In each box plot of the subsequent figures, boxes are defined by the 25th and 75th percentiles, the horizontal line within the box represents the median, whereas the whiskers represent the non-outlier range. Outliers are marked as circles (they are present only in some figures). Independent two-sample t test was applied for testing whether the differences between (temperature, precipitation and flow) change projections are statistically different between two sources.
The comparison of annual and seasonal change of projected temperature between the reference and near future (NF) and far future (FF) periods used in the HBV and SWAT models in eight studied catchments, shown as box plots across climate models, demonstrates that there is very little difference between the two sources (Fig. S1, Supplementary Material). The warming is ubiquitous and fairly uniform spatially. There is little difference between seasonal temperature increases in the near future, and a considerable difference in the far future, with the highest level of warming occurring in the winter season (mean across all catchments, climate models and sources was 4.2 °C), and the lowest in summer season (3.1 °C). In general, climate model spread is higher than the differences between catchments. Model spread varies across seasons and future horizons. For example, for winter season, it is higher for the near future than for the far future, whilst for spring and summer season, the opposite takes place. The lowest spread occurs in summer in the near future (below 0.5 °C for all catchments), and the highest for spring in the far future (above 2 °C for five catchments).
In order to compare temperature change signals from both sources in a more explicit manner, differences in temperature change projections between CHASE and CHIHE (cf. Eq. 1) were calculated (Fig. 2). The results for the far future are shown, as in the near future the differences were generally lower. There is a good agreement between projections from both projects for mean annual and summer and autumn season temperatures, in which case the differences in terms of the absolute values are lower than 0.5 °C for every catchment and model. However, larger discrepancies can be observed for the Wisła catchment in spring and summer, and the Nysa catchment in spring. In the worst case, CM1, CM2 and CM4 show the increases in temperature in the Wisła catchment larger by 0.8–1.0 °C in CHASE (SWAT) than in CHIHE (HBV). In contrast, projections used in HBV show a larger increase in spring temperature by 0.6–0.8 °C than projections used in SWAT for CM5 and CM7 in the Wisła and Nysa catchments.
Precipitation forcing data used in HBV and SWAT were not as consistent between each other as air temperature data (Fig. 3). In 95 out of 112 individual cases (eight catchments, seven climate models, two time horizons), projected changes in annual sums of precipitation used as HBV input were higher than the corresponding changes in the SWAT input. Mean projected change across these 112 cases is 12.7% in HBV and 9.3% in SWAT. In three catchments (marked with asterisks in Fig. 3), Narewka (near future), Wisła and Myśla (far future) the differences were statistically significant at the level = 0.05. Overall, projections from both sources agree that wetter conditions are expected for both horizons and all catchments. In the far future, a higher magnitude of increase in precipitation is projected for catchments situated on the Polish Plain than for mountainous catchments.
Projections of seasonal precipitation change also differ between the inputs used by the HBV and SWAT (Fig. 3). For all four seasons, the arithmetic mean of changes across 112 individual cases is higher for HBV input than SWAT input by 3–5%. Although statistically significant difference occurs only once (spring precipitation in the Narewka catchment in the far future), there are more cases with visually different results (e.g. winter precipitation in the Wisła catchment or summer precipitation in the Myśla catchment in the far future). Spatial differences in projected changes are much higher for precipitation than for temperature. However, patterns differ among seasons and time horizons. Climate model spread also varies: for example, in the near future in winter the mean spread across catchments is nearly the double of the corresponding value for summer (which is approximately 15%), for both sources of precipitation input. A further insight into the differences in precipitation change projections between the two sources is possible through analysing the differences between the projected changes used as SWAT input and projected changes used as HBV input for various catchments and models (Fig. 4). As in Fig. 2, only the results for the far future are shown, because the magnitude of difference for the near future is much lower. They strongly vary with temporal aggregation, catchments and climate models, but the last factor has the most pronounced effect. The main message is that, with very few exceptions, projections of changes used as SWAT input are lower than corresponding projections used as HBV input. At annual level, in the majority of cases the difference (Δ) is about 5%. In the lowland catchments (Oleśnica, Flinta, Myśla and Narewka) the projections of precipitation used in the HBV model are larger by more than 5% than the corresponding projections used in the SWAT model for some climate models and even reach 15% for Myśla for the CM2. Much higher differences can be noted at seasonal level, particularly for CM2 in spring and summer and for the Myśla catchment for spring, summer and autumn.
For each climate model, there exists at least one season-catchment combination for which the difference is higher than 15%. Due to the fact that in all catchments summer precipitation has the highest share in annual precipitation, the differences in summer contribute most to the differences observed at the annual level. In three cases, the difference between projections from both sources is large and consistent between climate models:
In three mountainous catchments, Dunajec, Wisła and Nysa, the difference in winter precipitation change is higher than 10% for six (Dunajec) and four (Wisła and Nysa) climate models.
In the Narewka catchment, five climate models, and in Myśla, one climate model, suggest that the change in spring precipitation used as SWAT input is lower by at least 15% than the corresponding change used as HBV input.
In summer, for all the lowland catchments except Narewka, the difference in precipitation is higher than 10% for a number of climate models.
In autumn the spread of differences larger than 10% is more equally distributed among the catchments, with only three models showing 15% difference for Myśla and Narewka.
Flow bias in the reference period
Prior to analysing projected changes in river flow in selected catchments, it is worth to analyse the bias of simulation of mean annual and seasonal flows by HBV and SWAT driven by the ensemble of climate models in the reference period (Fig. 5). Ensemble median bias in mean annual flow is within the range of −/+25% for both hydrological models and all catchments except Myśla. In the latter, ensemble median bias is over 40%, and for SWAT there are three climate models for which it is above 70%. The biases in seasonal flows vary among seasons and catchments. Both HBV and SWAT tend to underestimate mean winter flow in mountainous catchments, HBV underestimates mean spring flow in mountainous catchments, whilst SWAT overestimates it in all catchments except Wisła. The mean summer flow is overestimated in all mountainous catchments by HBV and underestimated by SWAT (except for the Biała catchment). Finally, HBV simulations of the mean autumn flow have low bias in mountainous catchments, whereas in SWAT it is underestimated in all catchments except Biała.
A more complex picture is visible for the lowland catchments (Fig. 5). For the Oleśnica catchment, HBV flow projections have lower bias than SWAT simulations for spring and summer seasons, but the opposite takes place in autumn and winter seasons. Fairly similar pattern exists for the Flinta catchment, with one exception: here, both SWAT and HBV overestimate mean summer flow. The Myśla catchment is the most challenging case for both models—SWAT considerably overestimates mean flow for all seasons, whereas HBV has totally different results for winter and spring (low bias) and summer and autumn (very high ensemble median bias of 184 and 110%, respectively). For the Narewka catchment, the HBV bias is low in all season but summer, whilst the SWAT model underestimates autumn and winter flow. In general, the SWAT spread of bias is much larger for all lowland catchments than the HBV spread for the same catchments.
River flow changes
In the second stage of analysis, the projected changes in mean annual and seasonal flows simulated by the HBV and SWAT models for two future periods were calculated. Both hydrological models agree on the dominating upward direction of change in all catchments and time periods, although four mountainous catchments: Biała, Dunajec, Nysa and Wisła, have variable directions of change in the near future (Fig. 6). In general, there is a high agreement among hydrological models on the magnitude of change in the near future (no statistically significant differences, at \(p = 0.05\)), whereas in the far future there is a high disagreement (all differences statistically different, at \(p = 0.05\)). Both, the magnitude of projected change, and the climate model variability grow with time for most catchments. In most cases, for the lowland catchments, mean annual projections obtained with the help of SWAT model for the far future provide increases lower by the factor of two to three than the corresponding increases obtained with the help of the HBV model. The disagreement between the hydrological models is lower for four highland catchments, for which the magnitude of change is also lower.
Figure 6 shows projected changes in mean seasonal discharge according to the HBV and SWAT models. As with the mean annual discharge, there is a relatively good agreement between hydrological models in the near future and a much weaker agreement in the far future. In winter season, increases are dominating throughout all combinations of catchments, models and periods. Despite the fact that SWAT gives lower increases in the mean annual discharge in the far future, it projects higher increases in the far future mean winter discharge in mountainous catchments than HBV. For seven out of 16 cases, SWAT flow projections are statistically higher than HBV projections in winter (\(p = 0.05\)).
For spring season in the near future, climate model variability dominates and both HBV-based and SWAT-based projections show changes in different directions (Fig. 6). The Flinta catchment is the only one for which there is a statistically significant difference between HBV and SWAT. In contrast, in the far future, mean spring discharge is projected to increase according to the HBV model in each catchment, whereas according to SWAT it is projected to increase only in lowland catchments (and by a lower percentage). In each case in the far future, the difference is statistically significant at the level \(p = 0.05\). In mountainous catchments SWAT projects changes in variable directions, although decreases are dominating.
Projected flow changes in summer and autumn season between the hydrological models in the near future are very similar but they considerably diverge in the far future (Fig. 6). Indeed, in none of the 16 cases is the difference in flow projections in the near future statistically significant (\(p = 0.05\)). Increases in summer flow projected by SWAT in the far future in all catchments are statistically significantly lower than corresponding increases projected by HBV. For some climate models in the lowland catchments, increases projected by the HBV model are three-fold, and in one case (summer flow in the Oleśnica catchment) they are almost six-fold. For the Dunajec catchment, summer flow simulated by SWAT is projected to decrease according to most of climate models, whereas changes in the opposite direction are simulated by the HBV model. On the other hand, there is a fairly good agreement among the two hydrological models on projected changes of flow in the autumn season in the mountainous catchments and the Narewka catchment.
Figure 6 shows changes in flow indices expressed as percentage. It should be noted that the comparison of changes between catchments would give different outcomes if they were expressed as differences between future horizons and the reference period (e.g. in mm). The reason for this is a large variability in runoff coefficients and precipitation among catchments (cf. Fig. 1; Table 2). For example, a 10% increase in mean annual flow in the Dunajec catchment corresponds to an increase in runoff by more than 70 mm per year. Such a magnitude of increase in runoff in the Flinta catchment would correspond to an increase in mean annual flow by more than 80%.
Some insight into the way that hydrological models transform the climate change signal can be obtained from a simple analysis of mean annual flow change (ΔQ) as function of mean annual precipitation change (ΔP) (Fig. 7) for particular combinations of climate models, hydrological models, catchments and projection horizons. There are large differences between the nature of this response for different catchments and models.
For the HBV model results, catchments have different sensitivity for the near future and far future. The catchment sensitivities increase for far future but also become much more scattered, in particular for the lowland catchments. Mountainous catchments show nearly the same response to changes in precipitation in near- and far future. However, in the lowland catchments the changes of flow in the response to changes in precipitation are doubled in far future.
The SWAT model results show similar variability of sensitivity as the HBV model for the near future and much smaller scatter of the sensitivity values for the lowland catchments in the far future. The mountainous catchments show the sensitivity for the far future stabilising around 20–30%, which is lower than for the same precipitation change in the HBV model. This is consistent with the fact that changes in mean annual flow projected by SWAT were found to be lower than those for HBV in the far future (Fig. 6). In the SWAT model, the nature of flow response does not strongly depend on the projection horizon.
Despite some differences between SWAT and HBV revealed by this analysis, there are also some similarities, even in the far future. For example, the four lowland catchments are positioned in the same zones of the scatter plots for HBV (Fig. 7b) and SWAT (Fig. 7d). More precisely, two models consistently show that the Narewka catchment has a weaker response in mean annual flow change to a relatively high precipitation change than the Flinta catchment, even though the latter is exposed to a lower magnitude of precipitation change than the Narewka. Two other lowland catchments, i.e. the Oleśnica and the Myśla, are situated in the middle zone of Fig. 7b, d, between points representing the Flinta and the Narewka catchments.
Flow change vs. runoff coefficients
In order to examine the spatial variability of projected changes in mean annual flow, they are plotted in Fig. 8, for each climate model against the runoff coefficients of each catchment (cf. Fig. 1 for the map of runoff coefficients). Results for the far future are only shown, as the magnitude of changes and differences are higher in this case than in the near future (cf. Figs. 6, 7). Although the number of catchments is rather small, some clear relationships are present. Both HBV- and SWAT-based projections show that in catchments with lower runoff coefficients, a higher magnitude of changes can be expected. Although regression lines were not plotted, for both hydrological models and each climate model (except for the SWAT-CM2 combination) the relationship has a convex parabolic shape, with a vertex situated for runoff coefficient values between 0.5 and 0.7. In the case of CM2, the results are consistent with Fig. 3: this is the climate model for which precipitation projections used as SWAT input are considerably lower in the lowland catchments than projections used as HBV input. This difference is now reflected in Fig. 8: mean annual CM2-driven flow change projected by SWAT in three catchments with the lowest runoff coefficients (Flinta, Myśla and Oleśnica) is equal to approximately 10%, whereas the corresponding change projected by HBV yields 80%. Certainly, this analysis neglects the effect of precipitation and temperature changes that vary across catchments and climate models.
Figure 8c shows the differences in mean annual flow change projections between SWAT and HBV as a function of runoff coefficient. This relationship also has a parabolic form, although this time concave. Four clusters of catchments characterised by decreasing values of runoff coefficients are also characterised by large discrepancy between SWAT-based and HBV-based projections. More specifically: (1) the Dunajec and Wisła catchments have mean runoff coefficient equal to 0.58, and mean difference in mean annual flow equal to 13%; (2) for the Biała and Nysa catchment, the same parameters are 0.41 and 20%; (3) for the Narewka catchment it is 0.27 and 36%; (4) for the last cluster composed of the Flinta, Myśla and Oleśnica catchments it is 0.17 and 71%.