni Bahdel and Chouly basin (43 and 34 events compared, respectively, to 14 and nine events in the reference period).

As regards the mean duration, we note that under RCP4.5 scenario only the RCA4-GFDL-ESM2M in the near future and four models (RCA4-CanESM2, RCA4-CNRM-CM5, RCA4-CSIRO-MK3 and RCA4-GFDL-ESM2M) in the far future project an increase at Hcaiba basin. For RCP8.5, the RCA4-CanESM2 and RCA4-CSIRO-MK3 models in the near future and RCA4-CSIRO-MK3, RCA4-CNRM-CM5 and RCA4-HadGEM2-ES models in the far future project an increase in the drought mean duration at Hcaiba basin. For the remaining two basins, the increase in the drought mean duration is projected only by the RCA4-CSIRO-MK3 model at Beni Bahdel basin in the far future under RCP8.5.

Concerning the maximum duration under the RCP4.5 scenario, all models except RCA4-CNRM-CM5 and RCA4-MIROC5 in Hcaiba basin and only the model CSIRO-MK3 in the Beni Bahdel project an increase in the near future. It is worth mentioning that all models project an increase in maximum duration in Hcaiba basin under the two scenarios in the far future. However, the maximum duration varies for the three basins between 41 and 44 months.

It must be emphasized, according to the simulations in the near future under RCP4.5, that the maximum severity remained unchangeable in Beni Bahdel and Chouly basin compared to the reference period, however, an increase is projected in Hcaiba basin by all models except RCA4-MIROC5. For the far future under RCP8.5, all models project an increase in maximum severity in Hcaiba and Beni bahdel basins except RCA4-MIROC5 and RCA4-GFDL-ESM2M which project a decrease in Beni Bahdel.

It is worth mentioning that in Chouly basin drought frequency is projected to increase by all models while all mean and maximum drought duration and severity are projected to decrease. This could be related to the exceptional event that affected this basin compared to the other events which make the maximum of duration and severity higher in the reference period. The two future scenarios do not project this event in the future period.

Using the bivariate copula functions for each basin, the joint return periods (10, 50 and 100 years) of the duration and severity series were computed for all models for the far future period (2058–2100) under the two studied scenarios. Afterwards, we make a comparison between the historical and future return periods. This comparison is based on the investigation of rate change of drought risk between the reference and future periods.

Figure 7 presents the decrease in future return periods (10, 50 and 100 years) under climate change (RCP4.5 and RCP8.5) in the future period 2058–2100 compared to the reference period. The results show that all simulations project a decrease in return periods and consequentially an increase in drought risk for the three basins, for example, for Beni Bahdel basin, and according to the RCA4-MIROC5 simulation, the return period of 10 years in the reference period likely will be about 6.1 years under RCP8.5 and 8.1 years under RCP4.5 under climate change. Similarly, for the return period 10, 50 or 100 years, the corresponding severity and duration likely will become more severe and longer than for the reference period. The results also show that the Beni Bahdel basin has the highest increase in terms of frequency and thus is the most vulnerable to climate change (Figure 7); the maximum rates, 40% under RCP8.5 and about 30% under RCP 4.5, are projected by the RCA4-CSIRO-MK3 and IPSL-CM5A models, respectively (Figure 7(a)). From Figures 5–7, the RCA4-CSIRO-MK3 is the most pessimistic model for the three basins as it predicts more serious extremes.

The SPI is used on various time scales. Each scale has its specific physical significance (i.e. irrigation, seasonal or inter-annual water resources); a longer-time scale may reflect impacts on water storage and hydrological response. In this study, we have mainly compared the characteristics of hydrological drought events based on only streamflow with those obtained by only precipitation. The characteristics of the SPI index at higher aggregation (SPI-12) have demonstrated the best correlation with streamflow (SDI-6) compared to the other time scales. Romano et al. (2013) found the same results although SPI depends only on precipitation without consideration of streamflow or other factors. For hydrological drought investigation, the seasonal cumulative streamflow, expressed by longer month aggregations, can show the hydrological behaviour regarding the memory effect of the basin (Madadgar & Moradkhani 2013; Lee et al. 2019). Barker et al. (2016) found that the linkage between meteorological and hydrological drought is well shown by using six-month lags between their indices. Lee et al. (2019) found that the difference between the SPI and SDI increases as the accumulation period decreases, and therefore the 12-month SPI and SDI have a similar pattern. Das et al. (2020) also found that the variability in the probability density of different drought properties is observed under the 12-month drought scale determined by SPI-12 in two Indian Himalayan states. From another aspect, in karst systems, SPI-12 showed a similar trend with spring discharge (Fiorillo 2009; Fiorillo & Guadagno 2010). In addition, Mo et al. (2019) found that the changes in runoff were mainly associated with rainfall in the karst area. SPI-12 can also show the hydro-geological conditions when groundwater data such as spring discharge are missing (Fiorillo 2009). Fiorillo & Guadagno (2010) have found that annual cumulative rainfall has similar trends with rivers and springs discharge. Hence, in our study, the monthly SPI-12 and SDI-6 time series are analyzed to investigate the superficial and groundwater availability and its connection with rainfall.

As far as the basin response is concerned, the Beni Bahdel basin has particular behaviour, where the response to dry conditions is faster compared to Chouly and El-Hcaiba basins. This difference refers to the existence of karst springs at Chouly and El-Hcaiba basins that feed the streams sub-continuously and that impede the launching of hydrological drought events. Laaha et al. (2017) also found that, at the regions affected by dry/wet conditions, hydrological drought events occur earlier/later and are expected to be more/less accentuated.

Furthermore, for drought propagation, most meteorological and hydrological drought events have occurred simultaneously for the Beni Bahdel basin while meteorological drought triggers the hydrological drought during the following year for Chouly basin or even more for El-Hcaiba basin. This delay can be referred to the memory effect that characterizes the karsts aquifer system and that can impede the hydrological drought occurrence in dry conditions (Tallaksen & Lanen 2004; Fiorillo 2009). In the literature, Fiorillo & Guadagno (2010) found similar results at two different karst systems in the Campania region of Italy, where long meteorological drought has generated shrinking in springs discharge and has led to a flat springs hydrograph during the following year in the case of severe events (SPI 12 less than –1,5)). Other research reported that this relationship depends on the lithology and characteristics of the watershed aquifer (López-Moreno et al. 2013; Barker et al. 2016), the altitudes and vegetation cover of the catchments (Peña-Gallardo et al. 2019), and that anthropogenic factors play an important role in the basin's response to climate droughts (Lorenzo-Lacruz et al. 2013; Wu et al. 2017).

From another aspect, we have also found that the most important drought events at the Beni Bahdel and Chouly basin occur in the same period, while El-Hcaiba basin is affected subsequently. This could be referred to its localization near to the high plains of Chellif basin east. However, it can also reveal that the region is affected by the same event but that it is more dominant/persistent in the extreme west of the country and that could be related to other factors. For example Zeroual et al. (2017) has highlighted that the rainfall variability in western Algeria is strongly correlated with large scale atmospheric circulations such as the Southern Oscillation and with the regional scale of western Mediterranean oscillation, similar results were found by Tramblay & Hertig (2018) for western Algeria, Morocco and Spain by investigating the association of dry spells and atmospheric circulations.

The risk that the three basins witness drought events determined by bivariate return periods (severity-duration) varies among the basins and is predicted to expand during the next 50 years. This occurrence is related to the rise of dry conditions in the study region and over the western Mediterranean basin (Spinoni et al. 2015; Tramblay & Hertig 2018; Santillán et al. 2019; Zeroual et al. 2019; Zkhiri et al. 2019).

Drought events in the Beni Bahdel basin have mostly short return periods, indicating a higher hydrological drought risk in this area, while it is shown to be longer moving to Chouly and El-Hcaiba basins, respectively. The short return periods in the Beni Bahdel can be particularly related to the high frequency of past events since the basin is the first dominated by meteorological drought events that are more persistent compared to the other basins. This is due to the basin's response to meteorological drought where the Beni Bahdel basin has faster propagation to hydrological drought events compared to Chouly and El-Hcaiba. Also, it is worth mentioning the importance of the determination of return periods for water resources management where historical events, especially those that have a short return period, have to be taken into account for an optimal regularization of reservoirs and for planning the future irrigation schedule.

As far as the future evolution (2021–2100) of the temporal variability of drought in the three basins is concerned, this study shows that all simulations predict that the drought event will rise in terms of frequency and duration. The increase in drought events projected for the RCP8.5 scenario and for the far future is higher than that projected by the other scenario and periods.

These findings confirm the studies of drought trends in the literature over the Mediterranean basin projected for the 21st century, which have revealed the dry spells in several regions surrounding the Mediterranean, namely those of the southern shore of the Mediterranean, Iberian Peninsula and southern Italy (Dubrovský et al. 2014; IPCC 2014; Founda et al. 2019). This region is considered by most projected scenarios for the 21st century as the most exposed to extreme events (droughts and floods) (Dubrovský et al. 2014; Tramblay & Somot 2018).

The exacerbation of hydrological drought in the region will cause negative impacts on water resources availability. We note that the dams in the region are intended primarily for irrigation of the western plains that contribute essentially to food production and provision in the country (Zeroual et al. 2019; Achour et al. 2020).

This study has highlighted the drought risk in terms of return period and drought projection under climate change. We note a consensus between hydrological drought risk presented by the joint return periods of historical events and projected hydrological drought events.

The identification of historical drought events recurrence in each basin can well describe the drought risk in the future. From other aspects, the investigation of future events return periods can reveal the impact of climate change in the semi-arid region of the Mediterranean on drought risk.

For example, for the Beni Bahdel basin, the severe hydrological event that occurred during 2002–2008 has caused an enormous drop in the Beni Bahdel dam storage levels (the storage level reached less than 20% of the total storage in 2007 and 2008). The total cereal production in the two years 2007 and 2008 dropped by 58 and 62% respectively compared to the year 2006. This drought event has a probability of occurrence of at least once in 65 years (which means a possible occurrence of at least once before 2067). It also corresponds to an antecedent meteorological drought with severity S = 64 and corresponding duration D = 59 months from June 1996 and concurrent meteorological drought with severity S = 40 and corresponding duration D = 49 months from December 2003. Similarly, we can determine according to future projections that a similar hydrological event is projected to occur once in 2066–2069 under RCP4.5 with severity S = 48.2 and corresponding duration D = 40, and twice (2020–2024, 2048–2052) under RCP8.5. These projected events will likely be more recurrent in the future with a return period of 42 years under RCP8.5 and 50 years under RCP4.5 according to future projections from the most pessimistic model RCA4-MIROC5. Therefore, this event likely has a high risk of occurring whether according to historical return period or future projections under RCP 4.5 or RCP 8.5.

As a brief summary, the potential drought events in the semi-arid region of the Mediterranean will become more severe, threatening the region with more frequent, longer and more severe drought events. According to Zkhiri et al. (2019), the future projections of drought events in the High Atlas basins, central Morocco, indicate a strong increase in the frequency of SPI events below −2, determined by five RCM climate simulations, under two emission scenarios (RCP 4.5 and RCP 8.5).

Ouhamdouch et al. (2019) found similar results by studying the impact of climate change on future flows in northern Morocco where the future flows show a downward trend of 38% by 2050. Similarly, Tramblay et al. (2018) found that a decrease of water availability over northwest African dams is projected by five studied RCM models under RCP4.5 and RCP8.5 scenarios.

This analysis has allowed us to determine the hydrological drought risk in the area under climate change in order to establish the best solution for water supply and food-processing production and for drought mitigation. The risk of hydrological drought in the studied basins requires artificial groundwater recharge or the use of renewable/unrenewable groundwater resources in case of multi-year drought exacerbation.

In this paper we have examined the potential changes in the hydrological drought risk at three basins located in the most extended karst massif of northern Algeria. We have also investigated the joint return periods of drought events in the reference period and their risk of occurrence in the future under two climate change scenarios. Finally, we have assessed the change rates in 10, 50 and 100 years drought return periods under climate change between reference and future drought events using copulas.

The meteorological and hydrological drought events expressed by SPI-12 and SDI-6 mostly agree well and highlight the memory effect of the basin in the response to dry conditions. The meteorological drought events at Beni Bahdel basin transform faster into hydrological events followed by Chouly and finally El-Hcaiba. The Beni Bahdel and the Chouly basins have the same climatic conditions where meteorological drought occurs simultaneously. Furthermore, the occurrence of hydrological drought sequences is related to the severity and duration of the antecedent/concurrent meteorological drought.

The frequency of drought differs in each basin. The Beni Bahdel basin has shorter return periods (less than 25 years) and thus is more prone to drought. Therefore, the damaging drought events that have short return periods have to be taken into account for water reservoir plans.

Drought is expected to exacerbate under climate change. Furthermore, the variation across the nine model projections is similar for hydrological drought where most of them project the same conditions in terms of frequency, severity and also duration. RCA4-CSIRO-MK3 is the most pessimistic and projects the most severe droughts among other climate models combined with a decrease in their return periods. All models' simulation under the two scenarios project higher drought severity and duration and higher return periods than the corresponding values determined using the reference period data.

Finally, the studied basins are all threatened by drought risk that tends to exacerbate under climate change, especially under RCP8.5 scenario and according to drought events' return periods. Thus, the future critical situation implies the need for establishing an appropriate management schedule to preserve the water availability and food-processing production in the region. This situation requires artificial groundwater recharge or use of renewable/unrenewable groundwater resources in the case of multi-year drought persistence. Such solutions are more adequate since this region is known as the largest natural groundwater reserve in west Algeria. Future studies are required in order to establish comprehensive risk-based quantitative design standards of water resource systems for drought mitigation across all north Algerian watersheds.

The authors wish to thank the National Agency of Water Resources for providing material help and data on which the reported analysis is based.