Drainage systems are currently implemented on agricultural plots subjected to temporary or permanent waterlogging issues. Drained plots account for 9 % of all arable soils in France. As such, the need for accurate hydrological modeling is crucial, especially in an unstable future context affected by climate change. The aim of this paper is to assess the capacity of the SIDRA-RU hydrological drainage model to represent the variability in pedoclimatic conditions within French metropolitan areas and to demonstrate the utility of this model as a long-term management tool. The model is initially calibrated using the
Subsurface drainage is an agricultural soil management technique that controls soil water content and increases aeration in soils subjected to temporary or permanent water saturation issues into the soil depth (Jamagne, 1968; Baize and Jabiol, 2011). Plot water conditions are stabilized, thus ensuring better crop yields (Broadhead and Skaggs, 1982; Armstrong et al., 1988; Nijland et al., 2005; Ibrahim et al., 2013), while reducing the flood risk on plots (Henine et al., 2014; Tuohy et al., 2018b). Drained soils often belong to the hydromorphic soil category and sometimes, in the French context, lie on a shallow and impervious layer that reduces the deep infiltration (Thompson et al., 1997; Lange et al., 2011).
In France, all artificially drained soils comprise more than 2.7 million ha of arable soils (RGA – Agreste, 2010), i.e., close to 10 % of all arable land, corresponding to about 20 % for cereal-type field crops. In practice, several techniques exist to drain soils, such as subsurface drainage and open ditch. However, in France, over 80 % of drainage practices are conducted by introducing perforated pipes lying on the impermeable layer. The drain depth, spacing, slope, and diameter of these pipes constitute the main characteristics of each design; they are constrained by the local study site conditions, such as soil characteristics and climate (Mulqueen, 1998).
Since the economic and environmental consequences of climate change are of increasing concern to stakeholders, proper drainage practices have become a major issue. Predicting the long-term behavior of these systems is even more crucial in this context of water resource protection and restoration, since drainage has an impact on water quality (Tournebize et al., 2012, 2017, 2020). The literature contain studies targeting the impact of climate change on drainage practices, with an emphasis on either the increase in annual drained water balance (Pease et al., 2017) or agricultural productivity on drained plots (Jiang et al., 2020b). These topics raise concerns over the sustainability of existing drainage systems, and their need to be redesigned has come to the fore (Deelstra, 2015; Abd-Elaty et al., 2019). A common theme across all these studies is the need to properly represent drainage systems within each study area.
In this context, hydrological modeling offers a widespread tool for predicting drainage discharge, with several models currently in use, e.g., DRAINMOD (Skaggs, 1981; Skaggs et al., 2012) in the USA. This spatially distributed model operates on various spatial scales (Konyha and Skaggs, 1992; Brown et al., 2013) and integrates many modules in order to represent different hydrological processes and solute transports (Breve et al., 1997). In Europe, the MACRO model (Larsbo et al., 2005; Jarvis and Larsbo, 2012) is currently used by the FOCUS group (Adriaanse et al., 1996; Boesten et al., 1997) to evaluate drainage system performance and contaminant transport (Jarvis et al., 1997; Beulke et al., 2001). These two models, despite demonstrating their effectiveness, have been designed using physically based modeling strategies with a large number of parameters. Their calibration on several study sites becomes difficult and time consuming (Beven, 1989). Given this complexity, the SIDRA-RU model offers an interesting alternative. The model is semi-conceptual (Beskow et al., 2011), being composed of one physically based part (the SIDRA module) coupled with a conceptual part (the RU module), and parsimonious, by virtue of requiring the calibration of only six parameters, hence making it easy to configure (Perrin et al., 2003). Initially intended to simulate drained discharge during flood periods (Lesaffre and Zimmer, 1987), the SIDRA model converts weather-dependent soil recharge into drainage discharge by solving a semi-analytical formula derived from the Boussinesq equation (Boussinesq, 1904). Various modules have been integrated so as to better represent infiltration (Kao et al., 1998), water flux in the unsaturated zone (Bouarfa and Zimmer, 2000), or pesticide leaching (Branger et al., 2009). The RU module was recently integrated in order to model water transfer in the unsaturated zone (Henine et al., 2021). This new version inputs a continuous recharge term into the SIDRA module, which then allows for the simulation of drainage discharge over the entire hydrological cycle.
Due to soil diversity within French drained areas, a model used for management purposes must initially be as general as possible and correctly calibrated to ensure that the model behavior matches the behavior of each studied site as closely as possible (Perrin, 2000). As such, a reliable calibration protocol often depends on the choice of objective function (OF), which serves as the numerical criterion to be optimized so that the simulation more accurately resembles reality. Many OFs can be used to calibrate a model, such as the Nash–Sutcliffe efficiency (NSE; Nash and Sutcliffe, 1970) or the Kling–Gupta efficiency (KGE; Gupta et al., 2009), depending on the purpose of the particular study. The model here is intended for use on future prediction data under a long-term management scenario. From this perspective, the model must be temporally robust, i.e., its performance and parameters must remain independent of the period chosen for calibration (Klemeš, 1986). Such an evaluation can be performed by means of various tests, which tend to depend on the model structure (Refsgaard and Storm, 1996; Refsgaard, 2001; Henriksen et al., 2003; Daggupati et al., 2015). Since SIDRA-RU is a simple model, the split-sample test (Klemeš, 1986) is considered to be sufficient (Refsgaard, 1997). However, the national-scale evaluation of a hydrological model requires a large database, which is not readily available in the drainage hydrology field. This lack of data is the reason for the paucity of studies in the current literature.
The aim of this study is to assess the ability of a hydrological drainage model to simulate observed drainage discharge across several representative sites spread out in France. An exhaustive database, composed of 22 experimental sites and encompassing the main drained regions in France, has been built to account for the large diversity of French drained soils on which the model was tested. Database completeness is one of this paper's main strengths and allows the generalization of our results on soil diversity. The hydrological model chosen for this work is SIDRA-RU, a parsimonious model that yields continuous simulations and can easily be run on the database. In addition, the SIDRA-RU model offers a novel tool for the hydrological drainage modeling community; this study, therefore, provides an opportunity to test its performance at the national scale, which raises another point of interest regarding this study. Moreover, the temporal robustness of the SIDRA-RU model is assessed in the aim of asserting whether or not the model can be used within the scope of a long-term management tool, i.e., one capable of incorporating climate change.
A multitude of materials constitute French soils, as defined by their geological context, textural evolution, and regional climate. All of the above characteristics serve to determine the uniqueness of a soil. Making generalizations about soil diversity then becomes a necessary step. Indeed, grouping them by soil category facilitates their modeling. Several official classifications serve to group soil types (FAO, 1988; Krogh and Greve, 1999; Driessen et al., 2000). In this study, we are proposing to classify them by texture, thus making it possible to sort the database into three categories (see Fig. 1 and Table 1). Let us note that we do not consider the geological context or the regional climate to classify soils here.
The pedological distribution of French soils, produced by Jamagne et al. (1977) and Lagacherie and Favrot (1989), overlain on the spatial distribution of French drainage (RGA, 2010; top panel). The spatial distribution of sites aggregated by texture with observed flow data is shown in the bottom panel.
Main characteristics of the 22 plots with observed discharges and associated
The Lagacherie and Jamagne classification (Jamagne et al., 1977; Lagacherie and Favrot, 1987; Richer-de-Forges et al., 2008) has been used to evaluate this strategy. According to Fig. 1 (top), three distinct soil types occupy most of the regions with the highest drainage ratio (i.e., percentage of a land area that has been drained, with drainage ratio values above 50 % of total arable area). First, the Glossic and Planolosic soils, belonging to the Luvisols and mainly located around the Paris Basin and in the Allier region (see Fig. 1), are defined by textural differentiation between the surface horizon, which is often silty and sometimes sandy–loamy, and a deep clayey horizon. Second, brown acidic and leached soils, mostly distributed in the western part of France (Fig. 1), lie on a magmatic and metamorphic substratum; they are often characterized by a silty–clay or loamy texture. Third, the Pseudogley soils are also substantially drained, yet they remain only slightly correlated with any specific soil texture. We assume here that, among the three studied soil textures, they are more related to silty–clay soils (Lévy, 1972). Figure 1 (bottom) shows the drainage ratio of arable soil in France. Most drainage systems (approximately 80 %) lie on a loamy texture, according to Lagacherie and Favrot (1987), except in the eastern part of France (Fig. 1), where drained soils are either predominantly silty–clay or heavy clayey soils.
A representative database of drainage discharge across France was specially assembled for this study. The data originate from the following sources: (1) the ORACLE research project (Tallec et al., 2015) and artificial Rampillon wetland (Tournebize et al., 2012, 2017; Lebrun et al., 2019), (2) the partnership with the ARVALIS Institute, which monitors the La Jaillière experimental site, and (3) data from reference drainage sites dating between the 1960s and 1980s (Lagacherie and Favrot, 1987; Jannot, 1988).
The data from this last source stem from monitoring experiments managed by INRAE (formerly Cemagref) that test drainage modalities, i.e., what depth, space, or pipes best fit to the field conditions. The combination of these sources yields a database of nearly 200 years of cumulative hydrological records on drained plots in diverse pedoclimatic contexts over a broad plot scale range (e.g., 0.8–700 ha). The resulting extensive data set compiled for hydrological modeling purposes encourages the transferability of this study's findings.
The drainage network of the various study plots is based on similar technical characteristics (see Table 1) composed of PVC perforated and corrugated pipes lying at a depth of 0.85 to 1 m, with an inter-drain spacing from 8 to 24 m. The most widely used method for monitoring drainage discharge consists of measuring the corresponding water level at the drainage collector outlet using a calibration curve fitted at each measurement site, by designing a control section where flow is hydraulically managed. Before the 1980s and 1990s, data were recorded on a paper sheet that followed the motion of a floater linked to the water level. Nowadays, water level sensors (floating systems equipped with ultrasonic measurements) are used, and the data are digitally recorded. To ensure data homogeneity, observed data have been manually assessed by expert judgment in order to highlight periods of suspect data quality; as deemed necessary, the data have been corrected or deleted. The 22 study plots are distributed over three distinct soil textures, namely silty, silty–clay, and clayey. A total of 15 of them are characterized by a silty soil texture that covers most French regions (Fig. 1). The database is more limited with regard to clayey soils, characterized by sites like Saint_Laurent_P2 or Courcival_P3. Some regions, e.g., eastern France, which are strongly characterized by a clayey texture yet with just one clayey site, are not well represented. The SIDRA-RU performance in this region will be estimated from the SIDRA-RU global performance on clayey soils from the database, comprising 44 years of observed discharges (Table 1). Moreover, some regions with a high drainage ratio, e.g., southwestern France, have no observation points and are therefore not covered by this study. Lastly, the Pseudogley soils are mostly correlated with the silty–clay soils, yet the database does not provide any relevant silty–clay plots. Model performance will thus be estimated by the global performance for all such sites. Each site was defined by the aforementioned technical characteristics (drain depth, mid-drain spacing corresponding to the half-space between two successive drains, and surface area), plus the length of available observed discharge logs and suitability to the split-sample test (Table 1).
Due to a lack of agronomic data, we assume here that growing practices do not affect the subsurface drainage hydrology on the study plots, except in the absence of a tillage technique (Dairon et al., 2017), a situation that is not widespread in France. This assumption is supported by four observations, namely that (1) the subsurface drainage is mainly effective during fall and winter, when the actual evapotranspiration is low; (2) our study investigates drained soils in winter primarily used for winter crops, such as wheat and maize (Zimmer, 1996), so we assume that the studied plots are cultivated every year throughout this season without a fallow period; (3) these crop types impact the subsurface drainage hydrology in a similar manner; hence, annual crop rotation does not add significant bias to the model calibration; and (4) the effect of cover crops (Meyer et al., 2018) has been neglected due to the fact that they were not widely used before 2012 and, for 19 of the 22 sites, the corresponding study periods ended before 2012.
The meteorological data were provided by the SAFRAN database (Vidal et al., 2010), a meteorological reanalysis covering France and supplying both precipitation and potential evapotranspiration (PET; based on the FAO-56 Penman–Monteith PET; Córdova et al., 2015) data on all database-referenced plots. These data are available from 1959 to 2019 at a daily time step and a spatial resolution of 8 km, which is 1000 times greater than the scale of the studied plots. This difference may introduce errors on model outputs, but all such errors are considered to be negligible.
Diagram presenting the various modeling stages of the SIDRA-RU model.
The SIDRA-RU model is a semi-conceptual, lumped model that describes the hydrological processes of artificial drainage systems. This model is based on the principle of rainfall–drainage discharge conversion and uses the rainfall,
First, an evapotranspiration module converts PET into an approximate value of actual evapotranspiration, called corrected evapotranspiration (CET), from the available water level
Second, the RU module, a conceptual storage, calculates the water table recharge term Stage 1: Stage 2: Stage 3:
Third, the calculated water table recharge
Parameter optimization (i.e., calibration) is commonly performed in a hydrological modeling context in order to adapt the hydrological model parameters to the specific study area context. This process is relevant for conceptual parameters since, by construction, they can neither be measured directly nor be easily correlated with any physical characteristics of the studied system. Some of the physically based parameters that are too difficult to measure can also be calibrated.
The model calibration herein has been based on the algorithm implemented in the airGR R package (Coron et al., 2017, 2020) and is composed of two parts. First, a systematic examination of the parameter space provides the most likely zone of convergence, on the basis of a grid-screening algorithm (Mathevet, 2005), according to a given performance criterion (i.e., OF). Each parameter space is defined by its specific distribution and intrinsic statistical characteristics with respect to the soil texture. Second, a steepest-descent local search procedure (Michel, 1991) seeks to improve the OF, beginning with the grid-screening part, and find a more accurate estimate of the parameter set, i.e., with higher model performance.
The hydraulic conductivity
Various OFs are commonly used in hydrological calibration processes, depending to a great extent on the primary aim of the study. The most widespread OFs are RMSE (Anderson and Woessner, 1992), MSE (mean squared error; Ye et al., 2020), NSE (Nash and Sutcliffe, 1970), and, more recently, KGE (Gupta et al., 2009). Our goal here is to evaluate model performance in order to simulate an entire hydrological cycle and represent the interannual variations in the study plot to produce long-term projections about future drainage hydrology. We have thus introduced the
The temporal robustness of the model proves to be of use when facing a time period different from that chosen for calibration, i.e., enhancing the model's capacity to perform equally well over different and contrasting time periods (Li et al., 2012). This point is particularly important when the model is intended for application under future climate change scenarios (Thirel et al., 2015b).
The choice of evaluation strategy mainly depends on model structure. For lumped conceptual models, a simple split-sample test (Klemeš, 1986) is sufficient to assess the robustness of such a model (Refsgaard and Storm, 1996; Daggupati et al., 2015). The split-sample test, as illustrated in Fig. 3, consists of splitting the data period into two sub-periods (P1 and P2) and then calibrating the model over both of them independently.
Functional diagram of the split-sample test.
Thus, two optimal parameter sets are obtained (one covering P1 and the other P2), with each being tested over the other sub-period (e.g., evaluation sub-period P2 for a calibration over sub-period P1). If the
This test was performed on the records from nine plots showing at least 10 years of time series data (Table 1). These series were split into two equal length periods, and the
Evaluating and highlighting model performance limitations during the calibration process using numerical scores, such as NSE, RMSE, or KGE criteria, are often recommended, in addition to graphical analysis (Moriasi et al., 2015). However, understanding their evolution is indeed difficult. To overcome this difficulty, a model performance classification by means of a range of numerical score values will help assess the calibration quality.
In subsurface drainage modeling, the bibliography is not exhaustive with respect to
The model has also been evaluated in terms of its capacity to reproduce annual cumulative discharges, with a direct comparison conducted between observed and simulated flow rates using both the linear correlation coefficient
Table 1 lists the performance over the entire calibration period obtained from all 22 sites, and Table 2 classifies the model performance from each soil texture according to the score ranges. Performance varies across the three soil textures, with both unsatisfactory
Hydrograph on La_Jaillière_P4 after calibration over the entire available record, plus a magnification of hydrological year 2002–2003 and the associated temporal and direct comparisons between observed (blue curve) and predicted (red curve) cumulative discharges.
The La_Jaillière_P4 plot is used as an example to illustrate the temporal comparison between observed and simulated discharges over 16 years (see Fig. 4). These same graphs are available in Appendix B to illustrate the case of a silty–clay soil at La Bouzule_P2 and in Appendix C for the case of a clayey soil at Saint-Laurent_P2.
Figure 4 shows that the simulated discharges are in good agreement with observations in terms of both seasonal dynamics (dry and wet season alternation) and cumulative distribution. Rainfall series are not directly correlated with discharge, as some rainfall events appearing from September to November do not systematically lead to subsurface flow. However, winter rains typically turn into discharge after a period of 1 or 2 days. A graphical analysis shows that simulated drainage discharges generally start in the same period as the observed discharge, with various delays depending on the soil type.
During hydrological year 2002–2003, the drainage season starts on the same day for both simulation and observation (i.e., 1 November). In 2002–2003, the maximum observed drainage discharge lies close to 15
The capacity of the SIDRA-RU model to represent the water balance has also been assessed. Figure 4 shows that the simulated cumulative drainage discharge over March 2002 lies close to the observed discharge, with a slight underestimation of 10 mm, i.e., below 3 % of the annual drained water balance relative to 350 mm. The linear regression between observed and simulated cumulative discharges yields an equation close to a
Hydrograph on Courcival_P3 after calibration over the entire available record, plus a magnification of the hydrological year 1992–1993 and the associated temporal and direct comparisons between observed (blue curve) and predicted (red curve) cumulative discharges.
Unlike at La_Jaillière_P4, the SIDRA-RU model at the Courcival_P3 plot (see Fig. 5), which lies on a swelling clayey soil, shows larger discrepancies between observed and simulated discharges on the plots. The red and blue curves do not coincide on a significant portion of the logs. From 1985 to 1995, simulated discharges often started later than the observed ones, with delays ranging approximately from 2 weeks to 2 months. The start of the drainage season is defined here when significant discharges appear.
The plots for the cumulative drained discharge in 1992–1993 reveal that the annual drained water balance diverges by
Comparison of the drained annual discharges between simulation and observation – each point represents the drained water quantity from a hydrological year for a given site, and all points have been classified by texture.
Figure 6 provides a comparison between the predicted and observed total cumulative discharges on each plot and for each hydrological year, as classified by soil texture. The linear regressions lie close to the
Differences between prediction and observation of low (
In order to use the SIDRA-RU model for a long-term prediction of agricultural drainage management in France, the model must first be able to reproduce both high and low drainage discharges. Figure 7a depicts the differences in the
Figure 7c shows that the box plot medians for the
The
Comparison between the
For starters, this figure shows that all points are located under the line
On silty soils, the maximum variation is observed at Melarchez (the largest studied site, with 700 ha), with
Hydrographs of observed and predicted discharges on calibration (parameter set derived from sub-period P2) and evaluation (parameter set from sub-period P1) recorded at La Jaillière in 2002–2003. Predicted cumulative discharges are directly compared to observations.
A graphical comparison between the predicted drainage discharges, calculated using the evaluation parameter set, and the observed discharges helps assess model robustness, as depicted in Fig. 9 at La_Jaillière_P4 during the 2002–2003 hydrological year. The calibration simulations use the parameter set calibrated over sub-period P2 (including the 2002–2003 season), while the evaluation simulations use the parameter set calibrated over sub-period P1. Figure 9 indicates that both calibration and evaluation drainage discharges lie close to the observed levels. The peak flows from both simulation curves have been superimposed on the main part of the drainage season; hence, the evaluation parameter set performs well over the studied time period.
Differences between the two simulations only exist at the beginning of the drainage season, as strongly controlled by the RU module. The drainage season appears to start on the same day, according to both simulation curves, but the discharge magnitude during the first few days differs from one simulation to another. This same graphical approach has been used on a clayey soil available in Appendix D for Saint_Laurent_P2 during the 1982–1983 season. These observations are similar to the previous ones on La_Jaillière_P4.
Similar observations have also been recorded on the cumulative discharges (Fig. 9), which display an identical behavior, except for a gap between two predicted flows appearing at the start of the drainage season, and have tended to remain so throughout the year. A direct comparison of the cumulative flows between observation and simulation yields linear regressions near the
The use of the split-sample test for model calibration raises the question of parameter set similarity between the two calibration periods across all sites studied. Figure 10 illustrates how each of the four parameters (
Comparison of parameter sets from both calibration periods (sites are referenced by their associated index, as shown in Table 1).
One of the two extreme outlier points corresponds to Melarchez (point S4), where
Soil reservoir parameters
The literature (Perrin, 2000; Coron et al., 2012; Montanari et al., 2013; Thirel et al., 2015a) suggest that, when assessing generalizable hydrological models, the aim typically consists of determining the ability of a model to reproduce the hydrological behaviors of various study areas. The larger the database, the more reliable the study because the model is being evaluated over a wider diversity of geological and climatic contexts (Gupta et al., 2014). To the best of our knowledge, however, in drainage modeling, the models are often evaluated on a short-term database, with just a few years on a few sites, e.g., DRAINMOD (Skaggs et al., 2012), MACRO (Jarvis and Larsbo, 2012), ADAPT (Gowda et al., 2012), and SWAT (Arnold et al., 2012). This lack of data limits the opportunities to test models in various soil types and, moreover, prevents the assessment of the relevance of models over a broad array of spatial scales.
The database used in this study was specifically built to assess the performance of a drainage discharge model on a larger data set. In total, 22 experimental sites were compiled, accounting for nearly 200 hydrological years spread over the main drainage areas in France, on contrasting soil types and in different pedoclimatic settings, featuring an extensive plot-scale range. This database has been classified according to three main soil textures (silty, silty–clay, and clayey) using available drainage discharge data. Not all regions with a high drainage rate are well represented; however, 80 % of French drained soils do have silty soil textures, as represented by the majority (15 of the 22 plots) in the database. The real advantage of this database is the ability it offers to apply the model to referenced sites that represent a large majority of France's drainage diversity. This topic accounts for both the originality and key contribution of this study.
The diversity of the database introduced requires a model that operates correctly and in accordance with each site's specific conditions, i.e., as generalizable as possible. In this context, a physically based model is theoretically better suited. However, this kind of model is generally composed of many parameters representing the complexity of a study site, such as current crop, root depth, saturated water content (i.e., unsaturated one), hydraulic parameters, or water-holding capacity. Using such a model on a large database requires providing all these characteristics for each site, which is in practice very difficult because the measurement technics are globally expensive. The calibration might be useful, but it is time consuming due to the large number of parameters. Conversely, a simple model like SIDRA-RU offers significant advantages, as it requires little information and faster calibration, thus becoming in practice suitable for generalizing. The model is lumped and, initially requiring the calibration of six parameters, the assumptions made in this study allow us to reduce this number to four parameters. Compared to the original SIDRA model (Lesaffre and Zimmer, 1987), the current version simulates continuous discharges over several hydrological years, thus simulating both wet and dry periods. The calibration process only necessitates brief rudimentary knowledge of the soil texture, used here as a priori input data to establish the model parameter distributions. Regarding input variables, the model requires rainfall and PET data at a daily time step to predict the drainage discharge. Managed in conjunction with the SAFRAN climate database to satisfy data needs across France, SIDRA-RU can easily be launched on all drained areas throughout the territory.
Once calibrated, the model shows a performance ranging from acceptable to very good on all sites, judging from the
Let us note, however, that some peak flows have been slightly underestimated. This behavior is generalizable and has been observed on most database sites. However, the phenomenon varies from one site to another without any clear relationship with any soil characteristic and mostly depending on the calibration quality. One explanation might be that the SIDRA-RU model is calibrated on both high and low drainage discharges, as represented by a unique parameter set, which consequently introduces biases. These biases become more significant on extreme discharges. Furthermore, the missed variations sometimes lead to missing the beginning of the drainage season. During some hydrological years, the observed drainage discharge starts before water level in the soil reservoir reaches the
Regarding the split-sample test, results first showed that evaluation
Another important analysis is the consistency of model calibration. The SIDRA module actually solves a simplified formula derived from the Boussinesq equation and requires a good estimation of both hydraulic parameters, namely hydraulic conductivity
Regarding hydraulic conductivity
Distribution of
As regards
On non-deformable clayey soils, such as Saint_Laurent_P2, SIDRA-RU produces a good performance. However, at Courcival_P3, a plot lying on a deformable swelling clayey soil, the SIDRA-RU performance is significantly poorer. As such, the latter does not constitute an exception in the drainage modeling community. Indeed, the literature identifies clayey soils as being a recurring problem for drainage modeling (Robinson et al., 1987; Snow et al., 2007), especially in mole drainage, as currently practiced on heavy clayey soils and swelling clays (Jarvis and Leeds-Harrison, 1987; Tuohy et al., 2016). This finding is mainly due to a difference in hydraulic characteristics between silty soils, on which the model design is primarily based, and heavy or swelling clayey soils. The latter are characterized by natural pedological deformations, like soil surface fracturing, which lead to preferential flow zones before saturation (Beven and Germann, 1982; Jarvis and Leeds-Harrison, 1987). The horizontal soil profile is no longer homogeneous, which contradicts one of the main hypotheses of SIDRA. Moreover, agricultural practices like plowing exacerbate this phenomenon and, therefore, affect soil porosity. One way to improve the results at Courcival_P3 would be to artificially locate the pipe at a depth of 30 cm instead of the current 90 cm.
Evolution of the water table shape between a silty and a heavy clayey drained soil, adapted from Bouarfa and Zimmer (2000) and Branger et al. (2009).
Another critical assumption of the SIDRA module is the elliptical shape of the water table; this assumption facilitates the numerical resolution of the Boussinesq equation. As regards heavy clayey soils, this hypothesis is no longer suitable, since the water table shape evolves towards a rectangular structure (Fig. 12). This phenomenon is due to the very low hydraulic conductivity (Robinson and Rycroft, 1999; Skaggs et al., 1999), which, for SIDRA-RU, is difficult to integrate. Furthermore, the water table level drops when approaching the drain because the soil has been turned over on this profile in order to bury the drain, like for the aforementioned natural pedological deformations.
Furthermore, the issues observed on clayey soils are specifically significant at the start of the drainage season. At Saint_Laurent_P2, delays occur more frequently than on silty soil, so the more the plot is defined by heavy or swelling clayey soil, the longer the delay. At Courcival_P3, these delays were of the order of 1 month. The RU module design partially addresses this problem by generating a soil profile recharge before saturation, but this issue remains a major limitation of the model, i.e., the more clayey the soil, the poorer the model performs. A soil is considered to be mainly clayey once the clay fraction exceeds 35 % (Richer-de-Forges et al., 2008), which can be defined as the limit beyond which good model performance is no longer guaranteed.
Some models show their efficiency when simulating drainage discharge on such soils. Among them, there is the MACRO model which showed its efficiency in simulating drainage discharge in Europe, including structured soils like heavy clayey soils (Köhne et al., 2009). However, these models are based on a more physically oriented approach than SIDRA-RU. Building on this type of concept would probably improve the latter in simulating drainage discharges on clayey soils but might undermine its generalist nature.
Another limitation of this model is that the RU parameters are slightly less temporally robust than the SIDRA parameters, showing fewer stable values between calibration sub-periods than the SIDRA parameters. This issue might originate from dependencies of the RU module on interannual meteorological variations. The
Another point we wish to discuss deals with the dependency of the aforementioned interpretations on the conditions and choices involved in conducting this study. We noted above the lack of data in the database used for certain French regions. Making assumptions, e.g., grouping soils by categories, to address this weaknesses would then offer a good alternative yet still be dependent on arbitrary decisions. Adding reference sites could complete and significantly improve the SIDRA-RU model robustness analysis.
Moreover, the model is based on some rather important assumptions, such as neglecting both the surface runoff and recharge to groundwater, or the ones made on the parameters
Furthermore, the choice of calibration process drives the calibration results based on study goals. We have implemented herein a grid-screening algorithm that assigns the best combination of parameters according to their respective distributions, coupled with a step-by-step algorithm. This approach has the advantage of being theoretically entirely automatic and, thus, eliminates the subjective aspect of calibration; however, external decisions influencing results are still necessary.
As an example, in order to track the main purpose of this study, SIDRA-RU is calibrated using the
The distribution functions of each model parameter are required for the calibration, which in this case depends on the decision to classify soils based on their texture. Regarding
Driving the distribution with realistic value ranges at each site can prove to be a relevant solution. For example, if we consider that, on a specific soil,
In the perspective of long-term management on drained plots, predicting flows in order to better monitor the use of agricultural pollutants is a major concern due to pollutant transfers occurring with drainage flow (Kladivko et al., 2001; Trajanov et al., 2018). Thus, a good model can be used as a decision-making tool, for example to restrict pollutants' application during flow period for the case of pesticides (Lewan et al., 2009; Zajíček et al., 2018; Kobierska et al., 2020). Indeed, using water content as an indicator to anticipate the start of drainage flow in order to reduce pesticide applications is a recommended strategy instead of restriction timing (Brown and van Beinum, 2009; Lewan et al., 2009). In this context, using SIDRA-RU may be quite relevant. However, the current form of the RU module is not optimal to accurately represent the fate of pollutant in soil profile as it is too simple to precisely represent the behavior of the water table inside the unsaturated zone. To overcome this problem, this model type is generally coupled with pedotransfer functions (Jury and Roth, 1992; Magesan et al., 1994) to transfer water and pollutant stock from the unsaturated zone to the saturated zone. Within this framework, the perspective of the PESTDRAIN module (Branger et al., 2009), coupled with the SIDRA-RU model, should allow the simulation of pesticide leaching by including two reservoirs, namely a fast reservoir to mimic preferential flow above the drain area and a slow reservoir through the soil matrix. Based on a similar approach, combining SIDRA-RU with a nitrate leaching module might also be useful in order to correctly assess water pollution on French drained plots.
The aim of this study has been to implement an large database characterizing the main drainage areas throughout France, so as to assess the performance and robustness of the SIDRA-RU model. A database comprising 22 drained sites was built to represent French soil diversity and describe the three main soil textures from France's main drained parts (i.e., silty, silty–clay, and clayey). Results indicate that the SIDRA-RU model yields satisfactory drainage discharge simulations for nearly all studied sites. Moreover, the model shows especially good performance on silty soils, which account for 80 % of all drained plots in France. Despite a number of limitations, particularly for clayey soils, the model was found to be temporally robust at the national scale, which enables us to conduct long-term impact studies. Once calibrated, this model can indeed be used to assess the resilience of drainage systems under climate change according to climate scenarios like those from the CMIP5 project (Eyring et al., 2007; Taylor et al., 2011) or the upcoming CMIP6 project.
Table of mean and standard deviation (SD) per soil texture used for the parameter distribution.
Hydrograph on La_Bouzule_P2 after calibration over the entire available record, plus a magnified plot of hydrological year 1982–1983 and the associated temporal and direct comparisons between observed (blue curve) and predicted (red curve) cumulative discharges.
Hydrograph on Saint_Laurent_P2 after calibration over the entire available record, plus a magnified plot of hydrological year 1982–1983 and the associated temporal and direct comparisons between observed (blue curve) and predicted (red curve) cumulative discharges.
Hydrographs of observed and predicted discharges on the calibration (parameter set derived from sub-period P2) and evaluation (parameter set from sub-period P1) recorded at Saint_Laurent_P2 in 1982–1983. Predicted cumulative discharges are directly compared to observations.
An intellectual property certificate has been issued by the Agency for Program Protection for the model code, under the reference IDDN.FR.001.260018.000.S.P.2018.000.30100. There is no public repository providing it. However, this paper provides an exhaustive description of the model, enabling readers to easily reproduce it. The code is written in FORTRAN language and then imported in R software, to be computed with the calibration algorithm “airGR” (Coron et al., 2017, 2020) available on CRAN (Comprehensive R Archive Network).
Climatic data for this work were provided by Météo-France and are not publicly available since they are part of a commercial product. Météo-France can freely provide them on request for research and non-profit purposes. Most of the flow data are available on request. However, some flow data are provided by private partners and are not available.
AJ, HH, and JT conceptualized the work. LC and GT provided the methodological guidelines. CC, HH, and JT collected data and rendered them sustainable. AJ drafted the paper. HH, CC, LC, GT, and JT all revised the paper and contributed to its analyses and discussions.
The authors declare that they have no conflict of interest.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This work benefited from the French state aid managed by the ANR under the Investissements d'avenir programme (grant no. ANR-16-CONV-0003; CLAND project). Our gratitude is also expressed to the Météo-France Meteorological Agency for providing the data used in this work from the SAFRAN database. Moreover, thanks are addressed to both the ARVALIS Plant Sciences Institute for data from the La Jaillière measurement site and the ORACLE Research Observatory for its data. The authors would like to thank the editor (Christian Stamm) and the two reviewers, for their constructive and kind comments and suggestions.
This research has been supported by the ANR under the Investissements d'avenir programme (grant no. ANR-16-CONV-0003; CLAND project).
This paper was edited by Christian Stamm and reviewed by Nicholas Jarvis and one anonymous referee.