The streamflow simulations for the past and the future are conducted in 38 watersheds in Cyprus. Cyprus is an island in the eastern part of the Mediterranean Sea, at latitude 35° N and longitude 33° E. The climate of the island is characterized by the typical variability in precipitation in terms of a clear seasonality pattern, as well as, the year-to-year variability in the distribution of precipitation in the wet period found in Mediterranean climates (Hoerling et al., 2012). December and January are the wettest months; about 80 % of total annual precipitation occurs between November and April. The Troodos massif is the water tower of Cyprus and it covers more than 25 % of the island's total area (9251 km²). The 38 watersheds of the study form a radial drainage system around Troodos (Fig. 1). The 38 streamflow gauges defining the areas of the studied watersheds were selected such that any large dams and major waterworks are located downstream of the gauges. Common land use changes occurring in the studied area include agricultural land abandonment and fires over shrubland and forests. Subsequently, a conversion of the abandoned and burned area occurs, through shrub expansion and a slow (re-)growth of, predominantly, pine trees in areas where soil and rain conditions favour their development. Geologically, Troodos is constituted by an ophiolite complex with faulted and highly fractured rocks, especially the gabbro on the upper hillslope, forming fractured aquifers or aquifer systems favoring infiltration (Udluft et al., 2006). Soils in Troodos have a stony gravelly texture and a high variability in soil depth from very shallow (0–10 cm) up to about 100 cm (Camera et al., 2017). Christofi et al. (2020) used isotope and hydrogeochemical sampling and modelling to identify regional groundwater flow through the diabase and basal-group units of the Troodos Fractured Aquifer, providing evidence of groundwater flow across surface watershed boundaries in the Troodos Mountains.

A summary of the long-term average hydroclimatic characteristics of the studied watersheds is shown in Fig. 1a. The map of Cyprus with the boundaries of the 38 watersheds and the trends in precipitation and streamflow in the 1980–2015 period are also shown in Fig. 1b. For precipitation, Sen's slopes ranged from -2.2 to 3.2 mm yr⁻¹, with no statistically significant (p<0.05) trends (Mann-Kendall test) among the 38 watersheds. For streamflow, Sen's slopes ranged from -0.22 to 0.01 Mm³ yr⁻¹ with statistically significant negative trends for two watersheds. The reference evapotranspiration (ET) trend (not shown) was statistically significant for all watersheds and Sen's slopes ranged from 1.3 to 3.8 mm yr⁻¹.

Daily precipitation and temperature observations for the area of Cyprus (CY-OBS) for 1980–2015 in gridded datasets at 1 km spatial resolution were used for forcing the hydrological model (Sect. 2.3) and the evaluation of RCM output (Sect. 2.4). The periods 1980–1998 and 1998–2015 were used for calibration and validation, respectively. These gridded observations were generated with the spatial interpolation methods described by Camera et al. (2014) and Sofokleous et al. (2021). The temperature data include daily minimum and maximum values, from which the daily average temperature was computed. The daily reference evapotranspiration, required as input for the hydrological model was computed from the daily minimum and maximum temperature with the Hargreaves equation (Hargreaves and Samani, 1985). The daily data on watershed-level were extracted from the gridded 1 km data, by averaging the gridded data over the watershed areas. For the hydrologic model calibration and evaluation, daily streamflow observations at the outlets of the 38 watersheds from the monitoring network of the Water Development Department of Cyprus (Fig. 1), in the period 1980–2015, were used. Missing streamflow data (six watersheds, 4–14 years) were filled, by fitting linear or exponential relationships with downstream or upstream stations in the same watershed. These data were used for the calibration and evaluation of the hydrological model. Streamflow observations from Kryos watershed in the southern slopes of the Troodos mountains (r9-6-2-90, see Supplement) were also corrected by taking into account monthly water diversions from Arminou Dam located on the Diarizos river (not included in the study area) to the Kryos watershed, which started in the hydrological year 1998–1999. The data correction periods spans outside the calibration period.

The conceptual, four-parameter, daily GR4J model was used for the streamflow simulations. A detailed description of the model is given by Perrin et al. (2003). The GR4J model has shown good performance for different environments (e.g., Broderick et al., 2016; Le Coz et al., 2016; Guo et al., 2020). The four model parameters represent the watershed's soil water or production storage (X1; mm), groundwater exchange acting on streamflow (X2; mm), streamflow storage (X3; mm), and a time parameter for the unit hydrograph (X4; days). The groundwater exchange coefficient can add a positive (inflow to stream) or negative (recharge losses) contribution to the water balance. The parameters are embedded in exponential equations, which are used to represent the hydrological processes. All input data, i.e., the forcing data of daily precipitation and reference evapotranspiration are expressed in mm over the watershed area. The GR4J model implementation in R (Coron et al., 2023) was used. The parameter optimization in the model code uses the steepest descent local search algorithm. The GR4J model was applied on the 38 watersheds of the study area with streamflow records for the hydrologic years 1980–1981 to 2014–2015 for model calibration and validation.

Daily precipitation, minimum and maximum temperatures were extracted for Cyprus from 18 CMIP5-driven RCMs from EURO-CORDEX (European hub of the international Coordinated Regional Climate Downscaling Experiment; Jacob et al., 2020), covering a historical (1980–2010) and a future period (2030–2060). This ensemble is based on a combination of six driving GCMs and seven RCMs used for dynamical downscaling. The historical period is used as the reference, to which future precipitation changes are computed. The RCM horizontal spatial resolution is 0.11° (∼12 km), and the temporal resolution is at the daily time step. The RCP8.5 scenario, corresponding to the upper range of the projected global-mean surface temperature increases (Meinshausen et al., 2011) was selected to quantify the effects of the least favorable among the plausible climate change scenarios (Pedersen et al., 2021) on surface water resources.

The EURO-CORDEX data were bias-corrected and downscaled using the quantile delta mapping (QDM) algorithm (Cannon et al., 2015). The QDM method preserves the relative (precipitation) or absolute (temperature) changes in quantiles simulated by the RCM. The R MBC Package of Cannon (2024) was used for the QDM application. The CY-OBS area was covered by 55 land-based RCM grid cells. For the downscaling of the climate data from 12 to 1 km resolution, the RCM values were re-gridded to the 1 km grid of the CY-OBS data, and the bias correction algorithm was applied with the 1 km CY-OBS and the re-gridded RCM data. To ensure that bias-corrected values of daily minimum temperature (Tmin) do not exceed those of daily maximum temperature (Tmax), the diurnal temperature range (DTR = Tmax–Tmin) and the Tmax were bias-corrected, following the suggestion by Thrasher et al. (2012), who found that bias-correcting Tmax and DTR resulted in smaller errors.

The 18 bias-corrected and downscaled RCMs were evaluated for their skill in reproducing observed precipitation indices and for their skill in reproducing observed streamflow, when these RCMs are used as forcing for the calibrated GR4J model in the 1980–2010 reference period. The precipitation indices were computed as long-term averages of annual index values. These are (i) the average and (ii) the standard deviation of annual precipitation, (iii) the ratio of the precipitation of the five wettest months to the annual precipitation (W5R; from November to March), (iv) the Simple Daily Intensity Index (SDII), which is the average precipitation rate on wet days (≥ 1 mm) and (v) the number of days with daily precipitation exceeding 10 mm (R10mm). For streamflow, the relative error of the simulated to observed annual streamflow in each watershed was computed. Based on this evaluation, a subset of RCMs best simulating the reference period precipitation and streamflow was selected to be used for modeling the future streamflow with the calibrated GR4J parameterization.

The changes in future precipitation and streamflow, relative to the reference period, were evaluated for the total period changes and the changes in the driest and wettest years, i.e., in the two consecutive driest and wettest years, as well as the five driest and five wettest years overall. The changes in precipitation, reference evapotranspiration (ET) and streamflow were also analyzed for groups of the 38 watersheds, defined by four flow regime types. The four flow regime types are: Permanent flow (P), Intermittent pools flow (I-p), Intermittent harsh flow (I-h), and Ephemeral flow (E). These types of flow regimes are derived from two indices of monthly streamflow data, including the flow permanence (Mf), i.e., the long-term mean annual relative number of months with flow, and the six-month seasonal predictability of dry periods (SD6), based on the ratio of multi-annual frequencies of the zero-flow months for the contiguous six wetter months of the year to the frequencies of the zero-flow months for the contiguous six drier months. A detailed description of the derivation of the two indices and the flow regime type is given in Gallart et al. (2012). The flow regime type for the reference period was assigned by the Water Development Department of Cyprus, and the future flow regime type was computed from the simulated flow, modelled with the climate projections of the different RCMs.

The comparison of the model performances achieved with the six objective functions based on the median values of the 38 watersheds (averaged for all 5-year periods) for each performance measure is shown in Table 1. In the calibration runs, the best evaluation measure values were achieved when the evaluation measure was used as the objective function, shown in bold in the table. These results corroborate that the optimization routine in the GR4J model implementation in R performs reasonably well.

The CSS, which normalizes the evaluation measure values and combines them into one relative performance score, shows that KGE results in the best model performance in the calibration period for 31 out of 38 watersheds, followed by NSEsqrt. Under the conventional split-sample method, the model parameterization obtained with the KGE (0.89) in the calibration would be selected and the validation, based on the full set of 182 5-year periods (14 5-year calibrations × 13 5-year validations), would confirm an acceptable performance (KGE 0.53). However, the evaluation matrices show that the parameterization obtained with the NSEsqrt in the calibration received higher scores for the validation (NSEsqrt 0.74, KGE 0.59) than the KGE-based parameterization (NSEsqrt 0.70, KGE 0.53). In the validation, the NSEsqrt-calibrated model also outperformed the calibrated models in 14 watersheds, whereas the KGE-calibrated model outperformed the other calibrated models in three watersheds only, which is the second worst performance out of the six optimizations. Interestingly, the NSE-based functions led to better model performance, based on CSS, for a larger number of watersheds in the validation compared to the KGE-based functions. The NSE-based functions also achieved a much smaller PBIAS than the KGE-based functions in the validation period, opposite to the lower PBIAS with KGE-based functions in the calibration.

The comparison of the objective functions, based on the CSS and across different long-term average precipitation of the 38 watersheds in Fig. 2, shows that the optimization using NSEsqrt is, on average, the best for all precipitation regimes in the validation for the 182 5-year validation periods, whereas the CSS based on KGE has no important difference from the CSS based on NSE, NSElog and KGElog functions. The patterns in Fig. 2 may be interpreted in relation to both the streamflow transformation and the formulation of the objective functions. The relatively stable or improving performance of the model calibrated with NSElog and KGElog with increasing watershed wetness suggests that logarithmic transformation reduces the influence of the larger and more frequent peak flows, characteristic of wetter catchments. This tendency is most apparent under wetter validation conditions (ΔP>15 %). However, this advantage is not observed during drying validation periods (ΔP≤-5 %), possibly because reduced-flow conditions increase the relative importance of bias and variability errors between simulations and observations. Since KGE explicitly incorporates correlation, bias, and variability components, KGE-based calibration may be particularly sensitive to these changes under drying conditions, even when logarithmic transformation is applied. In contrast, NSEsqrt appears more robust across precipitation regimes. The square-root transformation has a weaker effect in attenuation of peak flows, compared to the logarithmic, and thus it may retain a more balanced representation between moderate and high flows during evaluation across both wetter and drier conditions. In contrast, models calibrated with KGEsqrt show consistently weaker performance across precipitation conditions. These findings suggest that the interaction between transformation type and objective-function structure differs between NSE and KGE formulations, with NSEsqrt providing a more stable compromise across contrasting hydroclimatic conditions.

Subselecting calibration-validation periods with specific changes in climate conditions from the calibration to the validation reveals some model performance losses or gains (Table 1). Simulations in both wetter validation periods (ΔP>15 % and 5 % <ΔP≤ 15 %) achieved comparable performance with simulations in the validation period with small changes in precipitation (-5 % <ΔP≤ 5 %), whereas the evaluation for the drier validation (ΔP≤-5 %) showed the worst performance compared to the other three classes. This is evident in the reduction of the values of the six goodness-of-fit measures and the increase in the PBIAS values, indicating that the observed hydrographs and the total streamflow volumes are simulated with less accuracy in drier validation periods than in wetter validation periods. The NSEsqrt objective function led to the lowest increase, on average, in PBIAS in the driest validation periods, i.e., with a PBIAS of 9 %, compared to the other objective functions, where PBIAS ranged from 17 % to 39 %. The comparison of the objective functions based on the model performance with the CSS for the wettest validation periods (Fig. 2) shows the close distance of the objective functions across all precipitation regimes of the 38 watersheds. For the drying condition in validation, NSEsqrt-calibrated model stands out as the best-performing across all precipitation regimes, with also increasing distance from the calibrated model with the other objective functions for the wettest watersheds.

In the overall comparison of the performance of the model calibrated with six objective functions with the use of the unitless and objective CSS, the model runs calibrated with KGE rank first (0.74) in the calibration, followed by those calibrated with NSEsqrt (0.65). However, in the full set of 182 validation runs, the model calibrated with NSEsqrt ranks first (0.66), followed by those calibrated with NSElog (0.62), KGE (0.60) and NSE (0.59). The model calibrated with NSEsqrt ranks also first in all four precipitation change classes (CSS from 0.66 to 0.70), whereas models calibrated with KGE show more variable ranking. Due to its consistent performance for all climate change conditions, the objective function of NSEsqrt is selected for the model optimization for the future streamflow simulations with the GR4J model.

The optimized model parameters derived using different objective functions and calibration periods were plotted against the average precipitation of the respective calibration periods (Appendix Fig. A1). Although the model parameters do not directly correspond to measurable watershed properties, their calibrated values can provide insight into integrated watershed behavior. In particular, calibrated values for parameter X2 (groundwater exchange) ranged for all watersheds from negative to near zero. In GR4J, negative X2 values imply a loss from the water balance of a watershed. This could be losses to groundwater below the level of the stream (not returning as baseflow) or diversions from streamflow for irrigation, which are processes that are not explicitly represented by the model. The three watersheds with the most negative X2 values have a larger fraction of agricultural land cover (43 %–48 %), whereas the five watersheds with near-zero X2 have a larger fraction of forest cover (74 %–99 %) (Sofokleous et al., 2023). This pattern suggests that the calibrated model correctly accounts for water losses. In addition, the results in Fig. A1 also show that, for a given watershed, parameter values generally vary more between different objective functions than between different calibration periods. Therefore, understanding how different objective functions lead to different model outputs is important for hydrological model applications in areas with diverse watersheds conditions and high climate variability.

To select the optimal GR4J parameter set for the simulation of streamflow in the future period, the NSEsqrt values of the 14 5-year calibration windows and their respective NSEsqrt values in the five validation periods were extracted per watershed. Figure 3 shows the best NSEsqrt value of the 14 5-year calibration runs and the best average NSEsqrt value of the validation of each of these 14 model parameterizations over the 13 validation runs (all), and similarly over the validation runs for each of the four precipitation change classes. The best NSEsqrt values identify the best performing model parameterization (out of the 14) for different climate conditions, for each watershed. The median and interquartile range of the NSEsqrt values are also shown. For the future simulation experiments, the parameter set from the 5-year calibration run, with the highest validation performance in drying conditions, i.e., -5 % <ΔP≤ 5 % and ΔP≤-5 %, in line with the climate model projections for the study area, was selected for each watershed.

Interestingly, nearly all watersheds were simulated with acceptable performance by all calibration sets, with the best and median NSEsqrt of the 14 parameterizations on average equal to 0.88 and 0.83 for the calibration and 0.78 and 0.69 for all validation runs, respectively. The average interquartile range of NSEsqrt obtained with the 14 parameterizations, equal to 0.09 for the calibration and 0.15 for any validation experiment for all watersheds, shows the small variability in the performance of GR4J with most of the 14 parameterizations based on the 5-year calibration periods.

Examining the same results, with the watersheds compared by their long-term average precipitation, it is seen that lower NSEsqrt values, by about 0.1 and greater variability, up to 0.2, among the 14 parameterizations are obtained for the ten driest watersheds (average precipitation 373 to 517 mm yr⁻¹) compared to the ten wettest watersheds (average precipitation 674 to 868 mm yr⁻¹). The 12^th watershed (Gialia near Pano Gialia, 528 mm yr⁻¹ precipitation) stands out as the worst performing out of the 38 watersheds in all validation sets with a median NSEsqrt below 0.2. This particular watershed is only 16 km² in size and has a small dam upstream the streamflow gauge of the watershed. The best performing watershed in all validation sets (Stavros tis Psokas, 621 mm yr-1), with median NSEsqrt above 0.89 has a size of 86 km², which classifies the watershed as medium to large-sized (Fig. 1).

The bias correction of the RCMs with the QDM method yielded model indices very close to the observed values for total precipitation, with a maximum difference of 3 % and for the SDII and R10mm with minimal difference from the observed values for all 18 RCMS (Table 2). The first nine models listed in Table 2 comprise the subset of RCMs used to quantify future changes in streamflow. These nine models simultaneously achieved the lowest error in the simulation of the ratio of the precipitation in the five wettest months to the annual total (W5R) and in the simulation of annual average streamflow. W5R ranged between 0.79 and 0.84 for the selected RCMs and between 0.60 and 0.73 for the other nine models, compared to the observed 0.82 W5R value. The relative error in the simulation of streamflow ranged between 1 % and 11 % in absolute values for the selected RCMs and between 5 % and 40 % for the remaining models. The W5R, as an evaluator of bias-corrected precipitation, has the strongest correlation (0.86) with the error in streamflow simulations compared to the correlation of streamflow error with any of the other four precipitation indices (0.13–0.25).

Precipitation in 2030–2060 is projected to decrease by 16 % according to the driest model and by 6 % according to the median model, relative to the 562 mm yr⁻¹ over the 38 watersheds for the reference period (Table 3). Total streamflow of the 38 watersheds is projected to remain the same in the best model case or decrease up to 39 %, in the worst case, relative to the 1980–2010 reference period value of 173 Mm³ yr⁻¹. The median projected change in streamflow is a reduction of 17 %, and it is given by the NorESM-RACMO RCM. The projected change in streamflow magnitude (14 %–39 %) was nearly double or more than double the change in precipitation (6 %–16 %) for six out of the nine models that provide the largest changes.

Table 3 shows the analysis of the 30-year annual series of precipitation for the driest years in the reference period and the driest years projected in the future period. According to these results, the driest 2-year period in the future could be from 3 % up to 38 % drier that the historical reference, with a median change of 11 %. Streamflow in the two driest years could be 9 % to 70 % less in the future, with a median value of 36 %. The top-5 driest years could have, on average, a median reduction of 17 % in precipitation and 35 % in streamflow. These reductions during these driest future periods are significantly greater than the projected 30-year long-term reductions, indicating that dry years in future will cause even more significant drought conditions than in the past.

A similar analysis for the wettest 2-year period and the top-5 wettest years shows a small increase in precipitation, with a median value of 15 % and 4 %, respectively, and nearly no change in streamflow relative to the streamflow of the wettest years of the past. These results indicate that, while very wet conditions are also expected in the future in some years, as in the past, the magnitude of reduction of water resources in the driest years in the future is disproportionally larger than the positive change in the wettest years.

The future change in precipitation and reference evapotranspiration and the change in streamflow for the four flow type classes of the 38 watersheds are shown in Fig. 4a. Results are presented for the three models with the minimum, median and maximum streamflow projected change among the nine RCMs. The HADGEM/HIRHAM model simulates the maximum streamflow change in the future, from -12 % to -47 %, for the median changes of each of the four flow types (as shown by the medians in the boxplot in Fig. 4a). The same model also projects the highest precipitation reduction (11 %–17 %) and ET increase (7 %–9 %) among the three models. For the median model, the NorESM/RACMO, with streamflow change up to 20 % in the future for the four flow types, the ET increase is the least among the three models and equal to about 4 %. Overall, the streamflow changes in the future from different RCMs have a variability attributed to the variability of both precipitation and ET.

The watersheds with permanent flow (P) and intermittent pool (I-p) flow regime, which contribute 71 % in the total volume of streamflow of the 38 watersheds, are projected to exhibit the highest future reduction in streamflow, up to -47 % and -42 % respectively, relative to the reduction up to 20 % for intermittent harsh (I-h) and up to 12 % for ephemeral streams (E) according to HADGEM/HIRHAM RCM in Fig. 4a. The median model, NorESM/RACMO and the model with the least projected streamflow change, MPI/RACMO, also show that P and I-p flow types will be subjected to the highest reductions relative to I-h and E flow regimes. The flow regime type will also become subject to change, implying a negative impact on the ecological conditions in the area of the 38 watersheds. Figure 4b shows that two watersheds with P, three with I-p and one with I-h flow regime have the highest probability (five or more models out of the nine) for a shift in their flow regime, while another one watershed with P, eight watersheds with I-p are and three with I-h are also projected to exhibit a change in flow regime, according to fewer than five out of the nine models. Overall, 18 out of the 38 watersheds are indicated to exhibit a change in their flow regime according to at least one of the nine RCMs.

This study examined the changes in the hydrological model performance when the model is used at contrasting climate periods with multiple performance measures used for the calibration and validation. Although the Differential Split Sample Test approach has been used by several studies in the context of model transferability, with different lengths in the examined sub-periods, the transferability performance was evaluated with different or fewer number of measures than those used for the model calibration. For instance, Dakhlaoui et al. (2017, 2019) used 7–8 year long sub-periods or sub-periods comprised of discontinuous years with KGE as the calibration measure and NSE and Volume error as the transferability evaluators. Guo et al. (2020) used the KGE for calibration of sub-periods separated by 3-month windows and validated the model using KGE and the three components of KGE. The evaluation of the model calibrated with six objective functions in the current study provides a comprehensive evaluation of how KGE and NSE, with no transformation, the logarithmic and the square transformation of streamflow values affect model performance during calibration and validation under contrasting climate conditions. KGE was found to be the optimal measure for model optimization, for the calibration period only. The model calibrated with it ranked second in the validation periods with increased wetness (ΔP≥ 15 %) and third or fourth in the full validation set and the validations with a small change in precipitation or increasing dryness (-5 % <ΔP≤5 % and ΔP≤-5 %). This outcome for the 5-year calibration runs of the study with KGE, a widely used objective function and evaluation measure, could be further investigated by assessing the model calibration and validation performance using the three components of KGE as objective functions, separately. Fowler et al. (2018) found that models calibrated with a split-KGE, i.e., computed individually per year and averaged, performed better than when KGE was computed for the full period.

The value of the composite score CSS, which normalizes and combines the values of seven evaluation measures, was highest for the NSEsqrt objective function. This finding agrees with Seiller et al. (2017), who found that the square root transformation with NSE is better, on average, for all flow regimes than calibration on NSE with inverse transformation or no transformation when the calibrated model is transferred to warmer conditions for catchments in southeastern Canada. These authors also showed that the square root transformation is optimal at the expense of some loss of confidence for very high flows. In addition to the best relative model performance with NSEsqrt as objective function, in warmer conditions, according to these authors, the current study showed that NSEsqrt was the optimal measure for warmer and wetter, as well as for warmer and drier conditions in the validation. The model runs calibrated with NSEsqrt ranked also first with increasing distance from the second highest ranked objective function with increasing precipitation reduction (ΔP) in the validation period. This result suggests that NSEsqrt becomes more robust for increasing precipitation reduction in the validation period, relative to the calibration period. The comparison of the hydrological model performance for 14 GR4J parameterizations with NSEsqrt for the different validation experiments showed that drier watersheds, i.e., with lower precipitation than their wetter counterparts in the same Mediterranean study area, are simulated with lower values and higher variability of the performance measure, among the different parameterizations. These findings are in line with the study of Muñoz-Castro et al. (2023), who found that the choice of the objective function was more important for high-aridity and lower runoff coefficients than for wetter climates and that a higher model performance and higher parameter agreement was achieved for wetter basins in their study of 92 basins in continental Chile.

Dakhlaoui et al. (2017) found that a 25 % difference in precipitation and 1.75 °C in temperature were the thresholds for acceptable model transferability for streamflow simulations with changing climate conditions. The difference in climate conditions of the calibration-validation periods, used for the comparison of the objective functions in the present study, had a maximum temperature change of 1.5 °C and precipitation change range from -14 % to +25 %, except one calibration-validation set with precipitation change of 32 %.

The long-term average precipitation data from the EURO-CORDEX ensemble, evaluated before and after bias correction (Table 2), shows that this was an essential step for obtaining values close to the observations. The good performance of the Quantile Delta Mapping (QDM) method used here was highlighted in other studies (e.g., Meyer et al., 2019). However, 9 out of 18 RCMs showed poor reproduction of the seasonal cycle of precipitation. This leads to the conclusion that if an RCM fails to reasonably represent the seasonality of the reference period, it may also not be able to model future shifts. Mascaro et al. (2018) found that an ensemble of raw EURO-CORDEX RCMs underestimated winter precipitation in Sardinia, which has a similar climate and geographical extent as Cyprus. These authors' finding is in line with lower simulated precipitation in the five wettest months (WR5) by the RCMs in the current study, which was found to be correlated with the errors in simulated streamflow.

Comparison of CMIP5 and CMIP6 models in previous global and regional studies showed that uncertainty in precipitation is high in both generations of models (Cos et al., 2022; Wu et al., 2024). For the Mediterranean, the IPCC Working Group I Interactive Atlas (Gutiérrez et al., 2021) reports a median precipitation change ranging from -7.9 % to -18.3 % for CMIP5 and a change from -7.1 % to -17.8 % for CMIP6 models for different global warming levels and for pathways RCP8.5/SSP5-8.5 (relative to the pre-industrial period). For the Eastern Mediterranean and specifically Turkey, Bağçaci et al. (2021) found that the precipitation decline is 2.5 % smaller and the temperature increase is up to 0.35 °C higher in CMIP6 compared to CMIP5. Despite the differences reported for the latest CMIP6- relative to CMIP5-GCMs, regionally downscaled CMIP6 models are not produced at the time of preparation of this manuscript, which does not allow a direct comparison with CMIP5 RCMs (EURO-CORDEX) used here. This work showed that bias correcting the RCM of CMIP models is necessary for the quantification of regional impacts of climate change, as shown by the errors with and without bias correction of RCM outputs for annual totals and seasonal distribution of precipitation, described in Sect. 3.3.

The projected median change in future precipitation for 2030–2060 under RCP8.5, relative to 1980–2010, of the nine models (-6 %, with a range from 2 % to -16 %), is in line with the overall observed drying of the island. These changes are within the range reported by other studies for the Mediterranean as a whole, about 10 % in the first half and up to 20 %–40 % in the second half of the 21st century (Zittis et al., 2021a). The projected median change of -17 % for streamflow, with a range from 6 % to -39 % from the nine RCMs, indicates a pathway of decreasing freshwater resources by the mid-21st century for Cyprus. Projected streamflow changes for two catchments on the southern slopes of the Troodos mountains by 2050 (up to 24 % and 17 %), as presented by Ragab et al. (2010), are within the projections of the current study.

Dakhlaoui et al. (2019) reported for 2040–2070, relative to 1970–2000, median runoff changes up to -6.2 % under RCP4.5 (scenario not examined here) and from -13 % to -31 % under RCP8.5 for different RCMs and basins in Tunisia. For other Mediterranean locations and the end-of-the-century horizon, which was not analyzed here, runoff reductions seem to be close to 20 % under RCP4.5, i.e. 19 % in southern Italy (Senatore et al., 2022; reference period 1975–2005) and 16 %–19 % in Tunisia (Dakhlaoui et al., 2019). For the same horizon, relative to various 30-year periods in the second half of the 20th century, under the less optimistic, business-as-usual scenarios, i.e., B2 or RCP8.5, streamflow reductions will be exacerbated, i.e., 26 %–54 % in southern France (Lespinas et al., 2014), 50 %–60 % in central Spain (Sánchez-Gómez et al., 2023), and 37 %–57 % in Tunisia (Dakhlaoui et al., 2019).

The differences in the magnitude of precipitation and streamflow projected changes are comparable to the differences in the magnitudes of changes computed from two different periods in the past (1916–1969 and 1970–2000), as found in a study for the islands' water resources trends (Water Development Department, 2002). Part of this difference can be attributed to the increased reference ET in the watersheds in the future (4 %–8 %), driven by the robust warming (Zittis et al., 2021b; Lazoglou et al., 2024). The combined effect of increased ET and reduced streamflow suggests potential reductions in surface water availability, which may exceed the current natural replenishment of dams and groundwater. However, rainfall-runoff relations are non-linear because less runoff tends to be generated under drier watershed conditions. Additional pressures on streamflow include declines in groundwater levels, exacerbated by higher extractions for agricultural production in dry years (Zoumides et al., 2013; Leduc et al., 2017). Gutiérrez-Jurado et al. (2019) showed how the initial head in groundwater levels influences the onset of streamflow in Mediterranean climates. Previous modelling experience in Cyprus showed the sensitivity of simulated streamflow to the groundwater processes in hydrological modeling (Ragab et al., 2010; Camera et al., 2020; Sofokleous et al., 2023).

The change of flow regime type is most likely to occur for the wettest watersheds (Fig. 4). This was also found by Raymond et al. (2019) who related the shift from a continuous flow regime to intermittent to the lower precipitation and increased number of consecutive days without rain. Pascual et al. (2015) showed that from the comparison of flow changes in three watersheds in northeastern Spain, the two wettest watersheds were expected to experience larger reductions in streamflow (34 %) than the drier watershed (25 %). Schneider et al. (2013) found that the most remarkable flow regime alterations in Europe are expected in the Mediterranean. The results of this climate impact assessment on the fresh water resources in Cyprus contribute to the broadening of the knowledge of changes in streamflow at watershed level in the climate hotspot region of the eastern Mediterranean.

As shown with the range of changes for precipitation and streamflow presented above, uncertainty in future projections, particularly in climate change impacts on streamflow, is large. Given that the impacts of climate change are derived from a chain of modelling steps, the uncertainty in the future projections for streamflow could be identified for each step. In this study, multiple RCMs were used to drive the hydrological simulations. The emphasis was placed on multiple RCMs rather than multiple hydrological models, as previous research has shown that for mid- to high-flow and mean annual flow conditions, variability in rainfall–runoff model outputs is greater when runoff projections are based on a single rainfall–runoff model combined with multiple climate models (Teng et al., 2012; Petheram et al., 2012). Other studies also showed that a significant source of uncertainty comes from the RCMs for studying climate impacts on streamflow (Teutschbein and Seibert, 2012). RCMs exhibit systematic biases in both temperature and precipitation, influenced by the individual structures of the GCMs and RCMs. For this reason, a bias correction step is necessary (Christensen et al., 2008). In this study, 18 different RCMs were selected, representing combinations of six GCMs downscaled by seven RCMs.

A single hydrological model, GR4J, was used in this study, as the particular model is a well-established conceptual model structure for streamflow simulations across a wide range of hydroclimatic conditions. Its performance and robustness have been found to be comparable to those of a multimodel approach (Seiller et al., 2017). However, uncertainties also arise from the hydrological model structure used in impact studies. A particular limitation of “bucket-type” rainfall–runoff models, such as GR4J, is their tendency to underestimate multi-year drought conditions due to finite storage, which limits their ability to represent sustained groundwater decline (Fowler et al., 2020). The use of model parameters calibrated during observed dry conditions with trends similar to the projected climate, as suggested in this study, and structural modification of these models to account for catchment memory (e.g. Grigg and Hughes, 2018; Hughes et al., 2021) could help reduce this underestimation.