As the Seine basin is the most urbanized and industrialized French basin, with intensive crop areas, the impact of floods and low flow is very important. The aim of this study is to improve the understanding of the hydrological functioning of the Seine basin and to estimate the influence of the large aquifers that cover a great part of the domain, as well as the impact of the urbanized areas. The simulation is based on the SAFRAN-ISBA-MODCOU (SIM) coupled system, which includes the Analysis System for Providing Atmospheric Information Relevant to Snow (SAFRAN) analysis of the atmospheric forcing, two surface schemes that take into account the urban areas (Town Energy Budget (TEB)) and the natural areas (Interactions Soil-Biosphere-Atmosphere (ISBA)), and the distributed hydrological model Coupled Model (MODCOU). A 17-year simulation was performed and validated against the observations of about 100 river gauges and 130 piezometric stations. The simulated streamflows are comparable with the observations, especially for large river stations of the Seine and its main tributaries. The spatial and temporal variations of the aquifer levels simulated by MODCOU are also comparable to the observed piezometric data, although the local comparison can present differences. The impact of urban areas on the streamflows appears to be important at the local scale but not very marked at the basin scale, but their impact on the energy budget is strong. Moreover, it is shown that the aquifers play a very important role. The underground contribution to the discharge reaches more than 80% of the Seine river flow in summertime and ∼40% in winter. Even during the wintertime floods, the aquifers play a crucial role, since the larger part of the drainage is stored in the water tables, which significantly reduces the intensity of the floods and, thus, of their consequences. It is finally shown that the ability of SIM to reproduce the evolution of the aquifers and the partition between surface runoff and drainage leads to a good simulation of the main floods of the Seine.
 The exceptional heat wave and the severe drought that affected France and a large part of Europe in summer 2003 have highlighted the dependencies of the ecological system and of human activities on the availability of water. It has reinforced the need for a better understanding of the regional water cycle and of its possible changes. Such an issue is at the core of the Global Energy and Water Cycle Experiment (GEWEX) program, which aims at improving the understanding of the spatial and temporal evolution of the water cycle. The numerous regional-scale experiments that took place in GEWEX, over, for instance, large boreal forests (Boreal Ecosystem-Atmosphere Study (BOREAS) [Sellers et al., 1997]), tropical regions (Large-Scale Biosphere-Atmosphere Experiment in Amazonia (LBA) [Nobre, 1996]), and temperate areas (Hydrological Atmospheric Pilot Experiment-Modelisation du Bilan Hydrique (Hapex-Mobilhy) [Andre et al., 1986]), have resulted in a better knowledge of the processes and of their variability. Another part of GEWEX, the Project for Intercomparison of Land-Surface Parameterization Schemes (PILPS) [Henderson-Sellers et al., 1993], has been a powerful tool to improve the simulation of surface processes in the Soil-Vegetation-Atmosphere Transfer (SVAT) schemes implemented in climate and hydrological models. The PILPS experiments have also shown the interest in taking into account such physical processes as subgrid runoff (PILPS 2c) or frozen soil (PILPS 2d [Schlosser et al., 2000] and PILPS 2e [Bowling et al., 2001]). Moreover, the PILPS 2c [Wood et al., 1998] and the Rhone-Aggregation [Boone et al., 2004] experiments have demonstrated the interest in coupling SVAT schemes to a hydrological model, in order to validate part of the water and energy budgets at the catchment scale, through the comparison of observed and simulated streamflows.
 In the present study, focus is put on the coupled atmosphere-surface-hydrology modeling of a large French basin: the Seine basin. This basin covers about one fifth of France, including the Paris metropolitan area. It is also the most urbanized and industrialized French basin. Moreover, this basin is an important agricultural region, with intensive production of crops in the Beauce area. The high pressure of industry, agriculture, and population on the water resources and quality, as well as the serious possible damages of floods or drought in this basin, have led the French research community and the managers to work jointly to improve the knowledge of the functioning of the Seine basin, which is particularly complex because of several overlying widespread aquifers. The Interdisciplinary Program of Environmental Research on the Seine Basin (PIREN-Seine) is leading several studies about water quality [Riffard et al., 2002], ecology [Garnier et al., 2002], and diffusion of the pollution [Gomez et al., 2002]. In the framework of this project, the Coupled Model (MODCOU) hydrological model [Ledoux et al., 1984] was calibrated and validated for the Seine basin by Gomez .
 In this article, the MODCOU hydrological model is coupled to the Interactions Soil-Biosphere-Atmosphere (ISBA) [Noilhan and Planton, 1989] detailed SVAT scheme. The low-level atmospheric parameters needed by the ISBA surface scheme are produced by the Analysis System for Providing Atmospheric Information Relevant to Snow (SAFRAN) system [Durand et al., 1993]. Such a coupled system was already used and validated over two mountainous and snow-based French basins, the Rhone [Etchevers et al., 2001] and the Adour-Garonne [Morel, 2003].
 However, in these previous studies, the aquifers were not very large, and their impacts on the river flows were not as important as the ones over the Seine basin. It is therefore the first application of a detailed SVAT scheme coupled with such a complex modeling of the Seine aquifers and river routing system.
 Moreover, the study will also focus on the impact of the urban area owing to the use of the Town Energy Balance (TEB) surface scheme [Masson, 2000]. Special emphasis is placed on the impact of urban areas on the hydrological cycle throughout the entire year and during flood events.
 The SAFRAN-ISBA-TEB-MODCOU simulation is run over 17 years, from 1 August 1985 to 31 July 2002. It is validated using ∼100 daily river gauges and 130 piezometric observations.
 The rest of this paper is divided into five sections. Section 2 briefly presents the different models and the coupling approach. The Seine basin and the databases used for the simulation are presented in sections 3 and 4. The results of the simulation and the comparison to observed data are detailed in section 5, and finally, section 6 is an analysis of the floods of the Seine at Paris.
2. SAFRAN-ISBA-TEB-MODCOU Coupled Model
 The modeling system is composed of three different parts [Habets et al., 1999a, 1999b]. The atmospheric forcing is calculated using the Analysis System for Providing Atmospheric Information Relevant to Snow (SAFRAN) analysis system [Durand et al., 1993]. The processes at the surface are simulated by two surface schemes: the Interactions Soil-Biosphere-Atmosphere (ISBA) [Noilhan and Planton, 1989; Habets et al., 1999a] scheme computes the water and energy budgets over natural surfaces, and the Town Energy Budget (TEB) [Masson, 2000] scheme computes the water and energy budgets over urban areas. The last component is the MODCOU macroscale hydrological model [Ledoux et al., 1984], which simulates surface and underground water transfers and computes the river flows as well as the level of the aquifers.
 The SAFRAN system [Durand et al., 1993] analyzes the low-level atmospheric variables needed by the surface schemes, such as precipitation, incoming longwave and shortwave radiation fluxes, wind speed, air temperature, and humidity. The analysis takes into account all the available observations collected by Météo-France, as well as the operational analyses of the weather prediction model of Météo-France, and some climatological data. Over the Seine basin, ∼1000 rain gauges and ∼200 synoptic stations were available. The SAFRAN analyses are performed in homogeneous climatic zones and assume a vertical gradient of the variables in these zones. In each zone the atmospheric forcing is interpolated hourly and on a regular 8-km grid.
2.2. ISBA and TEB
 Two different schemes are used to simulate the surface water and energy budgets on the 8-km grid: ISBA over natural areas and TEB over urban areas. The fluxes are averaged over the cell according to the fraction of each land cover.
 The ISBA surface scheme is used in the numerical weather prediction model as well as in the climate model of Météo-France. In this study, the force-restore version of ISBA is used, with three soil layers [Boone et al., 1998], a detailed three-layer snow scheme [Boone and Etchevers, 2001], a subgrid runoff scheme [Habets et al., 1999b], and a drainage parametrization that includes a minimum gravitational drainage under dry conditions [Habets et al., 1999b]. This version of ISBA has 12 prognostic variables: the surface and deeper soil temperatures, the soil water content of the surface, root zone, and deeper soil, the water content of the interception reservoir, the ice water content of the surface and of the root zone [Boone, 2000], the snow albedo, and three variables for each layer of the snow pack: the snow water equivalent, the thickness, and the amount of heat inside the snow pack.
 The TEB urban scheme simulates the water and energy budgets of urban areas [Masson, 2000]. Towns are represented using a canyon geometry in which radiation can be trapped. Because of the complex shape of the city surface, three surface energy budgets are considered, over roofs, roads, and walls, and two additional energy budgets are computed for snow when present on roads and roofs. Two water budgets are also computed, on the roofs and roads. The TEB scheme has been used in a three-dimensional model for Paris and Marseille [Lemonsu et al., 2004]. It has also been tested in an off-line mode for a small urbanized catchment over a few years (A. Lemonsu et al., Hydrological validation of the TEB scheme on a suburban catchment, submitted to Journal of Hydrology, 2003). This article presents the first use of TEB in a hydrological study at the mesoscale and for the long term.
 The last component of the modeling system is the macroscale hydrological model MODCOU [Ledoux et al., 1984; Gomez, 2002], which simulates the river flow and the evolution of the aquifers. MODCOU routes the surface runoff to the hydrographic network, while the gravitational drainage computed by ISBA is transferred to the aquifers.
 MODCOU uses a surface layer and one or more underground layer(s), corresponding to the vertical succession of aquifers. The surface and underground domains are divided into cells that range from 1 to 8 km, the highest resolution (1 km) being used for river grid cells or at the border of subbasins. A grid cell is considered as a river cell when the corresponding subbasin is larger than 250 km2. For each surface cell a unique direction was defined from topographic information for surface water transfer. The time transfer between each grid cell is also determined from topography data, and the surface runoff is routed through the hydrologic network to the outlet of the basin using isochronous zones with a daily time step [Ledoux et al., 1984; Gomez et al., 2002]. The time of concentration, which is the maximum time transfer to reach the outlet of the basin, is usually calibrated, on the basis of the observations.
 MODCOU assumes that the transfers are bidimensional in the aquifer, and the evolution of the groundwater level is computed using the diffusion equation. Vertical exchanges between the different aquifers through semipermeable layers are also taken into account.
where h is the hydraulic head, is the transmissivity tensor, S is the storage coefficient, q is the pumped discharge, and qsup and qinf are the discharges vertically exchanged through the upper and lower semipermeable layers, respectively:
where K and l are the vertical permeability and the thickness of the semipermeable layer, respectively. The exchange of water between rivers and aquifers is computed according to simple relations depending on the hydraulic state of the system and using a transfer coefficient that represents the clogging of the river channel [Ledoux et al., 1984; Gomez et al., 2003].
where Qnap is the discharge exchanged between the river and the aquifer, Tp is a transfer coefficient between the aquifer and the river that depends on the clogging of the river channel, H is the hydraulic head, and H0 is the river level.
2.4. Coupling SAFRAN, ISBA, TEB, and MODCOU
Figure 1 presents the coupling method [Habets et al., 1999a, 1999b]. The atmospheric forcing assimilated by SAFRAN is an input for ISBA and TEB. The surface runoff computed by ISBA and TEB as well as the gravitational drainage computed by ISBA are transferred to MODCOU: Drainage is transferred to the underground layer and contributes to the evolution of the aquifers, while surface runoff is transferred to the surface hydrologic network. As the time step of MODCOU (1 day) is larger than that of ISBA and TEB (5 min), surface runoff and gravitational drainage are first accumulated daily and then transferred to MODCOU.
3. Seine Basin
 The Seine basin covers ∼20% of France (95,000 km2) (Figure 2a). The basin is quite flat, with an altitude lower than 300 m, except in the south of the basin (the Morvan), where it reaches 900 m (Figure 2b). The Seine river is 776 km long, and its main tributaries are the Oise, the Marne, the Yonne, the Eure, and the Aube rivers (Figure 2b). The Seine basin is the most urbanized basin in France, with 25% of the population and ∼2000 km2 of city areas. The fraction of town on the 8-km grid can reach 80% in the Paris metropolitan region (Figure 3).
 The Seine basin is represented by MODCOU by a surface layer that contains 35,698 grid cells (of which 6481 are river cells) (Figure 4a). The sedimentary basin of the Seine is characterized by a complex hydrogeology (Figure 4b), with many stacked aquifers and semipermeable layers. Only the first three aquifers are modeled [see Gomez, 2002; Gomez et al., 2003]: the Oligocene aquifer on the top (9912 km2), the Eocene aquifer in the middle (18,928 km2), and the Chalk aquifer below (65,468 km2).
 These three aquifers are vertically stacked (Figure 4b) and are separated by semipermeable layers. Only the center part of the Chalk aquifer is confined, covered either by the Oligocene only, by the Eocene only, or by both the Oligocene and Eocene aquifers. Since the Chalk aquifer is larger than the two others, it is the top aquifer over most of its surface. Water transfers can occur between the Oligocene and the Eocene, between the Oligocene and the Chalk, and between the Eocene and the Chalk. Since the three aquifers have an unconfined part, each of them can directly receive gravitational drainage computed by ISBA.
 The hydrologic properties of the aquifers (transmissivity and storage coefficient) have been calibrated by Gomez . The transmissivity ranges from 0.22 × 10−4 m2 s−1 to 0.14 m2 s−1, and the storage coefficient ranges from 0.99 × 10−4 to 0.1. The inflow to the Chalk groundwater table by the deeper Albien aquifer below (not modeled) is taken into account through an incoming flux of 600 L s−1 [Gomez, 2002].
4.1. Physiographic Data
 The surface parameters needed by ISBA and TEB are extracted from the ECOCLIMAP 1-km database [Masson et al., 2003]. This global database includes several existing data sets, including, over France, the CORINE land cover [Giordano et al., 1990], the Forest Information from Remote Sensing (FIRS) climate map [European Commission, 1995], and the National Institute for Agronomic Research (INRA) soil map, as well as the satellite archive of the normalized difference vegetation index (NDVI) from NOAA advanced very high resolution radiometer (AVHRR). For each vegetation type the corresponding parameters (soil depth, albedo, emissivity, fraction of vegetation, leaf area index, roughness lengths) are associated, and the parameters are aggregated on a 8-km grid according to the Noilhan and Lacarrere  method.
 The land cover of the Seine basin is characterized by a predominance of crops associated with a soil depth of ∼2 m (including 1.5 m for the root zone). In the eastern part of the basin, forests are associated with deeper soils (3 m). The clay fraction is quite homogeneous in the basin (ranging from 15 to 25%), whereas the sand fraction is more contrasted (from 10 to 60% in the Sologne region, in the southwest of the basin).
 Parameters like albedo and leaf area index (LAI) follow an annual cycle, with the minimum and maximum values prescribed for each vegetation type. This annual cycle was derived from the satellite archive of the VEGETATION/NDVI from SPOT/VEGETATION for the year 2000 and is assumed to be identical for each year. On an annual average the LAI ranges from 1.5 to 2.8 m2 m−2, with a maximum in spring (up to 4.5 m2 m−2). The albedo is quite homogeneous over the domain (∼0.15).
 Two parameters of the ISBA surface scheme are not derived from the soil and vegetation types: the b parameter of the subgrid runoff scheme and the value of wdrain included in the drainage parametrization. In this study, the default values for b = 0.5 and for the base flow wdrain = 10−3 were used [Habets et al., 1999b]. However, as the minimum drainage is set to represent the effects of groundwater tables that are not explicitly resolved by the hydrological model, it is set to zero where the aquifer is explicitly resolved (i.e., in the main part of the domain).
4.2. Atmospheric Forcing
 The mean annual temperature (Figure 5a) varies from 282 K on the eastern border of the domain to 285 K along the coast (northwestern border) and in the south of the basin. The impact of urban area on the screen-level atmospheric conditions can be seen, with a temperature gradient of 2 K between the Paris area and its surrounding regions.
 The accumulated precipitation map (Figure 5b) presents three distinct regions, with low values (∼600 kg m−2 yr−1) in the center of the basin and maximal values (∼1200 kg m−2 yr−1) on the coastal border as well as on the southern and eastern borders. The higher annual amounts of precipitation are found in the Morvan region (southeast of the basin), associated with the maximum altitude. The average annual amount of rainfall is ∼800 kg m−2, with a standard deviation of 150 kg m−2, with dry years with less than 600 kg m−2 and wetter years with ∼1000 kg m−2 (Figure 5c). Snowfall is negligible, with a maximal annual snow amount of 40 kg m−2. Monthly averaged precipitation over 17 years is remarkably homogeneous, with ∼60 kg m−2 month−1. Contrary to other French regions (Rhone, Adour-Garonne), precipitation is largely uniformly distributed, with no marked annual cycle. However, the interannual variability of the monthly rainfall is important with a standard deviation of 30 kg m−2.
5. Analysis of the 17-Year Simulation (1985–2002)
 The simulation starts on 1 August 1985 and ends on 31 July 2002. There is no reinitialization during the 17-year simulation. To initialize the simulation, a preliminary run of 2 years was made using the meteorological data of the first year (1985–1986) to reach the equilibrium state of the model. This solution provides a spatially coherent state compatible with the 1985–1986 atmospheric forcing.
 The results of the simulation are presented in this section and compared to the available observations. There were 104 gauging stations with daily observations of the streamflows used for the validation. Figure 2 shows the position of 8 of them, located on the most important rivers of the basin. More than 138 piezometric stations were also available, giving local observations of the levels of the three simulated aquifers.
5.1. Energy and Water Budgets
5.1.1. Water Budget
 Over the whole domain the mean annual amount of precipitation (snowfall and rainfall) reaches 818 kg m−2 yr−1. The mean annual amount of evaporation reaches 563 kg m−2 yr−1, and the amount of runoff reaches 256 kg m−2 yr−1 (drainage, 178 kg m−2 yr−1; surface runoff, 78 kg m−2 yr−1). On average, the evaporation represents 70% of the precipitation. This is more than in the Adour-Garonne and Rhone basins (only about 60% and 50%, respectively [Morel, 2003; Etchevers, 2000], because of the contribution of mountains where the evaporation is weak. However, the result obtained on the Seine is comparable to the ones obtained over the flat parts of the Rhone and Adour-Garonne basins.
 The spatial variability of the water budget is governed by the nature of the surface (vegetation, soil) and by the meteorological conditions. Runoff and evaporation (Figure 6) show the same pattern as the precipitation (Figure 2b), with small amounts in the center of the basin and large values near the eastern border of the domain. The evaporation/precipitation ratio varies from 50–60% in the rainiest eastern region to 60–80% in the driest center of the basin.
 Urban zones like the Paris area clearly stand out: They are characterized by strong surface runoff and low drainage (because they are impervious), as well as weak evaporation (generated by the small areas of garden and parks).
Figure 7 shows the mean monthly evolution of the water budget (1985–2002) as well as the monthly budget for two contrasted years, 1991–1992 and 1994–1995. On average, the evaporation is maximum in May, while the total runoff is larger in January. However, as illustrated in Figure 7, the interannual variability is very important in this basin (stronger than in the Adour-Garonne and Rhone basins), mainly because of the interannual variability of precipitation and because of the fact that there is no snowmelt effect in springtime. For instance, January is the wettest month in 1995 and the driest in 1992, with a total runoff of ∼100 kg m−2 in 1995 and only 10 kg m−2 in 1992 and an evaporation in 1992 (14 kg m−2) larger than the precipitation (11 kg m−2), which is quite unusual in wintertime.
5.1.2. Energy Budget
 The impact of the urbanized area is quite clear on the energy budget. Indeed, while the net radiation flux field is quite homogeneous (ranging from 50 to 70 W m−2), the sensible and latent heat flux fields are more contrasted, with, especially, the urbanized areas that clearly stand out. The cities are characterized by a low latent heat flux and a very high sensible heat flux (∼40 W m−2), whereas it ranges from 5 to 20 W m−2 elsewhere. The Bowen ratio (sensible over latent heat flux) ranges from 0.2 to 0.6 in natural areas but is clearly higher than 1 in city areas (Figure 8).
Figure 8b presents the temporal evolution of the energy budget. The sensible heat flux is negative in wintertime, which is consistent with the fact that the basin is located north of the 45th parallel. On average for the whole basin, the Bowen ratio is lower than 1, except for the driest years (as 1991), for which the latent heat flux is limited by the availability of water stored in the soil.
5.2. Comparison to Observed Streamflows and Piezometric Data
 The simulated daily streamflows are compared to the daily observations at the 104 gauging stations. To check the quality of the simulation, two statistical criteria are computed: the efficiency E [Nash and Sutcliff, 1970] and the ratio of simulated over observed annual discharge Qsim/Qobs. Good simulations are characterized by values of these criteria close to unity.
Figures 9 and 10 present the comparison between observed and simulated daily river flows at four main gauging stations: the Seine at Paris and at Poses (outlet of the basin), the Oise at Pont-Sainte-Maxence, and the Marne at Noisiel. The simulation is in good agreement with the observations, even during floods as well as low flow, and the statistical results are reasonable, with an efficiency that ranges from 0.67 to 0.87 and an error on the annual discharges lower than 5% for all the stations except the Aube at Arcis (Figure 11).
 The quality of the simulated river flows is correlated to the catchment size (Figure 12). The same problem was seen in previous studies on the Rhone and Adour-Garonne basins. Large basins (more than 10,000 km2) are well simulated, with an efficiency greater than 0.8 and an error on the discharge lower than 10%. For medium-size basins (between 1000 km2 and 10,000 km2), results are acceptable with an error on the discharge lower than 10% for only half of them. However, smaller basins (less than 1000 km2) obtain more contrasted results, with an efficiency ranging from negative value to above 0.8 and with lower average statistical criteria. The weak quality of the simulation over the small basins is certainly due to several components, including (1) errors in the spatial and temporal representation of the atmospheric forcing (and, especially, of the precipitation) (since the SAFRAN analysis is performed over homogeneous areas of ∼900 km2, it cannot reproduce the details at small scale), (2) errors in the description of the land cover, (3) approximations of the physics, and also (4) a local effect of the water management (small dams, pumping).
 For the whole domain the efficiency is greater than 0.5 in 57 stations out of 104 (55%) (Figure 12). The efficiency surpasses 0.7 in 26 stations (25%), which is comparable to the Rhone basin results [Etchevers et al., 2001] and better than for the Adour-Garonne basin (only 13% [Morel, 2003]).
 Whereas the discharge is well estimated at the outlet of the basin (Q = 0.98), it is on average over all the stations underestimated by 10%.
Figure 13 presents a comparison between observed and simulated piezometric surfaces in August 1998. There is a large variability of the piezometry, especially for the Chalk layer, with head ranging from 20 to 200 m. In the confined part of the aquifers the gradient of head is very weak (particularly for the Chalk). Over the unconfined part the piezometric surface is closer to the topography, with high levels in mountainous areas and low levels in valleys that drain the aquifers. The simulation is able to reproduce the main observed spatial variation over the three layers. Figure 14 presents the comparison of observed and simulated piezometric heads on average over an entire year. The simulation compares reasonably well with the observations, with a square correlation coefficient higher than 0.95 and an averaged bias lower than 3 m, though in some sites the simulation can present a larger bias.
5.3. Hydrology of the Basin
5.3.1. Role of the Eastern Part of the Domain
 The simulation has shown that the hydrology of the Seine basin is mainly driven by the eastern part of the Seine basin (the Morvan region), since this region generates 70% of the discharge of the Seine at Paris. Indeed, the Morvan receives the highest amounts of rainfall, with an average of 1300 kg m−2 yr−1. The mean annual amount of rainfall over this region represents ∼40% of the total rainfall of the basin. Moreover, the Morvan is characterized by small valleys with a large river network that limits the spatial extent of the aquifers and by a rapid transfer of the water to the riverbed.
5.3.2. Impact of Urban Areas on Streamflows
 Towns are characterized by a strong surface runoff, because of the small capacity of the water reservoirs and the fact that the soils are impervious. The effects of towns on streamflows can be locally very important and can generate local floods. The use of TEB to simulate the urban areas, as was mentioned before, has a large impact on the water and energy budgets. However, it also has a positive impact on the simulation of the river flow, because of an increase of the fast surface runoff. It leads to a slight improvement of the streamflow simulation for small and highly urbanized catchments such as the Orge and the Yerres, compared to a simulation where towns are prescribed as bare impervious soil (Table 1).
Table 1. Impact of TEB for Highly Urbanized Subbasins: 1985–1990 Average of the Efficiency of the Daily River Flows and the Ratio of Simulated Over Observed Annual Discharge Qa
ISBA and TEB
ISBA and TEB
In the ISBA-only run, the urban areas are taken into account by an impervious bare soil.
ORGE at Morsang (932 km2)
YERRES at Evry (779 km2)
SEINE at Poissy (62,806 km2)
SEINE at Poses (65,686 km2)
SEINE at Paris (43,507 km2)
 Nevertheless, at the scale of the whole Seine basin, urban areas seem to have a smaller impact. Their explicit simulation through the adjunction of the TEB scheme has no sensible impact on the quality of the streamflow simulation of the Seine and of its main tributaries, which is explained by the fact that towns represent less than 3% of the total area of the basin and that the cities are not located in the rainiest areas.
5.4. Analysis of the Evolution of the Aquifers
 The annual evolution of the groundwater tables is linked to the annual amount of precipitation (Figures 15 and 5). The aquifers stored water during the 7 rainiest years (e.g., 1987–1988 and 2000–2001). In contrast, during the other 10 years, which received lower amounts of rainfall, the level of the groundwater tables decreases or is constant. The long-term trend is a small decrease of the water stored in the aquifers (compared to the amplitude of the annual variations). At the end of the simulation (July 2002) the level of the aquifers is lower than at the beginning of the simulation (August 1985), with a mean deficit after 17 years of simulation of 5 × 109 m3.
 The annual amount of infiltration to the aquifers reaches 131 kg m−2 yr−1, with a high interannual variability and a strong standard deviation (84 kg m−2 yr−1). The amount of water transferred from the aquifers to the rivers is much more regular, with a mean annual amount of 136 kg m−2 yr−1 and a standard deviation of 27 kg m−2 yr−1.
 The groundwater table of the Chalk is the largest one in terms of area and volume (3 and 2 times larger than the volume stored in the Oligocene and Eocene, respectively). The Chalk receives 79.1% of the total amount of infiltration, while the Oligocene and the Eocene aquifers receive only 9.6% and 13.3%, respectively. The contribution of the Chalk to the total amount of water transferred from groundwater to rivers reaches 76.4% (6.2% and 17.4% for the Oligocene and the Eocene aquifers, respectively; Figure 15).
 This study showed that the aquifers play a major role in the hydrologic functioning of the Seine basin (Figure 16). The contribution of the groundwater tables represents on average 40% of the average discharge of the outlet of the Seine basin (Poses). It reaches 75% in summer and ∼25% in winter, at a time of year when most of the infiltration is stored by the aquifers (see section 6.1).
 The base flow is particularly high for the Aisne catchment (at Choisy; see Figure 16), the Eure catchment (at Charpont), and the Oise catchment (at Pont-Sainte-Maxence). On an annual average over these basins it surpasses 50%, ranging from 30 to 50% in winter and reaching more than 80% in summer. In contrast, in the Yonne and the Aube catchments, the aquifer contribution ranges from only 5% in winter to 30–50% in summer, and its annual average is only 15%. This can be explained by the fact that the aquifers are present in only a small part of these catchments. These rivers, and particularly the Yonne, are very sensitive to surface runoff and consequently have a more rapid response to rainfall.
 The importance of the underground part of the discharges emphasizes the fact that in the Seine basin a correct simulation of the groundwater levels is particularly important for a good simulation of river flows.
6. Floods of the Seine at Paris
 The Seine is characterized by long floods in wintertime, which last ∼20 days. In this section, the processes leading to the formation of a flood are investigated as well as the role of the aquifers during floods. Moreover, the simulation of the nine main floods that occurred in recent years (between 1998 and 2002) is presented.
6.1. Detailed Analysis of a Flood
 Two important rivers join the Seine upstream from Paris: the Marne and the Yonne. Since they represent 40 and 30%, respectively, of the discharge at Paris, the floods of the Seine at Paris depend on the timing of the floods of these tributaries.
 The floods of the Seine have been classified into three different types by Marti and Lepelletier , following the temporal distribution of rainfall: (1) The simple floods are generated by a few rainy days (2 or 3) over the whole basin. The combination of the floods of the Yonne, Marne, and Seine above Paris determine the intensity and duration of the flood in Paris. (2) The double floods are due to two rain spells, separated by ∼4–6 days. This situation leads to a single flood wave for the Marne and the Seine above Paris, but to two successive floods of the Yonne, which reacts more rapidly to rainfall because of the fact that a large part of its basin is located in a mountainous area. The second flood wave of the Yonne river coincides at Paris with the ones from the Marne and the Seine and leads to a larger flood. This scenario caused the centennial flood of 1910. (3) Finally, the multiple floods are caused by a succession of rainfall events during a few weeks, which leads to rises in the water levels and finally generates a flood.
 The simulations of the nine main floods occurring between 1998 and 2002 have been analyzed in detail and compared to the naturalized observed streamflows. In all cases, most of the surface runoff is generated by the southeastern part of the basin. The rainfall usually occurs on a soil that is already wet, which leads to a high surface runoff. The urban areas' contribution to floods is weak at the scale of the basin, since they generate only 4% of the runoff during the flooding period. This article will focus on the December 1999 flood, which is a double flood. Figure 17a shows the daily amount of rain (averaged over the catchment of the Seine in Paris) as well as the observed and simulated discharges in December 1999. Two rainy spells can be distinguished: a first one from 17 to 19 December and a second one from 25 to 28 December. The two events do not compare to the heavy rainfall events that occur in the south of France, but are well-spread important rainfalls (∼10 kg m−2 d−1) that last a few days and that finally represent ∼15% of the annual amount of precipitation. At Paris a small increase of the river flow is followed by a more important one, and the top observed discharge reaches 1550 m3 s−1, which corresponds to a 10-year return period.
 The aquifers appear to play a crucial role during the floods, since the larger part of the drainage is stored underground. The fraction of drainage stored in the aquifers ranges from 65 to 85% (Table 2). If the aquifer could not store this water, it would quickly run off to the rivers; consequently, the discharges of the rivers would increase by ∼500 m3 s−1 on average, and the maximum discharge in December 1999 would increase from 1550 to ∼3000 m3 s−1 (Figure 17b). Subsurface storage helped to reduce the intensity of the floods and to prevent catastrophic damages.
Table 2. Maximum Observed Discharge, Amount of Water Stored in the Underground Layer, and Mean Additional Discharge That Would Exist Without the Underground Layer, for Five of the Selected Floodsa
Top Discharge, m3 s−1
Storage, 109 m3
Mean Additional Discharge, m3 s−1
Storage is given in 109 m3 and in percentage of the drainage.
6.2. Statistical Analysis of the Simulation of Floods
 For each flood the two statistical criteria defined in section 5.2 were computed over a period of about one month. They are presented in Table 3. The processes that lead to the flood of the Seine at Paris seem to be well simulated by SAFRAN-ISBA-MODCOU (SIM), as the statistical results are quite good (Table 3).
Table 3. Statistical Criteria (Efficiency E and Ratio Between Simulated and Observed Discharges Over the Period Q) and Top Observed Discharges of Floods of the Seine at Paris From 1998 Until 2002a
Maximum Observed Discharge
Discharges are given in m3 s−1.
20 Feb. 1999/25 March 1999
1300 m3 s−1
8 Dec. 1999/10 Jan. 2000
1550 m3 s−1
8 Feb. 2000/12 March 2000
1100 m3 s−1
19 Nov. 2000/9 Dec. 2000
900 m3 s−1
20 Jan. 2001/25 Feb. 2001
900 m3 s−1
8 March 2001/8 April 2001
1550 m3 s−1
25 Nov. 2001/21 Dec. 2002
870 m3 s−1
25 Dec. 2001/15 Jan. 2002
1200 m3 s−1
1 Feb. 2002/16 March 2002
1300 m3 s−1
 Only the first two floods of 2001–2002 are poorly simulated by SIM: The flood of December 2001 is underestimated by 10%, and the one of January 2002 is underestimated by 30%. Two possible reasons for this underestimation have been considered: The first one is a bad estimation of the evolution of the vegetation. The observed vegetation cycle of the year 2000 is reproduced every year, but it does not necessarily correspond to the situation of 2001–2002. As a consequence, the evaporation could be overestimated, which would lead to an underestimation of soil water content and, consequently, to less runoff production. The second reason could be that the levels of the aquifers are underestimated and the amount of water stored in the aquifers could be too high. Because of the lack of data the vegetation parameters and the piezometric level could not be corrected to represent the conditions of the year 2000–2001. Instead, some sensitivity tests were done, from which it appears that the two sources of error can certainly be combined to explain the underestimation of the floods of 2001–2002.
 In the seven other studied floods the model obtains good results, with an efficiency that ranges from 0.68 to 0.93 and surpasses 0.9 (very good) four times. The ratio between the annual simulated and observed discharges is always greater than 0.9 and lower than 1.1. SIM is able to produce good simulations of the floods of the Seine, in terms of intensity as well as in terms of temporal correlation.
 The aim of the study was to improve the understanding of the hydrological functioning of the Seine basin. The Seine basin, with its industrial pole, intensive agricultural regions, and high-density population (25% of France), is particularly sensitive to water resources and quality. It is also characterized by the importance of the groundwater, with several aquifer layers that cover a large part of the domain. In this study, an off-line coupled surface hydrology model was used to simulate the water and energy budgets of the Seine basin over a 17-year period, from 1985 to 2002. The simulation is based on the SIM coupled system, which includes the SAFRAN [Durand et al., 1993] analysis of the atmospheric forcing, two surface schemes that take into account the urban areas (TEB [Masson, 2000]) and natural areas (ISBA [Noilhan and Planton, 1989]), and the distributed hydrological model (MODCOU [Ledoux et al., 1984]).
 In contrast to the Rhone and Adour-Garone basins, where the SIM model has already been applied, the Seine basin is quite flat (no high mountain range) and presents a smooth spatial variation of the atmospheric forcing. However, two regions stand out, with a drier and warmer center area and a wetter and colder eastern part. Average monthly rainfall is also very smooth, with no marked annual cycle (∼60 mm month−1), but presents a large interannual variability, with a standard deviation of ∼25 mm month−1.
 The simulation was validated using ∼100 river gauges and 130 piezometric stations. The simulated streamflows are comparable with the observations, especially for larger river stations at the Seine and its main tributaries. The average error in the annual discharge over all the river gauges reaches 11% and is lower than 10% in one third of the stations. The efficiency is classified as good (E > 0.7) for 25% of the stations and correct (E > 0.5) for 55%. Such results are comparable to the ones obtained over the Rhone [Etchevers et al., 2001] and Adour-Garonne [Morel, 2003] basins. The spatial variation of the aquifer levels simulated by MODCOU is comparable to the observed piezometric data, although the local comparison can present large differences.
 The comparison of the observations and simulations gives confidence to the simulation of the functioning of the Seine basin by the SIM system and has allowed a more detailed analysis of the hydrological cycle. It appears that most of the discharge (70%) is generated by a small part of the basin (20%), the eastern area, which receives 40% of the rainfall. This region, the Morvan, is characterized by small rainy valleys with a large river network that limits the spatial extent of the aquifers, leading to rapid access of the water to the riverbed.
 The spatial variation of the total evaporation over precipitation ratio is quite smooth with an average of 70%, which is close to the ratio obtained in the flat areas of the Rhone and Adour-Garonne basins. However, the lowest values in the Morvan (50%) are still higher than what was obtained over the snowy Alps and Pyrenees mountains (∼20%).
 The urban areas, with very weak evaporation, a large runoff, and strong sensible and ground heat fluxes, have a very different signature than natural areas. However, the impact of an explicit modeling of the cities (like Paris) appeared to be very small on the Seine streamflow, even if it is more significant on small highly urbanized basins.
 A large part of this work was dedicated to the analysis of the role of the groundwater on the hydrological cycle. The underground part of the discharge reaches more than 80% of the Seine river flow in summertime and ∼40% in winter. During the wintertime floods the aquifers play a crucial role, since the larger part of the drainage is stored in the water tables, which significantly reduces the intensity of the floods and, thus, of their consequences.
 Special attention was paid to the simulation of the long and slow floods of the Seine at Paris. Such floods are directly connected to the timing of the floods in the two main tributaries that join the Seine river less than 50 km from Paris and that represent 70% of the flooding discharge at Paris. Eight of the nine latest large floods were accurately simulated by the SIM system, with an efficiency higher than 0.9 in 40% of the cases.
 Such results demonstrate the ability of SIM to represent the main behaviors of the Seine basin, especially during low flow and floods. The SAFRAN-ISBA-MODCOU coupled system is now operationally used at Météo-France. It will be used to monitor the soil moisture, snow pack, and aquifer level over the Seine basin and all of France. Such real-time simulation could be used to provide initial state conditions to atmospheric or hydrologic models and will be used in a forthcoming study to test the ability of SIM to forecast the long floods of the Seine basin.