The ecological state of the Baltic Sea has been under pressure over the past few decades due to elevated nutrient inputs [1
]. Hypoxia, which has serious negative ecological implications, occurs frequently. Hypoxia in sheltered marine areas is often associated with increased nutrient levels, and nutrient loads are recognized as the main driver of hypoxia in the Baltic Sea [4
The elevated nutrient inputs originate predominantly from increased river nutrient loads [10
]. Nitrogen loads have primarily increased due to augmented use of agricultural fertilizers, compared with preindustrial levels [11
]. A source apportionment for the total riverine nutrient loading of the Baltic Sea in 2000 revealed that for total nitrogen (TN), natural background losses accounted for 28%, diffuse losses for 64%, and point source discharges for 8% of the riverine TN input to the Baltic Sea [12
]. Agriculture accounted for 70–90% of the diffuse TN load [13
]. For phosphorous, Mörth et al. [14
] estimated that 15% to 70% of the total phosphorous (TP) load originated from point sources in sparsely populated boreal catchments and densely populated temperate catchments respectively between 1996–2000. Atmospheric deposition of nitrogen from land-based combustion of fossil fuels amounts to between 25% and 33% of the total nitrogen input to the Baltic Sea [15
]. Ship traffic contributes 4–5% of the total atmospheric nitrogen deposition through fuel combustion [16
The rising incidence of eutrophication has led policy makers and researchers to focus on the development of tools to reduce the nutrient input to the Baltic Sea and thus prevent further deterioration of ecological quality [17
]. The most complex strategy developed is the Baltic Sea Action Plan [15
], which includes an eutrophication segment aiming “to have a Baltic Sea unaffected by eutrophication” and in “good ecological status” [18
]. However, knowledge-based assessment tools and management solutions are needed to achieve the objectives of the Baltic Sea Action Plan. With respect to agricultural nutrient losses, the potential impact of various management strategies needs to be assessed. According to the Helsinki Commission (HELCOM) pollution load compilation [20
], the required reduction in TN load (total TN load) is 118,000 tons N, compared to the 1997–2003 reference period. By 2010, approximately 69% of this reduction was archived (90,000 tons N), but with large variation between the seven marine basins of the Baltic Sea. The remaining requirements after the 2008–2010 period are varying between zero and approximately 62,000 tons N (26% of the 2008–2010 total load) to the Baltic Proper (the main central part of the Baltic Sea).
Development of strategies aiming at reducing the nutrient loads to the Baltic Sea must be multi-faceted and comprehensive as the success of the applied strategies depends on multiple processes. Surface runoff, soil erosion, and nutrient retention/transport between the root zone and the catchment outlet are complex processes that are influenced by several environmental and anthropogenic factors. Process-based models describing both the hydrology and the sediment and nutrient transport within catchments are useful tools for describing such complex processes, and for evaluating the impact of measures and management strategies. Such models are capable of handling large amounts of data, combined with theoretical and expert knowledge, to integrate all available information about the system of interest.
Several studies involving modelling of river water and nutrient loads to marine areas with different purposes, and using different types of models, have been published [21
]. Specifically addressing the Baltic Sea, the hydrological model HBV was set up for the entire Baltic Sea drainage basin to estimate runoff [29
], with a preliminary study on scale effects [30
]. The HBV model applied a rather crude spatial distribution approach, based on variability parameters, to model soil moisture dynamics and runoff in each individual sub-basin. The variability parameters were found to be relatively stable over a wide range of scales. Therefore, the HBV model could be applied at macro scale [30
]. The “Hydrological Predictions for the Environment” HYPE model was set up and tested for both Sweden and the entire Baltic Sea catchment, reporting a water balance error of <10% and <25%, respectively [31
Concerning the modelling of nutrient transport, a large-scale model “Catchment Simulation software” (CSIM) was set up for the entire Baltic Sea basin, aiming to describe the substantial differences in nutrient loads between the various catchments relative to geographical conditions, land use, population density, climate, etc. [14
]. This was the first modelling work of its type conducted for the Baltic Sea catchment, and the main challenge was to capture the huge scale differences in nutrient transport [14
]. Therefore, the model was kept relatively simple, focusing on key processes. The CSIM model describes inter-annual and seasonal variability of water and nutrient fluxes for 105 catchments. However, large-scale lumped models, such as CSIM, set up for entire regions are not capable of describing the impact of various mitigation measures applied at small/semi-distributed scale. Thus, to develop cost-effective adaptation strategies, models that can handle processes at small/semi-distributed scale, are needed.
In this study, the semi distributed (SWAT) was set up for six type catchments within the Baltic Sea watershed. Models like SWAT permit working on a fine spatial scale and addressing catchment and agricultural management in a detailed way [33
]. Several individual SWAT models have been set up for areas within the Baltic Sea watershed with a focus on different objectives [17
]. Ekstrand et al. [34
] calibrated the SWAT model to five river basins in Sweden (tributaries to Lake Mälaren) to improve the modelling of phosphorus losses relative to the rainfall-runoff coefficient-based Watershed Management System (WATSHMAN) model. Francos et al. [36
] applied the SWAT model to the Kerava watershed (South Finland) and found a good agreement between the measured and predicted values of runoff, total N, and total P concentrations at the outlet, especially when using precipitation data with fine spatial resolution. Lam et al. [38
] performed an assessment of point and diffuse source pollution of nitrate in the Kielstau catchment (North German lowlands). Marcinkowski et al. [42
] modelled combined climate, land use change, and fertilizer application scenarios for the Reda catchment in northern Poland. Abbaspour et al. [44
] constructed a continental scale model covering Europe and thereby also the Baltic Sea catchment. All these studies focused, however, on certain geographical areas and were not representative of the entire Baltic Sea watershed. Also, these studies dealt only with hydrological processes, or had little to no emphasis on agricultural management.
This study focuses on modelling the effect of changes in agricultural fertilization practices on nutrient loads in different river catchments around the Baltic Sea. Similar SWAT studies have previously been performed, for example, Santhi et al. [45
] evaluated the long-term effects of Water Quality Management Plans on non-point source pollution in a Texas catchment. Schilling and Wolter [46
] examined a suite of measures, among these fertilizer application reductions, to meet regulatory limits for public water supplies. White et al. [47
] modelled nutrient loads from six Oklahoma catchments, also identifying critical source areas for sediment and phosphorous. Thodsen et al. [41
] used SWAT to identify high risk and low risk areas for diffuse nutrient losses in the Odense Fjord catchment, Denmark.
We ran four agricultural fertilization scenarios. These were ±20% chemical fertilizer application and ±20% manure fertilization. In the southern part of the Baltic Sea catchment, in Poland, the Baltic countries, Russia, and Belarus, it is likely that agriculture will intensify with higher fertilizer and manure application rates as a consequence of an expansion in meat and dairy production, leading to increased diffuse nutrient losses [48
]. The scenarios increasing chemical and manure fertilization were based on this prediction, and on the possibility of water quality, not being the top priority of decisions made in all countries around the Baltic Sea. Along the northern and western rim of the Baltic Sea, in Germany, Denmark, Sweden, and Finland, application rates were likely to be stable or decline following political regulations. There is a continuous need to reduce nutrient loads to meet reduction needs in marine areas and lakes. The 20% reduction scenarios were based on this prediction.
We hypothesised that the effect of altered fertilizer/manure application would vary according to differences in the fraction of agricultural land use and agricultural fertilization practices in the Baltic catchments. The findings of our study will be of use for decision makers in targeting mitigation measures.
The aim of this study was to evaluate the effect of changed agricultural chemical fertilizer and manure application rates on nutrient loads in six type catchments within the Baltic Sea drainage basin.
In model comparison studies, it is important that the inputs to the different models are as comparable as possible. However, some inputs to the six SWAT models in this study differed, and for instance the soil data were derived from different sources because no uniform single source exists. The HWSD is a collection of national maps that vary in terms of spatial resolution; thus, the HWSD is detailed for Estonia and Lithuania but very crude for Sweden. Therefore, we decided that the best available soil map would be used for each catchment. MARS50 climate data were used in all catchments for all parameters except for precipitation (Odense and Pärnu) and temperature (Pärnu), which were in these cases available at a better spatial scale from other sources. For the Pärnu catchment, the replacement of the MAR50 precipitation with data of better spatial resolution resulted in noticeable improvements to the model; the Nash-Sutcliffe value was improved from −0.24 to 0.61 for runoff during the calibration period. The importance of good precipitation data for calibrating hydrological models is also emphasised by Chaibou et al. [76
] and by De Almeida Bressiani [77
], who tested a range of precipitation data. The Corine land use map is another aggregation of national maps, but as these are made from a common standard, harmonisation is greater than that of the HWSD, and the Corine map was therefore used in all catchments. The agricultural management and crop yield data used in setting up the models were obtained from national and EU statistics, and were therefore comparable. The expertise of the modellers with agricultural statistics differed as knowledge about local agricultural practices varied from extensive to sparse. Availability of calibration data differed in terms of the amount of data (spatial and temporal), quality of the data, and parameters of the data. All stages of the modelling process are subject to uncertainties and errors, and uncertainty is even stronger in a study comparing conditions in different geographical areas and different countries. As described above, some of the basic input data to the SWAT model, such as the soil maps, differed, increasing the uncertainty related to the soil maps, and implying higher uncertainty in our study than in a study dealing with just one homogeneously produced soil map. The same increase in uncertainty associated with a “total study” was added from all other input data. However, the six catchments were primarily chosen based on the extensive available data, and we therefore believe that the quality was the best possible for the aim of this study. Furthermore, the catchments represented the variations in geographical and agricultural management conditions in the Baltic Sea watershed. For modelling of river runoff, the quality of the climate forcing data, particularly precipitation, is important [69
]. The spatial scale of the precipitation data is highly significant. SWAT uses a time series of daily precipitation from the nearest station (in this case the centre point of a grid) and applies this precipitation to a sub-basin, implying that a single value is used for lumped areas of varying size. Potentially, this could induce certain scale problems. For example, a single-station value is applied to a large catchment, the problem being that, for example, a single thunder storm shower (>100 mm day−1
) affecting 1% of a large catchment, including the precipitation gauge location, would simulate an extreme high-flow event not occurring in reality, and would not be representative of the entire catchment when applied to the hydrological model. A similar problem may arise if gridded precipitation data is averaged over too few stations. On the other hand, the spatial scale of the gridded observations could be too large, and the variation in water discharge caused by differences in precipitation would therefore not be reflected in the model. In this study, a 50 km grid resolution (2500 km2
) was used for four of the six catchments, while a finer resolution was available for the latter two. The average sub-basin size in the four catchments ranged from 16 km2
for the River Plonia to 787 km2
for the River Kalix. Thus, it is obvious that the grid scale of the precipitation data fitted the average sub-basin size of the River Kalix better than that of the River Plonia. That is, if the River Plonia catchment had a substantial geographical precipitation gradient, the 50 km gridded data would presumably be too coarse to capture this. Potentially, the entire River Plonia catchment could be covered by only one grid cell.
By choosing SWAT to model all six catchments, the problem of comparing results from different models was avoided and a consistent modelling approach was applied, lending credibility to the comparison of calibration/validation and scenario results between the six watersheds.
The most sensitive parameters for each SWAT model at an early stage of each calibration step are shown in Table 8
. We chose not to quantify parameter sensitivity, as the sensibility of single parameters depends on a number of circumstances—for instance, choice of objective function, temporal resolution of calibration data, calibration procedure, spatial resolution of input data, conceptual model uncertainty, modeller’s knowledge of the modelled catchment, chosen initial calibration parameter range, climate of the chosen calibration, and validation period—at the stage of the calibration procedure during which the sensitivity analysis is performed as well as on the model output in focus [69
]. For example, the modeller’s knowledge about the catchment and experience in calibrating models for a particular area may strongly influence parameter sensitivity. In this study, the modelling team had extensive experience in modelling the River Odense catchment, leading to knowledge about parameter setting for the modelling of NO3
, where the denitrification threshold water content (SDNCO) should be between 0.75 and 0.99 and the denitrification exponential rate coefficient below 1 to produce realistic denitrification rates, thereby avoiding non-uniqueness problems [69
]. Extensive knowledge reduced the sensitivity of these parameters compared with a situation where calibration was initiated with a full range. Similar knowledge was not available for the River Kalix, Norrström, Nevezis and Pärnu, and the range of these parameters therefore had to be larger, producing a potentially greater sensitivity. Where calibration data were available with daily resolution, parameters addressing fast-responding parameters were more sensitive than where calibration data were available with monthly resolution.
Model uncertainties relative to reproduction of observational values of Q, NO3
and MinP are evaluated in Table 5
. Moriasi et al. [78
] states that Nash-Sutcliffe efficiency values > 0.75 for monthly time steps of runoff are “very good” (Rivers Odense, Kalix, Nevezis), values between 0.75 and 0.65 are “good” and values between 0.65 and 0.50 are “satisfactory” (Rivers Norrström, Pärnu, Plonia) (Table 5
). Nash-Sutcliffe values calculated from daily values are usually lower than values calculated from monthly values. Mass balance “percent BIAS” (PBIAS) (%) for N and P estimates of ±25% were considered “very good” (River Odense NO3
& MinP, River Kalix NO3
& MinP, River Pärnu NO3
, River Nevezis MinP and River Plonia NO3
& MinP), ±25% to ±40% as “good” (River Pärnu MinP) and the interval ±40% to ±70% as “satisfactory” (River Nevezis NO3
). Overall, the validation statistics were considered adequate.
The total water balance error of the six SWAT model validations was relatively good, with a maximum error of 11%, but when evaluated according to the 0.25 percentile (25% of all values are smaller than this) and the 0.75 percentile, the models had larger biases (Table 5
The scenario simulations showed that the effect of altered fertilizer application was strongest in catchments with a large fraction of agricultural land use and with intense agriculture like the River Odense and River Plonia catchments (Table 1
and Table 6
). River Norrström and River Nevezis exhibited a stronger response to the scenarios than River Pärnu, although the combined agricultural pressure in the three catchments was about the same. River Kalix showed a negligible response to the scenarios, which was expected, as agriculture only occupies 0.5% of its catchment area.
For the Kalix, Norrström, Pärnu and Nevezis catchments, the modelled effects of the scenarios on MinP were of the same magnitude as the effect on NO3
, and were an order of magnitude smaller on MinP than on NO3
for the River Odense and Plonia catchments. This reflected the higher NO3
input to these intensely farmed catchments. Besides this, the effects on MinP were rather linked to soil erosion processes than to leaching processes. Therefore, catchments with a relatively continental climate and, consequently, relatively large erosive spring snow melt events, showed a stronger response than catchments with a more Atlantic climate having milder winters and a smaller build-up of snow. Additionally, SWAT did not simulate leaching of dissolved phosphorous from the soil to the groundwater and into the river [54
The results presented in this paper for the Baltic Sea catchments suggest that the effect of changed agricultural nutrient applications on riverine nutrient loads will be strongest in areas with relatively intensive fertilizer application. Therefore, decision makers should focus on mitigation methods in these areas if they are aiming at maximum impacts per hectare of agricultural land. Large differences in nutrient loads in the high intensity agricultural areas remain, though, due to differences in, for instance, fertilizer application procedures and crop yield and notable differences in retention found primarily for nitrogen [79
]. A study running a +20% fertilization scenario for catchments in central Germany found NO3
river load increases from 2% to 6% [75
Along the southern rim of the Baltic Sea, in Poland (despite the fact that the Polish Plonia catchment has the highest chemical fertilizer application rates among the six catchments included in this study), the Baltic countries, Russia and Belarus, it is very likely that agriculture will be intensified, with higher fertilizer and manure application rates, as a consequence of an expansion of meat and dairy production, leading to increased diffuse nutrient losses [48
]. Along the northern and western rim of the Baltic Sea, in Germany, Denmark, Sweden and Finland, application rates are likely to be stable, or will decline following political regulations.
The loads of organic nitrogen and phosphorous make up a substantial part of the total nutrient input to the Baltic Sea, and a primary part of total nutrient inputs in northern boreal forest areas [14
]. The organic fractions were not considered in this study for two reasons: (1) data were not available for all six catchments; (2) the scenarios were not thought to substantially influence the load of organic N and P.