Data-driven regionalization of river discharges and emergent land cover–evapotranspiration relationships across Sweden

Authors

  • Ype van der Velde,

    Corresponding author
    1. Environmental Sciences, Copernicus Institute for Sustainable Development, Utrecht University, Utrecht, Netherlands
    2. Physical Geography and Quaternary Geology, Stockholm University, Stockholm, Sweden
    3. Bert Bolin Centre for Climate Research, Stockholm University, Stockholm, Sweden
    • Corresponding author: Y. van der Velde, Environmental Sciences, Copernicus Institute for Sustainable Development, Utrecht University, Utrecht, Netherlands. (y.vandervelde@uu.nl)

    Search for more papers by this author
  • Steve W. Lyon,

    1. Physical Geography and Quaternary Geology, Stockholm University, Stockholm, Sweden
    2. Bert Bolin Centre for Climate Research, Stockholm University, Stockholm, Sweden
    Search for more papers by this author
  • Georgia Destouni

    1. Physical Geography and Quaternary Geology, Stockholm University, Stockholm, Sweden
    2. Bert Bolin Centre for Climate Research, Stockholm University, Stockholm, Sweden
    Search for more papers by this author

Abstract

[1] Changes in river discharge and river water quality, due to climate change and other drivers such as land cover change, pose both societal and ecosystem threats. Analyses of measured terrestrial river fluxes are key for identifying the drivers and quantifying the magnitudes of such riverine changes. In this paper, we develop and apply a data-driven regionalization approach using the dense network of discharge measurements in Sweden. The developed regionalization approach facilitates detailed mapping of discharges (Q) and change trends in Q across Sweden. Combining these with estimates of precipitation (P) and change trends in P, we estimated actual evapotranspiration (AET) and change trends in AET via catchment-scale water balance constraints. We identified characteristic land cover–evapotranspiration relationships by plotting water use efficiency (AET/P) against energy use efficiency (AET/potential ET) for areas with unique land cover across Sweden. Our results show that wetlands have clearly lower water and energy use efficiencies compared to open waters, forests, and agriculture, and that agriculture has water and energy use efficiencies closest to those of open waters. We further compared the data-driven regionalization estimates of different water balance components with estimates of regional climate models (RCMs). The RCMs do not describe well the observed change trends in Sweden. In particular, for evapotranspiration, the observed change trends are not reproduced by any of the investigated 24 RCMs.

1 Introduction

[2] The hydrology of river basins is continuously changing driven by both natural and human-induced processes. Thorough analyses of available data on terrestrial water fluxes with use of water balance constraints can crucially contribute to the identification and quantification of these changes and their drivers. For instance, both for the dramatically changing Aral Sea drainage basin in Central Asia [Destouni et al., 2010; Jarsjö et al., 2012] and for much less dramatic yet still changing drainage basins in Sweden [Destouni et al., 2012], such analyses have been effective in distinguishing between climate-induced and land and water use–induced hydrological change. In this study, we focus our analysis on Sweden, where changes in terrestrial water fluxes are important for anticipating flood risks, and salinity and nutrient impacts on the vulnerable Baltic Sea ecosystem [Rodhe and Winsor, 2002; Conley et al., 2009; Dargahi and Cvetkovic, 2011].

[3] Coastal water ecosystems are particularly sensitive to changes in river discharge and river water quality, as most coastal waters do not mix well with the oceans. Hence, a large number of studies of spatial and temporal change patterns of river discharges and nutrient loads relate to coastal waters, such as the Baltic Sea [Darracq et al., 2005; Dargahi and Cvetkovic, 2011], the Gulf of Mexico [Basu et al., 2010], Hudson Bay [Déry and Wood, 2005; Déry et al., 2011], the Aegean Sea [Durdu and Cvetkovic, 2009], the Bay of Bengal [Asokan et al., 2010], the Aral Sea [Shibuo et al., 2007; Jarsjö et al., 2012], and the Arctic Ocean [Lammers et al., 2001; Arnell, 2005; Dyurgerov et al., 2010]. All these coastal regions experience river discharge and/or river water quality stresses. Relating river discharges to landscape characteristics, such as land cover and soil types, via a regionalization approach is crucial for understanding the cause of these stresses and the interactions between water quantity and quality stresses.

[4] Most often, a hydrological model is used to achieve regionalization of discharges (e.g., Sperna Weiland et al. [2012] on a global scale; Bengtsson [1995], Mörth et al. [2007], and Lindström et al. [2010] for the Baltic Sea region; Samaniego et al. [2010] and Livneh and Lettenmaier [2012] for parameter estimation techniques). The parameters of these models, which usually relate to physical properties, are fitted to measured river discharges and spatially extrapolated to ungauged basins. While each approach has various strengths and weaknesses, the common shortcoming of model-based regionalization is that the concepts and processes included in the hydrological model partly dictate the results, which can mislead assessments of spatiotemporal water flow patterns, especially with regard to marginally understood vegetation-atmosphere interactions [Lyon et al., 2008]. In this paper, we pursue instead a data-driven approach to regionalize river discharges and explore emerging land cover evapotranspiration characteristics. In this context, we define evapotranspiration as the sum of soil evaporation, open water evaporation, plant transpiration, and evaporation of intercepted water.

[5] Peel et al. [2010] and Oudin et al. [2008] analyzed between them more than 2000 river basins and found clear evidence for land cover–specific evapotranspiration fluxes. Williams et al. [2012] analyzed flux tower measurements of evapotranspiration and confirmed such relationships. From these studies, it is evident that land cover is an important control on the terrestrial water balance, but clear methods to quantify contrasting evapotranspiration characteristics of different land cover types are lacking. Furthermore, Peel et al. [2010] and Oudin et al. [2008] found that for catchments larger than 1000 km2, relationships between land cover and discharge break down due to heterogeneity of the landscape. A data-driven regionalization approach that downscales specific discharges from catchments with a wide range of scales, to a scale where land cover evapotranspiration relationships manifest (i.e., <1000 km2), may thus help to elucidate the role of land cover within the terrestrial hydrologic cycle.

[6] A data-driven regionalization approach for discharges can also bridge classical scale disparities between catchment hydrology and climate model results [Hostetler, 1994]. Many studies have developed and tested ways to downscale and bias correct results from global climate models (GCMs) or regional climate models (RCMs) [Terink et al., 2010; Teutschbein et al., 2011; Jarsjö et al., 2012; Visser et al., 2012]. These methods, however, are “one-way bridges” in that they make climate model results suitable for hydrological modeling but do not allow for full coupling and feedback between the terrestrial and atmospheric systems [Seuffert et al., 2002; Dirmeyer et al., 2006; Lyon et al., 2008]. These methods also cannot be used to corroborate measured river discharges that represent relatively small spatial scales with the larger scales of GCMs or RCMs. River discharges, however, contain valuable information on the separation of precipitation into runoff and evapotranspiration. Therefore, a regionalization approach that is able to upscale measured discharges and related evapotranspiration fluxes from individual river basins to typical scales of climate models is warranted.

[7] Discharge measurements are potentially sensitive to measurement errors, changes in measurement setup or location, catchment size, and local landscape features and changes thereof. This sensitivity makes it difficult to distill and quantify regional patterns of river discharge change directly from observations (for example, see Wilson et al. [2010], who analyzed discharge trends within Scandinavia). In analogy with the handling of point-scale meteorological measurements, we may be able to reduce uncertainty for river discharges by quantifying the spatial correlation of discharge measurements between gauging stations [e.g., Skøien and Blöschl, 2007]. Such a regionalization approach could thus improve estimates of water flow to coastal waters, help to establish land cover evapotranspiration characteristics, and improve the comparisons between climate models and discharge measurements.

[8] The main objective of this paper is to develop a data-driven regionalization approach for discharge measurements, using Swedish river discharges to the Baltic Sea during the past 50 years as a regional example, to understand the influence of land cover and climate change on the hydrology and the atmosphere-land interactions. Furthermore, the data-driven estimates of Swedish river discharges are compared with corresponding estimates of multiple RCMs, assessing the RCM ensemble potential to accurately describe the processes that drive discharges and their changes for the Baltic Sea region.

2 Material and Methods

[9] The challenge in regionalizing discharge data is that a discharge measurement typically represents a large area (i.e., the hydrologic catchment). Therefore, measured values at the catchment outlet cannot easily be appointed to a single location within the catchment. The measured value first has to be differentiated across the catchment based on meteorological and landscape characteristics before extrapolation to ungauged catchments can be performed. In the following sections, we describe the type and origin of the data we have used and the details of the analyses applied to these data in order to perform such a differentiation.

2.1 Data

2.1.1 Precipitation and Temperature

[10] We used the gridded PTHBV data base compiled by the Swedish Meteorological and Hydrological Institute (SMHI; luftweb.smhi.se), which provides daily estimates of precipitation and average temperature across Sweden for the period 1961–2010. Because we found considerably higher discharges than precipitation amounts reported in the PTHBV data base for the highest Swedish mountains along the Swedish-Norwegian border, the E-Obs data set [Haylock et al., 2008] was used to supplement the PTHBV data base for these regions. This European data set gives higher precipitation values for this mountain range as it interpolates between Norwegian and Swedish observations and therefore includes more high altitude stations. This correction is made for <1% of the total surface area of Sweden but shows the limitations of the gridding procedures underlying the PTHBV data base and our limited understanding of precipitation and evapotranspiration processes at high elevations. We did not use information on the uncertainty of the PTHBV data set.

2.1.2 River Discharge

[11] The SMHI maintains an extensive network of measurement stations for river discharges across Sweden (vattenweb.smhi.se). For more than 300 rivers with catchment sizes ranging from 1 to 46,000 km2, the SMHI provides daily discharge time series. No information on discharge uncertainty is available. We used only those catchments for which we could independently reproduce the given catchment area in the SMHI data base within 5% through our own delineation based on a digital elevation model. This gave a total of 270 catchments to be included in this current study.

2.1.3 Land Cover and Soil Type

[12] The Corine Land Cover 2006 data set (www.eea.europa.eu) was used to characterize land cover. This 100 × 100 m satellite image product was upscaled to 10 × 10 km grid cells by quantifying the percentages of artificial surfaces, agricultural land, forests, wetlands, open water, and open natural vegetation (all natural vegetation types not being forests or wetlands). The same approach was used for quantifying 10 × 10 km soil types. Soil types, derived from Swedish National Soil data base, were subdivided into five classes—peat, clay, sand, till, and rock—and for each class, an upscaled 10 × 10 km grid was created with the coverage of each soil type. Maps of land cover and soil type along with topographical information are shown in Figure 1. No spatially detailed information on land cover change during the past 50 years was available.

Figure 1.

Land cover (a), soil type (b), and meteorological variables: temperature (c), trend in temperature (d), precipitation (e), and the trend in precipitation (f) for the period 1960–2010. Figure 1a also shows the major cities, and Figure 2b shows the elevation contours of 200 and 500 m. In Figures 1a and 1b, a land cover or soil type is drawn whenever its coverage is more than 20%. Overlapping colors occur and create mixed colors.

2.1.4 Climate Model Data

[13] Monthly average values for precipitation, temperature, and evapotranspiration from 24 RCM runs were downloaded from the ENSEMBLES archive (http://www.ensembles-eu.org/). These 24 runs are combinations of 13 RCMs with boundary conditions from 10 GCMs as listed in Appendix A. Although these simulations cover the period 1961–2100, we compared them with hydrological observations for the period 1961–2010. Since uncertain future emission scenarios have not yet impacted the 1961–2010 period, all models should match observed discharges during this period.

2.2 Data Analyses

2.2.1 Averages and Trends

[14] The long-term average precipitation and temperature during the period 1961–2010 were calculated by taking the mean value over the entire period for each grid cell in the PTHBV data base. The average river discharge for each catchment was calculated by first creating yearly averages (from October to October). By taking this hydrological period, we expected minimal year-to-year variation in water storage within a catchment due to snow accumulation or groundwater depletion by evapotranspiration. All years with discharge values reported for at least 85% of the year were considered. Missing values were substituted with corresponding monthly averages from the 20 year period enveloping the missing value. The average yearly discharge was then calculated by taking the mean value of all the years in the data series when at least 20 years of data were available.

[15] Trends in precipitation, temperature, and discharge were calculated by creating 20 year moving averages for each of these time series. It is important to note that in this study, we assumed 20 year periods to represent climate, and thus, a trend in the 20 year moving average quantifies climate change during the period 1961–2010. Trends were characterized using an autoregressive–moving-average (ARMA) model with linear drift, which was fitted to these time series to account for autocorrelation. To actually fit the ARMA models to the various time series, an automated model fitting procedure was used to select the optimal model form based on the Bayesian Information Criteria (BIC). The BIC is a measure that weights the number of parameters against explanatory power of a model compared to other models describing the same objective variable (as implemented by R Development Core Team [2011]).

2.2.2 Regionalization of Specific Discharge and Actual Evapotranspiration

[16] Specific river discharge (Q), trend in specific discharge (dQ), actual evapotranspiration (AET), and trend in actual evapotranspiration (dAET) were the objective variables that needed to be extrapolated from measured catchments to ungauged (sub-)catchments. Because the specific discharge of an area is clearly dependent on the amount of precipitation (P), more meaningful results were likely obtained by regionalizing the runoff coefficient Q/P instead of the Q directly. The following steps summarize our data-driven regionalization approach:

  1. [17] Outliers in Q/P and dQ data sets were identified and removed by testing the impact of each data point on the root mean squared error of the optimal linear model describing Q/P or dQ. All possible combinations of linear models that explain Q/P and dQ as a function of any number of explaining variables were evaluated. Table 1 gives all the explaining variables considered. The BIC criterion was again used to select the best performing model (lowest BIC). No more than two points were removed from each data set.

  2. [18] Random samples containing 30% of the data were taken. Again, the optimal linear model (following the same procedure as in step 1) was determined. This model was then tested against the entire data set: if the explained variance of the linear model was within 5% of the maximum explained variance using all the data, the linear model was accepted. This step was repeated until we had an ensemble of 1000 linear models that all described the data approximately equally well.

  3. [19] The residuals for each of these models were mapped back into space, and each residual was appointed to the center of its catchment. Only catchments with an area less than 10,000 km2 were considered in this step. For larger catchments (which accounted for 17 out of the 270 catchments considered), appointing an average value to the center of a catchment would yield too large allocation errors (the residual value belongs to the entire catchment but is appointed to its center, causing an allocation error). Ordinary kriging with an automatically fitted spherical variogram model with nugget to account for the spatially uncorrelated part of the residuals was applied to create maps of residuals on a 10 × 10 km grid for each of the optimal linear models. The correlations between the object variable and explaining variables were again added to these maps of kriged residuals (approximating an external drift approach). This yielded an ensemble of 1000 maps of Q/P and dQ, which all were assumed to represent the observed values equally well.

  4. [20] Maps of the average and the standard deviation for Q/P and dQ were calculated based on these 1000 maps. The catchments with areas larger than 10,000 km2 were used to test the regionalization approach by integrating the average Q and dQ maps over these catchments and comparing these values with the measured values.

  5. [21] Average AET was calculated from the difference between precipitation and discharge assuming zero change in storage (i.e., using a catchment water balance approach to quantify AET). Similarly, dAET was calculated as the difference between the trend in precipitation (dP) and trend in discharge (dQ), assuming a zero change in the change of storage. This assumption and its potential implications will be further discussed in section 3.

Table 1. Explaining Variable Used in Regionalization of Q/P and dQ
Explaining VariableQ/PdQ
LongitudeXX
LatitudeXX
ElevationXX
Ln(Elevation)XX
Precipitation (P)XX
Temperature (T)XX
Degree days (DD)  
(Yearly sum of daily T > 5)XX
Land use (6 classes)XX
Soil type (5 classes)XX
Specific discharge (Q) X
Trend in precipitation (dP) X
Trend in temperature (dT) X
Trend in degree days (dDD) X

[22] Each linear model in step 2 can be seen as a simplistic empirical hydrological model. By creating an ensemble of models that all perform equally well, we were able to account for model structure uncertainty. Such an uncertainty analysis of model structure would not have been possible if a fixed hydrological concept was used to describe discharge or discharge change. Because we first regionalized the measured data via a linear model approach and thereafter added the spatially correlated part of the error to the linear model via ordinary kriging, we minimized the error stemming from allocating a single value representing the entire catchment to the center of the catchment. Although this type of error still exists, it only affects the spatially correlated part of the linear model residuals. The “allocation” error in our approach thus is smaller than it would have been if we had assigned the entire specific discharge to the center of each catchment and then extrapolated these values via ordinary kriging.

2.2.3 Comparison with Climate Models

[23] We compared our data-derived maps of temperature, precipitation, discharge, and evapotranspiration and their trends with the results of 24 RCMs for Sweden from the ENSEMBLES data archive (see Appendix A for a list of the RCMs). Means and trends from the climate models were calculated similarly to our procedures for precipitation and temperature. The average discharge of a climate model was calculated as the difference between precipitation and evapotranspiration for 20 year periods.

3 Results and Discussion

3.1 Regional Patterns of T and P

[24] Sweden has a spatial gradient in average yearly temperatures between +7°C and −7°C, which is related to elevation and latitude (Figure 1c). In contrast, the trend in temperature over the past 50 years shows a relative uniform increase over the entire country of 0.036°C per year on average (Figure 1d). The patterns of averages and trends for precipitation are more complex (Figures 1e and 1f). The most precipitation and the largest increases in precipitation occurred in the northwest mountain region and the southwest region around the city of Gothenburg. The southeast region around the capital Stockholm, however, receives relatively little precipitation, and also, the changes in precipitation for this region are minimal.

3.2 Regionalization of Q

[25] The explaining variables of Table 1 contain a lot of redundant information. For example, “natural open” vegetation prevails in the high mountains above the three line (i.e., no forest cover), where also P is high and T is low. All such direct correlations between variables are shown in Figure 2. Our ensemble approach was developed to distill as much information as possible from these variables, without overlooking variables that correlate slightly less than the highest possible correlation but are just as likely to explain variability in Q/P or dQ.

Figure 2.

Correlation matrix showing all correlations between catchment properties of 270 catchments.

[26] The maximum explained variance (R2) for any linear combination of variables describing the runoff coefficient Q/P is 0.86. Precipitation is present in all the 1000 selected linear models (Figure 3a) and is positively correlated to Q/P, indicating that areas with more precipitation tend to have higher runoff coefficients. The second most dominant variable is open water. Although the direct correlation between Q/P and open water is low (Figure 2), open water introduces unique information that is not included in any of the other variables. A negative relationship between the percentage of open water and Q/P was found for 62% of the models adopted. This is likely related to relatively high evapotranspiration fluxes occurring from open waters compared to other land cover types. This increases the total evapotranspiration from areas with relatively high fractions of open water, thereby reducing their Q/P compared to areas with a smaller percentage of open water.

Figure 3.

(a and b) Contribution of explanatory variables to the 1000 linear models for runoff coefficient Q/P and trend in specific discharge (dQ), respectively. Red lines indicate a positive correlation; blue lines indicate a negative correlation.

[27] The vegetation type natural open is the third most dominant variable. Natural open is most common in the high mountains (Swedish fjäll). These areas receive much precipitation and have relatively low evapotranspiration because of low temperatures. This leads to relatively high runoff coefficients. The found negative relationship probably is a correction term for nonlinear behavior of Q/P with elevation (and thus with P and T; Figure 2). Other dominant variables (albeit less so than the previously discussed) are latitude and temperature. Latitude is positively correlated, while T is negatively correlated with the runoff coefficient. This explains how Q/P generally increases as evapotranspiration decreases from the south to the north of Sweden. Because latitude and T are very strongly correlated (Figure 2), and both variables relate directly to evapotranspiration and thus to Q/P, all models contain either latitude (±50% of the models) or T (±50% of the models, Figure 3). The fraction of wetlands is also important, with Q/P being positively correlated to the presence of wetlands. Lyon et al. [2012] found similar results for wetlands in boreal Sweden, where the wetlands maintained high specific discharges during relatively dry periods of the year. The high Q/P in wetland areas is caused by a relatively low evapotranspiration. This is an interesting finding considering the negative relationship found between open water and Q/P, which was attributed to high evapotranspiration fluxes of open water. We will return to this apparent contradiction in the following sections. Lastly, soil types could not explain the variance in Q/P between catchments. The strong correlation between agriculture and clay and wetlands and peat (Figure 2) partly explains why soil types do not provide much additional information.

[28] The maximum R2 for a linear combination of variables (Table 1) describing the change trend in discharge, dQ, is 0.61. This value is considerably lower than the R2 for Q/P (0.86), reflecting the larger uncertainties in change trends than in averages. Figure 3b shows that dQ is strongly related to dP and the percentages of forest and open water. The negative correlation with forests could relate to an increase in forest productivity and corresponding forest evapotranspiration during the period 1961–2010 (as observed by Hellström and Lindström [2008] for Sweden and by Hember et al. [2012] in Canada) and to large-scale draining of wetlands and subsequent reforestation [Lohila et al., 2010]. About 30% of the dQ models show a negative correlation with agriculture. As agricultural productivity has increased during the past 50 years due to improved farming strategies, this has likely led to an increase in crop transpiration and hence a reduced Q. Note that no general information was available on land cover change for this 50 year period. Hence, the correlations between land cover and dQ do indicate that AET has changed, but we cannot infer from this analysis if this AET change is caused by land cover change or by a change in water use of the land cover type. A recent study of 20th century agricultural and hydropower developments in nine large Swedish catchments indicates that most likely both changes have occurred [Destouni et al., 2012]. That study showed that increased agricultural area (change in land cover) and/or production (change in water use of the same land cover type), as well as hydropower expansion (changing land cover to more open water, and likely also changing the water use of surrounding land cover by increasing water saturation in those areas), generally increased AET and decreased Q relative to original land cover conditions.

[29] The average map for Q (i.e., Q/P multiplied by P; Figure 4a) indicates high specific discharges for the mountain regions and low specific discharge for the big lake areas in the south of Sweden (note: Figure 1a shows the locations of major Swedish lakes). However, the uncertainty map (Figure 4b) also points out that the uncertainty of specific discharge for open water is relatively large (100–150 mm per year) compared to other areas. Figure 4c shows that open water is associated with larger increases in Q than their surroundings. This can be explained by the ability of vegetation to adapt its evapotranspiration in response to changing rainfall and temperature, whereas the lake itself cannot change its evaporation. The standard deviation map for dQ in Figure 4d shows large trend uncertainty in open water areas. This is because, although the signal of the fraction open water in the dQ observations is evident, its magnitude is uncertain due to the relative small open water fractions within catchments.

Figure 4.

(a) Average specific discharge (b) Standard deviation of specific discharge. Both maps were derived from the 1000 linear models describing specific discharge. (c) Average trend in specific discharge. The blue and red dots show significant trend (P < 0.05), while the gray locations do not have a significant trend. (d) Standard deviation of the trend in specific discharge.

[30] The average Q/P and dQ maps were tested by integrating the average Q and the dQ values in the maps over catchment area for each of the 17 catchments larger than 10,000 km2. These catchments were used for deriving the linear models but were not used for the kriging of the residuals. The integrated values compared rather well with measured average discharge (Figure 5a) and trend in discharge (Figure 5b) for these large catchments. For both variables, there is no apparent bias around the 1:1 line and it is clear that the variability associated with the smaller-scale measurements is substantially reduced for these large catchments, highlighting general consistency in the used approach.

Figure 5.

(a) Validation of specific discharge. (b) Validation of trend in specific discharge. The red dots are catchments larger than 10,000 km2, which were not used in creating the kriged maps and serve as a validation basis. The horizontal and vertical red lines through the red dots give an indication for the measurement and mapping uncertainty, respectively. The average sill of the nugget variogram model for the residuals is used as an indication for the measurement uncertainty. An indication for the mapping uncertainty is determined by combining the kriging variance of the residuals minus the sill of the nugget variogram, and the standard deviation from all linear models. The green dots correspond to catchments that are smaller than our cell size (10 × 10 km) and thus were expected to give large deviations between measurements and maps.

3.3 Evapotranspiration: Average and Change Trend

[31] The patterns and values of the AET map (Figure 6a), with low AET in the mountains and high AET in the south and around the big lakes, determined via catchment-wise water balance, correspond well with the map published by Halldin [1988]. Further, the values for lake evaporation in southern Sweden (500–650 mm/year) compare well with evaporation rates found by Halldin [1988] and Saxena et al. [1999] for Sweden and by Solantie and Joukola [2001] for Finland who reported values between 500 and 700 mm/year depending on lake depth. The change trend in evapotranspiration, dAET (Figure 6b), indicates that evapotranspiration has increased over the considered period in many parts of Sweden and particularly in the agriculturally rich southern and southeastern Sweden. This is consistent with the recent results of Destouni et al. [2012] for selected Swedish catchments and with Hellström and Lindström [2008] who reported large biomass increases for Swedish forests during the past 50 years, which are likely to increase forest evapotranspiration.

Figure 6.

(a) Average yearly AET. (b) Trend in AET.

[32] As noted before, the estimate of dAET is based on the assumption of zero change in average annual storage change, dΔS, as expressed by the water balance–based equation for change trends dΔS = dP −dQ − dAET. The assumption of dΔS ≈ 0 is stronger than that of average annual storage change ΔS ≈ 0 in the water balance equation for annual averages ΔS = PQ − AET. Hence, in theory, part of the increasing trend in evapotranspiration (positive dAET) could instead be attributed to increasing water storage change (positive dΔS). However, recent work by Brutsaert [2008] considering storage changes in the midwestern United States showed storage change trend values of about one order of magnitude smaller than the change trends determined in the current study. Furthermore, if storage would be increasing, this would likely be accompanied by an increase in discharge. However, for instance, in the Stockholm region with a relatively large evapotranspiration change trend, there is no discernible discharge change trend. We therefore do not expect dAET (Figure 6b) to instead be an increasing water storage change dΔS, but rather to be predominantly due to some other change mechanism, such as increased transpiration by vegetation, as also found in the study by Destouni et al. [2012].

3.4 Land Cover–Evapotranspiration Relationships

[33] There are marked differences in the energy and water use efficiencies of five major land cover types across Sweden (Figure 7). Here energy use efficiency is defined by the ratio of AET over potential evapotranspiration (PET), and water use efficiency is defined by the ratio of AET to P. For these efficiency definitions, PET was calculated using the Priestley-Taylor equation [Priestley and Taylor, 1972; Zhang et al., 2001] in which PET is a function of incoming radiation (downloaded from www.SMHI.se) and temperature. Additionally, we fitted the constants in the Priestley-Taylor equation to the 11 year average (2000–2010) PET of the MODIS-derived PET product [downloaded from http://www.ntsg.umt.edu/project/mod16; Mu et al., 2011]. Comparison of these efficiencies (similar to those considered in a Budyko analysis [Budyko, 1958]) has longstanding tradition for characterizing climatic controls on a catchment's annual AET. Under energy-limited conditions like in Sweden, many factors next to water limitation can cause AET to be smaller than PET (as calculated by the Priestley-Taylor equation), such as snow and ice melt, land cover differences in albedo, and differences in growing season length. This is reflected in Figure 7, which shows that the presented characterization distinguishes essential differences in AET between different types of land cover, even when there is no water limitation.

Figure 7.

The water and energy use efficiency of Swedish land cover types. Each point represents a 10 × 10 km grid cell in Sweden. The contour lines indicate points belonging to the five major land cover types. The contour line envelopes a smoothed 80% probability region of water and energy use efficiencies for grid cells with >60% area covered by a specific land cover. Dots that fall outside a contour line either belong to the 20% of points falling outside the 80% probability region or have no land cover type that covers more than 60% of the grid cell. The red line divides the plot area in locations that are more energy than water limited below the line and locations that are more water than energy limited above the line.

[34] Agricultural areas in Sweden are found to prevail in an AET/P and AET/PET range space of around 0.40–0.75 and 0.60–0.80, respectively (Figure 7). These ranges are high compared to those of wetlands and natural open environments, and much narrower than those of forest, implying that agricultural areas in Sweden are the land cover type closest to behaving like an open water environment with regard to its water and energy use efficiency. This result explains and further quantifies the findings of Destouni et al. [2012] that 20th century expansions of agricultural area and/or production have shifted AET and AET/P to higher levels.

[35] Swedish wetlands (of which a majority are boreal peatlands) also prevail in a narrow AET/P and AET/PET range space, but with remarkably lower values than agricultural areas and forests. Hence, these wetlands behave less like open water environments than agricultural areas do, using relatively little of available water and energy for evapotranspiration, and by doing so maintain their relatively wet environments. A possible explanation for the low energy and water use by wetlands is that wetland vegetation has physiological adaptations to prevent oxygen stress in the root zone [Bartholomeus et al., 2008; Bartholomeus et al., 2011]. These physiological adaptations are likely to reduce AET. For example, the development of a shallow root system to prevent oxygen stress makes the wetland vegetation prone to water stress under dry summer conditions. Additionally, wetland vegetation reduces air exchange between the atmosphere and the air directly in contact with the wet soil or water surface compared to open water surfaces and thus limits wetland AET. Moreover, Lohila et al. [2010] found that boreal peatlands have a higher albedo than either pine forests or open water [Nunez et al., 1972], thus providing an additional mechanism for low AET of wetlands. All these factors can explain the apparent contradiction noted above of wetlands having relatively low AET, while open water environments have much higher evapotranspiration.

[36] In contrast to agriculture and wetlands, forests prevail across a large AET/P and AET/PET range space (Figure 7), reflecting the many different forest types, forest densities, and forest development stages that prevail across Sweden, and possibly also indicating resilience of forests with regard to coping with environmental changes. From previous studies, we know that many wetlands in Sweden have been drained and planted with pine forests. Due to their much higher evapotranspiration than wetlands, the pine forests can sustain themselves in these previously wet environments even without maintenance of the drainage system [Lohila et al., 2010]. Just like Williams et al. [2012], Peel et al. [2010], and Oudin et al. [2008], we did not find clear evidence that forests evapotranspire more than agricultural land cover, as is a common assumption based on the deeper rooting depth of forests and their higher interception capacity. Hence, conversion of wetlands into forests or agriculture seems to impact the terrestrial water balance much stronger than conversion of agricultural land into forests or forests into agricultural land.

[37] The natural open land cover environments in Sweden prevail mostly high up in the mountains where precipitation is high and temperature is low (Figure 2). Figure 7 shows a clear transition in resource use efficiencies from forest and their relatively high AET/P values, through wetlands, to natural open environments with the lowest water use efficiencies, AET/P. Deforestation of forests on higher elevations and latitudes, where forests do not regrow quickly, agriculture is not profitable, and thus deforested areas turn into natural open, may significantly change the terrestrial water balance. Furthermore, future increases in temperature may trigger ecosystem shifts from natural open vegetation into wetlands or forests, potentially reducing river discharge due to increased evapotranspiration.

3.5 Comparing Discharge Observations with RCM Estimates

[38] We compared our maps for T, dT, P, dP, Q, dQ, AET, and dAET (Figures 1, 3, and 5) with the results of 24 RCM runs (Figure 8) in order to assess the ability of RCM ensembles to describe the observed averages and trends. On average, the RCMs underestimate T by about 1°C and overestimate P by about 100 mm/year. These deviations compare well to the results found by Lorenz and Jacob [2010] for T in Scandinavia. Average evapotranspiration seems to be more or less in agreement with our estimates. Since the RCMs overestimate precipitation and calculate comparable AET fluxes, and since Q is calculated by the long-term difference between P and AET, Q is also overpredicted by the climate models by about the same amount as the precipitation (~100 mm/year). Overall, the measured values lie well within the range of simulated averages by the RCMs.

Figure 8.

Evaluation of the ability of RCMs to describe observed averages and trends in temperature, precipitation, evapotranspiration, and discharge. The green line indicates the mean of the RCMs. The blue line indicates the value based on measurements.

[39] The change trends in precipitation and temperature show a different picture (Figure 8). Both dP and dT are underestimated by the RCM ensemble mean by a factor of 2.0 and 1.7, respectively, compared to the data-driven estimates presented in this study (Figures 1d and 1f). Further, dAET is underestimated by a factor of 4.3 compared to our water balance–based estimate (Figure 6b), with our estimate lying far outside the range of the RCM ensemble. For dQ, there were no strong differences between the RCMs and our estimates inferred from Figure 3b. However, closer inspection of dP and dAET (Figure 8) reveals that dQ of the RCMs is actually a result of underpredicting both dP and dAET, i.e., this is a case of getting the right answer for the wrong reason. Clearly, the processes underlying the observed dP and dQ are not well represented by the RCMs. This RCM problem is also reflected by the wide distribution of discharge change trends from the RCMs. Similar mismatches in observed and projected change trends have been found by Bring and Destouni [2011] and Jarsjö et al. [2012]. In addition, the large underestimation of the dAET indicates that even though AET may be adequately represented in the RCMs, the changes in vegetation-atmosphere interactions that are evident from discharge records are not adequately captured. As we have seen in the previous section, it is likely that key processes, such as increased primary production of agricultural areas driven by optimization of farming strategies, increased biomass of forests, drainage of wetlands, and generally conversions between different land cover types, can bring about regional change trends in evapotranspiration that are not (well) implemented in the current generation of RCMs. To improve projections of river discharge toward the Baltic Sea, a better characterization of vegetation-specific evapotranspiration fluxes and their change, for example, by using the land cover–specific ranges in water and energy use efficiencies presented in Figure 7, is necessary.

4 Conclusions

[40] The data-driven regionalization approach developed here is based on an ensemble of relationships between physical parameters and specific river discharge amounts and change trends. The approach is innovative in that it does not depend on predefined model concepts. In addition to the anticipated strong relationship between precipitation and discharge, we found clear relationships between river discharge and some key land cover types, in particular open water for both the average value and the change trend of discharge, natural open areas and wetlands for the average discharge, and forest and agricultural areas for the discharge change trend.

[41] By plotting water use efficiency (AET/P) against energy use efficiency (AET/PET) for the different land cover types, we distinguished specific evapotranspiration characteristics of each land cover type and could elucidate marked differences between them. For instance, agricultural areas were found to use much of available water and energy resources and, in this respect, behave most similarly to open water areas among the different other land cover types. In contrast, wetlands were found to behave more differently than open water areas in comparison with both agricultural and forest areas, and to use relatively little of available water and energy resources. Hence, a land cover change from wetlands to forest or agriculture is expected to have a larger impact on the terrestrial water balance than a land cover change from forest to agriculture or from agriculture to forest.

[42] Further, comparison between our data-driven discharge regionalization approach and estimates from RCMs highlighted inabilities of the current generation of RCMs to represent current conditions and project future changes in different components of the terrestrial hydrological cycle in Sweden. Such inability is particularly apparent for evapotranspiration, as the RCMs cannot resolve and represent well the distinct land cover evapotranspiration relationships and changes thereof. These considerations should be taken into account when using RCM projections for the Baltic Sea region.

Acknowledgments

[43] Support for this research was received from Stockholm University's Strategic Marine Environmental Research Funds through the BEAM Program, and the Swedish Research Council (VR, project 2009–3221). The ENSEMBLES data used in this work were funded by the EU FP6 Integrated Project ENSEMBLES (contract 505539) whose support is gratefully acknowledged. Steve W. Lyon acknowledges support from the Baltic Nest Institute at Stockholm University which is partially funded by the Swedish Agency for Marine and Water Management. Stefan Dekker, Ben Livneh, and two anonymous reviewers are thanked for their comments on the manuscript.

Appendix A

[44] We compared the results of 24 RCM runs with our data-driven approach. We only used RCM runs that have both precipitation and evapotranspiration as output variables. In total, 13 different RCMs and 10 GCMs (GCMs form the boundary conditions for the RCMs) have been used. Therefore, not all 24 runs are independent. To illustrate this, we give the matrix of the evaluated combinations of RCMs and GCMs (following the ENSEMBLES project; for more information on the used abbreviations, see http://ensemblesrt3.dmi.dk/participants.html).

Table A1. The 24 RCM runs evaluated are combinations of 13 RCMs and 10 GCMs.Thumbnail image of

Ancillary