Assessing N emissions in surface water at the national level: Comparison of country-wide vs. regionalized models

Science of the Total Environment 443 (2013) 152–162

Contents lists available at SciVerse ScienceDirect

Science of the Total Environment journal homepage: www.elsevier.com/locate/scitotenv

Assessing N emissions in surface water at the national level: Comparison of country-wide vs. regionalized models Rémi Dupas a,⁎, Florence Curie b, Chantal Gascuel-Odoux a, Florentina Moatar b, Magalie Delmas c, Virginie Parnaudeau a, Patrick Durand a a b c

INRA, UMR 1069, Soil Agro and hydroSystem, F-35000 Rennes, France Université François Rabelais – Tours, EA 6293, Géo-Hydrosystèmes Continentaux, Faculté des Sciences et Techniques, Parc de Grandmont, F-37200 Tours, France INRA, US 1106, InfoSol Unit, F-45075 Orleans 2, France

H I G H L I G H T S ► ► ► ►

We model N-nitrate emissions at national and regional scales. Terrestrial basin retention accounts for 49% of N surplus (median value for France). River retention represents 18% of incoming N discharge (median value for France). Regionalizing N load models improves fit, but precision of parameter estimates decreases.

a r t i c l e

i n f o

Article history: Received 27 July 2012 Received in revised form 1 October 2012 Accepted 1 October 2012 Available online 24 November 2012 Keywords: Nitrate Nutrient retention Nitrogen surplus Statistical modeling Load estimation Catchment

a b s t r a c t Many countries are developing models to estimate N emissions in rivers as part of national-scale water quality assessments. Generally, models are applied with national databases, while at the regional scale, more detailed databases are sometimes available. This paper discusses pros and cons of developing regionalized models versus applying countrywide models. A case study is used to support the discussion. The model used, called Nutting-N (NUTrient Transfer modelING-Nitrogen), relies on a statistical approach linking nitrogen sources and watershed land and river characteristics and aims to evaluate the risk of water bodies failing to reach quality objectives defined by national and federal policies. After calibration and evaluation at the national scale (France), the predictive quality of the model was compared with two regionalized models in a crystalline massif (Brittany, western France, 27,000 km2) and in a sedimentary basin (Seine, Paris basin, 78,000 km2), where detailed regional databases are available. The national-scale model provided robust predictions in most conditions encountered in France (efficiency=0.69). Terrestrial retention was related mainly to specific runoff, and its median value was estimated at 49% of the N surplus, whereas median river retention represented 18% of incoming N discharge. Regionalizing the model generally improved goodness-of-fit, as the root mean squared error was reduced by 6– 24%. However, precision of parameter estimates degraded when too few monitoring basins were available or when variability in land and river characteristics was too low in the calibration dataset. Hence, regional-scale models should be advocated only after the trade-off between improvement of fit and degradation of parameter estimates is examined. © 2012 Elsevier B.V. All rights reserved.

1. Introduction Nitrate is a major cause of surface water degradation, both in terms of ecosystem health and drinking water quality (Sutton et al., 2011). The European Union Water Framework Directive aims at reaching good status for all water bodies by 2015 (Official Journal of the European Communities, 2000). Therefore, all member states are required to analyze nutrient pressures and impacts to: i) evaluate the

⁎ Corresponding author. Tel.: +33 2 23 48 70 47. E-mail address: [email protected] (R. Dupas). 0048-9697/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.scitotenv.2012.10.011

risk of water bodies failing to reach good status and identify those watersheds where prioritized nonpoint-source control measures should be implemented; and ii) assess the contributions of different N sources. Complementing monitoring networks with the use of models helps estimate N loads in areas where sampling points are lacking or are unreliable due to sparse sampling frequency, and improves understanding of the processes controlling nutrient transfer. A number of models have been developed for estimating N transfer from agricultural land to aquatic ecosystems, ranging from detailed process-based models to statistical regression models (Billen et al., 2011). These models differ in complexity, data requirement and objectives, and therefore in scale of application. Statistical models are lumped

R. Dupas et al. / Science of the Total Environment 443 (2013) 152–162

models that link the observed N load to the sum of all N net inputs (called Net Anthropogenic Nitrogen Input (NANI), Billen et al. (2009)) via simple linear regressions (Billen et al., 2009; Howarth et al., 1996) or to several input types and multiple linear regressions (Alexander et al., 2002; Caraco and Cole, 1999; Seitzinger and Kroeze, 1998). Howarth et al. (2006) expressed the regression coefficient as a function of precipitation or discharge, as first highlighted by Behrendt and Opitz (1999). The export-coefficient approach is also a statistical approach that relates N load to the percentage of different land-use types at regional (Grimvall and Stalnacke, 1996; Johnes, 1996; Johnes and Heathwaite, 1997; McFarland and Hauck, 2001) or national scales (Worrall et al., 2012). At an increasing degree of complexity, we find simple conceptual models that are spatially referenced nonlinear regression models that predict N loads based on the quantification of diffuse sources and point sources, together with a few explanatory variables. The most widely used are the U.S. model SPARROW (Smith et al., 1997) and the European model GREEN (Grizzetti et al., 2008), but the same structure has been used in other models (Dumont et al., 2005; Mayorga et al., 2010; Pieterse et al., 2003; Seitzinger et al., 2005). Semi-empirical conceptual models, such as MONERIS (Behrendt et al., 2002) and POLFLOW (de Wit, 2001) have a higher degree of complexity by considering different flow pathways in addition to in-stream retention. Process-based models often require a large number of input data at a high spatial and temporal resolution, and therefore are often not adapted to large scales (Schoumans et al., 2009). However, a few process-based models have been applied to large watersheds or even at the country scale; most comprise elements taken from existing crop models such as ANIMO (Groenendijk et al., 2005), STICS (Brisson et al., 2003), and DNDC (Li et al., 1992), combined with a hydrological model. The resulting models (Beaujouan et al., 2001; Ledoux et al., 2007; Leip et al., 2008; Lindstrom et al., 2010; Schoumans et al., 2009; Wolf et al., 2003) display N transfer estimates at fine time and space resolutions, provided that reliable data are available and correct parameterization is performed (Schoumans et al., 2009). The model presented here, called Nutting-N (NUTrient Transfer modeling-Nitrogen), is inspired by SPARROW and GREEN. According to Preston et al. (2011), this type of model constitutes an ideal tradeoff between complexity and applicability at the large scale because the required databases are generally available. Both models have been applied at various scales, ranging from continental – the U.S. for SPARROW (Alexander et al., 2000; Smith et al., 1997) and Europe for GREEN (Grizzetti et al., 2008) – to regional: 6 of the 8 major U.S. river basins for SPARROW (Preston et al., 2011) and a variety of well-documented European watersheds, such as the Great Ouse (Grizzetti et al., 2005a),

153

the Wash, Vilaine and Zelvka rivers (Grizzetti et al., 2005b) for GREEN. Regionalization of the SPARROW national-scale model improved model accuracy (Schwarz et al., 2006), although uncertainty in coefficient estimates may increase because regional models encompass a narrower range of variability in basin characteristics (Schwarz et al., 2011). The objective of this study is threefold: i) develop an innovative model to simulate N loads in each water body at the national scale by computing N surplus as a diffuse source, instead of as N input in SPARROW and GREEN; ii) test potential explanatory variables specific to France and evaluate which are the main factors controlling N export; and iii) compare goodness-of-fit and uncertainty of parameter estimates for a national model and for regionalized models developed in France in two contrasting regions, regarding their physical environment and their farming systems. The two study regions are i) coastal watersheds in Brittany located on a crystalline massif, where the agricultural sector is devoted to intensive animal production, and ii) sub-watersheds of the Seine river basin, in the Paris sedimentary basin, where land use is dedicated mainly to cereals and industrial crops. 2. Methods 2.1. Model description 2.1.1. Overview of the model The Nutting-N model consists of a statistical approach linking N sources and watershed land and river characteristics. Total runoff is divided into shallow and deep components (Fig. 1), each of which transfers a distinct N source. N surplus (i.e. soil-surface agricultural N balance and atmospheric deposition) is transferred only by subsurface flow, whereas the N-nitrate concentration measured in aquifers is transferred only by deep groundwater flow. Surface runoff was not considered due to its low contribution to nitrate transfer. Partitioning total runoff into shallow and deep components addresses the issue of time lags: measured N-nitrate concentration in aquifers may differ from sub-root concentration in situations where aquifer concentration is not yet in equilibrium with current N surplus on agricultural lands (Billen et al., 2009). In other words, N concentrations measured in aquifers today do not necessarily reflect current agricultural management practices aggregated in the N-surplus calculation, but rather the fertilization practices of the past few decades (Van Drecht et al., 2003; Molenat et al., 2008). The partition into two runoff components relies on the heuristic method developed by Meinardi et al. (1994), which determines a base flow index in ungaged basins, accounting for the

Fig. 1. Pathways and processes in Nutting-N. L is the specific nitrate load at the outlet of each watershed; N sources (N surplus, [NO3−] aquifer and point sources) are represented in light gray ovals; shallow groundwater discharge to streams (Lsgw) and deep groundwater discharge to streams (Ldgw) are represented by dark gray arrows; catchment- and river-reduction factors (B and R) are represented by thin arrows.

154

R. Dupas et al. / Science of the Total Environment 443 (2013) 152–162

combined effects of aquifer type, topsoil texture, slope and land use. The name of the model, Nutting, symbolizes the model's construction from few datasets, which cannot completely represent the process involved in N transport from agricultural lands to rivers. The nitrate specific load (L) at the outlet of each river basin is expressed as:
 L ¼ R  B  Lsgw þ Ldgw þ PS −denit lake

ð1Þ

where is the N load coming from shallow groundwater discharge to streams (derived from the base flow index and N surplus in kg N.ha −1.yr −1), Ldgw is the N load coming from deep groundwater discharge to streams (derived from total runoff, the base flow index and deep groundwater N concentration), PS is point sources from domestic and industrial origins (kg N.ha −1.yr −1), R and B are river- and watershed-retention factors, respectively. They themselves combine observed variables and calibrated parameters (developed in Sections 2.1.5. and 2.1.6). denitlake is a retention factor for lakes and reservoirs.

Lsgw

2.1.2. N load coming from shallow groundwater discharge to streams Lsgw Shallow groundwater discharge to streams consists of subsurface runoff and interflow. The Lsgw term in Eq. (1) is deduced from the base flow index of Meinardi et al. (1994) and N surplus estimated by NOPOLU System2 (Solagro, 2010; Schoumans et al., 2009). NOPOLU System2 is a model that calculates the yearly N surplus on land surfaces in line with the OECD (2001) definition. NOPOLU System2 methodology is based on disaggregation of statistical data at various administrative levels to any relevant hydrological level using land-cover information. It requires data from agricultural censuses (e.g., livestock, crops), land-cover data and GIS information about the administrative and hydrological scales considered (Campling et al., 2005). A wide range of N processes are modeled, including symbiotic fixation, atmospheric deposition, application of synthetic and organic fertilizer, volatilization in livestock buildings/storing areas/fields, as well as all N exports. 2.1.3. N load coming from deep groundwater discharge to streams Ldgw The Ldgw term is calculated on the assumption that base flow can be expressed as a constant fraction of total runoff. Hence, data on base flow index (Meinardi et al., 1994) and on spatially distributed long-term runoff (Sauquet, 2006) are multiplied to estimate base flow in each of the study watersheds (Fig. 2.) prior to calibration of Nutting-N. N loads from deep groundwater discharge are deduced from this base flow calculation and nitrate concentrations in groundwater. When available in

Fig. 2. Watershed characteristics (France dataset).

R. Dupas et al. / Science of the Total Environment 443 (2013) 152–162

regional studies, measured data for runoff and base flow from flow stations are used (Institute of Hydrology, 1980). 2.1.4. Point sources PS The PS term is the sum of all domestic and industrial N emissions to streams in the study area. 2.1.5. Watershed N retention factor B Nutting-N is a static model, i.e. it assumes that under steady-state conditions there is no change in the soil N pool: N nitrification and immobilization balance each other. As a consequence, the N surplus corresponds to a soil N pool that is either retained or transferred within the interannual period of study. The partition of N surplus between retention and transfer is aggregated in the watershed-retention factor B, which is expressed as an exponential land-to-water delivery function: B ¼ expð∑i αi  XiÞ

ð2Þ

where Xi are the independent variables of the model, and αi are parameters to be calibrated. A wide range of factors influence the rate at which diffuse-source N is transferred to streams, including hydrological and meteorological conditions, topography, soil properties, land use, management practices and riparian wetlands. High runoff values increase N transfer to streams (Behrendt and Opitz, 1999; Dumont et al., 2005; Howarth et al., 2006) since they reduce N residence time in watersheds, decreasing the probability of retention processes occurring (Green et al., 2004). As Nutting-N aims to be applicable to ungaged basins, total runoff was approximated by effective rainfall, computed as the sum of P-ETP for the months when P-ETP > 0. Mean distance to stream, stream density and mean slope are topographic factors which also influence N residence time in the watershed and thus N retention. The soil factors expected to influence N leaching most are texture, depth and organic matter content (Velthof et al., 2009). Soil permeability and hydraulic conductivity are the prevalent soil variables in N transfer models (Grizzetti et al., 2008; Smith et al., 1997), since they influence both the vulnerability to N leaching and the probability of denitrification conditions occurring in soils. We derived hydraulic conductivity from soil textures described in soil maps using pedotransfer rules of Wosten et al. (2001). Forests and semi-natural areas can affect transport and transformation of N and act like water purifiers. Therefore, the percentage of forests and semi-natural areas was tested as a potential explanatory variable. Regarding management practices, fertilizer use is already encapsulated in the N surplus term. Other management practices, such as the use of cover crops, affect N leaching but could not be considered due to the lack of data at regional and national scales. Riparian wetlands play an important role in the control of N transfer by reducing N concentration in water, particularly through denitrification (Curie et al., 2009, 2011; Montreuil and Merot, 2006; Montreuil et al., 2010). The topographic wetness index (Beven and Kirkby, 1979) estimates the degree of water accumulation at a site (Merot et al., 2003) and is computed as:

155

main drivers of in-stream N retention (Seitzinger et al., 2002). In the European model GREEN, the river-retention factor is a function only of normalized river length (Grizzetti et al., 2008), whereas its U.S. counterpart SPARROW can integrate channel-geometry data that are available in the U.S.: travel time, streamflow, and whether the reach is part of a reservoir (Smith et al., 1997). Since recent models have been developed to estimate these hydraulic characteristics of rivers at the French national scale (Pella et al., in press), we integrated a river reduction-factor similar to that of Boyer et al. (2006), which requires data on mean residence time and channel depth. Hence, the mean annual rate of nutrient removal in streams is estimated as a first-order decay relationship: R ¼ expðαj  T=DÞ

ð3Þ

where αj is a mass-transfer coefficient to be calibrated (see Section 2.5.), T is the mean residence time (computed as the ratio of reach length over mean stream velocity) and D is the mean stream depth. The D/T ratio is hereafter called “hydraulic load”. The N retention factor for lakes and reservoirs was taken from the Euroharp nutrient retention tool EUROHARP-NUTRET. This tool was developed from a database of measured N inputs and outputs for 65 lakes and 113 reservoirs throughout the world (Kronvang et al., 2004). N retention was calculated by applying specific retention rates to the lakes and reservoirs in the studied watersheds, grouped by hydraulic residence time category. Hydraulic residence time was calculated as the ratio between hydraulic volume and mean interannual discharge, as estimated by the hydrological model LOIEAU (Folton and Lavabre, 2006). 2.2. Study areas A 2007 review of national monitoring data in France shows that the mean nitrate concentration in surface water was above 30 mg.L−1 in 21 out of 177 hydrographic sectors. In addition, 13 of the 1628 monitoring stations reviewed had an annual mean concentration above 50 mg.L −1, the maximum limit for drinking water. These nitrate hot spots were located mostly in northwestern France, where intensive agriculture is the dominant land use. Thus, the model was calibrated and evaluated for the whole of France, as well as for two contrasting regions located in the northwest: coastal rivers of Brittany and sub-watersheds of the Seine river basin.

where ‘a’ is the specific watershed area and ‘b’ is the local slope. It can be entered in the model as is, or after refinement of riparian wetland delineation via further techniques, such as remote sensing (Agence de l'eau Seine Normandie, 2006) or machine-learning methods integrating various soil and climate factors (Lemercier et al., 2012).

2.2.1. Seine The Seine river basin is located in northern/central France and covers an area of 78,000 km2 (14% of France) in the Paris geological basin, a sedimentary basin composed of a succession of consolidated and unconsolidated rock layers. The main soil textures are loams and clay. The Seine basin has a temperate climate, with a mean annual rainfall ranging from 620 mm (Beauce region, in the south) to nearly 1100 mm (Morvan region, in the southeast). Mean long-term annual rainfall over the entire basin is 745 mm, while mean evapotranspiration approaches 600 mm (Viennot et al., 2009). The Paris basin is rather flat and generally lower than 300 m above sea level, except in the southeast (Morvan), where maximum altitude reaches 900 m. Land use in the Seine river basin is mostly dedicated to cereals and industrial crops, supported by intense application of synthetic fertilizers. Mean N surplus equals 26 kg N.ha−1, and specific nitrate fluxes range from 3 to 24 kg N.ha−1.yr−1. This river basin has high population densities (nearly 200 inhab.km−2), mostly concentrated along the main branch of the Seine river between Paris and Le Havre.

2.1.6. River-retention factor R and lake- and reservoir-retention factor denitlake Channel flow conditions (e.g., water discharge, velocity, channel depth, width and slope) and water temperature are known as the

2.2.2. Brittany The Brittany region is located in western France and covers 27,000 km 2 (5% of France). Brittany belongs to the Armorican Massif, with geology dominated by crystalline bedrock overlain by superficial

TWI ¼ ln


a  tanb

156

R. Dupas et al. / Science of the Total Environment 443 (2013) 152–162

formations, such as Quaternary loess (aeolian silt) and alluvial and colluvial deposits. Silty loam is the dominant soil texture. Brittany has a temperate oceanic climate, with a mean annual temperature of 11.2 °C. Annual rainfall ranges from 700 to 1200 mm and mean evapotranspiration from 500 to 600 mm (Molenat et al., 2002, 2008). Due to the impervious substratum and wet climate, drainage density is high (about 1 km/km 2). Groundwater is shallow, causing waterlogging of soils along valleys. Valley-bottom wetlands account for 15–20% of the land surface. A large part of the agricultural sector is devoted to intensive animal production, representing 55%, 40%, and 25% of the national production of pigs, chickens, and cow milk, respectively. Mean N surplus reaches 45.5 kg N.ha −1, and specific nitrate fluxes reach 9–89 kg N.ha −1.yr −1, with high variability in space and time (Gascuel-Odoux et al., 2010).

load datasets (whole of France, Seine and Brittany) were created, based on five criteria: i) presence of a flow station and a quality station on the same reach; ii) at least one nitrate measurement per month from 2005 to 2009; iii) not more than two missing data points per year; iv) independent watersheds (though two nested basins could be taken together if the sub-basin surface area was b 20% of the basin containing it); and v) watersheds in which point sources represented b50% of N pressure (so model calibration did not rely too much on point-source estimates). Flow data were available daily, whereas nitrate sampling frequency was lower (generally monthly). The discharge-weighted concentration method (Moatar and Meybeck, 2007; Moatar et al., 2012) was used to estimate mean interannual nitrate load, i.e. the product of discharge-weighted mean concentration and mean interannual discharge. The set of watersheds used in the calibration encompass a wide range of variability regarding hydrology and nitrogen characteristics (Fig. 2).

2.3. Datasets 2.5. Variable selection and model parameterization Table 1 summarizes the datasets used for calibration/application of Nutting-N in the whole of France, Brittany and the Seine river basin. Since 2007, measurement of N emissions from all domestic and industrial point sources has become a legal obligation in France. Since data are generally available for 2008 and/or 2009, depending on the river basin district, the average value over these two years was used. If not, emission factors estimated by water agencies for the last available year (usually 2007) were used. 2.4. Calculation of mean interannual N loads for model calibration Calibration of the model requires observed mean interannual Nnitrate load data, which were derived from a network of stations monitoring water quality and flow within the three study areas (whole of France, Brittany and Seine). Nutting-N assumes that N inputs and outputs (including long-term N retention) are balanced within the interannual period of calibration, i.e. the weather represents longer-term climate conditions, and agricultural management practices remain quasi-stationary. Therefore, the calibration period must be long enough to smooth out interannual variations in hydrological conditions, but not too long, so as to comply with the latter assumption. Indeed, agricultural management is changing rapidly, especially in areas where water quality action programs are implemented. While most SPARROW models were calibrated with detrended data of long-term load estimates spanning 15- to 20year periods (Schwarz et al., 2006), our calibration period was only 4– 5 years to avoid complications due to the specification of trends. The 2005–2009 period was chosen, as it included wet and dry years. In addition, it was consistent with the cyclic hydrological pattern observed by Gascuel-Odoux et al. (2010) in Brittany. Three independent N-nitrate

Several watershed characteristics were identified as potential explanatory variables (Table 1) to be entered in the B and R terms of Eq. (1). Since calibration was performed considering a lumped model in a number of watersheds (Fig. 4), Xi factors (Eq. (2)) were taken as average values. The variable selection procedure aimed to select a small number of uncorrelated variables to optimize model accuracy and prevent over fitting. Therefore, calibration occurred in two steps. First, the model was linearized to identify variables to select and determine initial values for the parameters before performing a Gauss–Newton algorithm. The R factor was replaced by an in-stream specific denitrification factor, denitriver (Kronvang et al., 2004): L ¼ B  Lsgw þ Ldgw þ PS−denit lake −denit river :

ð4Þ

The denitriver factor was adjusted proportionally to each watershed's estimated N surplus. After log-transformation and replacing the B term by its expression in Eq. (2), Eq. (4) became a multiple regression model: L−Ldgw −PS þ denit lake þ denit river log Lsgw

! ¼ ∑i αi  Xi:

ð5Þ

Then, variable selection was performed according to the Bayesian Information Criterion statistic, and the linear model was fitted with an analysis of variance. In the second step, the initial model (Eq. (1)) was recalibrated by minimizing a weighted least-squared objective function with a modified Gauss–Newton algorithm. Variables were selected to

Table 1 Database used for calibration of the France, Brittany and Seine models.

Lsgw PS B

Input variable or potential explanatory variable

France model

N surplus N point sources Effective rainfall–runoff factor Topographic factors Soil factor

NOPOLU 2007 (SOeS) Industrial and domestic emissions (water agencies) P-ETP (Météo France Total runoff (Banque Hydro) P-ETP (Météo France SAFRAN database) SAFRAN database) Mean distance to stream, stream density (BD Carthage) and mean slope (50 m DEM) Soil texture (France Soil texture (Brittany 1:250,000 Soil texture (France 1:1,000,000 soil map) 1:1,000,000 soil map) soil map) Corine Land Cover 2006 class 3

Forests and semi-natural areas factor Riparian wetlands factor R River system data Denitlake Lakes and reservoir data Ldgw N concentration in aquifers Total runoff BFI

Brittany model

Seine model

Beven Index (DEM 50 m resolution map of hydromorphic Map of riparian wetlands (AESN, 2006) 50 m) soils (Lemercier et al., 2012) BD Carthage + hydraulic geometries as modeled by ESTIMKART (Pella et al., in press) BD carthage + hydraulic residence time as modeled by CEMAGREF Nitrate concentration (ADES database) Total runoff map (Sauquet, 2006) Base flow index as calculated by Meinardi et al. (1994) Base flow index as calculated by Meinardi et al. (1994) + base flow index as estimated by the Wellington method

R. Dupas et al. / Science of the Total Environment 443 (2013) 152–162

157

estimation by the models was assessed by comparing their statistical significance when they were recalibrated with the entire dataset for each study site. Performance of Nutting-N was first evaluated for a national-scale model (Section 3.1) followed by two regionalized models (Brittany, Seine). Regionalizing a model consisted of performing specific calibration for a region, considering either nationally available data (Section 3.2) or more detailed local data (Section 3.3).

3. Results 3.1. Model development and evaluation at the national scale

Fig. 3. Predicted versus observed N-nitrate specific load, national model. Horizontal error bars represent observed loads first and last deciles. The red line is a 1:1 line.

minimize root mean squared error of prediction during a crossvalidation procedure, and initial parameter values estimated in the modified linear model (Eq. (5)) were entered. In the final model, since all αi were expected to be negative, B factors were kept in the [0,1] interval. Hence, variables expected to positively influence N transfer were taken in their inverse form. 2.6. Model evaluation First, model goodness-of-fit was assessed with the root mean squared error (RMSE) and R2 (Moriasi et al., 2007). Both model-fit statistics were calculated according to a ‘leave one out’ cross-validation: from a dataset containing n watersheds, model accuracy was evaluated by comparing model predictions when calibrated on n-1 watersheds to the observed load of each watershed. Second, precision of coefficient

The national-scale model was calibrated with a basin retention factor including only effective rainfall as an independent variable, while the river-retention factor was calibrated using the inverse of the areal hydraulic load (i.e. the ratio between residence time and mean channel depth). Fig. 3 shows predicted versus observed nitrate-specific loads for 182 watersheds resulting from cross-validation: values of R2 and RMSE reached 0.69 and 6.34 kg N.ha −1.yr−1, respectively. All parameter estimates for the B and R factors were significant at p b 0.01 (Table 2). Based on these parameter estimates, France's median value of basin retention (1-B) was 49% of N surplus, and river retention (1-R) was 18% of N discharge to streams. The highest predicted N-nitrate loads were found in western and northern France, i.e. the regions where the highest nitrate concentrations are observed in streams (Fig. 4). This corresponds to areas with high effective rainfall and N surplus, which confirms that these two factors mainly control N-nitrate load in France. Spatial distribution of prediction error shows that the model tends to overestimate nitrate loads in mountainous regions dominated by forest and grassland, such as the Vosges, Alps, Pyrenees, and Massif Central. Model goodness-of-fit is best for flat areas dedicated to arable land, such as the Aquitaine basin, Parisian basin, and western France. A specific application of the national model was performed in the two latter regions (Fig. 5). Model efficiency (R 2) was lower than that for the whole-of-France cross-validation (R 2 = 0.36 in Brittany and 0.61 in the Seine basin), and highly productive watersheds were severely underestimated in Brittany.

Fig. 4. Map of predicted N-nitrate specific load (left) and its prediction error (right).

158

R. Dupas et al. / Science of the Total Environment 443 (2013) 152–162

Table 2 Results of parameter estimations. Significance codes: (***): P value b 0.001; (**): P value b 0.01; (*): P value b 0.05; (.): P value b 0.1.

a. France model b. Seine model (national data) c. Seine model (regional data) d. Brittany model (national data) e. Brittany model (regional data)

Model parameters

Units

Estimate

Standard error

t value

P value

Effective rainfall Hydraulic load Effective rainfall Hydraulic load Specific runoff Hydraulic load Effective rainfall Hydraulic load Specific runoff %hydromorphic soils Average distance to stream Average slope Correction factor Hydraulic load

mm s−1.ha.m mm s−1.ha.m mm s−1.ha.m mm s−1.ha.m mm %−1 m−1 %−1 – s−1.ha.m

2.77e + 02 2.25e − 03 1.08e + 02 6.49e − 03 2.43e + 02 1.68e − 03 3.19e + 02 3.02e − 04 1.31e + 02 3.20e − 02 1.88e − 03 9.26e − 2 1.69e + 00 3.30e − 03

4.28e + 01 8.11e − 01 6.347e + 01 2.12e − 03 5.910e + 01 7.225e − 04 8.076e + 01 1.398e − 03 5.311e + 01 7.088e − 03 8.197e − 04 2.610e − 02 3.260e − 01 1.618e − 03

6.429 2.769 1.70 3.06 4.109 2.329 3.952 0.216 2.475 4.518 2.287 3.550 5.172 2.041

9.018–10 6.2e − 03 9.65e − 02 3.85e − 03 1.80e − 04 2.48e − 02 2.20e − 04 8.30e − 01 1.74e − 02 3.63e − 05 2.63e − 02 8.28e − 04 3.76e − 06 4.64e − 02

3.2. Model regionalization with the national database For this part of the study, the same database was used as for the national calibration to evaluate the relevance of regionalizing the model without introducing new data. For both Brittany and Seine regionalized models, the basin-retention factor was a function only of effective rainfall, and the river-transfer factor a function of areal hydraulic load. Fig. 6 shows predicted versus observed nitrate specific loads for the 58 Brittany watersheds and the 44 Seine sub-watersheds, resulting from the cross-validation. R2 reached 0.51 in Brittany and 0.66 in the Seine basin, a slight improvement compared to the simple application of the nationally calibrated model, since RMSE was reduced by 12% in the Brittany watersheds but by only 6% in the Seine sub-watersheds. According to this calibration, median basin and river retention were, respectively, 38% and 48% in the Seine region and 49% and b1% in Brittany. However, significance of at least one parameter estimate for this calibration in each region exceeded 0.05 (Table 2); this suggests that these estimates are not reliable and should not be used in predictive models. 3.3. Model regionalization with local data Regional datasets were used to perform a second regionalization of the model (Table 1). In the Seine basin (Fig. 8), detailed hydrologic attributes derived from hygrograph analysis were considered: measured total runoff was used instead of previous long-term spatially explicit

*** ** . ** *** * ** * *** * *** *** *

runoff estimates by Sauquet (2006), and BFi was estimated by the Wallingford method (Institute of Hydrology, 1980) instead of the method of Meinardi et al. (1994). The same variables were entered in the model as in the first regionalized model, but both goodness-of-fit (R2 = 0.77, Fig. 7) and precision of parameter estimates were further improved (Table 2). The Brittany model was calibrated without deep groundwater discharge to streams, since contribution of this flow component is considered minor in crystalline bedrock. This change in the original model allowed the use of several variables in the basin-retention factor: specific runoff, percentage of hydromorphic soils, mean distance to streams and mean slope (Table 2). A regression constant, called the correction factor, was added to counterbalance the fact that N loads were higher than N surplus in 4 out of the 58 Brittany watersheds, which contradicts the steady-state hypothesis. This can be explained by discharge of N stored in the weathered bedrock (Molenat et al., 2002, 2008), but more likely by poor estimation of the N surplus in some watersheds, such as those situated in vegetable producing areas. The latter model improved both accuracy and precision of parameter estimates compared to a regional application of the national model or a regionalization with national datasets. Model efficiency (R2) was 0.64. Ultimately, regionalization with local data improved RMSE by 24% in the Brittany watersheds and 23% in the Seine sub-watersheds, compared to a simple application of the national model. Additionally, all parameter estimates were significant at p b 0.01, which means that these models can be used for prediction in other watersheds of these regions.

Fig. 5. Predicted versus observed N-nitrate specific load, the former from the national-scale model in Brittany coastal watersheds (left) and Seine sub-watersheds (right). Horizontal error bars represent observed loads first and last deciles. The red line is a 1:1 line.

R. Dupas et al. / Science of the Total Environment 443 (2013) 152–162

According to this calibration, median basin and river retention were, respectively, 66% and 16% in the Seine region and 23% and 27% in Brittany. 4. Discussion Although the variables selected in the watershed-retention factor B differed from one regionalized model to another, the general structure of Eq. (1) did not change. Hence, the predictive quality and the variable entered in each regionalized model can be compared and discussed. 4.1. Comparative evaluation of national-scale and regionalized models Overall, the nationally calibrated model provided good estimates of nitrate loads in most of the 182 study watersheds. R 2 and RMSE reached 0.69 and 6.34 kg N.ha −1.yr −1, respectively, which falls within the range of R 2 values recorded for regional and national SPARROW models, i.e. 0.60–0.86 (Preston et al., 2011). However, a clear bias was observed for mountainous regions, for which the model tended to overpredict nitrate loads. Several reasons can account for this bias: i) estimation of N agricultural surplus is more complicated for pastures, particularly when transhumance exists; ii) hilly watersheds are underrepresented in our dataset; and iii) the proportion of natural and semi-natural areas is high in these watersheds, contrary to the majority of our study watersheds. As too few monitoring basins exist in these regions, it was impossible to establish an acceptable dataset for specific calibration in mountainous areas. Application of the national model in the Seine basin did not reveal any significant bias, and R 2 was of the same order of magnitude as that for the national cross-validation (R 2 = 0.61). The model predicted well in Brittany, except for underestimating highly productive watersheds (i.e. > 40 kg N.ha −1.yr −1), due to such watersheds being underrepresented in the national calibration dataset. Additionally, these watersheds, located mostly in northern and western Brittany, contain specific crops such as vegetables, for which estimation of N surplus is arduous, and thus underestimated (Fig. 9). Regionalization of the model with nationally available data only slightly reduced the RMSE of predicted nitrate loads (by about 12% in Brittany and 6% in the Seine river basin). However, it decreased precision of parameter estimates (Table 2) because regionalized models encompassed a narrower range of variability in basin characteristics. Furthermore, parameterizing a statistical model with effective rainfall (or total runoff) and hydraulic load was difficult, since both factors were correlated in our dataset. In-stream retention functions that have been calibrated independently from land-to-water delivery models do

159

exist (Boyer et al., 2006; Lepisto et al., 2006), but their application in France has yet to be seriously discussed (Billen et al. 2011). From this initial calibration, the use of regionalized models seems unwarranted when the same databases are used for national calibration, because the national model exhibited no significant bias in the two regions we studied; thus, the only effect of developing regionalized models was to increase uncertainty in parameter estimates. Schwarz et al. (2011) concluded the opposite in the U.S., because their national model exhibited bias in some regions, such as those with a semi-arid climate, and because they benefited from more monitoring basins for calibrating regionalized models, which increased precision of parameter estimates. When regionalizing the locally detailed model, goodness-of-fit was further improved (RMSE decreased by 24% in Brittany and 23% in the Seine basin), and parameter estimates remained significant. Hence, the use of regionalized models should be advocated in France only when regional databases are expected to improve model fit and when enough monitoring basins are available for calibrating. 4.2. Analysis of basin characteristics controlling nitrate transfer Looking at basin and river reduction factors provides insight into the main processes that govern nitrate transfer and retention. As highlighted by Preston et al. (2011), N transfer from soils to streams is controlled mostly by runoff of precipitation. In Brittany, mean distance to streams, percentage of hydromorphic soils and mean slope were also identified as important variables to include in the land-to-water delivery factor. A long distance to streams increases basin retention, probably due to the longer residence time in the watershed, and watersheds with a high percentage of hydromorphic soils transfer less nitrate, due to denitrification in riparian wetlands. In Brittany, high slopes are correlated with low basin retentions, which contradict our initial hypothesis (Section 2.1.5). We chose to keep this variable in the model, although its effect is not physically interpretable. For the whole of France, median basin retention reached 49% of N surplus, and median river retention represented 18% of N discharge to streams. The latter value is the same order of magnitude as the 16% estimated for U.S. rivers (Van Breemen et al., 2002) and the 12% for U.K. rivers (Worrall et al., 2012). Application of the model of Seitzinger et al. (2002) in the U.K. (Worrall et al., 2012) resulted in estimates of river retention ranging from 6 to 27% between 1974 and 2005, which is also consistent with our estimates. A national assessment in Denmark (Windolf et al., 2011) demonstrated similar trends as in France: 54% basin retention and 21% retention in surface waters. Considering the parameter estimates based on the national calibration, the median river

Fig. 6. Predicted versus observed N-nitrate specific load, the former from regionalized models using nationally-available data in the Brittany (left) and Seine regions (right). Horizontal error bars represent observed loads first and last deciles. The red line is a 1:1 line.

160

R. Dupas et al. / Science of the Total Environment 443 (2013) 152–162

Fig. 7. Predicted versus observed N-nitrate specific load, the former from regionalized models using detailed regional data in the Brittany (left) and Seine regions (right). Horizontal error bars represent observed loads first and last deciles. The red line is a 1:1 line.

retention rate was 21% in the Seine river basin, which is within the 20– 30% range of estimates reported by Billen et al. (2007). In Brittany, the median in-stream retention rate was 19%, which is higher than the 11% reported by Montreuil et al. (2010) in the Scorff River of Brittany. According to the national model, median basin retention in Brittany was similar to that for the whole of France (46%) but much smaller than that in the Seine basin (71%), suggesting that Seine watersheds are important N sinks. Regionalized models exhibited the same trend in median basin retention: 23% in Brittany and 66% in the Seine basin, although the balance between basin and river retentions differed slightly due to problems of equifinality. 4.3. Limits of the model and potential improvements A number of watershed characteristics were tested as potential explanatory variables to include in the watershed-retention factor B

(Table 1), but some of the databases used at the French national scale have a poor spatial resolution (e.g. soil data), or the methods used to generate them are questionable (e.g. delineation of riparian wetlands with a topographic index). We demonstrated that high-resolution and good-quality soil data improved model predictions in Brittany coastal watersheds, which suggests that developing national databases for environmental characteristics (e.g. soil, delineated wetlands, geology) would help improve this type of approach. The current version of Nutting-N does not route N-fluxes along rivers because a national drainage network and an associated drainage basin model remains lacking in France. Such a model, including the latest delineation of Water Framework Directive water bodies, is expected to be available in the future and should be coupled with Nutting-N. The N load coming from deep groundwater discharge to streams is roughly accounted for in Nutting-N. Representation of this flow component should improve soon because of an upcoming hydrogeological

Fig. 8. Map of predicted N-nitrate specific load; Seine model with detailed hydrological data.

R. Dupas et al. / Science of the Total Environment 443 (2013) 152–162

161

Fig. 9. Map of predicted N-nitrate specific load, Bretagne model with detailed regional data.

model for estimating nitrate concentration in groundwater; its predictions could be input into Nutting-N instead of the spatially distributed nitrate concentration data. Furthermore, Nutting-N results could be refined by using total runoff and base flow estimated from a highresolution hydrological model instead of data of long-term spatially explicit runoff and heuristic methods for base flow estimation. 5. Conclusion N-nitrate loads were estimated within all watersheds in France with a model linking N sources and watershed land and river characteristics. Regionalizing the model increased goodness-of-fit but also increased uncertainty in parameter estimates. In this study, when regionalizing the model with the same data as the national model, parameter estimates were not significant, whereas regionalization with local data enabled development of models with both increased goodness-of-fit and acceptable parameter estimates. Since regionalized models may decrease precision of parameter estimates, a hybrid approach is advocated, combining a nationally calibrated river-retention factor and a regionally calibrated basin-retention factor. In this study, N-nitrate transfer was shown to be related mostly to runoff of precipitation; however, other variables had influence in one of the regional models. This suggests that developing national databases for environmental characteristics (e.g. soil, wetlands), which is already done in some countries, would help improve this type of approach. For the whole of France, predicted median basin retention equaled 49% of N surplus, and median river retention reached 18% of incoming discharge; basin retention in a sedimentary basin was higher (71%), indicating a net sink of N. Conversely, some watersheds on crystalline bedrock seem to be net sources of N, as suggested by N loads >N surplus, but this can only be confirmed by studying the evolution of N loads and surplus over several decades. Acknowledgments This work was funded by ONEMA, the French National Agency for Water and Aquatic Environments. Nicolas Domange and Gaëlle Deronzier contributed to management of the project. Brittany water quality data was provided by CPER 2007–2013 Bretagne/GP5. Soil data was provided by INRA Infosol Orléans (Christine Le Bas) and Sols de Bretagne programs (Blandine Lemercier, Lionel Berthier). River hydraulic geometry data was provided by Hervé Pella and Nicolas Lamouroux of IRSTEA. Meteo data come from Météo France. The authors would like to thank Josette Launay and Pierre Aurousseau from Conseil Scientifique de l'Environnement en Bretagne (CSEB) for their contribution during the first phase of this

study. We also thank Michelle and Michael Corson for their advice in writing this paper and correcting its English style.

References Agence de l'eau Seine Normandie. Cartographie des zones à dominante humide du bassin Seine – Normandie. Rapport Final. Groupement Gallia Sana/I-Mage Consult/Cere; 2006. Alexander RB, Smith RA, Schwarz GE. Effect of stream channel size on the delivery of nitrogen to the Gulf of Mexico. Nature 2000;403:758–61. Alexander RB, Johnes PJ, Boyer EW, Smith RA. A comparison of models for estimating the riverine export of nitrogen from large watersheds. Biogeochemistry 2002;57: 295–339. Beaujouan V, Durand P, Ruiz L. Modelling the effect of the spatial distribution of agricultural practices on nitrogen fluxes in rural catchments. Ecol Model 2001;137:93-105. Behrendt H, Opitz D. Retention of nutrients in river systems: dependence on specific runoff and hydraulic load. Hydrobiologia 1999;410:111–22. Behrendt H, Kornmilch M, Opitz D, Schmoll O, Scholz G. Estimation of the nutrient inputs into river systems - experiences from German rivers. Regional Environ Change 2002;3:107–17. Beven KJ, Kirkby MJ. A physically based, variable contributing area model of basin hydrology. Hydrol Sci Bull 1979;24:43–69. Billen G, Garnier J, Nemery J, Sebilo M, Sferratore A, Barles S, et al. A long-term view of nutrient transfers through the Seine river continuum. Sci Total Environ 2007;375: 80–97. Billen G, Thieu V, Garnier J, Silvestre M. Modelling the N cascade in regional watersheds: the case study of the Seine, Somme and Scheldt rivers. Agric Ecosyst Environ 2009;133: 234–46. Billen G, Silvestre M, Grizzetti B, Leip A, Garnier J, Voss M, et al. In: The European Nitrogen Assessment, Sutton MA, Howard CM, Erisman JW, Billen G, Bleeker A, Grennfelt P, van Grinsven H, Grizzetti B, editors. Nitrogen flows from European regional watersheds to coastal marine waters. Cambridge University Press; 2011. p. 271–97. Boyer EW, Alexander RB, Parton WJ, Li CS, Butterbach-Bahl K, Donner SD, et al. Modeling denitrification in terrestrial and aquatic ecosystems at regional scales. Ecol Appl 2006;16: 2123–42. Brisson N, Gary C, Justes E, Roche R, Mary B, Ripoche D, et al. An overview of the crop model STICS. Eur J Agron 2003;18:309–32. Campling P, Terres JM, Walle SV, Van Orshoven J, Crouzet P. Estimation of nitrogen balances from agriculture for EU-15: spatialisation of estimates to river basins using the CORINE Land Cover. Phys Chem Earth 2005;30:25–34. Caraco NF, Cole JJ. Human impact on nitrate export: an analysis using major world rivers. Ambio 1999;28:167–70. Curie F, Ducharne A, Sebilo M, Bendjoudi H. Denitrification in a hyporheic riparian zone controlled by river regulation in the Seine river basin (France). Hydrol Processes 2009;23:655–64. Curie F, Ducharne A, Bendjoudi H, Billen G. Spatialization of denitrification by river corridors in regional-scale watersheds: case study of the Seine river basin. Phys Chem Earth 2011;36:530–8. de Wit MJM. Nutrient fluxes at the river basin scale. I: the PolFlow model. Hydrol Processes 2001;15:743–59. Dumont E, Harrison JA, Kroeze C, Bakker EJ, Seitzinger SP. Global distribution and sources of dissolved inorganic nitrogen export to the coastal zone: results from a spatially explicit, global model. Global Biogeochem Cycles 2005;19. Folton N, Lavabre J. Regionalization of a monthly rainfall–runoff model for the southern half of France based on a sample of 880 gauged catchments; 2006.

162

R. Dupas et al. / Science of the Total Environment 443 (2013) 152–162

Gascuel-Odoux C, Aurousseau P, Durand P, Ruiz L, Molenat J. The role of climate on inter-annual variation in stream nitrate fluxes and concentrations. Sci Total Environ 2010;408:5657–66. Green PA, Vorosmarty CJ, Meybeck M, Galloway JN, Peterson BJ, Boyer EW. Pre-industrial and contemporary fluxes of nitrogen through rivers: a global assessment based on typology. Biogeochemistry 2004;68:71-105. Grimvall A, Stalnacke P. Statistical methods for source apportionment of riverine loads of pollutants. Environmetrics 1996;7:201–13. Grizzetti B, Bouraoui F, de Marsily G, Bidoglio G. A statistical approach to estimate nitrogen sectorial contribution to total load. Water Sci Technol 2005a;51:83–90. Grizzetti B, Bouraoui F, de Marsily G, Bidoglio G. A statistical method for source apportionment of riverine nitrogen loads. J Hydrol 2005b;304:302–15. Grizzetti B, Bouraoui F, De Marsily G. Assessing nitrogen pressures on European surface water. Global Biogeochem Cycles 2008;22. Groenendijk P, Renaud LV, Roelsma J. Prediction of nitrogen and phosphorus leaching to groundwater and surface waters. Process descriptions of the ANIMO 4.0 model, Report 983. Wageningen, The Netherlands: Alterra; 2005. Howarth RW, Billen G, Swaney D, Townsend A, Jaworski N, Lajtha K, et al. Regional nitrogen budgets and riverine N&P fluxes for the drainages to the North Atlantic Ocean: natural and human influences. Biogeochemistry 1996;35:75-139. Howarth RW, Swaney DP, Boyer EW, Marino R, Jaworski N, Goodale C. The influence of climate on average nitrogen export from large watersheds in the Northeastern United States. Biogeochemistry 2006;79:163–86. Institute of Hydrology. Low flow studies report; 1980. Wallingford, UK. Johnes PJ. Evaluation and management of the impact of land use change on the nitrogen and phosphorus load delivered to surface waters: The export coefficient modelling approach. J Hydrol 1996;183:323–49. Johnes PJ, Heathwaite AL. Modelling the impact of land use change on water quality in agricultural catchments. Hydrol Processes 1997;11:269–86. Kronvang B, Hezlar J, Boers P, Jensen JP, Behrendt H, Anderson T, et al. Nutrient retention handbook. Software Manual for EUROHARP-NUTRET and Scientific Review on Nutrient Retention, EUROHARP Report 9-2004, NIVA Report SNO 4878/2004; 2004. Oslo, Norway, 103 pp. Ledoux E, Gomez E, Monget JM, Viavattene C, Viennot P, Ducharne A, et al. Agriculture and groundwater nitrate contamination in the Seine basin. The STICS-MODCOU modelling chain. Sci Total Environ 2007;375:33–47. Leip A, Marchi G, Koeble R, Kempen M, Britz W, Li C. Linking an economic model for European agriculture with a mechanistic model to estimate nitrogen and carbon losses from arable soils in Europe. Biogeosci 2008;5:73–94. Lemercier B, Lacoste M, Loum M, Walter C. Extrapolation at regional scale of local soil knowledge using boosted classification trees: a two-step approach. Geoderma 2012;171:75–84. Lepisto A, Granlund K, Kortelainen P, Raike A. Nitrogen in river basins: sources, retention in the surface waters and peatlands, and fluxes to estuaries in Finland. Sci Total Environ 2006;365:238–59. Li CS, Frolking S, Frolking TA. A model of nitrous-oxide evolution from soil driven by rainfall events .1. Model structure and sensitivity. J Geophys Res -Atmos 1992;97:9759–76. Lindstrom G, Pers C, Rosberg J, Stromqvist J, Arheimer B. Development and testing of the HYPE (Hydrological Predictions for the Environment) water quality model for different spatial scales. Hydrol Res 2010;41:295–319. Mayorga E, Seitzinger SP, Harrison JA, Dumont E, Beusen AHW, Bouwman AF, et al. Global nutrient export from WaterSheds 2 (NEWS 2): model development and implementation. Environ Model Software 2010;25:837–53. McFarland AMS, Hauck LM. Determining nutrient export coefficients and source loading uncertainty using in-stream monitoring data. J Am Water Resour Assoc 2001;37:223–36. Meinardi CR, Beusen AHW, Bollen MJS, Klepper O. Vulnerability to diffuse pollution of European soils and groundwater. RIVM Report 350490991. Bilthoven: National Institute of Public Health and Environment; 1994. 10 pp. Merot P, Squividant H, Aurousseau P, Hefting M, Burt T, Maitre V, et al. Testing a climato-topographic index for predicting wetlands distribution along an European climate gradient. Ecol Model 2003;163:51–71. Moatar F, Meybeck M. Riverine fluxes of pollutants: towards predictions of uncertainties by flux duration indicators. C R Geosci 2007;339:367–82. Moatar F, Meybeck M, Raymond S, Birgand F, Curie F. River flux uncertainties by hydrological variability and riverine material behaviour. Hydrol Processes 2012. http://dx.doi.org/ 10.1002/hyp.9464. Molenat J, Durand P, Gascuel-Odoux C, Davy P, Gruau G. Mechanisms of nitrate transfer from soil to stream in an agricultural watershed of French Brittany. Water Air Soil Pollut 2002;133:161–83. Molenat J, Gascuel-Odoux C, Ruiz L, Gruau G. Role of water table dynamics on stream nitrate export and concentration in agricultural headwater catchment (France). J Hydrol 2008;348:363–78.

Montreuil O, Merot P. Nitrogen removal in valley bottom wetlands: assessment in headwater catchments distributed throughout a large basin. J Environ Qual 2006;35:2113–22. Montreuil O, Merot P, Marmonier P. Estimation of nitrate removal by riparian wetlands and streams in agricultural catchments: effect of discharge and stream order. Freshw Biol 2010;55:2305–18. Moriasi DN, Arnold JG, Van Liew MW, Bingner RL, Harmel RD, Veith TL. Model evaluation guidelines for systematic quantification of accuracy in watershed simulations. Transactions of the Asabe 2007;50:885–900. OECD, Organisation for Economic Co-operation and Development. Environmental indicators for agriculture, vol. 3. ; 2001. Methods and Results, Paris, France. Official Journal of the European Communities. Directive 2000/60/EC of the European Parliament and of the Council of 23 October 2000 establishing a framework for community action in the field of water policy; 2000. Pella H, Lejot J, Lamouroux N, Snelder T. The theoretical hydrographical network (RHT) for France and its environmental attributes. Géomorphie : relief, processus, environnement, in press. Pieterse NM, Bleuten W, Jorgensen SE. Contribution of point sources and diffuse sources to nitrogen and phosphorus loads in lowland river tributaries. J Hydrol 2003;271: 213–25. Preston SD, Alexander RB, Schwarz GE, Crawford CG. Factors affecting stream nutrient loads: a synthesis of regional sparrow model results for the continental United States. J Am Water Resour Assoc 2011;47:891–915. Sauquet E. Mapping mean annual river discharges: geostatistical developments for incorporating river network dependencies. J Hydrol 2006;331:300–14. Schoumans OF, Silgram M, Groenendijk P, Bouraoui F, Andersen HE, Kronvang B, et al. Description of nine nutrient loss models: capabilities and suitability based on their characteristics. J Environ Monit 2009;11:506–14. Schwarz GE, Hoos AB, Alexander RB, Smith RA. The SPARROW surface water-quality model—theory, applications and user documentation. U.S. Geological Survey, Techniques and Methods 6-B3; 2006. Schwarz GE, Alexander RB, Smith RA, Preston SD. The regionalization of national-scale sparrow models for stream nutrients. J Am Water Resour Assoc 2011;47:1151–72. Seitzinger SP, Kroeze C. Global distribution of nitrous oxide production and N inputs in freshwater and coastal marine ecosystems. Global Biogeochem Cycles 1998;12:93-113. Seitzinger SP, Styles RV, Boyer EW, Alexander RB, Billen G, Howarth RW, et al. Nitrogen retention in rivers: model development and application to watersheds in the northeastern USA. Biogeochemistry 2002;57:199–237. Seitzinger SP, Harrison JA, Dumont E, Beusen AHW, Bouwman AF. Sources and delivery of carbon, nitrogen, and phosphorus to the coastal zone: an overview of global Nutrient Export from Watersheds (NEWS) models and their application. Global Biogeochem Cycles 2005;19. Smith RA, Schwarz GE, Alexander RB. Regional interpretation of water-quality monitoring data. Water Resour Res 1997;33:2781–98. SOLAGRO. NOPOLU volet Agricole – Guide méthodologique; 2010. p. 89. Sutton MA, Howard CM, Erisman JW, Billen G, Bleeker A, Grennfelt P, van Grinsven H, Grizzetti B, editors. The European Nitrogen Assessment. Published by Cambridge University Press; 2011. Van Breemen N, Boyer EW, Goodale CL, Jaworski NA, Paustian K, Seitzinger SP, et al. Where did all the nitrogen go? Fate of nitrogen inputs to large watersheds in the northeastern USA. Biogeochemistry 2002;57:267–93. Van Drecht G, Bouwman AF, Knoop JM, Beusen AHW, Meinardi CR. Global modeling of the fate of nitrogen from point and nonpoint sources in soils, groundwater, and surface water. Global Biogeochem Cycles 2003;17. Velthof GL, Oudendag D, Witzke HR, Asman WAH, Klimont Z, Oenema O. Integrated Assessment of Nitrogen Losses from Agriculture in EU-27 using MITERRA-EUROPE. J Environ Qual 2009;38:402–17. Viennot P, Ledoux E, Monget JM, Schott C, Garnier C, Beaudouin N. La pollution du bassind de la Seine par les nitrates. Programme PIREN-Seine; 2009. 44 pp., Paris, France. Windolf J, Thodsen H, Troldborg L, Larsen SE, Bogestrand J, Ovesen NB, et al. A distributed modelling system for simulation of monthly runoff and nitrogen sources, loads and sinks for ungauged catchments in Denmark. J Environ Monit 2011;13: 2645–58. Wolf J, Beusen AHW, Groenendijk P, Kroon T, Rotter R, van Zeijts H. The integrated modeling system STONE for calculating nutrient emissions from agriculture in the Netherlands. Environ Model Software 2003;18:597–617. Worrall F, Burt TP, Howden NJK, Whelan MJ. The fluvial flux of nitrate from the UK terrestrial biosphere - An estimate of national-scale in-stream nitrate loss using an export coefficient model. J Hydrol 2012;414:31–9. Wosten JHM, Pachepsky YA, Rawls WJ. Pedotransfer functions: bridging the gap between available basic soil data and missing soil hydraulic characteristics. J Hydrol 2001;251:123–50.