Four catchments located in different parts of Finland were selected as study catchments (Fig. 2). These represent different hydrological regions in Finland. Loimijoki (Maurialankoski observation station, catchment area 2650 km2; lake percentage 3.1) is a medium sized river with high proportion of cultivated area on clay soils. Nilakka (catchment area 2160 km2; 18 % lake percentage) and Lentua (2050 km2; 13 %) observation stations are located at lake outlets in central Finland characterized by numerous lakes. Ounasjoki (Marraskoski observation station, 12 300 km2; 2.6 %) is a large river in northern Finland (Fig. 2) (Korhonen and Kuusisto, 2010). All the study catchments have long water level and discharge observation series, the longest from 1912 onwards (Lentua) and the shortest from 1935 onwards (Loimijoki).

Annual mean runoff in the study catchments varies from 280 to 370 mm. Runoff has a distinct seasonal variation with low values during winter and summer and a maximum in spring due to snowmelt. The average maximum snow water equivalent varies from 80 to 100 mm in the southern catchment (Loimijoki) to 180 mm in the northern Ounasjoki catchment (Perälä and Reuna, 1990). Annual soil and lake evaporation gradually decrease from southern Loimijoki (soil 400 mm; lake 540 mm) to northern Ounasjoki (soil 220 mm; lake 310 mm) (Hyvärinen et al., 1995). Autumn precipitation causes a second runoff peak, which is usually smaller than the spring peak. The spring floods are more pronounced in northern and central Finland (Ounasjoki, Lentua, Nilakka), while in southern Finland (Loimijoki) heavy rains in summer and autumn or rains with snowmelt in winter may cause major floods as well.

Bias corrections were calculated for the entire Finland including transboundary watershed areas in Norway, Sweden and Russia. The gridded data sets needed for the bias correction were calculated using observations from approximately 190 stations with daily temperature measurements at 2 m height and 250 stations with daily precipitation measurements from the Finnish Meteorological Institute (FMI). Additional observations from 11 temperature and 16 precipitation observation stations in Norway, Sweden and Russia were provided by the Norwegian Meteorological Institute, the Swedish Meteorological and Hydrological Institute (SMHI) and the Hydrometeorological Centre of Russia. Observations from 1961 to 2000 were used although the observation network varies during this period.

Gauge precipitation observations especially for snowfall contain various systematic measurement errors (Førland et al., 1996; Taskinen, 2015), which need to be corrected before they can be used for bias correction of RCM data. The correction of precipitation measurements consisted of the exposure method for aerodynamic correction as well as wetting and evaporation corrections (Taskinen, 2015). The areal values of the meteorological observations are calculated for each sub-basin of the hydrological model from three closest observation stations by inverse distance weighting taking into account the elevation differences. The areal values were converted to the same regular 0.25∘lat ×0.25∘long grid as the RCM data.

The observations of relative humidity at 2 m and wind speed at 10 m are used in the simulation of lake evaporation, which is an important hydrological variable for catchments in the lake area. The areal values are calculated in a similar way as temperature and precipitation and the effect of fetch to the wind speed on a lake is calculated as in Resio and Vincent (1977).

Five climate scenarios are used from four different RCMs forced with four different GCMs as given in Table 1 with acronyms used hereinafter. Selected RCM projections are the same as used in Veijalainen et al. (2012), excluding RCA-ECHAM5 (see Table 4), to enable comparison of results. The data were retrieved from the ENSEMBLE (ENSEMBLE-based Predictions of Climate Changes and their Impacts) project's research team 3 database (ensemblesrt3.dmi.dk; van der Linden and Mitchell, 2009). The GCMs were run under historic (1961–2000) and with A1B scenario (2001–2100) forcing. The GCM output was then used as boundary conditions to force RCMs over a common European domain in a regular 0.25∘lat ×0.25∘long grid (van der Linden and Mitchell, 2009).

The distribution based scaling (DBS) method described, e.g., in Yang et al. (2010) and Teutschbein and Seibert (2012) was used to scale temperature and precipitation time series to better represent observed distributions. The correction procedures using cumulative distribution functions (CDFs) are shown in Fig. 3. In this study CDFs are constructed on a daily basis for temperature and for all days with certain months for precipitation. The method of maximum likelihood is used to estimate distribution parameters.

Temperature (T) is described by a Gaussian (normal) distribution with daily mean (μ) and standard deviation (σ). The DBS approach for temperature included four steps: (1) to take into account the dependence between precipitation and temperature, the temperature data were divided into wet and dry days resulting in two sets of parameters; (μw, σw) for wet days and (μd, σd) for dry days, hereafter referred to as (μw–d, σw–d). The separation was conducted after excessive drizzle days were removed (described below; Eqs. 5, 6). In this study we also use the distribution parameters without wet–dry state separation (μ, σ). (2) To take into account seasonal variations, daily mean and standard deviation were calculated using a 15-day moving window and (3) were further smoothed with Fourier series with five harmonics on a daily basis over the control period 1961–2000, as in Yang et al. (2010). (4) These smoothed daily mean and standard deviation for each grid point were then used to calculate the daily (d) CDFs for observations (μobs, σobs) and RCMs (μcontr, σcontr) for the control period (Fig. 3). DBS parameters for the control period were used also to adjust the scenario (scen) runs. DBS procedure expressed in terms of the Gaussian CDF without wet–dry separation: Tcontr(d)=F-1FTcontr(d)μcontr,σcontr2μobs,σobs2,Tscen(d)=F-1FTscen(d)μcontr,σcontr2μobs,σobs2. DBS procedure expressed in terms of Gaussian CDF with wet–dry separation: Tcontr, w–d(d)=F-1⋅FTcontr, w–d(d)μcontr, w–d,σcontr, w–d2μobs, w–d,σobs, w–d2,Tscen, w–d(d)=F-1⋅FTscen, w–d(d)μcontr, w–d,σcontr, w–d2μobs, w–d,σobs, w–d2. For precipitation (P) single and double gamma distributions were used in four steps. In contrast to Yang et al. (2010) where the DBS parameters (shape α and scale β) were estimated seasonally, we estimated DBS parameters on a monthly basis. Single CDF for certain month is used for the whole time slice (1961–2000). Also seasonally optimized parameters were tried out, but these produced too high monthly precipitation sums for Finland (not shown) and thus were not used. (1) For both distributions, excessive drizzle days in the RCM data were first removed by defining a cut-off value (Pth,contr,m) that reduced the percentage of wet days in the RCMs to that of the observations on a monthly (m) basis. In this study only days with observed precipitation larger than 0.1 mm (Pth,obs,m) were considered wet days, and the rest dry days. A monthly precipitation threshold value for each RCM control run (Pth,contr,m) was then set to the cut-off value so that the percentage of RCM simulated and observed wet days matched (Eq. 5). Due to the stationary assumption the same threshold value was used to reduce the drizzle days for a future period to enable the scenario run to have different wet day frequency than the control run (Eq. 6). Precipitation amounts smaller than the threshold value were not redistributed to the remaining wet days. Pcontr(d)=0,ifPcontr(d)<Pth, contr,mPcontr,otherwisePscen(d)=0,ifPscen(d)<Pth, contr,mPscen,otherwise (2) The remaining daily precipitation was adjusted to match the observed frequency distribution using single gamma distribution (Eq. 7). (3) To better capture the extreme precipitation events a double gamma distribution was also used, then the observed and RCM generated precipitation distributions were separated into two by the 95th percentile of CDF (Pobs,95th, Pcontr,95th), resulting into two sets of parameters: (α1, β1) for below the 95th percentile precipitation and (α2, β2) above it. (4) These monthly parameters for each grid point were then used to calculate the CDFs for observations (αobs, βobs) and RCMs (αcontr, βcontr) during the control period (Eqs. 9, 10; Fig. 3). Monthly DBS parameters for the control period and the 95th percentile threshold (Pcontr,95th) were also used for the scenario (scen) runs (Eqs. 8, 11, 12). The DBS procedure expressed in terms of single gamma CDF: Pcontr(d)=F-1⋅FPcontr(d)|αcontr,m,βcontr,m|αobs,m,βobs,m,Pscen(d)=F-1⋅FPscen(d)|αcontr,m,βcontr,m|αobs,m,βobs,m, The DBS procedure expressed in terms of double gamma CDF: Pcontr,1(d)=F-1⋅FPcontr(d)|αcontr1,m,βcontr1,m|αobs1,m,βobs1,m,ifPcontr(d)<Pcontr,95th(m),Pcontr,2(d)=F-1⋅FPcontr(d)|αcontr2,m,βcontr2,m|αobs2,m,βobs2,m,ifPcontr(d)≥Pcontr,95th(m),Psken, 1(d)=F-1⋅FPscen(d)|αcontr1,m,βcontr1,m|αobs1,m,βobs1,m,ifPscen(d)<Pcontr,95th(m),Psken,2(d)=F-1⋅FPscen(d)|αcontr2,m,βcontr2,m|αobs2,m,βobs2,m,ifPscen(d)≥Pcontr,95th(m). Wind speed and specific humidity of the RCM data were corrected by adding the monthly mean differences between the observations and the RCMs. The same corrections were used in the scenario periods. Since the wind speed and specific humidity affect only the calculation of lake evaporation in the hydrological model, it is assumed that this simple bias correction works sufficiently well to achieve corresponding water level and discharge distribution as with observed input variables.

The hydrological model used in this paper was from the watershed simulation and forecasting system (WSFS). It is a conceptual hydrological model developed and operated at Finnish Environment Institute (SYKE) (Vehviläinen et al., 2005). The WSFS is used as the national hydrological forecasting and flood warning system (Finnish Environment Institute, 2015) as well as for research purposes (e.g. Veijalainen et al., 2012; Jakkila et al., 2014; Huttunen et al., 2015). The conceptual rainfall–runoff model in the WSFS is based on the HBV (Hydrologiska Byråns Vattenbalansavdelning) model structure developed at SMHI (Bergström, 1976), but the models differ from each other, e.g., in the river routing, catchment description and in some process models such as the snow model (Vehviläinen, 1992; Vehviläinen et al., 2005). HBV-type models have been used in several climate change impacts studies in different parts of the world (e.g. Steele-Dunne et al., 2008; van Pelt et al., 2009), most commonly in Scandinavia (e.g. Andréasson et al., 2004; Beldring et al., 2008)

The WSFS hydrological model consists of small sub-basins, numbering over 6000 in Finland with an average size of 60 km2 (20–500 km2) (Vehviläinen et al., 2005). The water balance is simulated for each sub-basin, and sub-basin are connected to produce the water balance and simulate water storage and transfer in the river and lake network within the entire catchment. The sub-models in WSFS include a precipitation model calculating areal value and form for precipitation, a snow accumulation and melt model based on the temperature-index (degree-day) approach, a rainfall–runoff model with soil moisture, sub-surface and groundwater storages, and models for lake and river routing.

The WSFS was calibrated against water level, discharge and snow line water equivalent observations from 1981 to 2012. The Nash–Sutcliffe efficiency criterion R2 (Nash and Sutcliffe, 1970) for the control period 1961–2000 in the four case study catchments was 0.78 for Loimijoki, 0.80 for Nilakka, 0.87 for Lentua and 0.87 for Ounasjoki. The R2 values within the calibration period 1981–2000 are considerably better than in the validation period 1961–1980: 0.84 and 0.71 for Loimijoki, 0.91 and 0.68 for Nilakka, 0.92 and 0.81 for Lentua and 0.87 and 0.88 for Ounasjoki, respectively, for calibration and validation periods. The reasons for remarkably lower values in the validation period are the possible changes in rating curves in Loimijoki and Nilakka and the change of the rain station gauges from Wild to Tretjakov-type gauges. The measurement errors for different gauge types are done separately (Taskinen, 2015), but the uncertainty range of wind effect on snowfalls is much larger for Wild than Tretjakov.

The five RCMs used in this study are able to capture the annual cycle of temperature in the control period quite well, but monthly temperatures are commonly underestimated throughout the year except in winter by RCA and REMO and in autumn by HIRHAM-A (Fig. 4). The CDFs show that all RCMs cumulate too much below 0 ∘C temperatures and too little above 0 ∘C temperatures especially in spring, although also in winter and autumn (Fig. 5).

There are prominent differences in the ability of RCMs to capture the annual cycle of precipitation during the control period (Fig. 4). All models in this study heavily overestimate precipitation accumulation almost throughout the year with some exceptions in summer and winter. In particular, HIRHAM-A and HIRHAM-B produce too much precipitation in spring and autumn and are too dry in summer. The overestimation in accumulated precipitation is relatively largest in spring, varying from 2.6–61 % in Nilakka to 24–81 % in Ounasjoki (Table 2). All RCMs show a higher percentage of wet days than observed, which is caused by too high percentage of light precipitation (≤ 1 mmday-1; Fig. 6). Occurrence of torrential (> 20 mmday-1) precipitation events is overestimated in RCMs in every catchment and season.

After applying the DBS method, biases in seasonally calculated daily mean temperatures in uncorrected RCM data are significantly reduced (Figs. 4, 5), from -8.7–5.3 to -0.2–0.5 ∘C. Also the standard deviation of the DBS-adjusted values is closer to observed values than that of uncorrected RCM data (not shown). DBS scaling preserves the RCM temperature variability in CDFs. The strong temperature increase around 0 ∘C found in the uncorrected RCM data is reduced after DBS scaling but can still be found from the CDFs (Fig. 5), although shifted towards observed values and higher temperatures. Daily temperatures adjusted with wet–dry separation produce more frequently higher winter maxima (> 5 ∘C) and lower minima (< -30 ∘C) than adjustment without the separation (Fig. 7). These extrema are originated from the separation of days to dry and wet which especially affects the CDF of dry days due to the small number of days (approx. 7–16 daysmonth-1) available. Otherwise there are no distinct differences between the two DBS approaches (Figs. 4, 5, 7), both give distributions that are similar to the observations. Due to the cases where daily winter maxima were excessively too high (e.g. > 15 ∘C in January) in DBS with wet–dry state separated data, we decided to use the DBS method without separation in further analysis of hydrological simulations.

Both single and double gamma DBS approaches for precipitation are able to reduce biases in seasonal precipitation accumulation from -22–81 to -3.0–1.7 % (Figs. 4, 6; Table 2) in all catchments. Distribution of drizzle and torrential precipitation is shifted towards observations and the number of dry days is forced to match observed values (Fig. 6).

There are no considerable differences in monthly mean accumulated precipitation between single and double gamma DBS. The largest differences are found in the treatment of heavy (> 95th percentile of CDF) precipitation (Figs. 6, 8). Considering daily mean precipitation amounts in the heavy precipitation distribution, DBS with double gamma overestimates daily mean heavy precipitation amounts in July by 0.2–6.5 % and DBS with single gamma by 12.0–21.7 % in Loimijoki and in Ounasjoki by -0.3–1.3 and by 3.4–14.8 %, respectively, compared to observed values. Due to a longer tail in the single gamma distribution in the heavy precipitation end of the distribution, the high values are in many cases larger and more frequent with single gamma than with double gamma DBS. In some cases the single gamma DBS approach even increases heavy precipitation values compared to observed values. Nevertheless, single gamma distribution was slightly better than double gamma, e.g., in winter and spring in northern Finland (root mean square error (RMSE) 2.78–3.10 in single gamma and 3.07–3.10 in double gamma in January in Ounasjoki). Still, in most cases the double gamma distribution produces heavy precipitation values closer to observed values than single gamma.

Finland is expected to experience a warmer and wetter climate towards the end of this century. Future changes in seasonal precipitation and mean temperature in Loimijoki catchment are shown in Table 3. After DBS adjustment, seasonal temperature increase varies from 1.4 to 5.1 ∘C in Loimijoki and from 1.3 to 6.6 ∘C in Ounasjoki in the latter part of this century, being the largest in winter. As for the control period, the DBS approach with wet–dry day separation produces higher temperature maxima for the scenario period compared to DBS approach without separation. Thus, it also produces higher seasonal mean values than DBS scaling without wet–dry separation. No distinct differences between the single and double gamma DBS approaches can be found for monthly and seasonal mean precipitation sums. Again, the greatest differences can be found from torrential precipitation, which are more frequent and intense in single gamma than in double gamma DBS-adjusted values. Future changes in seasonal precipitation sums vary more than temperature depending on RCM as well as season and area of investigation, and can even decrease by the end of this century. After DBS adjustment the change in seasonal precipitation sums varies from 1.7–39 % in Nilakka to -7.5–37.7 % in Loimijoki by the end of this century, being largest in winter.

The DBS method preserves the temperature trend of the uncorrected RCM data during 1961–2100 relatively well (Table 4; Fig. 9). The projected temperature trends in uncorrected RCM data vary between 0.3 and 0.5 ∘Cdecade-1 in the used scenarios. The differences between uncorrected RCM and DBS-adjusted seasonal trends are mainly less than ±0.1 ∘Cdecade-1 (Table 4). The largest differences between temperature trends in uncorrected and DBS-adjusted data can be seen in the scenarios of REMO and RCA, which produce more than 0.1 ∘Cdecade-1 larger temperature rise after DBS (Fig. 9). This is probably due to a too narrow temperature distribution (low standard deviation) in the control period compared to observed values (not shown). In the scenario period the standard deviation decreases even further, with increasing daily temperatures, causing more pronounced warming after DBS adjustment. Other climate models in this study do not produce any prominent decrease in standard deviation during the scenario period and thus the trends are better preserved.

Also trends in precipitation are preserved sufficiently well among RCMs after DBS adjustment and no distinct differences between RCMs or the two DBS methods can be found. In Loimijoki and Ounasjoki catchments most of the uncorrected scenarios show positive precipitation trends from 1.1 to 4.2 mmdecade-1 (Table 4). Only HIRHAM-A in Loimijoki and REMO in Ounasjoki do not show significant trends. The differences between RCM and adjusted seasonal trends are mainly from -0.6 to +0.3 mmdecade-1 (Table 4). The largest differences between trends of uncorrected and DBS-adjusted RCM data can be seen in seasonal precipitation simulated by HadRM in Ounasjoki (from -1.9 to -1.6 mmdecade-1) (Fig. 10). The trend simulated by HIRHAM-B is largest in spring in all catchments, which causes the large increase in precipitation accumulation (Table 3). Even though the trends are largest in winter or spring, the summer and autumn remain the wettest seasons of the year.

The discharges simulated with uncorrected RCM values (Fig. 11) show large differences compared to the observed discharges and discharges simulated with observed meteorological input values in the control period (hereinafter referred to as “control simulation”). The differences in simulated mean discharges in the control simulation and using RCM data with and without DBS adjustment for Loimijoki and Ounasjoki test sites are shown in Table 3. In the four test sites the annual mean discharges simulated with uncorrected RCM inputs were 16–104 % larger than annual mean discharges of the control simulation. The higher annual mean discharges are mainly caused by overestimation of precipitation in RCMs.

The seasonal differences are more pronouncedly affected by temperature biases in the RCM data. The HadRM and HIRHAM-B have negative temperature biases during winter, which cause smaller winter discharges in southern and central Finland. The negative temperature biases in spring (HIRHAM-B) cause a delay in the spring flood peak (Fig. 11). This delay causes negative biases to mean spring discharges in northern Finland even though the snowmelt floods are larger due to greater snow accumulation caused by positive precipitation and negative temperature biases. Summer mean discharges become larger with all uncorrected RCM outputs due to positive precipitation biases and larger recession flows caused by greater and delayed spring floods.

Using single gamma or double gamma precipitation corrections and temperature corrections without wet–dry separation, the biases in simulated mean discharges can be effectively reduced (Table 5). The differences in annual mean discharges decreased to less than 12 % in all test sites with DBS-adjusted RCM outputs. The difference is at the same level as the difference between control simulation discharges and observed discharges (less than 13 %), which indicates that biases in annual mean discharges are partly explained by the model sensitivity on input variables and partly by the residual biases in corrected RCM outputs.

The differences in seasonal mean discharges between simulations with DBS-adjusted RCM data and control simulation are in many cases larger than differences between observed discharges and discharges in the control simulation. Differences larger than 30 % are only found in winter and summer, when the discharges are low. But the remaining biases larger than 20 % during high flow season in Loimijoki found in REMO and RCA and larger than 50 % during the low flow season in HadRM and HIRHAM-B may have significant effect on the seasonal changes and changes in extreme discharges in climate change projections. The main reason for large and in some cases even larger remaining biases in winter discharges than in uncorrected data is the sensitivity of the hydrological model on near-zero temperatures. Even though the DBS method corrects the mean temperatures efficiently close to observations, the remaining biases in winter temperature extremes, which in control period are slightly above zero, cause remarkable biases in winter discharges and snow accumulation in the hydrological simulation. However, the seasonal variations in mean discharges after the DBS adjustment are remarkably closer to variations of control simulation (Fig. 11), highlighting the fact that the bias correction is required for RCM data used in studies of climate change effects on hydrology.

In addition to biases in RCM temperature and precipitation data, the biases in wind speed (WS) and specific humidity (SH) also affect the WSFS discharge simulations for catchments with high lake percentages. Biases in WS and SH of RCMs affect the lake evaporation in the hydrological model and typically cause a 5–45 % bias in the annual lake evaporation sums. In most of the study catchments the bias is largest in the RCA scenario giving 25–35 % negative bias caused by positive bias of SH and negative bias of WS. The bias in lake evaporation can be effectively decreased to 0–13 % by the simple mean bias correction method (Fig. 12).

The uncorrected WS and SH of RCMs cause a 0–11 % bias in annual mean discharges, and a 0–20 % bias in autumn mean discharges in the outlet of Nilakka, which has the highest lake percentage of the study catchments (18 %). In the catchments of Loimijoki and Lentua the biases in mean discharges (0–2 and 0–4 %) and autumn discharges (0–7 and 0–8 %) are smaller, and in the most northern located catchment of Ounasjoki the bias is insignificant.

The effect of different correction methods on annual and seasonal discharges as well as on the changes in discharges by the 2051–2090 period are shown in Fig. 13. The deviations of the simulated discharges with RCM data compared to control simulations in four test sites using all five scenarios without corrections, only with temperature corrections, precipitation corrections and with both temperature and precipitation corrections are shown in the upper part of the box plot. The lower figure shows the results of the climate change impacts on mean discharges with different corrections. The results show that the effect of precipitation correction affects more the annual discharges and the temperature correction affects more the seasonal discharges. However, without temperature correction the annual discharges still have positive biases due to cold biases, which decrease evapotranspiration. All four combinations of DBS temperature and precipitation correction methods used in this study produce similar results and none of the different DBS approaches are found to be superior with respect to mean discharges. Thus, the selection of the best methods is based on the performance of the correction method in decreasing the extreme temperature and precipitation biases, in which the temperature correction without wet–dry separation and double gamma for precipitation work significantly better.

Because of the biases in uncorrected RCM data, the mean discharge peaks caused by snowmelt (Fig. 11) are significantly larger than the control simulation discharge peaks, and the seasonal variation of discharges is also altered. Without effective bias correction the results of climate change impact studies could easily lead to false conclusions. The effect of DBS adjustment on changes in seasonal mean discharges is more pronounced than on annual discharges, because the temperature biases of uncorrected data have significant influence on seasonal discharges. The changes in mean winter and spring discharges may be 2 or even 3 times larger than without temperature correction (Fig. 13). If only temperature bias is corrected, the relative changes are close to the changes in temperature and precipitation corrected data, but the absolute changes are much larger due to wet bias in RCM data.

The temperature correction is essential especially when the high and low flows are studied. The difference between the changes in mean high discharges (MHQ) and mean low discharges (MNQ) by using uncorrected RCM data can be even to the other direction when the uncorrected RCM data are used compared to changes in the DBS-adjusted data. This can also be seen in summer mean discharges with HIRHAM-B scenario. The uncorrected scenario shows 35 % decrease in summer discharges in Loimijoki due to large recession flow after spring flood in the control period, which caused over 300 % wet bias in mean summer discharges (Tables 5, 6). The DBS-adjusted data of HIRHAM-B show a slight increase in summer discharges because a large precipitation increase compensates the increased evapotranspiration in this scenario.

The ability of the DBS method to preserve the precipitation and temperature trends (Figs. 9, 10) in most cases leads to similar changes in simulated annual mean discharges with uncorrected and DBS-adjusted RCM data (Fig. 13 and Table 6). In the HadRM-scenario, the DBS-adjusted data produce a lower increase than the uncorrected scenario in northern Finland, due to a smaller increase in precipitation trends after DBS adjustment. In northern Finland the differences between the results from simulations with uncorrected and DBS-adjusted data are clearest in spring, when the absolute biases in mean discharges in the control period are highest. The uncorrected HIRHAM-A and HIRHAM-B produce negative bias in mean spring discharges in the control period due to delayed spring floods. Thus, without bias corrections these scenarios produce too high increases in mean spring discharges.

The results show that climate change will have a significant impacts on seasonality of discharges in Finland due to increasing precipitation and shorter wintertime, which influence snow accumulation and increase evapotranspiration (Fig. 14). The springtime snowmelt floods will occur earlier and the average wintertime discharges will increase because the temperature will rise more often above zero in winter increasing rainfall and causing occasional snowmelt. The summer discharges will decrease due to earlier snowmelt and increased evapotranspiration, while the changes in autumn depend on the climate scenario, location and hydrological characteristics such as lake percentage of the study catchments. The DBS method influences significantly the projected changes of the seasonal discharges and in some cases even the annual discharges of the scenarios with large temperature biases.

The changes in annual mean discharges between the control and 2051–2090 periods in all study catchments are between -15 and 26 % (Table 6). For the period 2051–2090 HIRHAM-B produces the largest increases in annual mean discharges in all study catchments due to the largest increases in annual mean precipitation. Most of the scenarios show an increase in annual discharges, but especially for southern and central Finland some scenarios project decrease because the longer and warmer summers will cause a larger increase in evapotranspiration than the projected increase in precipitation.

In the study catchments all DBS-adjusted scenarios predict on average 2–4 weeks earlier snowmelt discharge peaks in spring for the 2051–2090 period compared to the control period 1961–2000. Figure 14 shows the results for three scenarios producing the largest variation of changes in mean discharges out of five scenarios used in this study. Because the snowmelt discharge peaks occur earlier, the recession flows in summer season decrease. The summer discharges decrease 20–50 % in all scenarios except in Nilakka and Loimijoki in the HIRHAM-B-scenario, which predicts a greater increase in precipitation than the other scenarios. The decrease in mean summer discharges is caused by the increase of the annual evapotranspiration by 10–40 % and lake evaporation by 10–80 %.

In addition to earlier spring discharge peaks and a decrease in summer discharges, all scenarios predict increase in winter discharges. The increase is more pronounced in the catchments of Loimijoki and Ounasjoki (40–150 %), which have lower lake percentage than Nilakka and Lentua, in which the winter discharges increase 10–70 %, depending on the used scenario.

The results show an increase in autumn mean discharges in northern Finland, where the autumn runoff peaks – typical in southern Finland at present – become more frequent. In the catchments with large lake percentages in southern and central Finland, the autumn mean discharges decrease in all scenarios due to an increase in evapotranspiration and larger soil moisture deficit in the beginning of autumn. In the southern catchments with low lake percentages, the change in mean autumn discharges depends on the scenario. Different autumn precipitation changes between the scenarios are the main reason for different changes in autumn discharges, but also the soil moisture content after summer has an influence and varies depending on temperature and precipitation changes during summer.

The relative changes in mean discharges, MHQ and MNQ together with changes in mean maximum snow water equivalent (SWE), mean maximum soil moisture deficit (SMD), mean evapotranspiration (ET) and mean runoff (R) in four test sites are shown in Fig. 15. The changes in annual high flows are mostly negative, due to decreased maximum SWE and consequently decreasing spring snowmelt floods. Only in the HIRHAM-B scenario the MHQ increase or remains the same in most test sites due to large increase in precipitation. The annual low flows decrease in southern Finland due to increased ET and maximum SMD, due to decrease in low flows in summer season. In northern Finland the annual MNQ increase, because the annual low flows normally occur in winter in the control period.