Variational Data Assimilation To Improve Subsurface Drainage Model Parameters

Variational Data Assimilation to improve subsurface drainage
model parameters
Samy Chelil1 , Hind Oubanas1 , Hocine Henine1 , Igor Gejadze1 , Pierre-Olivier Malaterre1 ,
and Julien Tournebize1
1
INRAE
November 24, 2022
Abstract
Variational data assimilation (VAR-DA) has been implemented to estimate the unknown input parameters of a new agricultural
subsurface drainage model (SIDRA-RU) through assimilating discharge observations. The adjoint model of SIDRA-RU has
been successfully generated by means of the automatic differentiation tool (TAPENADE). First, the adjoint model is used to
explore the local and global adjoint sensitivities of the valuable function defined over the drainage discharge simulations with
Posted on 24 Nov 2022 — CC-BY 4.0 — https://doi.org/10.1002/essoar.10507081.1 — This a preprint and has not been peer reviewed. Data may be preliminary.
respect to model input parameters. Next, the most influential parameters are estimated by applying the VAR-DA embedded
into a simple stochastic procedure in order to achieve the global minimum. The performed sensitivity analysis shows that
the most influential parameters on drainage discharge are those controlling the dynamics of the water table; the second most
influential parameters manage the starting date of the drainage season. Compared to an alternative gradient-free calibration
performance, the estimation of these governing parameters by the VAR-DA method improves the overall quality of the drainage
discharge prediction, in particular in terms of the cumulative water volume. Improved parameters generate less than 5 mm
(1%) of the discrepancy between simulated and observed water volume, based on the five years of daily discharge observations
on the Chantemerle agricultural parcels (36 ha). Preliminary numerical tests allow identifying the potential presence of local
minima as well as equifinality issues. The latter can be highlighted by the self-compensation of both the physical soil parameters
and the main conceptual parameters. Moreover, the proposed techniques may be applied to a panel of hydrological and water
quality models.
1
Water Resources Research
1 Variational Data Assimilation to improve subsurface drainage model parameters
2 Samy CHELIL1, Hind OUBANAS2, Hocine HENINE1, Igor GEJADZE2, Pierre Olivier
3 MALATERRE2 and Julien TOURNEBIZE1
1
4 University of Paris-Saclay, INRAE Jouy-en-Josas - Antony, UR HYCAR, Antony, France
2
5 G-EAU, Univ Montpellier, AgroParisTech, BRGM, CIRAD, IRD, INRAE, Institut Agro,
6 Montpellier, France
7 Corresponding author: Samy CHELIL (samy.chelil@inrae.fr)
8 Key Points:
9  The most influential parameters of the SIDRA-RU model have been identified using the
10 global adjoint sensitivity analysis
11  Variational data assimilation significantly reduces the water volume discrepancy in
12 comparison with a free-gradient calibration method
13  A gradient-based stochastic procedure allows the identification of the globally optimal
14 solution without the need for the prior/background
15 Abstract
16 Variational data assimilation (VAR-DA) has been implemented to estimate the unknown input
17 parameters of a new agricultural subsurface drainage model (SIDRA-RU) through assimilating
18 discharge observations. The adjoint model of SIDRA-RU has been successfully generated by
19 means of the automatic differentiation tool (TAPENADE). First, the adjoint model is used to
20 explore the local and global adjoint sensitivities of the valuable function defined over the drainage
21 discharge simulations with respect to model input parameters. Next, the most influential
22 parameters are estimated by applying the VAR-DA embedded into a simple stochastic procedure
23 in order to achieve the global minimum. The performed sensitivity analysis shows that the most
24 influential parameters on drainage discharge are those controlling the dynamics of the water table;
25 the second most influential parameters manage the starting date of the drainage season. Compared
26 to an alternative gradient-free calibration performance, the estimation of these governing
27 parameters by the VAR-DA method improves the overall quality of the drainage discharge
28 prediction, in particular in terms of the cumulative water volume. Improved parameters generate
29 less than 5 mm (1%) of the discrepancy between simulated and observed water volume, based on
30 the five years of daily discharge observations on the Chantemerle agricultural parcels (36 ha).
31 Preliminary numerical tests allow identifying the potential presence of local minima as well as
32 equifinality issues. The latter can be highlighted by the self-compensation of both the physical soil
33 parameters and the main conceptual parameters. Moreover, the proposed techniques may be
34 applied to a panel of hydrological and water quality models.
35 1. Introduction
36 Agricultural drainage is designed to remove the excess water within soils subjected to frequent
37 and continuous waterlogging, in particular during the winter season. The water flowing through
38 soil porosities is drained by gravity towards buried drains or ditches. Three stages can be
39 distinguished during the agricultural drainage year (Lesaffre, 1988). The first, called “pre-flow
40 drainage season”, consists of a gradual saturation of the soil and a limited flow into drains. Next
41 comes an “intensive drainage season” (IDS) featuring a fast continuous discharge of water via the
42 drains. The last stage encompasses the “drying season”. The latter is characterized by the gradual
43 continuous decrease in water flow until the drainage flow stops. The main objectives of drainage
44 are to ensure better crop yields (Mohtadullah, 1990) and facilitate grazing and access to cultivated
45 fields by farmers during the IDS.
46 The hydrological behavior of agricultural drainage networks is highly dependent on
47 hydrometeorological data (e.g. rainfall, evapotranspiration and temperature), soil hydro-physical
48 properties (e.g. hydraulic conductivity, porosity) and water holding capacity (WHC). To better
49 understand the hydrology of drained agricultural land, several existing numerical models
50 approximate the underlying dynamics and physical processes. A complete model of the perched
51 water table formed in drained soils would be complex, time-consuming and still require multiple
52 input data and parameters. In contrast, some simplified models allow for a reasonable monitoring
53 of the drainage discharge and water table level from a limited input dataset; they are not however
54 able to describe the physical evolution of hydrological and hydraulic processes (Tournebize et al.,
55 2004).
56 Several numerical models are reported in the literature (Gurovich & Oyarce, 2015) for designing
57 drainage networks and monitoring hydraulic patterns, including: DRAINMOD (Skaggs et al.,
58 2012), EnDrain (Valipour, 2012), CSUID (Alzraiee et al., 2013), and SIDRA (Lesaffre, 1988).
59 The latter was initially developed at INRAE (formerly Cemagref) by Lesaffre (1988) to describe
60 the hydrological and hydraulic behavior of drained agricultural land. More specifically, the model
61 simulates water table variations at the mid-drain and calculates discharge at the drain outlet. Since
62 that time, several improvements have been proposed (Sami Bouarfa, 1995). Recently, a new
63 conceptual module has been introduced to manage the quantity of water seeping through the
64 vadose soil zone (Jeantet et al., 2021). Other existing models describe the hydrological process at
65 the catchment scale and, as an option, integrate a module managing the water, pesticides and
66 nutriment fluxes evacuated by drains; examples of such models include: HYPE (Lindström et al.,
67 2010), SWAT (G. Arnold et al., 2012), ADAPT (Gowda et al., 2012), and PESTDRAIN (Branger
68 et al., 2009). Model performance depends on both the simplifications made to the underlying
69 equations and the quality of the input information. The latter can be improved by assimilating
70 available observations of the model outputs through either existing calibration techniques or
71 variational DA.
72 Therefore, it is important to analyze the influence of input information on model simulations in
73 order to identify the most influential inputs to be estimated. This step can be achieved by means
74 of Sensitivity Analysis (SA), which is considered a key preliminary step to the calibration/data
75 assimilation DA (Stange et al., 2000; Migliaccio & Chaubey, 2005). Various deterministic or
76 stochastic methods are available to perform SA. In practice, the deterministic approach is
77 commonly used in local SA, while the stochastic approach is more appropriate for global SA.
78 While the local SA reveals the model response to a small perturbation around a given control
79 parameter value, the global SA focuses on the variance of the model output due to input variability.
80 Several approaches can be employed in local and global SA, including variance-based approaches
81 (Sobol, 1993; Saltelli, 2002) and derivative-based methods, e.g. the adjoint sensitivity analysis
82 (Cacuci, 1981), which involves a simultaneous computation of all sensitivities using adjoint
83 operators within a single model run (Goutal et al., 2018; Gejadze et al., 2019). In (Kucherenko,
84 2009) and (Lamboni et al., 2013), inequality links have been established between the derivative-
85 based global sensitivity measures (DGSM) and Sobol’s indices, thus yielding SA for a large
86 number of input parameters. In the same context, Gejadze et al. (2019) derived a generalized
87 relationship between the global sensitivity indices and the Polynomial Chaos coefficients, in
88 making SA suitable for high-dimensional models.
89 Several studies have focused on improving model performance through the estimation of optimal
90 input parameters in the aim of closely replicating the available observations. The numerous
91 approaches identified range from commonly used calibration techniques in hydrological
92 applications, e.g. PAP-GR Michel’s calibration algorithm developed in (Michel, 1989) and
93 highlighted by Mathevet (2005), to more sophisticated data assimilation (DA) methods. DA
94 methods seek to solve the inverse problem by assimilating all available information; the Kalman
95 filter and its ensemble variants, Particle filters and the variational approach would all be
96 appropriate candidate DA methods. Variational DA has been preferred in operational geophysical
97 applications, e.g. numerical weather prediction (NWP) and ocean modeling, due to its scalability,
98 robustness and ability to handle nonlinear systems under high uncertainty (Courtier et al., 1998; P.
99 Gauthier et al., 1999; Rabier et al., 2000; Fischer et al., 2005; Pierre Gauthier et al., 2007).
100 References on the application of variational DA to hydrological models remain rather scarce. In
101 recent studies, the optimal estimates of the unknown model variables and/or parameters have been
102 obtained via minimization of a well-defined cost function in accordance with gradient-based
103 methods (Oubanas et al., 2018; Ghorbanidehno et al., 2020; Jay-Allemand et al., 2020). This
104 procedure requires developing the adjoint counterpart of the forward model. This could at times
105 prove to be difficult, which is one reason why variational DA has not gained more popularity
106 within the hydraulic/hydrological research community. One can note a relatively recent review (X.
107 Zhang et al., 2009) proposing an objective assessment of five global calibration algorithms,
108 excluding DA methods, that involve the distributed hydrological model SWAT (G. Arnold et al.,
109 2012). The performance of each algorithm is assessed using both the Nash-Sutcliffe criterion
110 (Nash & Sutcliffe, 1970) and the number of simulations required to reach the optimal solution.
111 Despite the convergence of these algorithms towards a satisfactory solution, a large number of
112 iterations is still needed to get a reliable solution, especially when handling a large number of data
113 and/or input parameters.
114 A common issue when solving ill-posed problems in a variational formulation using gradient-
115 based minimization is the likely presence of multiple local minima in the cost function (Dennis &
116 Moré, 1977). A comparison between different optimization algorithms applied to hydrological
117 models was presented by Arsenault et al. (2014), highlighting their convergence speed,
118 performance and the local minima issue. Some of these methods can be highly sensitive to the
119 choice of the initial guess, known as the “background” in DA or the “prior” in statistics. This
120 finding was supported by Pan and Wu (1998), in which an original approach based on the simplex
121 method was used to reveal and avoid convergence to local minima.
122 Skahill and Doherty (2006) proposed a simple methodology to address the local minima issue.
123 Based on a local search algorithm, their methodology consists of starting the minimization process
124 with different parameter values in order to increase the probability of reaching the global minima.
125 The same concept has been deployed when calibrating 3D morphodynamic model data
126 (Shoarinezhad et al., 2020) in using the PEST parameter estimation package (Welter et al., 2012),
127 thus showing that the local optimization process is substantially affected by the local minima issue.
128 This paper focuses on improving the performance of the SIDRA-RU drainage model (i.e. the latest
129 version of SIDRA), which integrates a new conceptual module in addition to the existing physical-
130 based one. Consequently, SIDRA-RU enables calculating the water table recharge by taking into
131 account the water pathway through the vadose zone of soils. SIDRA-RU is run with a daily time
132 step and requires knowledge of the rainfall, potential evapotranspiration and a set of input
133 parameters. The relative simplicity of SIDRA-RU facilitates generation of the adjoint code. Let us
134 note that SIDRA-RU will be coupled with a recently developed conceptual nitrate model (beta
135 release), which simulates nitrate concentrations at a daily time step by using both the drainage
136 discharge and variation of the nitrate pool in drained soils. Hence, the present work can be
137 considered as a test bench for implementing various techniques on a significantly more complex
138 and upgraded model encompassing the nitrate transport module.
139 For starters, the sensitivity analysis is performed so as to identify the most influential model inputs.
140 This investigation entails the adjoint local and global sensitivity analysis, in accordance with the
141 methodologies presented in Goutal et al. (2018) and Gejadze et al. (2019). The purpose here is to
142 implement and test a general methodology that allows selecting a sufficient control subset
143 (extracted from the full set of model inputs), while keeping in mind a possible extension to a
144 distributed parameter formulation in the subsequent upgraded versions. Next, the variational DA
145 method is applied to estimate the parameters included in this subset, which involves all available
146 information. Our choice of DA method has been motivated by the anticipated presence of
147 distributed variables and heterogeneous observations (i.e. discharge and nitrate concentrations) in
148 the upgraded model versions. Lastly, the local minima issue is investigated through the deployment
149 of ensemble-based approaches.
150 The remainder of this paper is organized as follows. Section 2.1 presents the study area and related
151 data. The SIDRA-RU drainage model is introduced in Section 2.2, while the methodology of the
152 adjoint sensitivity analysis, variational DA and handling of the local minima issue are described
153 in Sections 2.3, 2.4 and 2.4.1, respectively. The results of the experiments are displayed and
154 discussed in Section 3, with main findings being summarized in Section 4.
155 2. Materials and Methods
156 2.1. Study area description and data

157 The Chantemerle agricultural area has been considered for this study. The field includes a set of
158 artificially drained agricultural parcels and comprises a land area of 36 ha. It accounts for part of
159 the 9% of drained agricultural land in France. The study area is located in Aulnoy (Figure 1) within
160 the “Seine et Marne” department (70 km east of Paris).
161
162 Figure 1. Location of the Chantemerle agricultural parcels
163 These parcels are monitored through the GIS-ORACLE framework https://gisoracle.inrae.fr/. The
164 soil is of the luvisol type (Tournebize et al., 2015) with a silty texture (mixture of sand and clay).
165 This type of soil causes fast and continuous water flows during wet periods, when rainfall can
166 exceed 20 mm.day-1 (Figure 2). Since September 2008, the Chantemerle sub-catchment is
167 monitored by means of discharge and rainfall measuring devices at the collector outlet, as
168 presented by Tournebize et al. (2015). The data are collected at hourly intervals and then
169 transformed into the daily time step to ensure consistency with the model design. Daily potential
170 evapotranspiration (PET) data are extracted from the SAFRAN database (source: Météo-France).
171 PET values are calculated according to the Penman-Monteith formula. Available data cover the
172 period from September 1st, 2008 to August 31st, 2013, which spans five hydrological years in all
173 (Figure 2). On average, the drains are laid at a depth of 0.9 m and spaced by a distance of 10 to 12
174 meters.
175 Among all recorded data, two wet hydrological years, 2010/2011 and 2012/2013, can be
176 distinguished, accumulating respectively 172 mm and 217 mm of drained water. The three
177 remaining years account for a combined 209 mm, which is nearly equal to the drained water
178 accumulated in just a single wet season. This contrast in quantity of drained water (i.e. succession
179 of dry and wet seasons) lends more reliability to our results.
16 0
14 10
12
20
Q, PET (mm.d-1)
10
P (mm.d-1)
Potential evapotranspiration (PET)
30
Discharge (Q)
8
Rain (P)
40
6
50
4
2 60
0 70
Date (days)
180
181 Figure 2. Discharge data used as output data and rainfall, potential evapotranspiration
182 used as inputs in SIDRA-RU
183 2.2. The SIDRA-RU model
184 SIDRA-RU is a semi-analytical and semi-conceptual model, which is run at the scale of
185 agricultural plots with a daily time step. SIDRA-RU uses as inputs rainfall data (P) and potential
186 evapotranspiration data (PET), both considered as forcing variables, along with model parameters
187 and simulates water level variations (h) at the mid-drain and drainage discharges (Q) as outputs.
188 The SIDRA-RU model input parameters are listed in Table 1. Note that accurate values of the soil
189 and reservoir parameters are difficult (and sometimes impossible) to obtain at the field scale. These
190 values include: saturated hydraulic conductivity (𝐾𝑠𝑎𝑡 ), drainage porosity (µ), intermediate water
191 level threshold in the reservoir(𝑆𝑖𝑛𝑡𝑒𝑟 ), and reservoir storage capacity during the intensive drainage
192 season (𝑆𝑆𝐷𝐼). These parameters therefore must be estimated using discharge observations (Q).
193 The remaining parameters are considered as either given invariants, regardless of field
194 characteristics, or typical invariant characteristics at the field scale. The latter category includes
195 the drainage design parameters (drained area “S”, drain spacing “l”, depth “P”). It should be
196 pointed out that the presence of an impermeable layer, on which the drains are laid, limits deep
197 infiltrations (aspect not considered herein). In addition, (Augeard et al., 2005) showed, under the
198 same field conditions, that surface runoff becomes negligible since the agricultural plots are being
199 extensively drained, which leads to infrequent soil overflow.
200 Table 1. Classification of SIDRA-RU parameters

Structure Type Parameters Notes
Physical Soil Saturated hydraulic conductivity (Ksat) Most often
inaccessible
Drainage porosity (µ)
Values usually
Drainage design Parcels area (S)
known and
Drain spacing (l) dependent on the
Drain depth (P) studied field
Water table shape Coefficients (a1 and a2) Internal

parameters,
Conceptual Evapotranspiration Crop coefficient (Kc) assumed to be
Exponential factors (β1 and β2) invariants
Reservoir Water share coefficients (α1 and α2)
Intermediate threshold (Sinter) Inaccessible
Reservoir storage capacity (SSDI)
201 SIDRA-RU is divided into two modules. The first module (RU) is based on managing water
202 saturation in the conceptual soil reservoir, which is assumed to be comparable with the WHC of a
203 soil. In a first step, PET is transformed into corrected evapotranspiration (CET) using the 𝐾𝑐 crop
204 coefficient, as follows:
205 𝐶𝐸𝑇(𝑡) = 𝐾𝑐 ∗ 𝑃𝐸𝑇(𝑡) (1)
206 The 𝐾𝑐 coefficient is then calculated by means of an empirical approach based on a decomposition
207 of the hydrological year into two periods. The first period is defined by the wet months, ranging
208 from September to February, when no crops or only crops in the early growth stage are present;
209 the second period, considered as dry months, extends from April to August, when crops are present
210 at an intermediate/advanced growth stage:
𝐾𝑐 = 𝐾𝑐𝑚𝑎𝑥 (wet months)
211 { −(𝑆𝑖𝑛𝑡𝑒𝑟 −𝑆(𝑡−1)−𝑃(𝑡))
⁄
(2)
(𝑆𝑖𝑛𝑡𝑒𝑟 +𝑃(𝑡))
𝐾𝑐 (𝑡) = 𝛽1 + 𝛽2 ∗ 𝑒 (𝑑𝑟𝑦 𝑚𝑜𝑛𝑡ℎ𝑠)
212 where 𝛽1 and 𝛽2 are two parameters set respectively to 0.25 and 1, 𝑆 corresponds to the water
213 storage at time step (𝑡 − 1), 𝑆𝑖𝑛𝑡𝑒𝑟 is a calibration parameter indicating the threshold level at which
214 drain flow can be observed, and 𝐾𝑐𝑚𝑎𝑥 (set to 0.86) corresponds to the maximum crop coefficient
215 value.
216 𝑆𝑖𝑛𝑡𝑒𝑟 and 𝑆𝑆𝐷𝐼 both depend on the agricultural plot location (e.g. climate, soil type) and serve to
217 manage water level variations inside the conceptual reservoir. The sum of 𝑆𝑆𝐷𝐼 and 𝑆𝑖𝑛𝑡𝑒𝑟
218 equals 𝑆𝑚𝑎𝑥 , defined as the maximum conceptual reservoir capacity; its value is assumed to lie
219 close to the water holding capacity (WHC).
220 Next, variables P(t) and CET(t) are transformed into the net rainfall 𝑃𝑛𝑒𝑡 in order to update the
221 water storage 𝑆 of the previous time step, as follows:
𝑃𝑛𝑒𝑡 (𝑡) = 𝑃(𝑡) − 𝐶𝐸𝑇(𝑡)

222 { (3)
𝑆(𝑡) = 𝑆(𝑡 − 1) + 𝑃𝑛𝑒𝑡 (𝑡)
223 Recharge 𝑅(𝑡) and water storage 𝑆(𝑡) are determined according to the diagram presented in Figure
224 3.
Stock: 𝑆(𝑡 − 1)
0 < 𝑆(𝑡 − 1) < 𝑆𝑖𝑛𝑡𝑒𝑟 𝑆𝑖𝑛𝑡𝑒𝑟 < 𝑆(𝑡 − 1) < 𝑆𝑚𝑎𝑥 𝑆(𝑡 − 1) > 𝑆𝑚𝑎𝑥
𝑅(𝑡) = 0 𝑅(𝑡) = 𝛼 ∗ 𝑃𝑛𝑒𝑡 (𝑡) 𝑅(𝑡) = 𝑆(𝑡 − 1) − 𝑃𝑛𝑒𝑡 (𝑡) − 𝑆𝑚𝑎𝑥

pnet≥ 0
𝑆(𝑡) = 𝑆(𝑡 − 1) + 𝑃𝑛𝑒𝑡 (𝑡) 𝑆(𝑡) = 𝑆(𝑡 − 1) + (1 − 𝛼) ∗ 𝑃𝑛𝑒𝑡 (𝑡) 𝑆(𝑡) = 𝑆𝑚𝑎𝑥
Pnet< 0 𝑅(𝑡) = 0 & 𝑆(𝑡) = 𝑆(𝑡 − 1) + 𝑃𝑛𝑒𝑡 (𝑡)
225
226 Figure 3. Evolution of recharge with respect to the water storage level and the net rainfall
227 where 𝛼 is a coefficient introduced when the water level in the conceptual reservoir lies between
228 𝑆𝑖𝑛𝑡𝑒𝑟 and 𝑆𝑚𝑎𝑥 , in order to determine the net rainfall fraction either supplying the recharge or the
229 reservoir storage. Its value depends on the season period. Hence, 𝛼 assumes a value of 𝛼1 = 0.33
230 during the wet season and 𝛼2 = 0.07 during the dry season, when 𝑃𝑛𝑒𝑡 is essentially used to update
231 the water storage value.
232 The second part of the model, i.e. the SIDRA module (Lesaffre, 1988), which is physically based,
233 simulates both discharge at the drainage outlet and the water table level at the mid-drains. The
234 approach adopted is based on a horizontal integration (from drain to mid-drain) of the Boussinesq
235 Equation (S Bouarfa & Zimmer, 2000), as given by Equation 4:
µ 𝜕ℎ 𝜕ℎ 2 𝜕2 ℎ 𝑅(𝑡)
236 = (𝜕𝑥 ) + ℎ 𝜕𝑥 2 + 𝐾 (4)
𝐾𝑠𝑎𝑡 𝜕𝑡 𝑠𝑎𝑡
237 where µ [-] denotes the drainage porosity of the soil, 𝐾𝑠𝑎𝑡 [L.T-1] the saturated hydraulic
238 conductivity, h [L] the water level from the impermeable layer, and R [L] the recharge uniformly
239 applied between drain and mid-drain.
240 The spatial and temporal analytical integration of the Boussinesq Equation (Boussinesq, 1904) is
241 possible when considering two main assumptions: (1) the drainage porosity provides for a match
242 between water table fluctuations and the quantity of water stored in the surface soil layer; and (2)
243 the water flow direction is assumed to be horizontal in the soil compartment located between the
244 drain and mid-drain (known as the Dupuit-Forchheimer assumption). The water table level and
245 drainage discharge can then be defined in Equation 5 (see (S Bouarfa & Zimmer, 2000) and
246 (Lesaffre, 1988) for further details).
ℎ(𝑡−1)2
ℎ(𝑡) = ℎ(𝑡 − 1) + (𝑅(𝑡) − 𝐾 )⁄(µ ∗ 𝑎2 )
𝐿2
247 𝑄(𝑡) = 𝑅(𝑡) ∗ (1 − 𝑎1 ) + 𝐽(ℎ(𝑡)) ∗ 𝑎1 (5)
ℎ(𝑡)2
{ 𝐽(ℎ(𝑡)) = 𝐾 𝐿2
248 where Q [L.T-1] denotes the discharge simulated at the drainage outlet, 𝑎1 [-] and 𝑎2 [-] the water
249 table shape coefficients, and L [L] the half-drain spacing.
250 The simulated discharge at the drainage system outlet is divided into two parts: 𝑠1 + 𝑠2 = 1, where
(1−𝐴)𝑅(𝑡)
251 𝑠1 = 𝑄(𝑡) is the part generated by the recharge term R(t) due to rainfall and peak flow, and
𝐻(𝑡)2
252 𝑠2 = 𝐴𝐾 𝐿2 /𝑄(𝑡) is the part established by water table recession, as calculated according to the
253 Hooghoudt Formula (Hooghoudt, 1940). For a complete description of the latest version of the
254 SIDRA-RU model, the reader is referred to (Jeantet et al., 2021).
255 2.3. Adjoint sensitivity analysis

256 In addition to the climate data of rainfall and potential evapotranspiration, SIDRA-RU features 14
257 input parameters that are difficult to assess directly. In order to identify the parameters to estimate,
258 a sensitivity analysis (SA) is performed as an initial step to highlight those parameters resulting in
259 the model's most sensitive response. Next, the identified parameters form the control vector to be
260 estimated using the data assimilation method (see Section 2.4).
261 The present paper investigates both local and global adjoint SA. The latter are derivative-based
262 approaches that use the SIDRA-RU adjoint model to compute the gradient of a valuable function 𝐽,
263 as defined based on the model response. Note that the formulation of 𝐽 depends on the purpose of
264 the study.
265 2.3.1. Local adjoint sensitivity analysis (LASA)

266 Let us consider a model M that transforms input vector 𝐼 into model output vector 𝑋:
267 𝑀(𝐼) = 𝑋 (6)
268 In the particular case of the SIDRA-RU model, input vector 𝐼 includes the vector of parameters 𝑈𝑖
269 and the climate forcing datasets (i.e. rainfall and potential evapotranspiration):
270 𝐼 = {𝑃(𝑡), 𝑃𝐸𝑇(𝑡), 𝑈} (7)
271 In this study, rainfall P and evapotranspiration PET are considered to be known with high accuracy;
272 their values will therefore not be estimated. The input parameter vector 𝑈 is defined as follows:
𝑈 = {𝑆, 𝑙, 𝐾𝑠𝑎𝑡 , µ, 𝐴, 𝐶, 𝑃𝑑𝑟𝑎𝑖𝑛 , 𝑆𝑖𝑛𝑡𝑒𝑟 , 𝑆𝑆𝐷𝐼, 𝛽1 , 𝛽2 , 𝛼1 , 𝛼2 , 𝐾𝑐𝑚𝑎𝑥 }
273 { (8)
𝑆𝑆𝐷𝐼 = 𝑆𝑚𝑎𝑥 − 𝑆𝑖𝑛𝑡𝑒𝑟
274 The model response 𝑋 represents the simulated temporal evolution of drainage discharge 𝑄:
275 𝑋 = 𝑄(𝑡) tϵ[0,T] (9)
276 Since in the current SA framework the model output is directly and completely observed, no
277 observation operator is required and the corresponding observations are hereafter denoted 𝑋 ∗ .
278 In the context of local SA, the sensitivity of the model response with respect to inputs is quantified
279 locally around a given point. Since the true value of 𝑈 is unknown, it is typically substituted by a
280 prior 𝑈𝑏 , where index “b” stands for the “background” information in DA. A reasonable initial
281 guess of 𝑈 can be provided by an expert evaluation or through introducing a pre-calibration step.
282 Adjoint SA methods compute the sensitivity of model response 𝑋 to model inputs 𝑈 via the
283 gradient of the valuable function 𝐽(𝑈).
𝜕𝐽(𝑈) 𝜕𝐽(𝑈) 𝜕𝐽(𝑈) 𝜕𝐽(𝑈) 𝑇
284 = ( 𝜕𝑈 , ,…, ) (10)
𝜕𝑈 1 𝜕𝑈2 𝜕𝑈𝑛
285 Note that all sensitivities are computed simultaneously using a single adjoint model run, which
286 constitutes the main advantage of adjoint-based approaches compared to direct or statistical
287 methods. Moreover, the derivative obtained is exact (up to machine accuracy) rather than a
288 numerical approximation. The adjoint model is generated using the Tapenade automatic
289 differentiation tool (Hascoet & Pascual, 2013), developed at Inria
290 (https://team.inria.fr/ecuador/en/tapenade/).
291 Since the input variables are of different natures, the corresponding sensitivities might not be
292 physically comparable; hence, we define a background error covariance matrix 𝐵 = 𝐸(𝜉𝑏 𝜉𝑏𝑇 ),
293 where 𝜉𝑏 = 𝑈𝑡 − 𝑈𝑏 and 𝑈𝑡 is the truth. Each component of the gradient is then weighted by the
294 respective standard deviation, as follows:
𝑇
1/2 𝜕𝐽(𝑈) 1/2 𝜕𝐽(𝑈) 1/2 𝜕𝐽(𝑈) 1/2 𝜕𝐽(𝑈)
295 𝐵𝑈 = (𝐵𝑈1 , 𝐵𝑈2 , … , 𝐵𝑈𝑛 𝜕𝑈𝑛 ) (11)
𝜕𝑈 𝜕𝑈1 𝜕𝑈2
296 In this study, the Kling-Gupta Efficiency (KGE, Gupta et al. (2009)) has been chosen as a valuable
297 function 𝐽(𝑈). KGE is widely used to assess the performance of hydrological models and improve
298 simulation quality (Pechlivanidis et al., 2011; Patil & Stieglitz, 2015; Haas et al., 2016; Santos,
299 2018). The KGE metric ranges from -∞ to 1, where 1 corresponds to the best match between
300 simulations and observations.
301 𝐽(𝑋, 𝑋 ∗ ) = 1 − √(1 − 𝑟)2 + (1 − 𝑑)2 + (1 − 𝑚)2 (12)
302 where:
𝑋
303 -𝑚= : the ratio between the mean of the simulated and observed data.
𝑋∗
∑𝑛 ∗ ∗
𝑖=1(𝑋𝑖 −𝑋)(𝑋𝑖 −𝑋 )
304 -𝑟= 2
: the Pearson correlation coefficient.
√∑𝑛 (𝑋𝑖 −𝑋) √∑𝑛 (𝑋 ∗ −𝑋 ∗ )2
𝑖=1 𝑖=1 𝑖
2
𝑋 ∗ √∑𝑛
𝑖=1(𝑋𝑖 −𝑋)
305 -𝑑= : the ratio between the standard deviation of simulated and observed
𝑋 √∑𝑛 ∗ ∗ 2
𝑖=1(𝑋𝑖 −𝑋 )
306 values.
307 Moreover, the use of KGE allows the performances of adjoint-based and Sobol methods (Sobol,
308 1993) to be compared. The latter had recently been explored in Henine et al. (Submitted).
309 2.3.2. Global adjoint sensitivity analysis (GASA)

310 While the local SA answers how a small perturbation around the background value 𝑈𝑏 influences
311 model response 𝑋, the global SA focuses on the variance of model output or, more precisely, on
312 how input variability influences output variance. The global SA reveals which parts of the output
313 variance are due to different inputs through an estimation of Sobol indices, which are a central tool
314 (and key point) since they provide a quantitative and rigorous overview of the influence of inputs
315 on model response. The commonly used global SA method is the variance-based ANOVA
316 decomposition. This paper introduces the adjoint-based global SA described in (Gejadze et al.,
317 2019), featuring a methodology that derives a generalized relationship between the global
318 sensitivity indices and the Polynomial Chaos coefficients.
319 The variability of inputs is represented by a sample of random vectors obeying a given distribution.
320 One can consider for example the Gaussian distribution using the background vector 𝑈𝑏 as its
321 mode. However, since prior information on 𝑈 is not always available and/nor accurate, we use the
322 uniform distribution to define a random sample within a defined interval [𝑎𝑖 , 𝑏𝑖 ] for each element
323 of 𝑈𝑖 , as follows:
𝜉𝑖 ~𝑢(𝑎𝑖 , 𝑏𝑖 )
324 { , 𝜀𝑖 ~𝑢(0,1) (13)
𝜉𝑖 = (𝑏𝑖 − 𝑎𝑖 ) ∗ 𝜀𝑖 + 𝑎𝑖
325 The bounds 𝑎 and 𝑏 are specified for all SIDRA-RU parameters. The choice of the uniform
326 distribution overcomes the need for background 𝑈𝑏 and, subsequently, for the pre-calibration step.
327 Only the minimum and maximum values for each parameter are therefore needed.
328 The sample size is chosen by taking the computational requirements and model complexity into
329 account. Since the SIDRA-RU is relatively inexpensive, it is feasible to use a relatively large
330 sample size, which can be chosen to reach a compromise between the stability of the sensitivities
331 and CPU time.
332 In this paper, we only consider the first-order ‘main effect’ 𝑆𝑖 indices as well as the ‘total effect’
333 𝑆𝑖𝑇 indices, both defined in (Saltelli, 2002), as follows:
𝑉𝑎𝑟(𝐸[𝐽(𝜉)|𝜉𝑖 ])
𝑆𝑖 = 𝑉𝑎𝑟(𝐽(𝜉))
334 (14)
𝑉𝑎𝑟(𝐽(𝜉))−𝑉𝑎𝑟(𝐸[𝐽(𝜉)|𝜉−𝑖 ])
𝑇
{𝑆𝑖 = 𝑉𝑎𝑟(𝐽(𝜉))
335 where 𝐽(𝜉)|𝜉𝑖 represents the random value of 𝐽, with input component 𝜉𝑖 being fixed at its generic
336 value; 𝜉−𝑖 stands for ‘all, but i’, and 𝑉𝑎𝑟(𝐽(𝜉)) is the total variance of 𝐽(𝜉).
337 While the main effect indices, which are commonly used in ANOVA analysis, quantify the single
338 influence of individual variables or some groups of variables, the total effect indices measure the
339 influence of a variable jointly with all its interactions with other variables. Alternatively, these
340 indices can be viewed as a measure of the wiggle room remaining to the output when all but one
341 input variables are fixed. If the total sensitivity index of a variable is sufficiently small, this variable
342 can be removed from further analysis, because neither the variable nor any of its order interactions
343 have an impact on the valuable function. Thus, the total sensitivity index can be used to detect the
344 essential variables for the calibration process.
345 The methodology presented in Gejadze et al. (2019) suggests a generalized approximation for the
346 upper bound of the total effect indices 𝑆𝑖𝑇 based on a computation of the derivative of the valuable
347 function 𝐽. We limit ourselves here to the first-order analysis. Future work will investigate the
348 approximation of 𝑆𝑖𝑇 in considering second order interactions. Given that 𝜉 is a vector of
349 independent and identically distributed normal random variables, the following estimate holds:
𝜕𝐽(𝜉) 2
350 𝑆𝑖𝑇 𝑉𝑎𝑟(𝐽(𝜉)) ≤ 𝐸 [( ) ] (15)
𝜕𝜉𝑖
351 In the case of independent uniformly distributed variables defined over the interval [𝑎𝑖 , 𝑏𝑖 ]
352 (Lamboni et al., 2013), the following can be written:
𝜕𝐽(𝜉) 2 (𝑏𝑖 −𝑎𝑖 ) 2
353 𝑆𝑖𝑇 𝑉𝑎𝑟(𝐽(𝜉)) ≤ 𝐸 [( 𝜕𝜉 ) ] ( ) (16)
𝑖𝑛 𝜋
354 Similarly, we employ the same formulation of the valuable function introduced in Section 2.3.1.
355 The gradient of 𝐽 within the GASA framework is computed using automatic differentiation. A
356 detailed algorithm of the global adjoint SA applied to the SIDRA-RU model is described in
357 Figure 4.
358
359 Figure 4. Simplified version of the global adjoint SA algorithm
360 2.4. Variational data assimilation algorithm
361 Data assimilation methods are widely used in the geophysical sciences (Reichle, 2008). These
362 methods provide the best possible estimates of unknown model inputs (initial and driving
363 conditions, parameters) by combining all available information: observations, a priori data
364 (coming from expertise or preprocessing), and knowledge of the physics of the underlying
365 phenomena incorporated into the model.
366
367 Figure 5. Scheme describing the operating mode of the direct problem (blue lines) and inverse
368 problem (red lines) methods for the case of the SIDRA-RU model
369 In the ‘forward’ or ‘simulation’ mode, the forcing terms (rainfall and PET) and the SIDRA-RU
370 model parameters are provided as known inputs in order to simulate the drainage discharge. The
371 ‘backward’ or 'inverse' mode seeks to estimate the unknown input parameters by assimilating
372 available information from the observed output discharge (Figure 5).
373 Most of the previous studies using SIDRA-RU and similar models have only investigated
374 calibration techniques that utilize a direct simulation mode. The recent paper written by (Jeantet
375 et al., 2021) presents an example that uses the “Michel calibration (PAP-GR)” algorithm to
376 estimate the parameters (Ksat, µ, SSDI and Sinter). This algorithm was originally developed at
377 INRAE (formerly IRSTEA) to calibrate the GR (Génie Rural) models. An operational version of
378 the algorithm is available in an R package (Coron et al., 2017). This method is initially based on
379 assigning a probability distribution law for each calibrated parameter in order to target the optimal
380 parameter variation domain. The calibration procedure is then refined by searching for the optimal
381 parameter values within the pre-selected variational bounds.
382 The variational DA method yields optimal estimates of the unknown parameters by minimizing a
383 cost function 𝐽, defined as follows:
1 𝛼 2
𝐽(𝑈) = ‖(𝑀(𝑈)−𝑋 ∗ )‖2 + ‖𝐵 −1/2(𝑈−𝑈𝑏 )‖
2 2
384
𝐵 = 𝐸(𝜉𝑏 𝜉𝑏𝑇 )
(17)
𝑎≤𝑈≤𝑏
385 where 𝜉𝑏 is the background error and 𝛼 the regularization parameter. Then, Equation 17 is re-
386 formulated using the change of variables 𝑉 = 𝐵1/2 (𝑈 − 𝑈𝑏 ) and implemented in an 'iterative
387 regularization' fashion, as described in (Oubanas et al., 2018).
388 In the present study, a limited memory LBFGS-B algorithm (Zhu et al., 1997) has been used for
389 the constrained minimization step. As explained above, the choice of DA method is largely
390 motivated by the need to investigate and optimize the upgraded version of SIDRA-RU (to be
391 reported in a future paper). This model will be able to simulate the daily nitrate concentration
392 values by integrating the unknown distributed variable, in addition to the model input parameters.
393 The distributed variable represents the potentially leachable quantity of the nitrate initially trapped
394 in the soil surface layer. This quantity is estimated once a year, generally at the end of the autumn
395 season. The control vector will then be extended accordingly. For this reason, the choice of
396 variational DA is justified since the algorithm sought needs to be robust to the present uncertainties
397 and heterogeneity of the variables involved. Hence, this paper offers the first implementation of
398 the variational DA to SIDRA-type models, which will be adapted to the new upgraded version.
399 2.4.1. Toward avoiding the local minima issues

400 Due to ill-posedness of inverse problems and the expected high uncertainty levels in the modeled
401 dynamics and observations, many corresponding minimization problems suffer from the presence
402 of local minima, which prevents reaching the global minimum. Various approaches have been
403 investigated to mitigate this difficulty. Inspired by the works of Duan et al. (1992) and Skahill and
404 Doherty (2006), we suggest to use an ensemble minimization framework. The idea here is to start
405 each individual minimization process from a different first guess, which might result into different
406 minimization trajectories, thus limiting the dependence of the solution derived from the prior
407 knowledge and its uncertainty. The ensemble of priors/backgrounds (used as initial points for
408 minimization) is generated using a uniform distribution and limited to the sufficient subset
409 elements, as identified through the GASA method (see Section 3.1). Two configurations have been
410 tested herein: 1) the input parameters are perturbed one by one, setting the remaining ones at their
411 optimal values; and 2) parameters are perturbed simultaneously. The ensemble size is then chosen
412 as a compromise between target accuracy and available computational resources.
413 2.5. Evaluation of the temporal robustness of the model

414 An ensemble of operational tests proposed by KlemeŠ (1986) have been widely used to evaluate
415 the spatiotemporal robustness of hydrological models (Henriksen et al., 2003; Guo et al., 2020;
416 Jeantet et al., 2021). In the following, we have selected the split sample test to evaluate the
417 temporal robustness of the SIDRA-RU model. This method consists of splitting the study period
418 into two subperiods (e.g. 𝑃1 and 𝑃2 ). The first window 𝑃1 is considered as an assimilation period.
419 Hence, the daily discharge is assimilated using variational DA in order to estimate the control
420 vector 𝑈. The parameter estimates obtained are then validated by comparing the corresponding
421 model predictions with observations over the period 𝑃2 . Next, the assimilation and validation
422 periods 𝑃1 and 𝑃2 are swapped in order to validate the temporal robustness of the estimator.
423 Overall, four metrics have been selected to evaluate the model's predictive performance. On can
424 recall the KGE metric (Section 2.3.1; Equation 12) and the cost function (Section 2.4; Equation
425 17). In addition we use the root mean square error (RMSE (Kenney, 1939)), as defined in Equation
426 18,
∑𝑁 ∗
𝑖=1(𝑋𝑖 −𝑋𝑖 )
2
427 𝑅𝑀𝑆𝐸 = √ (18)
𝑁
428 and the Nash-Sutcliffe Efficiency (NSE (Nash & Sutcliffe, 1970)), as defined in Equation 19,
429 which is introduced to evaluate the performance of several hydrological and water quality models.
∑𝑁 (𝑋𝑖∗ −𝑋𝑖 )2
430 𝑁𝑆𝐸 = 1 − ∑𝑖=1
𝑁 ∗ )2
(19)
𝑖=1(𝑋𝑖 −𝑋
431 In addition, we are quantifying the water balance conservation by calculating the discrepancy on
432 cumulative drained water volumes.
433 3. Results and discussion
434 3.1. Adjoint local and global sensitivity analyses

435 The LASA method was first applied to assess the influence of SIDRA-RU input parameters on the
436 KGE metric. First, the local sensitivity around 'pre-calibrated prior' (Table 2) was computed (as
437 illustrated in Figure 6 with red crosses). Results show that the highest sensitivities are to soil
438 parameters µ and 𝐾𝑠𝑎𝑡 and reservoir parameters 𝑆𝑖𝑛𝑡𝑒𝑟 and 𝛼1 .
439 Second, in order to investigate the impact of the choice of prior on resulting sensitivities, a set of
440 10 input values were randomly chosen using the uniform distribution defined over the interval
441 [a,b] (Equation 13 and Table 2). The corresponding sensitivities have been computed using LASA.
442 The boxplots in Figure 6 show the range of impact of each parameter. The highest range is
443 observed for parameters 𝐾𝑠𝑎𝑡 , µ, α1 and α2 , whereas the smallest range was identified for
444 parameters S, 𝑃𝑑𝑟𝑎𝑖𝑛 and 𝛽1. One can notice that the choice of prior could change the parameter
445 ranking which emphasizes the importance of the global analysis.
446
447 Figure 6. Sensitivities of the “KGE metric” using a single 'pre-calibrated prior', in red crosses;
448 using ensemble of priors, in boxplots
449 Further analysis has therefore been performed using the GASA method over the entire range of
450 parameter variation, as defined within bounds [a,b]. Obviously, no priors are required in this
451 approach. A random ensemble of SIDRA-RU input parameters has been generated from the
452 uniform distribution. The ensemble size (n=2000) was chosen to maintain the balance between
453 stability of results and low CPU time (~2.6 s).
454 Table 2 lists the values of the boundaries (a) and (b) assigned according to both the prior
455 knowledge of SIDRA-RU model parameters and the data collected from the Chantemerle field.
456 Table 2. Minimum (a) and maximum (b) values defined for each parameter
Input S l Ksat µ a1 a2 Pdrain SSDI Sinter 𝛂1 𝛂2 β1 β2 Kcmax
(Ha) (m) (m/day) (-) (-) (-) (m) (mm) (mm) (-) (-) (-) (-) (-)
a 35.4 3.5 0.02 0.010 0.80 0.85 0.6 10.00 55.00 0.00 0.00 0.0 0.50 0.10
b 37.4 6.0 1.00 0.130 0.90 0.95 1.2 55.00 225.00 1.00 1.00 1.0 0.99 1.40
Pre-calibrated 36.4 5.0 0.27 0.042 0.86 0.89 0.9 46.48 92.58 0.33 0.07 0.4 0.90 0.86
Prior
457 The obtained ‘total effect’ sensitivities using the GASA approach are shown in Figure 7. Results
458 indicate that the saturated hydraulic conductivity 𝐾𝑠𝑎𝑡 is the most influential input parameter,
459 followed by drainage porosity µ and then the intermediate water level threshold in conceptual
460 reservoir 𝑆𝑖𝑛𝑡𝑒𝑟 , followed by the other parameters.
461 This finding means that only 𝐾𝑠𝑎𝑡 , µ and subsequent 𝑆𝑖𝑛𝑡𝑒𝑟 and 𝐾𝑐𝑚𝑎𝑥 parameters exert the
462 strongest impact on KGE values. In fact, the most influential parameters Ksat and µ, which govern
463 the physically based part of the model, describe the nervousness of the perched water table. In fact
464 𝐾𝑠𝑎𝑡 and µ are responsible for the water table dynamics. Parameter 𝑆𝑖𝑛𝑡𝑒𝑟 , ranked third, controls
465 the beginning of the intensive drainage season, which generally occurs during winter, while 𝐾𝑐𝑚𝑎𝑥
466 is dependent on the type of vegetation cover. Since information on crop rotation is not always
467 available, parameter 𝐾𝑐𝑚𝑎𝑥 is set by default equal to 0.86. Future work will investigate the
468 possibilities of integrating a crop rotation module should data be available. Lastly, although 𝑆𝑆𝐷𝐼
469 is classified as a slightly impacting parameter, it will be included in the control vector to be
470 estimated, using variational DA, due to its strong link with 𝑆𝑖𝑛𝑡𝑒𝑟 parameter (Section 2.2).
471
472 Figure 7. The influence of input parameters on the valuable function J (KGE)
473 The results obtained are consistent with the analysis performed using Sobol’s total and main-order
474 sensitivity indices in (Henine et al., Submitted). This output confirms that 𝐾𝑠𝑎𝑡 and µ have the
475 highest impact on KGE, whereas 𝑆𝑆𝐷𝐼 and 𝑆𝑖𝑛𝑡𝑒𝑟 parameters practically determine the starting
476 date of the drainage season. These four parameters constitute the sufficient control subset of
477 parameters, or the 'control vector'.
478 It should be noted that one limitation of the Sobol method is the significant computing resources
479 required (Iooss & Lemaître, 2015; X. Y. Zhang et al., 2015). This constraint accounts for the main
480 reason why the analysis was conducted on a limited input vector in (Henine et al., Submitted)
481 paper (CPU time > 10 min). Such a limitation can be overcome using the adjoint-based sensitivity
482 analysis approaches, which allow considering all SIDRA-RU inputs in a single model run.
483 3.2. SIDRA-RU parameters estimation using the variational DA method

484 Based on the results of GASA (Section 3.1), the variational DA method was implemented to
485 estimate the control vector 𝑈 = {𝐾𝑠𝑎𝑡 , µ, 𝑆𝑖𝑛𝑡𝑒𝑟 , 𝑆𝑆𝐷𝐼}. Two experimental set-ups have been
486 considered:
487 1- First, in order to assess (validate) the model's predictive performance with estimated
488 parameters, the split sample test was applied in considering the two time windows 𝑃1 ,
489 ranging from 2008 to 2010, and 𝑃2 covering the period 2010-2013.
490 2- Another experiment was performed to estimate the input parameters by assimilating
491 discharge data over the entire observation period 𝑃1 ∪ 𝑃2 from 2008 to 2013, with the
492 resulting simulated discharge being compared to that when parameters were obtained
493 according to the PAR-GR calibration method.
494 3.2.1. Validation of the variational DA framework

495 Table 3 shows the results of the split sample test. The parameter estimations were carried out by
496 minimizing the cost function given in Equation 17. One can note the good performance in terms
497 of KGE, NSE and RMSE achieved for both the assimilation and validation periods (KGE > 0.66;
498 NSE > 0.53; RMSE < 2.44 l/s). Moreover, swapping the assimilation and validation periods 𝑃1
499 and 𝑃2 leads to comparable estimates of the control vector (e.g. Smax estimated at 112 mm and 121
500 mm over periods 𝑃1 and 𝑃2 , respectively) as well as to similar performance metric values.
501 Table 3. Split simple test results based on variational assimilation of the drainage discharge data
Estimation period Validation period
Simulations Estimated parameters
P1 (2008-2010) P2 (2010-2013)
Ksat µ SSDI Sinter
RMSE NSE KGE Cost RMSE NSE KGE Cost
(m/d) (-) (mm) (mm)
P1  P2 1.62 0.56 0.66 0.74 2.44 0.71 0.77 2.50 0.19 0.042 40.8 71.2
Validation period Estimation period
Estimated parameters
P1 (2008-2010) P2 (2010-2013)
P2  P1 1.66 0.53 0.7 0.78 2.41 0.72 0.77 2.45 0.23 0.042 35.2 85.8
502 More than 100 observed values of 𝐾𝑠𝑎𝑡 and µ, corresponding to the silty soil (to be consistent with
503 the Chantemerle soil texture), have been extracted by Jeantet et al. (2021) from the drainage
504 reference sector reports (Lagacherie, 1987). The probability density functions (PDF) of the
505 saturated hydraulic conductivity (𝐾𝑠𝑎𝑡 ) and drainage porosity (µ) have been calculated using their
506 lognormal distributions, with the most frequently observed values (modes) corresponding to 0.235
507 m.day-1, 0.017 for 𝐾𝑠𝑎𝑡 and µ, respectively. The same soil texture could be found at the Arrou
508 agricultural catchment (located 50 km south of Chartres, France), where the measured value of
509 𝐾𝑠𝑎𝑡 at subsoil depth equals 0.41 m.day-1 (Zimmer et al., 1995). Therefore, the published data
510 values of 𝐾𝑠𝑎𝑡 and µ are close and consistent with the estimated values output by the variational
511 DA method.
512 Furthermore, Jamagne et al. (1977) and Goulet et al. (2004) sought to determine the water holding
513 capacity (WHC) function of certain soil parameters (e.g. water content at field capacity and wilting
514 point, apparent density). In these studies, the WHC ranged between 1 mm.cm-1 and 1.75 mm.cm-1
515 for silty soil. Based on the drain depth of the Chantemerle field (90 cm), these values correspond
516 to 90 mm and 157.5 mm, respectively. The estimates using variational DA are therefore consistent
517 with these data and allow concluding that they lie within the WHC variation interval. The estimated
518 SIDRA-RU parameters according to the variational DA are listed in Table 4 together with the
519 field-scale observed data.
520 Table 4. Estimated SIDRA-RU parameters compared with those found in the literature
Source 𝑲𝒔𝒂𝒕 (m.day-1) µ (-) 𝑺𝒎𝒂𝒙 (mm)
Variational calibration 0.19; 0.23 0.044 112; 121
(Lagacherie, 1987) 0.235 0.017 -
(Zimmer et al., 1995) 0.41 - -
Jamagne et al. (1977) & - - [90, 157]
Goulet et al. (2004)
521 Results indicate that the variational DA is a relevant methodology to improve predictive
522 performance of the SIDRA-RU model and its forthcoming upgraded versions. Since the adjoint
523 model is generated by automatic differentiation using Tapenade, the task of generating the adjoint
524 for any new SIDRA-RU model version is straightforward provided clean coding rules are being
525 respected.
526 3.2.2. Variational DA vs. Gradient-free PAR-GR method

527 Below, the variational DA method will be compared to the gradient-free PAR-GR calibration
528 method by estimating the control vector values over the entire observation period (2008-2013).
529 Three experimental configurations are considered with the following sets of input parameters:
530  (a) Non-informative inputs (priors/backgrounds) based on expertise (open-loop)
531  (b) Calibrated inputs from Michel’s calibration algorithm (PAP-GR)
532  (c) Estimated inputs from the variational DA method, taking the background values from (a)
533 Table 5 shows the values of input parameters obtained by configurations (a), (b) and (c), along
534 with the corresponding performance metrics. Note that the improvement achieved by variational
535 DA compared to the PAP-GR calibration algorithm is indicated in green brackets.
536 Table 5. Performance criteria corresponding to the application of two minimization methods
Control vector values Performance criteria
Ksat µ SSDI Sinter RMSE NSE KGE
a 0.9 0.06 18.0 30.0 4.66 -0.46 0.08
b 0.272 0.043 46.48 92.58 2.13 0.70 0.76
c 0.228 0.044 41.93 84.84 2.07 (-0.06) 0.71 (+0.01) 0.78 (+0.02)
537 In the experimental set-up (c), the variational DA is used starting from a non-informative
538 background from (a) in order to estimate input parameters 𝐾𝑠𝑎𝑡 , µ, 𝑆𝑖𝑛𝑡𝑒𝑟 𝑎𝑛𝑑 𝑆𝑆𝐷𝐼
539 simultaneously. Since the bounds of the physical parameter range have not been integrated into
540 SIDRA-RU, a constrained cost minimization is performed using the LBFGS-B method to
541 eliminate all non-physical solutions. Table 6 summarizes the lower and upper bounds assigned to
542 each parameter. The choice of bounds is based on the type of soil (i.e. silty texture). The estimated
543 control vector (c) is comparable to that derived from the split sample test in the first experiment
544 (Table 3). This result confirms the temporal robustness of the SIDRA-RU model.
545 Table 6. Assigned lower and upper bound values of each calibrated input parameter
Parameter Ksat (m/day) µ (-) SSDI (mm) Sinter (mm)
Lower bound 0.03 0.02 10.0 55.0
Upper bound 1.5 0.07 60.0 225.0
546 Figure 8 illustrates the observed (reference) and simulated discharges, over the total target period,
547 according to the three experimental configurations (a, b, c). In addition, the respective cumulative
548 water volume has been plotted in (d), thus making it possible to compare the performance obtained
549 using the variational calibration method and that resulting from application of the PAP-GR
550 calibration algorithm.
551 The model performance after variational DA is slightly better in terms of KGE, NSE and RMSE,
552 as compared to the PAP-GR calibration (Michel, 1989). Excluding the first year of data, which is
553 generally used as a warming period to adjust the saturation degree in the conceptual reservoir, the
554 discharge values are well represented in both cases (Figure 8b and 8c). However, when analyzing
555 the simulated cumulative drained water over the total target period (five drainage seasons) in
556 comparison with the measured reference discharge, Figure 8d shows only a 1% ratio between
557 measured and simulated total drained water volume using variational DA (∆𝑉 = 4.8 mm over five
558 years of data) versus 10% when using the PAP-GR calibration algorithm (∆𝑉 = 58.7 mm). In
559 addition, it is worth mentioning that a recent drainage modeling study has demonstrated the benefit
560 of using a multi-objective minimization approach to improve the estimation of HYDRUS
561 parameters. Yet a significant dispersion of 25% was observed between the cumulative measured
562 and simulated water volumes, which corresponds to 50 mm in just one drainage season (Turunen
563 et al., 2020). This result confirms the strong ability of variational DA to simulate accurately the
564 volume of water exported from drained agricultural fields.
565 In conclusion, the variational DA method implemented with the SIDRA-RU model has been
566 successfully validated in Section 3.2.1 using the split sample test. It allows for a robust estimation
567 of model parameters consistent with the in situ values found in the literature. Moreover, the
568 estimation of SIDRA-RU parameters using variational DA instead of the PAP-GR calibration
569 approach improves discharge simulations as well as cumulative water volume predictions.
570
571 Figure 8. Graphical comparison between experimental configurations (a), (b) and (c);
572 (d) shows the corresponding cumulative drainage discharge
573 3.3. Investigating the local minima issue

574 The second part of this study focuses on the issue of local minima, which is commonly encountered
575 when solving ill-posed problems in variational formulations involving a gradient-based
576 minimization. Two experimental set-ups have been considered. First, we analyzed the convergence
577 process for a given single parameter in assuming that the remaining parameters were known and
578 set to their optimal values. Second, we investigated the convergence process when all parameters
579 were being updated simultaneously.
580 - Single parameter estimation
581 In this experiment, each parameter 𝑈𝑖 was estimated separately while assuming that all other inputs
582 were known and set to their optimal values (Table 5, Row C). Hence, the control vector was limited
583 to one parameter at a time. For each parameter 𝑈𝑖 , an ensemble of priors/backgrounds of size
584 n=100 was generated and an ensemble estimation using the variational DA was performed. Figure
585 9 illustrates the convergence process for each parameter, starting from different initial values
586 within the given ensemble. The red color represents successful minimization cases, where a cost
587 function minimization has been reached. The corresponding ratios are presented in the legend,
588 together with the unsuccessful cases shown in gray.
589
590 Figure 9. Convergence trajectories from prior/background values towards the optimal values
591 during the single parameter estimation process
592 It can be noticed that the ensembles of parameters 𝐾𝑠𝑎𝑡 and µ converge towards unique values of
593 0.22 m.d-1 and 0.047, respectively, regardless of the initial background value. These two
594 parameters contribute to the physical SIDRA-RU module throughout the drainage season.
595 Consequently, convergence towards a global minimum can be explained by the strong consistency
596 of the physically based SIDRA-RU module. The convergence pattern for parameters 𝑆𝑆𝐷𝐼 and
597 𝑆𝑖𝑛𝑡𝑒𝑟 differs: dependent on the priors, the minimization process converges to different cost
598 function values, which underscores the presence of local minima or the equifinality issue. In fact,
599 these two parameters manage the conceptual part of the SIDRA-RU model; their contribution is
600 limited to a short period, essentially during the first few days of each drainage season. Our results
601 suggest that their values may require a dynamic year update during the minimization process.
602 Additional information on the reservoir capacity or beginning of the drainage season may be
603 needed to constrain further the minimization protocol as regards these reservoir parameters. The
604 optimal values of 𝑆𝑆𝐷𝐼 and 𝑆𝑖𝑛𝑡𝑒𝑟 , corresponding to the minimum cost function, are 42 mm and
605 82 mm, respectively, which are still consistent with the result obtained in Section 3.2.
606 - Simultaneous minimization
607 Similarly to the first set-up, this second experiment an ensemble of size n=100 of prior/background
608 vectors was generated. The simultaneous minimization of all parameters was then performed using
609 the variational DA method. Figure 10 shows the parameter sets that converges toward minimum
610 cost function values, as depicted here in red. Four main clusters (red dots) have been identified.
611 Similar cluster positions for the two physically based parameters 𝐾𝑠𝑎𝑡 [0.19-0.25] m.d-1 and µ
612 [0.040-0.047] can be detected.
613
614 Figure 10. Final cost values according to their corresponding parameter values
615 Note that the existence of a potential relationship for different soil textures between 𝐾𝑠𝑎𝑡 and µ, as
616 previously highlighted by Chossat and Saugnac (1985). Interestingly, the conceptual-based
617 parameters 𝑆𝑆𝐷𝐼 [47-61] mm and 𝑆𝑖𝑛𝑡𝑒𝑟 [65-81] mm display opposite cluster positions. This
618 finding suggests that these two groups of parameters are complementary and compensate for one
619 another, which typically leads to the occurrence of local minima and equifinality issues (nearly the
620 same prediction obtained with different parameter values).
621
622 Figure 11. Convergence trajectories from prior/background values towards optimal values
623 (case of simultaneous minimization)
624 Similarly to the first experiment, Figure 11 shows the minimization trajectories. The red color
625 refers to successful convergence cases reaching the minimum cost function value. Dashed lines
626 correspond to the optimal solution of each parameter in terms of the minimized cost function,
627 corresponding to: 𝐾𝑠𝑎𝑡 = 0.22 m/day, µ= 0.044 (-), 𝑆𝑆𝐷𝐼 = 48.68 mm, and 𝑆𝑖𝑛𝑡𝑒𝑟 = 78.76 mm.
628 It can be observed that several solutions are possible when estimating all parameters and no global
629 minimum can be clearly identified. The best control vector estimates, which correspond to the
630 minimum cost function value, are noted in Figure 11 (dashed black lines). The equifinality issue
631 can be mitigated when assimilating additional information by the variational DA, e.g. water level
632 variations ℎ(𝑡). We suggest herein that complementary information (namely nitrate concentrations
633 [NO3-] and fluxes) will additionally constrain the solution in the upgraded SIDRA-RU version.
634 The featured methods have been implemented in order to improve both the understanding and
635 estimation of the SIDRA-RU input parameters. The present work can be extended to other drainage
636 models as long as their adjoint is able to be generated.
637 4. Conclusions
638 Our primary objective has been to evaluate the influence of input parameters on the SIDRA-RU
639 model output. This step was accomplished using both local and global adjoint sensitivity analyses.
640 The identified parameter sets were then corrected using the variational data assimilation method.
641 The assimilated discharge data were collected from the 36-ha Chantemerle agricultural field for
642 the period between 2008 and 2013.
643 In addition to the local adjoint SA, the new global adjoint SA method was investigated on the
644 SIDRA-RU model. Based on an estimation of the upper bound of the “total effect” Sobol indices,
645 the GASA method showed consistent results compared to the classical Sobol method (applied to
646 the same model). The results of this analysis demonstrate that the two most significant parameters
647 are 𝐾𝑠𝑎𝑡 and µ. The GASA method has proven to be capable of estimating the impact of all input
648 parameters at a very low computational time (≈ 2.6 s). Out of all SIDRA-RU parameters, four were
649 preselected for the calibration step.
650 The variational DA method has been successfully implemented on the conceptual-analytical model
651 SIDRA-RU. This method was validated using the ‘split sample test’ and compared to Michel’s
652 step-by-step calibration method (PAP-GR). A slightly better performance could be noted with the
653 variational DA method, which allowed estimating the best parameter set and reproducing a
654 satisfactory daily drainage discharge simulation (KGE = 0.76 for PAP-GR vs. 0.78 for VAR-DA).
655 Another advantage of the variational DA method was demonstrated by comparing total cumulative
656 drainage volumes. The parameters estimated using this method yielded a cumulative volume close
657 to the observed volume at the field outlet (not exceeding 4.8 mm and less than 1% discrepancy)
658 when compared to values obtained with PAP-GR (58.7 mm, for a 10% error). Moreover, the
659 estimated parameter values are consistent with those found in the literature under similar soil
660 conditions (i.e. texture, climate). A planned forthcoming work will assimilate the nitrate
661 concentration observations using the upgraded version of SIDRA-RU, which will integrate a newly
662 developed nitrate transport model. We expect that the advantages of the implemented approach
663 will be particularly beneficial for the coupled drainage/nitrate model, since it requires yearly
664 estimations of a given input variable.
665 Lastly, the local minima issue has been explored. Our results here are in line with findings reported
666 in the literature. It has been confirmed that using gradient-based minimization may lead to
667 suboptimal solutions. This issue has been resolved by running an ensemble of minimizations,
668 starting from randomly generated priors (uniformly distributed on the interval [a,b]). In this
669 manner, it becomes possible to determine the globally optimal solution without needing the
670 prior/background.
671 Acknowledgments
672 This research was supported by a phD Grant from INRAE water department. The authors would
673 like to thank Irina Ginzburg for giving precious comments to improve this paper. We also thank
674 farmers of the Chantemerle parcels. Data archiving is underway. An uploaded within OZCAR
675 network is scheduled (ORACLE https://gisoracle.inrae.fr/).
676 References
677 Alzraiee, A. H., Gates, T. K., & Garcia, L. A. (2013). Modeling Subsurface Heterogeneity of
678 Irrigated and Drained Fields. II: Multivariate Stochastic Analysis of Root-Zone
679 Hydrosalinity and Crop Yield. Journal of Irrigation and Drainage Engineering, 139(10),
680 809-820. doi:doi:10.1061/(ASCE)IR.1943-4774.0000587
681 Arsenault, R., Poulin, A., Côté, P., & Brissette, F. (2014). Comparison of stochastic optimization
682 algorithms in hydrological model calibration. Journal of Hydrologic Engineering, 19(7),
683 1374-1384.
684 Augeard, B., Kao, C., Chaumont, C., & Vauclin, M. (2005). Mechanisms of surface runoff genesis
685 on a subsurface drained soil affected by surface crusting: A field investigation. Physics and
686 Chemistry of the Earth, Parts A/B/C, 30(8), 598-610.
687 doi:https://doi.org/10.1016/j.pce.2005.07.014
688 Bouarfa, S. (1995). Drainage of irrigated schemes Taking into account the evaporation in
689 drainage modeling. Université de Strasbourg. Retrieved from https://tel.archives-
690 ouvertes.fr/tel-00685486 Cirad database.
691 Bouarfa, S., & Zimmer, D. (2000). Water-table shapes and drain flow rates in shallow drainage
692 systems. Journal of Hydrology, 235(3-4), 264-275.
693 Boussinesq, J. (1904). Recherches théoriques sur l'écoulement des nappes d'eau infiltrées dans le
694 sol et sur le débit des sources. Journal de mathématiques pures et appliquées, 10, 5-78.
695 Branger, F., Tournebize, J., Carluer, N., Kao, C., Braud, I., & Vauclin, M. (2009). A simplified
696 modelling approach for pesticide transport in a tile-drained field: The PESTDRAIN model.
697 Agricultural Water Management, 96(3), 415-428.
698 doi:https://doi.org/10.1016/j.agwat.2008.09.005
699 Cacuci, D. (1981). Sensitivity Theory for Nonlinear Systems. I. Nonlinear Functional Analysis
700 Approach. Journal of Mathematical Physics, 22, 2794-2802. doi:10.1063/1.525186
701 Chossat, J. C., & Saugnac, A. M. (1985). Relation entre conductivité hydraulique et porosité de
702 drainage mesurées par la méthode du puits et des piézomètres. Sciences du sol(3), 151-167.
703 Coron, L., Thirel, G., Delaigue, O., Perrin, C., & Andréassian, V. (2017). The suite of lumped GR
704 hydrological models in an R package. Environmental Modelling & Software, 94, 166-171.
705 Courtier, P., Andersson, E., Heckley, W., Vasiljevic, D., Hamrud, M., Hollingsworth, A., . . .
706 Pailleux, J. (1998). The ECMWF implementation of three‐dimensional variational
707 assimilation (3D‐Var). I: Formulation. Quarterly Journal of the Royal Meteorological
708 Society, 124(550), 1783-1807.
709 Dennis, J., John E, & Moré, J. J. (1977). Quasi-Newton methods, motivation and theory. SIAM
710 review, 19(1), 46-89.
711 Duan, Q., Sorooshian, S., & Gupta, V. (1992). Effective and efficient global optimization for
712 conceptual rainfall-runoff models. Water Resources Research, 28(4), 1015-1031.
713 doi:https://doi.org/10.1029/91WR02985
714 Fischer, C., Montmerle, T., Berre, L., Auger, L., & Ştefănescu, S. E. (2005). An overview of the
715 variational assimilation in the ALADIN/France numerical weather‐prediction system.
716 Quarterly Journal of the Royal Meteorological Society: A Journal of the atmospheric
717 sciences, applied meteorology and physical oceanography, 131(613), 3477-3492.
718 G. Arnold, J., N. Moriasi, D., W. Gassman, P., C. Abbaspour, K., J. White, M., Srinivasan, R., . .
719 . K. Jha, M. (2012). SWAT: Model Use, Calibration, and Validation. Transactions of the
720 ASABE, 55(4), 1491-1508. doi:https://doi.org/10.13031/2013.42256
721 Gauthier, P., Charette, C., Fillion, L., Koclas, P., & Laroche, S. (1999). Implementation of a 3D
722 variational data assimilation system at the Canadian Meteorological Centre. Part I: The
723 global analysis. Atmosphere-Ocean, 37(2), 103-156. doi:10.1080/07055900.1999.9649623
724 Gauthier, P., Tanguay, M., Laroche, S., Pellerin, S., & Morneau, J. (2007). Extension of 3DVAR
725 to 4DVAR: Implementation of 4DVAR at the Meteorological Service of Canada. Monthly
726 Weather Review, 135(6), 2339-2354. doi:10.1175/mwr3394.1
727 Gejadze, I., Malaterre, P. O., & Shutyaev, V. (2019). On the use of derivatives in the polynomial
728 chaos based global sensitivity and uncertainty analysis applied to the distributed parameter
729 models. Journal of Computational Physics, 381, 218-245.
730 doi:https://doi.org/10.1016/j.jcp.2018.12.023
731 Ghorbanidehno, H., Kokkinaki, A., Lee, J., & Darve, E. (2020). Recent developments in fast and
732 scalable inverse modeling and data assimilation methods in hydrology. Journal of
733 Hydrology, 591, 125266. doi:https://doi.org/10.1016/j.jhydrol.2020.125266
734 Goulet, E., Morlat, R., Rioux, D., & Cesbron, S. (2004). A calculation method of available soil
735 water content : application to viticultural terroirs mapping of the Loire valley. OENO One,
736 38(4), 231-235. doi:10.20870/oeno-one.2004.38.4.913
737 Goutal, N., Goeury, C., Ata, R., Ricci, S., Nabil, E. M., Rochoux, M., . . . Malaterre, P.-O. (2018).
738 Uncertainty Quantification for River Flow Simulation Applied to a Real Test Case: The
739 Garonne Valley (pp. 169-187).
740 Gowda, P., Mulla, D., Desmond, E., Ward, A., & Moriasi, D. (2012). ADAPT: Model use,
741 calibration and validation. Transactions of the ASABE (American Society of Agricultural
742 and Biological Engineers), 55, 1345-1352. doi:10.13031/2013.42246
743 Guo, D., Zheng, F., Gupta, H., & Maier, H. R. (2020). On the Robustness of Conceptual Rainfall-
744 Runoff Models to Calibration and Evaluation Data Set Splits Selection: A Large Sample
745 Investigation. Water Resources Research, 56(3), e2019WR026752.
747 Gupta, H. V., Kling, H., Yilmaz, K. K., & Martinez, G. F. (2009). Decomposition of the mean
748 squared error and NSE performance criteria: Implications for improving hydrological
749 modelling. Journal of Hydrology, 377(1), 80-91.
750 doi:https://doi.org/10.1016/j.jhydrol.2009.08.003
751 Gurovich, L., & Oyarce, P. (2015). Modeling agricultural drainage hydraulic nets. Irrig Drain Syst
752 Eng, 4, 149-158.
753 Haas, M. B., Guse, B., Pfannerstill, M., & Fohrer, N. (2016). A joined multi-metric calibration of
754 river discharge and nitrate loads with different performance measures. Journal of
755 Hydrology, 536, 534-545. doi:https://doi.org/10.1016/j.jhydrol.2016.03.001
756 Hascoet, L., & Pascual, V. (2013). The Tapenade Automatic Differentiation tool: principles,
757 model, and specification. ACM Transactions on Mathematical Software, 39(3).
758 doi:10.1145/2450153.2450158
759 Henine, H., Jeantet, A., Chaumont, C., Chelil, S., Lauvernet, C., & Tournebize, J. (Submitted).
760 Coupling of a subsurface drainage model with a soil reservoir model to simulate drainage
761 discharge and drain flow start. Agricultural Water Management.
762 Henriksen, H. J., Troldborg, L., Nyegaard, P., Sonnenborg, T. O., Refsgaard, J. C., & Madsen, B.
763 (2003). Methodology for construction, calibration and validation of a national hydrological
764 model for Denmark. Journal of Hydrology, 280(1), 52-71.
765 doi:https://doi.org/10.1016/S0022-1694(03)00186-0
766 Hooghoudt, S. (1940). General consideration of the problem of field drainage by parallel drains,
767 ditches, watercourses, and channels. Contribution to the knowledge of some physical
768 parameters of the soil, 7.
769 Iooss, B., & Lemaître, P. (2015). A review on global sensitivity analysis methods Uncertainty
770 management in simulation-optimization of complex systems (pp. 101-122): Springer.
771 Jamagne, M., Betremieux, R., Begon, J. C., & Mori, A. (1977). Quelques données sur la variabilité
772 dans le milieu naturel de la réserve en eau des sols. Bulletin Technique d'Information, 324-
773 325, 627-641.
774 Jay-Allemand, M., Javelle, P., Gejadze, I., Arnaud, P., Malaterre, P. O., Fine, J. A., & Organde,
775 D. (2020). On the potential of variational calibration for a fully distributed hydrological
776 model: application on a Mediterranean catchment. Hydrol. Earth Syst. Sci., 24(11), 5519-
777 5538. doi:10.5194/hess-24-5519-2020
778 Jeantet, Henine, H., Chaumont, C., Collet, L., Thirel, G., & Tournebize, J. (2021). Robustness of
779 a parsimonious subsurface drainage model at the French national scale. Hydrol. Earth Syst.
780 Sci. Discuss., 2021, 1-33. doi:10.5194/hess-2021-168
781 Kenney, J. F. (1939). Mathematics of statistics: D. Van Nostrand.
782 KlemeŠ, V. (1986). Operational testing of hydrological simulation models. Hydrological Sciences
783 Journal, 31(1), 13-24. doi:10.1080/02626668609491024
784 Kucherenko, S. (2009). Derivative based global sensitivity measures and their link with global
785 sensitivity indices. Mathematics and Computers in Simulation, 79(10), 3009-3017.
786 Lagacherie, P. (1987). Synthèse générale sur les études de secteur de référence drainage.
787 Retrieved from https://hal.inrae.fr/hal-02603572
788 Lamboni, M., Iooss, B., Popelin, A.-L., & Gamboa, F. (2013). Derivative-based global sensitivity
789 measures: general links with Sobol’indices and numerical tests. Mathematics and
790 Computers in Simulation, 87, 45-54.
791 Lesaffre, B. (1988). Fonctionnement hydrologique et hydraulique du drainage souterrain des sols
792 temporairement engorgés : débits de pointe et modèle SIDRA : extension des principes
793 théoriques de Boussinesq et Guyon. Available from http://www.theses.fr/1988PA066362
794 Lindström, G., Pers, C., Rosberg, J., Strömqvist, J., & Arheimer, B. (2010). Development and
795 testing of the HYPE (Hydrological Predictions for the Environment) water quality model
796 for different spatial scales. Hydrology research, 41(3-4), 295-319.
797 doi:10.2166/nh.2010.007
798 Mathevet, T. (2005). Quels modèles pluie-débit globaux pour le pas de temps horaire?
799 Développement empirique et comparaison de modèles sur un large échantillon de bassins
800 versants. ENGREF (Paris): Cemagref (Antony), France, 463p.
801 Michel, C. (1989). Hydrologie appliquée aux petits bassins versants ruraux. Cemagref, antony.
802 Migliaccio, K., & Chaubey, I. (2005). Sensitivity Analysis, Calibration, and Validations for a
803 Multisite and Multivariable SWAT Model. JAWRA Journal of the American Water
804 Resources Association, 41, 1077-1089. doi:10.1111/j.1752-1688.2005.tb03786.x
805 Mohtadullah, K. (1990). Interdisciplinary planning, data needs and evaluation for drainage
806 projects. Paper presented at the Land drainage - Drainage agricole : actes du 4ème
807 séminaire international sur le drainage, Le Caire, EGY, 23-24 février 1990, Antony.
808 http://irsteadoc.irstea.fr/cemoa/PUB00004103
809 Nash, J. E., & Sutcliffe, J. V. (1970). River flow forecasting through conceptual models part I—
810 A discussion of principles. Journal of Hydrology, 10(3), 282-290.
811 Oubanas, H., Gejadze, I., Malaterre, P.-O., & Mercier, F. (2018). River discharge estimation from
812 synthetic SWOT-type observations using variational data assimilation and the full Saint-
813 Venant hydraulic model. Journal of Hydrology, 559, 638-647.
814 Pan, L., & Wu, L. (1998). A hybrid global optimization method for inverse estimation of hydraulic
815 parameters: Annealing‐simplex method. Water Resources Research, 34(9), 2261-2269.
817 Patil, S. D., & Stieglitz, M. (2015). Comparing spatial and temporal transferability of hydrological
818 model parameters. Journal of Hydrology, 525, 409-417.
819 doi:https://doi.org/10.1016/j.jhydrol.2015.04.003
820 Pechlivanidis, I., Jackson, B., Mcintyre, N., & Wheater, H. (2011). Catchment scale hydrological
821 modelling: a review of model types, calibration approaches and uncertainty analysis
822 methods in the context of recent developments in technology and applications. Global
823 NEST journal, 13(3), 193-214.
824 Rabier, F., Järvinen, H., Klinker, E., Mahfouf, J.-F., & Simmons, A. (2000). The ECMWF
825 operational implementation of four-dimensional variational assimilation. I: Experimental
826 results with simplified physics. Quarterly Journal of the Royal Meteorological Society,
827 126(564), 1143-1170. doi:https://doi.org/10.1002/qj.49712656415
828 Reichle, R. H. (2008). Data assimilation methods in the Earth sciences. Advances in Water
829 Resources, 31(11), 1411-1418. doi:https://doi.org/10.1016/j.advwatres.2008.01.001
830 Saltelli, A. (2002). Making best use of model evaluations to compute sensitivity indices. Computer
831 physics communications, 145(2), 280-297.
832 Santos, L. (2018). Perspectives for hydrological supermodels: evaluation of a method based on a
833 dynamical combination of rainfall-runoff models. Doctorat en Hydrologie, AgroParisTech.
834 Retrieved from https://hal.inrae.fr/tel-02609262 Inrae database.
835 Shoarinezhad, V., Wieprecht, S., & Haun, S. (2020). Comparison of Local and Global
836 Optimization Methods for Calibration of a 3D Morphodynamic Model of a Curved
837 Channel. Water, 12(5), 1333.
838 Skaggs, R. W., Youssef, M., & Chescheir, G. M. (2012). DRAINMOD: model use, calibration,
839 and validation. Transactions of the ASABE, 55, 1509-1522. doi:10.13031/2013.42259
840 Skahill, B. E., & Doherty, J. (2006). Efficient accommodation of local minima in watershed model
841 calibration. Journal of Hydrology, 329(1-2), 122-139.
842 Sobol, I. M. (1993). Sensitivity estimates for nonlinear mathematical models. Mathematical
843 modelling and computational experiments, 1(4), 407-414.
844 Stange, F., Butterbach-Bahl, K., Papen, H., Zechmeister-Boltenstern, S., Li, C., & Aber, J. (2000).
845 A process-oriented model of N2O and NO emissions from forest soils: 2. Sensitivity
846 analysis and validation. Journal of Geophysical Research: Atmospheres, 105(D4), 4385-
847 4398. doi:10.1029/1999jd900948
848 Tournebize, J., Chaumont, C., Fesneau, C., Guenne, A., Vincent, B., Garnier, J., & Mander, Ü.
849 (2015). Long-term nitrate removal in a buffering pond-reservoir system receiving water
850 from an agricultural drained catchment. Ecological Engineering, 80, 32-45.
851 Tournebize, J., Kao, C., Nikolic, N., & Zimmer, D. (2004). Adaptation of the STICS model to
852 subsurface drained soils. http://dx.doi.org/10.1051/agro:2004030, 24.
853 doi:10.1051/agro:2004030
854 Turunen, M., Gurarslan, G., Šimůnek, J., & Koivusalo, H. (2020). What is the worth of drain
855 discharge and surface runoff data in hydrological simulations? Journal of Hydrology,
856 125030.
857 Valipour, M. (2012). Effect of Drainage Parameters Change on Amount of Drain Discharge in
858 Subsurface Drainage Systems. IOSR Journal of Agriculture and Veterinary Science (IOSR-
859 JAVS), 1, 10-18. doi:10.9790/2380-0141018
860 Welter, D. E., Doherty, J. E., Hunt, R. J., Muffels, C. T., Tonkin, M. J., & Schreuder, W. A. (2012).
861 Approaches in highly parameterized inversion - PEST++, a Parameter ESTimation code
862 optimized for large environmental models (7-C5). Retrieved from Reston, VA:
863 http://pubs.er.usgs.gov/publication/tm7C5
864 Zhang, X., Srinivasan, R., Zhao, K., & Liew, M. V. (2009). Evaluation of global optimization
865 algorithms for parameter calibration of a computationally intensive hydrologic model.
866 Hydrological Processes, 23(3), 430-441. doi:10.1002/hyp.7152
867 Zhang, X. Y., Trame, M. N., Lesko, L. J., & Schmidt, S. (2015). Sobol Sensitivity Analysis: A
868 Tool to Guide the Development and Evaluation of Systems Pharmacology Models. CPT:
869 pharmacometrics & systems pharmacology, 4(2), 69-79. doi:10.1002/psp4.6
870 Zhu, C., Byrd, R. H., Lu, P., & Nocedal, J. (1997). Algorithm 778: L-BFGS-B: Fortran subroutines
871 for large-scale bound-constrained optimization. ACM Transactions on Mathematical
872 Software (TOMS), 23(4), 550-560.
873 Zimmer, D., Lorre, E., & Lesaffre, B. (1995). Parameter sensitivity and field evaluation of SIDRA
874 model. Irrigation and Drainage Systems, 9(3), 279-296. doi:10.1007/BF00880868

Variational Data Assimilation To Improve Subsurface Drainage Model Parameters

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Variational Data Assimilation To Improve Subsurface Drainage Model Parameters

Uploaded by

Copyright:

Available Formats

Variational Data Assimilation to improve subsurface drainage

November 24, 2022

1 Variational Data Assimilation to improve subsurface drainage model parameters

7 Corresponding author: Samy CHELIL (samy.chelil@inrae.fr)

155 2. Materials and Methods

156 2.1. Study area description and data

183 2.2. The SIDRA-RU model

200 Table 1. Classification of SIDRA-RU parameters

Water table shape Coefficients (a1 and a2) Internal

𝑃𝑛𝑒𝑡 (𝑡) = 𝑃(𝑡) − 𝐶𝐸𝑇(𝑡)

𝑅(𝑡) = 0 𝑅(𝑡) = 𝛼 ∗ 𝑃𝑛𝑒𝑡 (𝑡) 𝑅(𝑡) = 𝑆(𝑡 − 1) − 𝑃𝑛𝑒𝑡 (𝑡) − 𝑆𝑚𝑎𝑥

Pnet< 0 𝑅(𝑡) = 0 & 𝑆(𝑡) = 𝑆(𝑡 − 1) + 𝑃𝑛𝑒𝑡 (𝑡)

255 2.3. Adjoint sensitivity analysis

265 2.3.1. Local adjoint sensitivity analysis (LASA)

309 2.3.2. Global adjoint sensitivity analysis (GASA)

359 Figure 4. Simplified version of the global adjoint SA algorithm

360 2.4. Variational data assimilation algorithm

399 2.4.1. Toward avoiding the local minima issues

413 2.5. Evaluation of the temporal robustness of the model

433 3. Results and discussion

434 3.1. Adjoint local and global sensitivity analyses

483 3.2. SIDRA-RU parameters estimation using the variational DA method

494 3.2.1. Validation of the variational DA framework

526 3.2.2. Variational DA vs. Gradient-free PAR-GR method

573 3.3. Investigating the local minima issue

You might also like