Journal of Hydrology 591 (2020) 125331

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Research papers

Estimation of multiple inflows and effective channel by assimilation of T

multi-satellite hydraulic signatures: The ungauged anabranching Negro river
L. Pujola, , P-A. Garamboisa,b,c, P. Finaud-Guyota,d,e,f, J. Monnierg,h, K. Larnierg,h,i, R. Moséa,f,
⁎

S. Biancamariaj,k, H. Yesoua,n, D. Moreiral, A. Parism, S. Calmantj,k

a
Laboratoire des Sciences de l’ingenieur, de l’informatique et de l’imagerie (ICUBE), Fluid Mechanics Team, CNRS, Universite de Strasbourg, France
b
Now at: INRAE (Irstea), Aix Marseille Univ, RECOVER, Aix-en-Provence, France
c
INSA Strasbourg, Strasbourg, France
d
Now at: HSM, Univ Montpellier, CNRS, IRD, Montpellier, France
e
Now at: LEMON, INRIA, Montpellier, France
f
ENGEES, Strasbourg, France
g
Institut de Mathematiques de Toulouse (IMT), France
h
INSA Toulouse, France
i
CS Corporation, Business Unit Espace, Toulouse, France
j
Laboratoire D’ Etudes En Geophysique et Oceanographie Spatiales (LEGOS, UMR 5566 CNES CNRS IRD UPS), France
k
Universite de Toulouse III Paul Sabatier, OMP, Toulouse, France
l
UFRJ/CPRM, Av. Pasteur 404, 22290-040 Rio de Janeiro, Brazil
m
Collecte Localisation Satellite (CLS), Toulouse, France
n
SErvice Régional de Traitement d’Image et de Télédétection (SERTIT), Université de Strasbourg, Pole API, Illkirch, France

ARTICLE INFO ABSTRACT

Keywords: With the upcoming SWOT satellite mission, which should provide spatially dense river surface elevation, width
1D Saint–Venant model and slope observations globally, comes the opportunity to assimilate such data into hydrodynamic models, from
Hydrology couplings the reach scale to the hydrographic network scale. Based on the HiVDI (Hierarchical Variational Discharge
Variational assimilation Inversion) modeling strategy (Larnier et al. (2020)), this study tackles the forward and inverse modeling cap­
Satellite altimetry
abilities of distributed channel parameters and multiple inflows (in the 1D Saint–Venant model) from multi­
SWOT
satellite observations of river surface. It is shown on synthetic cases that the estimation of both inflows and
Hydraulic visibility
Ungauged river distributed channel parameters (bathymetry-friction) is achievable with a minimum spatial observability be­
tween inflows as long as their hydraulic signature is sampled. Next, a real case is studied: 871 km of the Negro
river (Amazon basin) including complex multichannel reaches, 21 tributaries and backwater controls from major
confluences. An effective modeling approach is proposed using (i) WS elevations from ENVISAT data and dense
in situ GPS flow lines (Moreira (2016)), (ii) average river top widths from optical imagery (Pekel et al. (2016)),
(iii) upstream and lateral flows from the MGB large-scale hydrological model (Paiva et al. (2013a)). The cali­
brated effective hydraulic model closely fits satellite altimetry observations and presents real-like spatial vari­
abilities; flood wave propagation and water surface observation frequential features are analyzed with iden­
tifiability maps following (Brisset et al. (2018)). Synthetic SWOT observations are generated from the simulated
flowlines and allow to infer model parameters (436 effective bathymetry points, 17 friction patches and 22
upstream and lateral hydrographs) given hydraulically coherent prior parameter values. Inferences of channel
parameters carried out on this fine hydraulic model applied at a large scale give satisfying results using noisy
SWOT-like data at reach scale. Inferences of spatially distributed temporal parameters (lateral inflows) give
satisfying results as well, with even relatively small scale hydrograph variations being inferred accurately on this
long reach. This study brings insights in: (i) the hydraulic visibility of the signatures of multiple inflow hy­
drographs at a large scale with SWOT; (ii) the simultaneous identifiability of spatially distributed channel
parameters and inflows by assimilation of satellite altimetry data; (iii) the need for prior information; (iv) the
need to further tailor and scale network hydrodynamic models and assimilation methods to improve the fusion
of multisource information and potential information feedback to hydrological modules in integrated chains.

⁎
Corresponding author.
E-mail address: [email protected] (L. Pujol).

https://doi.org/10.1016/j.jhydrol.2020.125331
Received 8 April 2020; Received in revised form 15 July 2020; Accepted 19 July 2020
Available online 09 August 2020
0022-1694/ © 2020 Elsevier B.V. All rights reserved.
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

1. Introduction (2017) and Larnier et al. (2020) for the present inverse flow problem.
Crucial aspects of this difficult inverse problem are (i) the spatio-
Hydrographic networks represent major flowpaths for freshwater in temporal sparsity of altimetric observations regarding flow controls – as
the water cycle and an interface with the space of human societies. It is analyzed in Brisset et al. (2018) for inferable hydrographs frequencies
of prior importance in a context of climate change to improve the with the introduction of identifiability maps and in Garambois et al.
knowledge and representation of continental water fluxes, including (2020) for inferable channel parameters patterns; (ii) the sensitivity of
river discharge, defined as an essential physical variable (see Global the triplet inference to good prior guesses on the sought parameters as
Climate Observing System (World Meteorological Organization, 2011; highlighted in a SWOT context by Garambois and Monnier (2015),
World Meteorological Organization, 1997)). However, modeling flows Yoon et al. (2016), Larnier et al. (2020), Tuozzolo et al. (2019) and
structure in the different compartments of a catchment remains a hard Garambois et al. (2020). The latest is highlighted by recent discharge
task (see (Schuite et al., 2019) and references therein) especially at estimates (in a triplet setup) from synthetic SWOT data on the Pô,
poorly gauged locations. In complement of in situ sensors networks, Garonne and Sacramento Rivers in Larnier et al. (2020) (see also
which are declining in several regions (e.g. (Fekete and Vorosmarty, (Oubanas et al., 2018b)), from AirSWOT airborne measurements on the
2002)), new generations of earth observation satellites and sensors Willamette River in Tuozzolo et al. (2019) or from ENVISAT altimetric
provide increasingly accurate and dense measurements of water surface data on an anabranching portion of the Xingu River (Garambois et al.,
variabilities. 2020). Using a biased prior hydrograph results in a biased estimate of
The Surface Water and Ocean Topography (SWOT) satellite, to be inflow hydrograph despite a correct temporal variability at observation
launched in 2021, will bring observations of water surface (WS) with an times – see Larnier et al. (2020) for detailed analysis. A hierarchical
unprecedented spatio-temporal coverage (Alsdorf et al., 2007; Durand modeling strategy HiVDI (Hierarchical Variational Discharge Inversion)
et al., 2010; Rodríguez, 2012; Biancamaria et al., 2016; Rodríguez is proposed in Larnier et al. (2020) including low complexity flow re­
et al., 2018). This will yield greater hydraulic visibility (see definition lations (Low Froude and locally steady-state) for providing robust prior
in (Garambois et al., 2017; Montazem et al., 2019; Rodríguez et al., guesses to the VDA process by taking advantage of databases or re­
2020)) of hydrological responses through WS signatures from the local gional hydrological models.
scale to the hydrographic network scale, hence an opportunity to better Most studies mentioned above tackle the estimation of a single
characterize hydrological fluxes and potentially constrain local to in­ upstream inflow discharge hydrograph from WS observations on rela­
tegrated hydrodynamic models and inverse problems. However, esti­ tively short river reaches regarding the spatio-temporal sparsity of
mating river discharge Q from “geometric” observables of flow surface (satellite) observations sampling and without complex flow zones –
(elevation Z, width W and slope S) remains a difficult inverse problem confluences, multichannel portions (except (Garambois et al., 2020)),
particularly in case of poor knowledge on river bathymetry and friction floodplains. Moreover, few recent studies address the effective mod­
(see (Garambois and Monnier, 2015; Larnier et al., 2020) and refer­ eling of (ungauged) river channels using multisatellite data (Garambois
ences therein). et al., 2017; Schneider et al., 2017; Garambois et al., 2020; O’Loughlin
Hydraulic inverse problems with various model complexities, data- et al., 2019).
unknowns types and amounts are investigated by recent studies in a The present study investigates the challenging inference of multiple
satellite data context (see (Biancamaria et al., 2016) for a review). A inflows and channel parameters patterns from hydraulic signatures in a
few studies started to test the benefit of assimilating (synthetic) SWOT SWOT context. Particular attention is paid to the difficult inference of
WS observations with sequential methods in simplified hydraulic hydraulic controls with correlated effects on WS signatures including
models, for estimating inflow discharge assuming known river friction overlapping backwater effects. Moreover, we presentan effective hy­
and bathymetry (Andreadis et al., 2007; Biancamaria et al., 2011) or draulic modeling approach based on multi-satellite observations of WS
inferring bathymetry assuming known friction (Durand et al., 2008; and accounting for hydrological model inputs. It is appliedto a long
Yoon et al., 2012). Next, methods based on low-complexity models river reach including confluences with tributaries and strong backwater
have been proposed for estimating river discharge from WS observables effects in the Amazon basin. The computational inverse method, based
in case of unknown bathymetry b and friction K, based on the low on the full 1D Saint–Venant equations, is that presented in Brisset et al.
Froude model (Durand et al., 2014; Garambois and Monnier, 2015), (2018) and Larnier et al. (2020) with a spatially distributed friction
hydraulic geometries (Gleason et al., 2014) or empirical algebraic flow power law in water depth and a simple piecewise linear channel
models (Bjerklie et al., 2018). The intercomparison of low complexity bathymetry (Garambois et al., 2020). It is adapted here to account for
methods in Durand et al. (2016) highlights the difficulty of estimating lateral inflows/offtakes and is weakly coupled to the large scale MGB
the so-called unknown triplet (Q, K , b) from WS observables as well as hydrological model (Collischonn et al., 2007; Pontes et al., 2017; Paiva
the importance of good prior guesses on the sought parameters. et al., 2013b). Numerical investigations of the resulting WS signatures
The combined use of dynamic flow models of river systems and and identifiability tests are presented along with sensitivity analysis to
optimization methods enables to solve hydraulic inverse problems, as the parameters of both the (forward) hydraulic model and the inverse
shown for upstream flood hydrograph(s) estimation by Roux and Dartus method. The challenging inference of multiple inflows and channel
(2006) from WS width time series and a 1D Saint–Venant model or by parameters patterns is investigated with various observations densities
Honnorat et al. (2006), Hostache et al. (2010) and Lai and Monnier including the assimilation of synthetic SWOT ones.
(2009) using variational assimilation of flow depth time series in a 2D The paper is organized as follows. Section 2 presents the modeling
shallow-water model. The variational data assimilation (VDA) approach approach with the 1D Saint–Venant flow model and the inverse com­
(see e.g. (Cacuci et al., 2013) and references therein) is suitable to putational method. Section 3 investigates the capabilities of the inverse
address the present hydraulic inverse problem from WS observations method for identifying spatially distributed inflows with and without
(see (Gejadze et al., 2017; Brisset et al., 2018; Oubanas et al., 2018a; unknown channel parameters given observation patterns of WS sig­
Garambois et al., 2020; Larnier et al., 2020) and references therein – natures including overlapping backwater effects. Section 4 presents the
single upstream hydrographs in all studies except multiple “stepwise” effective modeling approach from multisatellite data applied to 871 km
offtakes on synthetic and densely observed irrigation-like cases in of the Negro river (Amazon basin) and the analysis of flow propagation
Gejadze et al. (2017)). It consists in fitting the modeled flow features to features against SWOT observability. Section 5 proposes inference tests
observations through the optimization of control parameters in a var­ for spatially distributed inflows with and without unknown parameters
iational framework. To be solved efficiently, such an ill-posed inverse on the Negro case in the presence of strong backwater effects.
problem needs to be regularized: see Kaltenbacher et al. (2008) for the
theory of regularization of such inverse problems and Gejadze et al.

2
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

2. Modeling approach discrepancy is classically evaluated with the observation cost function
term jobs (c ) = 2 (Z (c) Zobs ) O2 computed on the observation spatial
1

2.1. The flow model and temporal grids, see details in A.

The control vector c contains temporally and spatially distributed
The Saint–Venant equations (de Saint-Venant, 1871) consist in the unknown “input parameters” of the 1D Saint–Venant model: a friction
unidirectional form of the shallow water equations and are commonly used law (Garambois et al., 2020) and lateral inflows, unlike in Larnier et al.
to describe open channel flows (see e.g. (Chow, 1964; Ancey, 2018; Guinot, (2020), where there is a spatially uniform friction law K (h) without
2010) for detailed assumptions including the long wave one). In what fol­ lateral flows. It reads:
lows, x denotes the curvilinear abscissae from upstream to downstream
c = (Qin0 ,…, QinP; Qlat
0 P 0 P
,1 ,…, Qlat ,1, Qlat ,2 ,…, Qlat , L; b1 ,…, bI ; 1 ,…, N ; 1 ,…, N)
T
along a reach of length L (usual simplifying hypothesis are used) and
t [0, T ] denotes the time. In this representation, let A (x , t ) be the flow (4)
cross sectional area [m2] and Q (x , t ) the discharge [m3. s 1] such that where Qinp is the upstream discharge (the superscript p [0. .P ] denotes
U = Q / A represents the longitudinal cross-section averaged velocity the observation time), Qlat , l is the lateral discharges injected in the in­
p

[m. s 1]. The Saint–Venant equations in (A, Q ) variables at a flow cross flow cell l [1. .L] (note that Qlat , l = qlat , l x ), bi the river bed elevation
section read as follows: (i [1. .I ] denotes the computational cross section index in space) and,
for each patch n [1. .N ] with N I , the spatially distributed para­
tA + xQ = klat qlat
meters n and n of the friction law (Eq. (3)) depending on the flow
tQ + x ( ) + gA
Q2
A xZ = gASf + klat Uqlat
(1)
depth.
The inversion consists in solving the following minimization pro­
where Z (x , t ) is the WS elevation [m] and Z = (b + h) with b (x ) the river blem: c = argminc j(c ) starting from the so-called prior cprior in the
bed level [m] and h (x , t ) the water depth [m], Rh (x , t ) = A/ Ph the hy­ parameter space. This minimization problem is solved using a first
draulic radius [m], Ph (x , t ) the wetted perimeter [m], g is the gravity order gradient-based algorithm, more precisely the classical L-BFGS
magnitude [m. s 2], qlat (x , t ) is the lineic lateral discharge [m2 . s 1], and quasi-Newton algorithm (see Appendix A).Note that the sought para­
klat is a lateral discharge coefficient chosen equal to one here since we meters have a correlated influence on the modeled flow lines, therefore
consider inflows only. In DassFlow, the friction term Sf is classically para­ leading to an ill-posed inverse problem. In order to be solved efficiently,
meterized with the empirical Manning–Strickler law established for uniform the optimization problem is “regularized” as detailed in Appendix A.
flows: The main steps of the method are illustrated in Fig. 1.
QQ
Sf = 3. Inference capabilities from WS signatures: synthetic test cases
K 2S 2Rh4/3 (2)
The Strickler friction coefficient K [m1/3. s 1] is defined as a power In order to calibrate the parameters of a hydraulic model (Eq. (1))
law in h: from WS observables, one has to identify and understand the influence
K (x , h (x , t )) = (x ) h ( x , t ) (x )
(3) of these parameters on the observable(s): in our case the WS profile.
Fluvial flows are studied here in the context of satellite altimetry (see
where and are spatially distributed parameters. This spatially dis­ (Garambois and Monnier, 2015)). Following Montazem et al. (2017),
tributed friction law enables a variation of friction effects in function of the influence of the parameters on the modeled flow lines is referred to
the flow state (see effective modeling of multichannel flows in as their “hydraulic signature” and a reach is defined inbetween two
Garambois et al. (2020)). fluvial HCs. Fluvial HCs can be defined in steady state (see (Montazem
Inflow hydrographs Qin (t ) and qlat , l (t ) at l [1. .L] are classically et al., 2017)) as “local maximal deviations of the flow depth from the
imposed respectively upstream of the river domain and at known in­ normal depth hn (equilibrium), imposing the upstream variation of the
jection cells, that is inbetween two computational cross-sections along water depth profile h (x ) over the so-called control length (Samuels, 1989)”.
the river channel. Let us recall the Froude number definition Fr = U / c They can stem from a change in either the hydraulic resistance, cross-
comparing the average flow velocity U to pressure wave celerity
c = gA/ W where W is the flow top width [m]. Considering subcritical
flows (Fr < 1) in a satellite observability context (see (Garambois and
Monnier, 2015)), a boundary condition is imposed at the downstream
end of the model using the Manning–Strickler equation depending on
the unknowns (A, Q , K ) out . The initial condition is set as the steady state
backwater curve profile Z0 (x ) = Z (Qin (t0 ), qlat ,1.. L (t0 )) for hot-start.
This 1D Saint–Venant model (Eq. (1)) is discretized using the classical
implicit Preissmann scheme (see e.g. (Cunge et al., 1980)) on a regular
grid of spacing x using a double sweep method enabling to deal with
flow regimes changes, t is precised in numerical cases. This is im­
plemented into the computational software DassFlow (Data assimila­
tion for free surface flows, 2019).

2.2. The computational inverse method

The paper studies the estimation of spatially and temporally dis­

tributed flow controls from WS observables using the inverse method
presented in Larnier et al. (2020) (see also (Brisset et al., 2018)) with an
augmented composite control vector c; the method is detailed in A. The
principle of the inverse method is to estimate (discrete) flow controls by
minimizing the discrepancy between observed and simulated flow lines, Fig. 1. Flowchart of the method using the HiVDI inverse method (Larnier et al.,
Zobs and Z (c ) respectively, the latter depending on the unknown para­ 2020) for variational calibration, adapted from Monnier et al. (2016, 2014) and
meters vector c through the hydrodynamic model(Eq. (1)). This Garambois et al. (2020).

3
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Fig. 2. TOP:Academic test cases configurations. Rectangular channels of length L = 1000 m and constant cross section width W = 300 m , constant bottom slope of
10 3 m/m for Ch1, 2 and varying between 10 4 and 10 2 m/m for Ch3 – the bottom b (x ) is defined by linear interpolation between the 4 bathymetry points
(diamonds, b = {2, 1.88, 1.28, 1.12} m ) – and friction is defined by constant values on 3 patches ( = {30, 12.5, 30} m1/3. s 1 ). Upstream inflow at x = 0 ; for Ch1, 2, 3
respectively lateral injections at abscissae (in m ): x = 300, x = {300, 700} , x = {350, 700} , and observations at {xS1 ,…, xS5} = {150, 500, 850, 450, 550} ,
{xSb1 ,…, xSb5} = {150, 325, 450, 600, 800} , {xSc1 ,…, xSc 4} = {0, 300, 600, 1000} . BOTTOM: Sample waterlines with visible upstream and downstream controls and sig­
natures. For the sake of clarity here, upstream and injected flow are set at 100 m3s 1 (Fr~0.12 0.3). Using the identifiability index Iident = Twave/ tobs introduced in
(Brisset et al., 2018) with Twave = L / ck and the kinematic wave velocity for a rectangular channel ck = 5/3U (ck = 1.16 m. s 1 considering average speed
U = 0.69 m. s 1) and a high observation frequency ( tobs = 20 s ), gives a high identiability index Iident = 43 for the present flow observation configuration.

section shape, bottom slope or total flow variation through lateral ex­ They are set up as twin experiments, where a forward run of the flow
changes. model (Eq. (1)) is used to generate perfect WS elevation observations
This section studies the influence of inflows on hydraulic signatures, which are then used to infer an unknown parameter vector c (Eq. (4))
the capabilities of the inverse method described above to infer multiple with the inverse method described in Section 2.2 and Appendix A. The
inflows and channel parameters (either spatially constant or not), with inferences are started from erroneous prior guesses c (0) that verify
a focus on the influence of the spatial observability of those hydraulic Manning–Strickler law for hydraulic consistency, that is unbiased priors
signatures. (see investigations in (Larnier et al., 2020; Garambois et al., 2020));
hydrograph priors are constant values equal to the average value of the
target hydrographs.
3.1. Test case design
Increasingly challenging inverse problems are considered, with in­
creasing number of unknowns sought simultaneously and various ob­
Three test cases configurations representing typical hydraulic-ob­
servations densities. Cases Ch1 and Ch2 are used to infer temporal
servations setup of increasing complexities involving lateral inflows are
parameters only, given a channel of constant slope and friction. Case
presented (see Fig. 2). Cases Ch1 and Ch2 are designed to study the
Ch3 is the most challenging case with all inflows and non constant
effect on the inference of the overlapping signatures triggered by the
channel parameters sought simultaneously.
propagations of, respectively, one or two lateral hydrographs, con­
commitantly with the one of the upstream inflow hydrograph. Case Ch3
is a complexification of Ch2 through the introduction of a non-flat 3.2. Informative content of hydraulic signatures: single/multiple inflows
bottom and a variable friction pattern K = (x ) as needed in a real river inferences
case in the next sections ( = 0 in Eq. (3) – see investigations on spa­
tialized friction laws with multiscale bathymetry controls in Garambois The fluvial signature from a single lateral inflow is divided in two
et al. (2020)). parts (see Ch1 on Fig. 2, bottom): (i) in the reach downstream of the
For all three channels the boundary conditions (fluvial) consist in: injection point, the cumulative flow (Q = Qin + Qlat,1) is uniform with a
(i) a normal depth (equilibrium) imposed downstream and (ii) sinu­ water depth corresponding to the normal depth imposed downstream,
soidal hydrographs (see Table 1) imposed upstream and at lateral in­ (ii) in the reach upstream of the injection point an M1 backwater curve
jection cells. The simulation time step is set to t = 20 s for all cases. profile (see (Chow, 1959; Montazem, 2018; Montazem et al., 2017) in
the present “altimetry context”) is obtained given the upstream flow Qin
Table 1 and the water depth imposed downstream of this reach as the normal
depth corresponding to the cumulative flow. In the case of two distinct
Parameter values for sinusoidal hydrographs Q x , t = Q0 (x ) + aQ (x )sin ( t)
2
T
lateral injections (Ch2 ), WS signatures overlap in the most upstream
used in synthetic channels; resulting modeled Froude ranges. Flows in m3s 1, reach because of the stronger backwater effect created by two down­
time T in s . stream inflows, which represent a more challenging inference problem.
Inference trials in case Ch1 with control vector c1 = (Qlat
0
,1 ,…, Qlat ,1 ) ,
P T

assuming a known constant Qin (x = 0, t ), show that a single observation

point in space with a dense sampling in time, placed either upstream
(S1, in Fig. 2, top) or downstream (S2) from the lateral inflow, is suf­
ficient to infer one lateral inflow hydrograph perfectly – noiseless twin
experiments – (not presented). Indeed, the hydraulic signature of a
lateral inflow is visible and fully informative either upstream from it
because of its downstream control on the upstream flow line or

4
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

downstream from it, in the signature of the cumulative flow. This (S3) the propagation of the total discharge without downstream con­
means that as long the river is well temporally-observed regarding its trol.
response time (see (Brisset et al., 2018) without lateral inflows) and Variant Ch2b: Assimilation is more difficult than in Ch2a but in­
that the temporal variations of the observed system stem from a single ferred hydrographs (Fig. 3, red line) are still accurate (Table 2). This
control, only one spatial point is needed to infer this parameter. testifies to the ability to discriminate multiple sources of various am­
In the case of two distinct lateral injections (Ch2 ), WS signatures plitudes given observations of hydraulic signatures at higher frequency
overlap in the most upstream reach because of the stronger backwater and at pertinent locations (S1, S 2 and S3).
effect created downstream by the two inflows, which represents a more Interestingly, this case highlights the expected misattribution be­
challenging inference problem considering the unknown control vector haviour between inflow sources as shown by an intermediate iteration
c2 = (Qin0 ,…, QinP; Qlat ,1 ,…, Qlat ,1, Qlat ,2 ,…, Qlat ,2 ) . Several variants of Ch2 (Fig. 3, orange line) and remaining to a lesser extent at convergence
0 P 0 P T

are considered to study the possible misattibution of flow controls (red line): Qin and Qlat,1 are respectively over- and underestimated). This
(locations, amplitudes and frequencies) in case of identical inflow hy­ may be due to the relatively higher contribution of Qin to the observed
drographs (Ch2a ), the backwater influence of inflow hydrographs on signature (it impacts WS elevation at S1, S 2 and S3) and consequently
Qin downstream signature observed at S1 given 4 times larger inflow its contribution in the cost function (observation part).
amplitude (Ch2b ) or 10 times higher frequency (Ch2c ), different ob­ Note that the final overestimation of Qin in Ch2b is slightly greater
servations samplings “mixed” inflows signatures (see Fig. 2 and than in Ch2a . This is likely due to greater WS elevation variation at S1
Table 1). caused by backwater from Qlat,1 , which is first attributed to Qin since it
has more impact on the cost function. Remember that, with perfect
observations of WS signatures, at the end of the optimization process,
3.2.1. Inference of multiple inflows nearly perfect hydrographs are inferred. However, the small flow mis­
For all cases, using perfect and dense observations in space (1 every attributions during this optimization shows the difficulty of inferring
10 m ) and also in time leads to quasi perfect inferences. The influence of multiple controls using an observation located in a strong backwater
a sparser sampling and of the observability patterns of overlapping WS signal.
signatures on the identifiability of multiple inflows with the present Variant Ch2c : Perfect inferences are obtained. An intermediate
inverse method is studied here – without a priori weighting of the iteration (Fig. 3, orange line) shows that the expected misattribution of
parameters in the inverse method, that is equal and unadjusted frequencies for all 3 inflows is present, though it disappears at con­
values (see Appendix A). The inferred hydrographs are summed up in vergence (Fig. 3, red line). This testifies to the ability to discriminate
Fig. 3. Scores are given in Table 2, including cost function values and multiple sources of various frequencies given observations of hydraulic
iterations number at convergence. signatures at higher frequency and at pertinent locations (S1, S 2 and
Variant Ch2a : Given only one observation station by reach S3).
(S1, S 2, S3) very satisfying inferences of the 3 inflows are obtained Variant Ch2d : Convergence is achieved but the flow upstream of S4
(Fig. 3, red line). Hence sufficient information is provided by those is misattributed between Qin and Qlat,1 . Signatures of Qin and Qlat,1 are
three stations observing distinct signatures in each reach from upstream only observed mixed, downstream of Qlat,1 (at S4 and S5) and down­
to downstream: (S1) propagation of the inflow Qin (x = 0, t ) in presence stream from both Qlat,1 and Qlat,2 (at S3). Given that all stations are lo­
of the overlapping backwater effects due to Qlat,1 (x = 300, t ) and cated in the downstream infuence of both inflows, the distribution of
Qlat,2 (x = 700, t ) ; (S2 ) propagation of Qin (x = 0, t ) + Qlat,1 (x = 300, t ) in flow between them makes little difference on the observed WS
presence of the overlapping backwater effect due to Qlat,2 (x = 700, t ) ;

Fig. 3. Inflows inferences from WS observations for all Ch2 variants. Intermediate iteration in the assimilation process are represented for Ch2b and Ch2c ; they are
hand-picked to illustrate “intermediate” behaviours before convergence (“inferred”).

5
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Table 2
Inferred parameters misfits to the truth for Ch2 variants. The RMSE [m3. s 1] and rRMSE represent the misfit of the inferred
parameters, while the cost function used in the assimilation process represents the misfit of
n n
variables.RMSE = n i = 1 (Qtarget .
1 i i i
Qinfered )2 , rRMSE = RMSE / i = 1 Qtarget

dynamics. This confirms the need to have at least one observation values used in the inverse problem regularization, related to the sought
station between each sought inflow in order to be able to “separate” parameters (see Section A) and denoted as weights, are given in Table 3.
them. Inference results are presented in Fig. 4. In red, the final estimate of
c3 for Ch3a with the “default” weights set (see Table 3). In green, final
inferences for variant-specific parameter weights adjusted through trial
3.2.2. Synthesis
and error. In orange, intermediate inferences with the “default” set of
These first tests showed that for inferring multiple inflows, i.e.
parameter weights. Equal values of 1, corresponding to “no weighting”,
spatially distributed temporal controls, a minimal spatial observability
were also tested: they lead to inaccurate inferences (not shown) and
of their WS signature is required with one observation point between
thus the “default” weights producing more interpretable results are
each inflow here. In case of observation stations affected by backwater
preferred. In further iterations, after the ones plotted in orange, beha­
influence, the potential difficulty of separating multiple inflows from
viours similar to the Ch3a “default” weights inferences (Fig. 4, in red)
their “mixed signature” is highlighted; using a higher spatial density of
appear (not shown), i.e. a shift of inferred hydrographs and Strickler
(simultaneous) observations leads to improved inferences in the present
coefficients away from the target. Also note that the inferred flow os­
configuration. Moreover, using observations with high temporal density
cillation in the first time step stems from the influence of the in­
(with regards to the response time in the considered river system) and
itialization scheme (see Section 2.1) in the optimization on this quickly
low spatial density, different frequencies can be correctly attributed to
responding channel.
multiple inflows (as highlighted for a single upstream inflow in Brisset
Variant Ch3a Qin is underestimated while the local friction is over­
et al. (2018)). Furthermore, note that if a supercritical regime occurs in
estimated, denoting a local tendency to equifinality. This is linked to a
a reach between inflows, their hydraulic signatures are disconnected
strong backwater influence, created by both Qlat,1 and the increase in
(not shown), effectively reducing the assimilation problem to that of
friction at x = 300 m . This local inflow error leads to compensation in
case Ch1.
downstream hydrographs. By adjusting parameter weights through trial
and error, accurate inferences are obtained (Fig. 4, in green). This
3.3. Multiple and composite controls inference means that dense observations of the WS elevation are not sufficient for
inferring all flow controls contained in c3 and that spatially distributed
In this section multiple inflows are sought simultaneously with channel regularization parameters, acting as weights in the parameter search,
parameters on case Ch3. Three friction patches are consistently applied to are required.
sub-reaches inbetween the 4 sought bathymetry points. The control vector is Variant Ch3b and Ch3c With sparses observations, the “default”
c3 = (Qin0 ,…, QinP; Qlat ,1 ,…, Qlat ,1, Qlat ,2 ,…, Qlat ,2; b1, b2 , b2 , b4; 1, 2, 3 ) .
0 P 0 P T
weight set leads to worse inferences. However, the existence of a set of
Searching both inflows and channel parameters creates a configuration adjusted weights that lead to good inferences (Fig. 4, in green) is en­
(intendedly) prone to equifinality problems on the sought parameters ough to show that the minimum observation spatial density of 1 station
having correlated influence in the water surface signal. Three observation between each inflow can be sufficient to infer the extended control
configurations (see Fig. 2) are studied: one with a high station density vector c3 . Note that adjusted weight for Ch3b and Ch3c are different
(Ch3a : 100 stations, 1 every 10 m ), another with fewer stations (Ch3b: 9 from adjusted weights for Ch3a (see Table 3).
stations, Sb1..5 and Sc1..4 ) and a third one with even fewer stations (Ch3c : 4 Using less observation points in space, the influence of spatial
stations, Sc1..4 ). Priors for inflows are those defined for case Ch2 (subSection parameters decreases without loss of meaningful information and thus
3.2.1), priors for channel parameter are hydraulically consistent with flow the relative influence of inflows increases. This simple test highlights
priors and initial flow line. For this equifinality prone configuration, the

Table 3
Parameter weight sets in Ch3 variants.

6
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Fig. 4. Inflow, bathymetry and friction patch inferences from WS observations for all Ch3 variants. In red, final inference with “default” parameter weights (see
Table 3). In green, final inference with adjusted parameter weights. In orange, intermediate inferences with “default” parameter weights.

Fig. 5. Study zone on the Negro river. ENVISAT Virtual Stations are numbered from 1 to 18 starting from upstream. The boundaries of the studied reach are defined
by stations 1 and 18. The 21 tributaries are divided between actual rivers and inflow points from the hydrological model. SWOT swaths give an almost full spatial
coverage of the reach. In gray is the average water extent map used to extract width values, from Pekel et al. (2016).

the weighting influence of the parameters in the regularization WS observation patterns relatively to those of spatio-temporal controls,
method in the present flow configuration. The spatial distribution and satisfiying inferences are obtained with the present inverse method. A
density of WS observations along with the weights change the hydraulic real and complex river case is considered hereafter.
representativity of spatially distributed parameters in the optimization
process. 4. Effective hydraulic model of the Negro river
The main difficulty uncovered with these academic cases is the
challenge presented by simultaneous inferences of multiple inflows and After addressing increasingly challenging hydraulic inverse pro­
channel parameters from their potentially overlapping hydraulic sig­ blems on synthetic test cases in the previous section, a real complex
natures. However, in the case of unbiased prior parameters and dense river flow case is now considered. It consists in 871 km of the Negro

7
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Fig. 6. Effective river channel bottom and width for spatially sparse, G 1 = {benv , W1}r [1.. R], and spatially dense, G 2 = {bGPS , W2 }G [1.. G] , model geometries along with a
low-flow GPS waterline from Moreira (2016).

river, including several confluences with tributaries and anabranching effective bottom elevations at ENVISAT VS resolution. Next, in view to
flow zones. The reach is located upstream of the Solimoes-Negro con­ test the additional constraints brought by spatially dense satellite data
fluence which is responsible for significant backwater effects (see e.g. (synthetic SWOT), a more spatially detailled effective channel geometry
(Montazem et al., 2017)). This section presents the elaboration of ef­ G 2 is built using a high resolution water mask and an in situ flow line as
fective flow models in view of performing forward and inverse flow explained below.
modeling from WS observations of varying sparsity in the next section.
The modeling approach consists in (i) a 1D hydraulic model (full 4.2.1. Effective geometry G 1 from altimetry and optical data
Saint–Venant equations, see subSection 2.1) (ii) based on effective cross An effective 1D channel with effective rectangular cross sections is
sections defined from multisatellite and in situ data and (iii) weakly set up from available multisatellite data (altimetry, optical) and a large
coupled to the large scale hydrological model MGB (Collischonn et al., scale hydrological model following (Garambois et al., 2017, 2020).
2007; Pontes et al., 2017; Paiva et al., 2013b). The idea is to build an According to Latrubesse and Franzinelli (2005), high width to depth
effective river flow model both in coherence with the main hydrological ratios make the rectangular channel a pertinent effective modeling
signals (inflows) propagations along with observable flow surface sig­ approach of the true geometry, even in highly anastomosed reaches –
natures and hydraulic controls (see (Montazem et al., 2017)). As shown where an error on the actual hydraulic perimeter Ph hence Rh (see
in what follows, this 1D approach allows for a fair representation of subSection 2.1) is expected. This is supported by a qualitative analysis
flow propagation and longitudinal signatures, which are the core focus of some additional ADCP measurements of river flow and cross-sec­
of this paper. tional bathymetry.

4.1. Study zone • The river centerline from Allen and Pavelsky (2015), formed by
30 × 30 m pixels, is used to calculate the river length and to project
The study domain corresponds to the main stream of the Negro all spatial objects, such as VS, widths and inflow points, on a single
River, a major “left-bank” Amazon tributary draining the north part of one-dimensional reference.
the basin, with an average discharge of 28 400 m3. s 1 (Agência Nacional • A longitudinal profile of cross sectional WS width W is calculated
de Âguas, 2020). The reach covers the 871 km upstream of its con­ from the average river extent map derived from 31 years
fluence with the Solimoes and presents singular channel morphologies (1984–2015) of optic landsat imagery by Pekel et al. (2016). A
such as multichannel flow zones mainly located in two large grabens single width value per centerline point is extracted in order to build
(Latrubesse and Franzinelli, 2005). Part of the reach is strongly influ­ a 1D rectangular geometry. For multi-channel reaches, the effective
enced by the control imposed by the Solimoes river at its confluence width is the sum of the widths of all channels. This underestimates
(average discharge of 100 819 m3. s 1 according to ORE HYBAM gauge the actual hydraulic perimeter. Specific hand-filtering based on
data (Cochonneau et al., 2006), their confluence gives birth to the hydraulical expertise was necessary in some anabranching parts of
Amazon river). This hydraulic control is due to higher discharge and a the model where the water extent may include inactive flow zones
consequently lower slope of the Negro River in its lower reach when not accounted for in the present 1D effective model. Note that Park
compared to the Solimões River near to the confluence (Filizola et al., and Latrubesse (2017) concurs to the necessity of reach-scale flow
2009; Callède et al., 2013). The reach of interest has been crosscut by zone evaluation in the Amazon river catchment.
18 ENVISAT ground tracks every 35 days from 2003 to 2010 (see (Da • An effective channel bottom elevation benv is obtained at each VS
Silva et al., 2012)), representing 68 to 79 measurements of WS eleva­ (Fig. 6, in red) from altimetric rating curves (RC) from Paris et al.
tions at each of the 18 Virtual Stations (VS). Note that the measure­ (2016). Its slopes range from 7.1 × 10 5 to 2.0 × 10 4 m/m with
ments are not simultaneous for each station. an average of 7.0 × 10 5 m/m . RCs were obtained by adjusting the
parameters ( , ) of a stage discharge relationship
Q = (Zsat b) Ssat 0.5
using WS elevations Zsat and slopes Ssat gained
4.2. Effective models construction by satellite altimetry and discharge Q simulated with the large scale
hydrological model MGB (Collischonn et al., 2007; Pontes et al.,
This section presents the elaboration of effective flow models from 2017; Paiva et al., 2013b) on the temporal window of interest.
multisatellite data. First, a G 1 “sparse” channel geometry is built from

8
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Effective rectangular cross sections geometries are defined at the above (Fig. 6, in purple), are calibrated following (Garambois et al.,
R = 18 VS using the above defined effective bottom elevations 2020) who use observations of the same nature. Friction patterns cG 1 ,
{benv }r [1.. R] and river widths {W1 }r [1.. R] . The final model geometry cG 2a and cG 2b found with the inverse method are shown in Fig. 7. Most
G 1 = {benv , W1}r [1.. R] is obtained by linear interpolation between those differences in calibrated friction from M1 (Fig. 7, in red) to M 2a (Fig. 7,
18 effective cross sections on the model grid with x = 200 m . in blue) correspond to their difference in bathymetry at the virtual
station point (Fig. 7, gray line), i.e. a lower slope in M 2a leads to a
4.2.2. Effective geometry G 2 at increased spatial resolution higher inferred Strickler parameter in order to match WS observations
Spatially dense WS elevation data is introduced in the form of an in (e.g. in patch 2 and 6. Inferred parameters for M 2b roughly match those
situ GPS flow line with G = 579 spatial points. It was collected by of M 2a , with some discrepancies in patch 2, 15 and 17. Using a dif­
survey ship along the whole studied reach over 7 days during the low- ferent BC influences WS sensitivity to parameters and the relative
flow period in December 2010 (Moreira, 2016); it provides local WS contribution to the cost function of local WS misfits, which explains
elevations Z every 1.4 km on average and WS slopes S for every 25 km differences in patch 15 and 17; the one in patch 2 stems for the high
reach (ranging between 2.0 × 10 5 and 8.11 × 10 5 m/m , averaging at friction values, hence lower WS sensitivity as analyzed after.
3.4 × 10 5 m/m ). Under the hypothesis of a wide rectangular cross
section and a steady uniform flow, the Manning equation writes: 4.3.1. Water levels analysis
The following presents a detailed analysis of the effective hydraulic
Q = K (Wh)5/3 (W + 2h) 2/3 S (5)
model for configuration M1, along with an analysis of changes obtained
The water depth writes h = (Z b) and the bottom elevation is for configurations M 2a and M 2b .
sought using (i) the fixed WS width pattern W2 from imagery, (ii) the The simulated WS elevation are compared to observed WS elevation
WS elevation ZGPS and slope SGPS given by the GPS profile and (iii) the at each ENVISAT virtual station in Fig. 8 – other time series are avail­
discharge Q from the hydrological model (see subSection 4.1) on the able in Appendix B. For the 3 models calibrated above, the modeled WS
river domain at the corresponding time t . We invert an effective are fairly close to observed WS given the limited modeling complexity
bathymetry bGPS using Eq. (5) by minimizing the square sum of misfits and data uncertainties. More precisely, the fit to the altimetric WS
to benv at ENVISAT stations through the modification of M = 14 friction elevation time series is fairly good, as shown for M1 in Fig. 8, and
values (( m, m = 0), m [1. .M ], friction law Eq. (3)). They are nearly unbiased as shown in Fig. 9 (left). The WS elevation global RMSE
simply spatialized into M “hydraulic” patches consistent with large is at 0.936 m for M1 ; similar results are found with M 2a (see Table 4).
scale morphological features classified as follows: single channels, Errors are greater in low and high flows, with consistent under­
multiple channels (from 2 to 3), lightly anastomosed and heavily ana­ estimations of flow amplitude upstream (VS 1 4 ) which turns into
stomosed (Fig. 6, in purple). The friction coefficient values are coherent overestimation downstream (VS 9 13), before disappearing closer to
with the physical properties of the classified reaches. the BC (VS 14 18). VS 5 to 8 are particularly accurate. Error metrics
The new bathymetry bGPS is coherent with the best available re­ are coherent with those from current state of the art models using sa­
ference data and its corresponding set of physically distributed Strickler tellite data (see e.g. (O’Loughlin et al., 2019) on the Congo river).
patches. The final model geometry is G 2 = {bGPS , W2 }G [1.. G] . The analysis of the time series for M1 gives insight on the behaviour
In the following, using either geometry G 1 or G 2 , the hydraulic of the 1D model regarding the real flow physics sampled with the sparse
model is inflowed with time series at a daily time step upstream of the nadir altimetry data and dense in situ low flow line. Modeling errors
river domain and at 21 tributaries (both river tributaries and runoff can stem from either an (expected) improper representation of the
inflows) corresponding to the 21 catchment cells feeding into the Negro channel and flow complexity or uncertain (ungauged) inflows and data.
river cells in the large scale hydrological model MGB (Pontes et al., Concerning the hydraulics, from downstream to upstream, relative
2017; Collischonn et al., 2007). The largest of these tributaries is the errors are lower in anabranching reaches outside of the backwater in­
Branco river at 657 km . fluence starting at the Branco tributary ( x = 657 km up to around
x = 350 km ) and in the backwater influence of the (known “perfect”)
4.3. Effective models calibration against altimetry downstream BC. Overall, relative errors are higher upstream, in single
channeled, low water height reaches and in the Branco backwater in­
The friction of the hydraulic model (Eq. (1)) is calibrated against fluences. Note that 2D complex lateral flows in floodplains or retention
altimetric WS elevation time series following (Garambois et al., 2020), behaviours from “igarape” rivers may happen in high flow periods (see
i.e. {Zsobs
, p }S = 16, P [68..79] at ENVISAT VS, the most downstream VS being
env (Fleischmann et al., 2019; Fassoni-Andrade et al., 2020)). These un­
used as BC (see subSection 2.1). The friction law is distributed using accounted phenomenons may decrease flood wave velocities and cause
N = 17 “ENVISAT” patches with constant ( n, n = 0), n [1. .N ] va­ hydrograph skewness (Fleischmann et al., 2016; Collischonn et al.,
lues for each reach between two successive VS. This choice is made to 2017; Alsdorf et al., 2007).
avoid spatial “overparameterization” in the calibration process re­ The 1D modeling of water levels compared to altimetry observations
garding the spatial sparsity of ENVISAT observations of WS signatures. (Fig. 8) can first be analyzed as follows:
The aim of parameter calibration is to obtain a “real-like” model as
close as possible to the sparse observation set. Three models are con­ • Stations 14 to 18 are located in reaches with different morphological
sidered, to assess the impact of the bathymetry refinement and of the properties. Stations 14 and 15 are located in a densely anastomosed
downstream BC on the modeled hydraulic signatures and on inverse reach upstream of the Branco river confluence, a major tributary.
problems: a “sparse” model (M1) using channel geometry G 1 and the Stations 16 and 17 are in single channel reaches, upstream from the
WS elevation time series from VS 18 as BC, a refined model (M 2a ) with confluence with the Solimoes river. Station 18 is in a densely ana­
channel geometry G 2 including all the spatial variability from multi­ stomosed reach at the location of the BC forcing on WS elevation.
source data described above while keeping the same BC and a further Their low relative misfits do not testify to the absence of complex
changed refined model (M 2b ) where the BC is changed to an altimetric hydraulic behaviours in this area but rather to the dominating in­
RC which is of interest for “operational-like” applications in other rivers fluence of the BC.
and basins. • Stations 5 to 13 are located in mostly homogeneous anastomosed
The inverse method presented in Larnier et al. (2020) and described reaches, with stations 5 to 8 in a less densely anastomosed region
in subSection 2.2 and Appendix A is used here, without regularization than stations 9 to 13. This spatial division corresponds to two trends
terms, for friction calibration. Effective Strickler patches, starting from in relative misfit, where lower misfit is seen in the less anastomosed
priors corresponding to average values of the “hydraulic” patches used reaches. This testifies to the difficulty of modeling potentially 2D

9
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Fig. 7. Friction patches after calibration against ENVISAT WS elevation observations. Inferred Strickler coefficient values are very close for all configurations for
patches 1, 3 to 5, 9 to 14 and 16. Patches 2, 6 and 15 are especially sensitive to model variations.

Fig. 8. Modeled and observed WS elevation at ENVISAT VS after friction calibration at all stations for M1.

hydrodynamics using 1D approach. Indeed, the more channels there 2, this is partly due to effective width estimation errors being more
are, the further away the simulated wetted perimeter is from the prevalent in the relatively narrow channel (around 2 km in width).
true wetted perimeter (and so the hydraulic radius). Note that Furthermore, note that effective channel bottom elevation for these
parameterizing the Strickler coefficient as described in Eq. (3) and stations are respectively 37.3 m and 36.3 m while the lowest
including (x ) in the control vector during the calibration process, ENVISAT WS elevation observation are respectively 36.6 m and
instead of the simpler (x ) = 0 used here, does not yield a better fit 35.8 m . This corresponds to low-flow water heights of 0.7 m and
in this complex case modeled with a single rectangular channel. 0.5 m which do not fit field measurements. Consequently, relatively
• Stations 1 to 4 are located on single channel reaches. Although the high friction coefficients are inferred between station pairs 1–2 and
area seems the most suitable to be modeled in 1D, it still has the 2–3 to fit low water depth. This misfit might be due to data error,
highest relative misfit to ENVISAT observations. For stations 1 and including effective width errors for stations 3 and 4 located in areas

10
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Fig. 9. Relative misfit between modeled and observed WS elevation at ENVISAT for M1 (left, base model) M 2a (middle, WS elevation at downstream BC) and M 2b
(right, rating curve at downstream BC). Crosses are average values, horizontal bars are median values.

of sharp width variations. Note that the higher the friction values, of the true hydrographs are used as erroneous values for sensitivity
the less sensitivity of the modeled WS elevation, which explains the trials and are referred to as QFG30 and QFG
+30
respectively.
highest spread of Strickler coefficient (K = 40 to 55 m1/3s 1) in reach • Friction: two Strickler repartitions, with coefficient values worth
2 found during calibration for the 3 models (Fig. 7). respectively 70% and 130% of the truth are used as erroneous values
for sensitivity and are referred to as K FG30 and K FG
+30
respectively.
The introduction of the refined geometry G 2 in M 2a (for generating • Bathymetry: the inflow sets QFG30 and QFG +30
and the true Strickler
spatially distributed SWOT data and to perform inference tests here­ values are used to dig two bathymetries as described in Section 4.3.
after) has low impact on WS elevation bias and errors at ENVISAT VS The bathymetry derived from underestimated flows is referred to as
(see Fig. 9), with only stations 1, 2 and 3 showing significant change. bFG30 (it overestimates the true bathymetry), and the other is referred
Using a rating curve as downstream BC in M 2b mostly impacts the to as bFG+30
.
downstream part of the model where some misfit to altimetry data
appears. Indeed, it is more difficult, using a simple power law de­ Identifiability map SWOT will provide spatially distributed observa­
pending on the local flow variables, to capture the influence of the tions with interesting revisit frequencies at the scales of the current
confluence with the Solimoes River – not modeled. The latest having river domain and hydrological signal propagations. Fig. 10 shows the
strong discharge variations out of phase with the one of the Negro River evolution of the simulated WS elevation anomaly during the yearly
itself (e.g. (Montazem et al., 2019)). peak flow (red-blue heatmap) as well as its multiple SWOT ob­
servability (in black). Based on the modeled flow, accounting for sev­
4.3.2. Effective model analysis eral inflows, the propagation of an intumescence corresponding to the
As a preliminary to hydraulic parameters inference from WS ob­ annual flood wave signature is represented along the river through the
servables, this subsection studies the spatio-temporal features of the maximum WS elevation in time (following (Montazem, 2018)) (Fig. 10,
simulated hydraulic signatures, their sensitivity to model parameters top, blue points). This intumescence propagation is visible on the up­
and their observability given a SWOT sampling. First, an analysis of a stream 400 km of the river from day 164 to day 173. It is detected by a
flood wave propagation, resulting from multiple inflows, and its hy­ SWOT swath at t = 166 d and another one at t = 170 d . It is more dif­
draulic signature visibility is performed using identifiability maps fol­ ficult to detect this signature in the downstream part of the river
lowing (Brisset et al., 2018). The latter consist in a space–time re­ ( x > 400 km ) affected by the strong downstream control imposed by
presentation of the WS signal and flow propagation features against the high water depths at the Negro-Solimoes confluence; a downstream
observability pattern. These maps, inspired by the theory of char­ control due to the Branco tributary also overlaps from x = 657 km to
acteristics (see (Thual, 2010; Guinot, 2010)), enable to read how the around x = 400 km . This control can be seen through the tracked WS
sought upstream discharge information is sampled in the downstream elevation maximum (Fig. 10, top, in gray), where an early rise in WS
WS deformations and help to estimate inferable hydrograph fre­ elevation originates from x = 657 km , and through the extreme water­
quencies. Next, a numerical sensitivity evaluation of the flow model is lines (Fig. 10, bottom, in blue), which highlights the change in length of
carried out. this influence in low and high flows. As a consequence, WS observations
In the context of regional hydrological modeling including river on the downstream part may contain combined information due to the
networks representation, the sensitivity of the present flow model is upstream hydrographs propagation but also to the expression of
studied by using erroneous inputs. These inputs are also used in Section downstream controls.
5 as erroneous priors for various assimilation setups. The maximum WS elevation is tracked for simulations with erro­
neous parameters as defined above (Fig. 10, top, in red, green and
• Inflow: two hydrograph sets (containing lateral inflows and the cyan). They are not plotted where the flow displays “pool behaviour”
(gray points). They highlight the sensitivity of propagation to model
upstream BC inflow) corresponding respectively to 70% and 130%

Table 4
RMSE and bias over 8 years for the M1, M 2a and M 2b models. Upstream metrics are calculated for stations 1 to 9 only,
which are outside of the BC’s backwater influence. The high global RMSE for M 2b comes from the known dephasing of the
Solimoes and Negro peak flow, which is not reproduced by the RC.

11
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Fig. 10. TOP: Identifiability maps and flood wave propagation during the yearly peak flow (May-June) in the Negro river model. The WS anomaly (heatmap) is given
by Zano (x , t ) = Z (x , t ) Z (x ) , where Z (x ) is the average local WS elevation from day 160 to 190. Blue points: tracking of maximum WS elevation value
Zm (x ) = max t [0,365] Z (x , t ), x [0, L]. Gray points: tracking of maximum WSE in the downstream pool. Dashed blue lines: fictitious trajectory at kinematic speed
ck = 5/3U (sparse dashes) and at U + c (dotted dashes), starting at x = 0 , at the time of the local WS elevation peak. The speeds are calculated from the simulated
flow speed U and water height h and updated every x = 200 m , such as t p + 1 = t p + p . BOTTOM: Extreme flow forcings and flow model variables over a 2 year
x
ck

period. Blue lines: Extreme simulated waterlines. Red lines: corresponding extreme Froude values. Green lines: corresponding cumulative injected flows. Vertical
black dashes are lateral inflow locations. Bold vertical dashes are inflows inferred in subSection 5.2.

parameters which is also an important point when they are varied upstream and the Riemann invariants are modified along the wave due
during an optimization process as featured in Section 5. The propaga­ to the source term effects. The wave celerities obtained on the Negro
tion time from 0 to 400 km can be evaluated to around 10 days, and is River model are given by reach in Table 5, relatively high wave speeds
estimated as follows for the rest of the river domain. are obtained, hence propagation of information both upstream and
The conservative part of the Saint–Venant equations (i.e. without downstream, with spatio-temporal variability. The WS signature (and
source terms) is hyperbolic: some quantities depending on the water the discharge) thus reflects the nonlinear combination of information
depth and velocity (known as the Riemann invariants) are transported coming from both upstream (due to inflows variations) and down­
by waves at speeds different from the flow speed (see e.g. (Thual, 2010; stream (due to local hydraulic controls or downstream BC – see the
Guinot, 2010)). The wave celerities are U + c and U c with c = gh method of characteristic in (Guinot, 2010)). This highlights the diffi­
for rectangular cross sections (see analysis of propagation features in culty of inferring multiple inflows from sparse observations of WS sig­
(Brisset et al., 2018)). For the fluvial regime of interest here nature, especially given uncertain channel parameters and backwater
(Fr = U / c < 1), information propagates both downstream and effects.

12
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Table 5
Identifiability indexes between each pair of inflow at low and high flow (see Fig. 10, bottom). Speeds are given in m. s 1. Iident is given for a reach of length L and an
observation time step tobs = 1 d by Iident = 5 .
L
u¯ tobs
3

Nevertheless, interesting frequential information can be gained are both temporally and spatially distributed can represent a very
from the identifiability map as introduced in the case of a single inflow. challenging inverse problem, as previously highlighted on synthetic
cases.
5
Using the kinematic wave speed 3 U (Fig. 10, top, dashed blue line)
which compares fairly well to the intumescence speed on the upstream Inferences of L = 21 inflow hydrographs from 2 years of SWOT
part of the reach (Fig. 10, top, x < 400 km ). This gives an approximate synthetic observations are studied here. The channel geometry, friction
propagation time Twave = 26 d on the whole domain, greater than the and BCs are assumed to be known, hence the control vector reduced to
SWOT observation cycle period of 21 days. This brings the reach ,1 ,…, Qlat ,1, Qlat ,2 ,…, Qlat , L ) . The inferences are started from a
0 P 0 P T
c = (Qlat
identifiability index to Iident = 1.23 (defined as Iident = Twave / tobs , i.e. the prior guess c consisting in true hydrographs affected by uncertainties
(0)

average number of time a wave is observed, see (Brisset et al., 2018)). of ± 30%, that is QFG +30
and QFG30 as defined above. Note that the in­
However, in the present case, the notion defined by Brisset et al. (2018) ference is started from a hydraulically consistent initial state using an
accounts for a single upstream inflow, not spatially distributed lateral unbiased prior in the first time steps (see investigations in (Larnier
inflows with potential upstream backwater controls. Actual identifia­ et al., 2020; Garambois et al., 2020)); the prior values of regularization
bility indexes for reaches in between each lateral inflow would be much parameters Qlat correspond to inflows magnitudes.
lower (estimated identifiabilities in between each inflow pair are given The inferred hydrographs from inflow prior QFG30 are presented in
in Table 5 considering a fictious tobs = 1 d full domain observability). Fig. 11 for DenseSet (green lines) and SWOTNoiseSet observations (orange
Furthermore, SWOT swaths observations consist in WS snapshots on lines). Results from prior QFG +30
are available in Appendix C. SWOTSet and
different parts of the river domain at given times, hence containing SWOTNoiseSet give almost identical inferences, therefore only the SWOT­
various and mixed signatures (in the sense introduced in Section 3) of NoiseSet inferences are presented. For under- and overestimated priors, the
both several inflows and channel parameters – the more downstream, assimilation of dense and SWOT observations enables to infer the true hy­
the more aggregated is the inflow information. Inferences of multiple drographs fairly well. RMSE ranges from 8.86 m3s 1 at x = 465 km up to
inflows and frequential analysis are presented in the next section given 578.31 m3s 1 for the Branco tributary at x = 657 km . RMSE for all in­
known or uncertain channel parameters, spatio-temporally dense or ferences presented in Fig. 11 can be found in Section C. Some inferences
sparse (SWOT) observations. show global under- or overestimations (e.g. x = 216, 388, 789 km ). These
biases are linked to the prior bias. Strong and numerous overlapping
5. Inferences from satellite observables backwater signals may also influence flow misattribution, as discussed in
the academic cases (Section 3) for a small scale model. As tested in nu­
This section studies the challenging inference of ungauged channel merical experiments (not shown), increasing a scalar value Qlat, l can give
parameters and multiple inflows on the Negro River case, which re­ more effective weight to an hydrograph Qlat , l in the inference and it can be
presents a real and complex large scale problem. Typical inverse pro­ found further away from its prior guess, which highlights the role of the
blems in hydrological-hydraulic modeling are studied here considering covariance matrix used for regularization.
SWOT WS observations. The inference of channel parameters or/and Note that temporal oscillations appear on the inferred hydrographs
inflows in the 1D Saint–Venant model is addressed using the inverse when using SWOTNoiseSet which is “temporally sparse” observation
method presented in subSection 2.2 (see also Appendix A). The patterns compared to flow propagation, which is not the case of
downstream BC is set as a known altimetric rating curve. Three ob­ DenseSet. These oscillations are especially present in downstream in­
servation sets are generated: spatially and temporally dense flows, which may link them to particular hydraulic responses in the BC
( t = 600 s ) observations (DenseSet), SWOT observations from the hy­ influence zone, although they can be seen in upstream inflows as well.
draulic model outputs masked by SWOT swaths (SWOTSet) and noisy They tend to be prevalent in declining limbs of hydrographs (e.g. in
SWOT observations using the large scale simulator (CNES, n.d.) to add Fig. 11, at 789 km , from day 120 to 300).
realistic measurement noise (SWOTNoiseSet). We first present in­ Note that, regardless of oscillations, inferences tend to be further from
ferences of inflows only, then of channel parameters, and finally of all the truth in decreasing hydrographs. These oscillations are not the effect of
those spatio-temporal controls simultaneously. signal misattribution, as they are present with any number of inferred hy­
drographs (not shown), nor are they caused by the shape of the prior, as
5.1. Multiple hydrographs inferences filtered priors also lead to oscillations (not shown). Instead, the oscillations
seems to stem from the combination of the low observation frequency
Depicting flow structure within a river network and a catchment is a compared to the spatially distributed inflow hydrographs and the nonlinear
key issue in hydrological modeling, especially in ungauged basins. hydraulic response. Keep in mind that we track flow information through
Seeking to infer, from distributed WS observations, flow controls that WS elevation deformations caused by the nonlinear propagation of

13
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Fig. 11. Lateral hydrographs inferences from SWOTNoiseSet and DenseSet, using the QFG30 inflow prior.

parameter signatures (see subSection 4.3.2). windows”, the infered values are not necessarily representing reality
A sample illustration of those oscillations on the simulated WS (see related WS misfit in Fig. 12). They are the optimal solution cor­
elevation is presented in Fig. 12, at 870 km , downstream for three os­ responding to the considered priors of the optimization problem. In
cillating inferred hydrographs (at x = 738, 754 and 789 km ). The in­ practice, this means that introducing an additional regularization term
ferred waterline from SWOTNoiseSet is compared to the truth at all t 2h
jreg , like 0 2 dt at observation points, would smooth (as following a
t
simulated times and at SWOT pass times only. The misfit is logically spline curve) between the identifiable windows instead of the obtained
lowest at SWOT pass times (goal of the optimization), while unobserved oscillations. This smooth discharge curve would not be more physical
periods exhibit a slightly oscillating (unconstrained) misfit. Higher than the present oscillations and we made the choice to not hide this
frequency observations, such as DenseSet, prevent this behaviour well understood phenomena. It is a logical consequence of the disparity
through a more complete spatio-temporal observability of the WS sig­ between the samplings of observations and parameters and does not
natures, hence constrain the spatio-temporal parameters inference impede interpretations of hydraulic signatures and identifiability.
further. Some model configurations where temporal parameters are Seeking to infer a control that is both temporally and spatially dis­
discretized at a greater time step than observation one do no exhibit tributed represents a challenging assimilation problem. In the present case:
such behaviours (e.g. results with DenseSet, (Garambois et al., 2020)). (i) the observation frequency now plays a role in identifying the hydraulic
As already shown in Brisset et al. (2018) for the identification of a signature, on top of its spatial density and resulting flow propagation: (ii)
single hydrograph, the identification is possible only in time windows varying nonlinear flow propagation, and so WS signatures, can result in
representing the wave propagation time ( Twave ~26 days in the present different inferences depending whether they are performed from observa­
case), yet we have multiple inflows and observation samples (see sub­ tions of rising/declining hydrographs propagations (local Q (Z ) hysteresis)
Section 4.3.2). As a consequence, outside the “identifiable time and (iii) indirect contributions to parameter weight in the inverse method

14
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Fig. 12. Difference between target and inferred WS elevation at 870 km , as observed by DenseSet and SWOTSet. The dotted line represents the inferred waterline
inferred from SWOTSet (with QFG30 as inflow prior), but observed by DenseSet. The difference between this waterline and the target waterline is the misfit to target. At
SWOT pass times, the misfit is low as expected from an inference from SWOTSet. It only displays WS elevation oscillations at unobserved times.

appear, as successive hydrographs influence the whole downstream water

+2.6m
benv does not contain a priori information on target bathymetry shape
line (established in subSection 3.2), which gives greater “effective weight” – such as a coarse DEM prior. The friction prior is K FG30 for both in­
to upstream hydrographs as the cause of the misfit is observed in more ferences.
stations and thus accounted for multiple times in the cost function. The Using bFG30 and benv
+2.6m
as priors, the inference of channel parameters
inferred flow variations may be compensating for errors made at upstream (friction and bathymetry) respectively reach a RMSE of
stations with different SWOT pass times, impacting their WS elevation ei­ (4.362 m1/3. s 1, 1.231 m) and (4.762 m1/3. s 1, 1.913 m) . Upstream
ther through backwater control or through the modification of the BC and ( x = 0 110 km ), irregularities appear in both inferred bathymetries
its own backwater effect. and correspond to underestimated Strickler coefficients. The high fric­
tion partially hides the hydraulic signature of the bathymetry and en­
5.1.1. Spatial parameters inference ables inference error inconsequential to the cost function. Downstream
The inference of effective channel parameters is studied here con­ ( x = 600 870 km ), in the strong backwater control of the downstream
sidering a control vector composed of all friction coefficient values and BC, inferences tend to stay close to their prior values. Furthermore,
bathymetry points. The bathymetry is composed of I = 436 free points inferences from the unrefined prior benv +2.6m
are smoother than those from
the refined prior bFG30 . This testifies to the role of the a priori bathymetry
(1 every 2000 m ) between each of which it is obtained by linear inter­
shape knowledge contained in the prior when the WS elevation sensi­
polation. SWOTSet is used, with a spatial observations sampling (1
tivity to bathymetry is low. Strong backwater effects make the inference
point every 200 m ), i.e. 10 times greater than the sought bathymetry for
of channel parameters more difficult and parameter compensation ap­
observed reaches. Two inferences from hydraulically consistent priors
pears due to the lessened sensitivity of the simulated WS (e.g. in green,
are presented, one with the refined bathymetry bFG30 introduced in
subSection 4.3.2 (Fig. 13, in green) and another one with benv +2.6m
, a at the last friction patch). The refined bathymetry prior bFG30 leads to
shifted bathymetry from the M1 model in subSection 4.2.1, in red. The inferred bathymetry and friction that are closer to the truth. It will be
2.6 m correspond to the spatially averaged shift of bFG30 compared to benv . used in the extended control vector inferences below.

Fig. 13. Bathymetry and friction pattern inferences from SWOTSet.

15
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

5.2. Inference of channel parameters and inflows correlation between sought inflows at x = 502, 657 and 754 km appears.
The Branco river flow, at x = 657 km , is better inferred and its well-fitted
This section investigates the simultaneous inference of both un­ peaks are also found in the two smaller rivers (e.g. at 520 days), which are
known inflows and channel parameters on the large scale Negro River in its upstream and downstream influences zones (see Fig. 10, left). In all
case; it combines all previously mentioned difficulties and corresponds inferences, the total flow at the downstream BC closely matches that of the
to an ungauged configuration. In the following, the aim is to determine: truth, which means that only hydraulic signatures are misattributed across
1) if SWOT data are sufficient to infer the extended control vector given the 4 inferred temporal parameteAccess yearrs, not the total flow. In cfilter ,
unbiased prior parameters; 2) how the added spatial complexity from more accurate inferences are obtained, with a smaller influence of the
lateral inflows impacts spatial parameter inference. In addition, further Branco river on other inflows in its influence zone and more accurate in­
investigations on the impact of lateral inflow prior shape, representing ference of small scale behaviours. The filtered prior QFG filter
introduces in­
for example hydrological modeling uncertainty in a simple manner, are formation on low frequency behaviours of the sought inflows, helping the
presented. The considered extended control vector is the following: assimilation process converge to the target inflows. This configuration al­
lows for a better fit of small scale variation in the controls.
cext = (Qlat , x = 65 (t ), Qlat , x= 502 (t ), Qlat , x = 657 (t ),
Qlat , x= 754 (t ); b1, …, bH ; 1, …, N) (6)
6. Conclusion
The inferences are performed from DenseSet and SWOTNoiseSet.
The bathymetry and friction priors are bFG30 and K FG30 respectively. Four This paper investigated the inference of river channel parameters and
major lateral inflows located at x = 65, 502, 657 and 754 km (Fig. 10, multiple inflows from water surface signatures in the context of satellite
bold dashed bars) are considered. Their reduced number facilitates the altimetry with the forthcoming SWOT mission and using water extents from
analysis of their spatial impacts. The other inflows are set to their target optical data as well. It uses the HiVDI inverse method presented in Larnier
values. Two inflows prior types are used: QFG flat
, that gives no a priori on et al. (2020), based on the 1D Saint–Venant equations and a variational
hydrograph shapes and QFG , hydrographs obtained by applying a
filter
assimilation scheme adapted to account for lateral inflows (mass and mo­
80 days moving average filter to the true hydrographs. Prior flow values mentum injections). Given hydraulically consistent prior guesses and reg­
in QFG
flat
are set to the target flow values from the first time step up to ularization weights, the method is successfully applied to synthetic test cases
120 days for the sake of initial hydraulic consistency. Inferences of all and a long reach of the anabranching Negro River in the Amazon basin
parameters for these inflow priors are presented in Figs. 14 and 15. The using multisatellite data.
inferred control vectors are referred to as cflat and cfilter . Inferred para­ Through low Froude synthetic cases, it is shown that the signature
meter scores can be found in Table 6. of a lateral inflow is visible downstream from the inflow point through
Inferred spatial parameters patterns are similar to those obtained pre­ the total flow signature and can be visible upstream in case of down­
viously without unknown inflows in subSection 5.1.1. cflat features a fair stream control at the injection. Following this analysis and using the
bathymetry fit downstream (x = 600 870 km ) while cfilter stays close to HiVDI variational assimilation method (global in time and space), a
the prior value. This may be due to the different range of the simulated study of the minimum spatial density of water surface observations
hydraulic responses in the first iterations: using QFG flat
leads to an increase in necessary to infer lateral inflows from their hydraulic signatures is
WS elevation sensitivity to bathymetry. Upstream (x = 0 110 km ), in­ carried out. Synthetic twin experiments yield the following results: (i)
creased bathymetry irregularities in cflat are linked to the erroneous prior given high observation temporal frequency relative to model hydraulic
flat
QFG leading to bathymetry errors in the first iterations, coupled with lower response, perfect inflows inferences can be obtained; (ii) to correctly
inferred Strickler coefficients, hence a lessened impact of bathymetry on the attribute signatures between multiple lateral inflows, a minimum of 1
water surface and the inability to correct the “initial” errors. observation point between each injection cell is necessary; (iii) when
In terms of temporal behaviours, both priors give fair estimates of hy­ simultaneously inferring inflows and/or channel parameters, a sensi­
draulic controls for DenseSet. Inferences from SWOTNoiseSet are close to tivity to parameter weights (see Section A) appears; (iv) given a priori
those from DenseSet with the presence of oscillations and the rising part of parameter weights, accurate inferences of inflows and channel para­
hydrographs are better fitted than decreasing ones, as observed in meters are achievable even with the minimum spatial observability.
subSection 5.1. In both cfilter and cflat and for both observation sets, a A method for building effective river models in coherence with

Fig. 14. Inflow, bathymetry and friction patch inferences from SWOT synthetic data: cflat , inferred control vector without a priori hydraulic behaviour.

16
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Fig. 15. Inflow, bathymetry and friction patch inferences from SWOT synthetic data: cfilter , inferred control vector with a priori hydraulic behaviour.

multisatellite data and including realistic spatial variations is introduced Those oscillations related to model observations time scales could be
based on multisource data of water surface elevation, width and slopes. This overcome by introducing additional regularizations – not done here for the
method makes use of (i) multimission altimetric rating curves (see (Paris sake of hydraulic analysis. Inference of purely spatial parameters (bathy­
et al., 2016)) or equivalently a distributed hydrological model and altimetry metry/friction) were carried out as well, leading to some complementary
data and (ii) water surface width like those from current databases (see remarks: (i) channel parameters equifinality is most present in the down­
(Allen and Pavelsky, 2018)); it should be applicable to rivers from the future stream part of the model, where the waterline is influenced by the strong
SWOT database. It is applied here to build a simple effective 1D model of backwater effect of the boundary condition (Solimoes River) which di­
the Negro river upstream from its confluence with the Solimoes river. It fits minishes water surface sensitivity to other parameters; (ii) bathymetry prior
currently available satellite water surfaces signatures and contains real-like shape influences the inferred bathymetry. Finally, simultaneous inference of
spatial variabilities and flood wave propagation features. channel parameters and spatially distributed inflows was achieved with
The inference capabilities of spatially distributed channel para­ satisfying accuracy. We show that, with the present method, large scale
meters and inflows from synthetic SWOT observations are highlighted temporal parameter variations can be found from synthetic SWOT ob­
on the Negro River case given hydraulically coherent priors. The in­ servations even without a priori knowledge of the shape of the hydrological
ference of temporal parameters in the form of 21 spatially distributed response, but that small scale variations can be better inferred with a priori
lateral inflow hydrographs leads to accurate estimates and low water hydrograph shape knowledge.
surface misfit at observation times. High frequency observations give Recall that the estimation of discharges and channel parameters
good inferences, with an expected sensitivity to both prior bias (see from (SWOT) WS observations is a difficult inverse problem because of
(Larnier et al., 2020; Garambois et al., 2020)) and prior shape. the correlated influence of flow controls on the observable water sur­
SWOT-like observations lead to comparable inferences, with slight os­ face signatures – non-uniqueness/equifinality issues. It is therefore
cillations due to the frequential disparity between observations temporal necessary to use hydraulically consistent priors as investigated in
controls combined to their spatial distribution and the resulting nonlinear Larnier et al. (2020), Tuozzolo et al. (2019) and Garambois et al. (2020)
flow propagation on the domain, as analyzed with identifiability maps. with HiVDI method that contains low complexity flow relations for

Table 6
Inferred parameter scores for extended control inferences.

17
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

deriving robust prior guesses from databases and hydrological models, function terms) should help maximizing the use of various information
or even in situ depth/discharge data when available (see Larnier et al., sources and enable smooth discharge estimates and better signal attri­
2020) – not the scope of this paper. As already discussed in Larnier et al. bution, given uneven and heterogeneous satellite data in combination
(2020), the VDA solution depends on the priors, which are the first with other complementary databases/knowledge. This could help
guess values and the covariances matrixes. Ongoing research efforts in leveraging better inferences of hydrological responses and flow struc­
the SWOT community, in view of global discharge estimates, focus on ture within a river basin and eventually enable information feedback to
the determination of priors through the construction of a SWOT a priori rainfall-runoff modules and ultimately regionalization issues.
database based on Andreadis et al. (2013), Allen and Pavelsky (2018)
and global/regional model outputs (see (Durand et al., 2016; Larnier
et al., 2020)), constrained with available in situ gauge measurements. Author contributions and acknowledgments
Note that a priori estimations/databases could be enriched or re­
processed during or after the SWOT mission lifetime and HiVDI would This work is a part of the PhD thesis work of LP.
enable refining discharge estimates (see (Larnier et al., 2020)). More­ Research plan: LP, PAG, JM, PFG; computational software
over, as shown in Larnier and Monnier (2020), priors obtained by deep DassFlow1D and satellite data curation toolbox adapted from their
learning can greatly improve global estimation. previous versions by LP; Numerical investigations by LP with PAG, PFG,
More generally, reaching unbiased estimates of discharge, from KL, JM for analysis. JM is the principal designer of the inverse com­
downstream to upstream of river networks with varying densities of in putational method and its analysis.
situ discharge data, hence ungauged river portions/basins, is a crucial Data: The water surface widths were derived by SB and HY con­
challenge in hydrology that could benefit from the fusion of com­ tributed to their analysis using satellite images. The SWOT synthetic
plementary in situ and remotely sensed data in integrated models. The observations were generated by KL. The multisatellite dataset was
present study brings insight in lateral inflows inference from hydraulic provided by SC and altimetric rating curves and MGB outputs by AP.
signatures and paves the way for further research on integrated hy­ The GPS flow lines were provided by DM and SC.
drological-hydraulic assimilation chains for river networks and in co­ Fundings: Most authors have been co-funded by CNES Tosca re­
herence with multisatellites observables (of local hydrodynamic sig­ search project 2014–2019; project supervision: JM, PAG. PhD of LP is
natures) to benefit from them in a regionalization perspective. co-funded by CNES and ICUBE.
Searching for distributed channel parameters and inflows given
temporally sparse SWOT data and a global assimilation method brings
the issue of signal attribution to the forefront, especially at the scale of a Declaration of Competing Interest
river network. Further research should focus on tributaries that could
be amenable to the use of SWOT and multisatellite observations to The authors declare the following financial interests/personal re­
better constrain estimates of lateral inflows and, next, distributed fluxes lationships which may be considered as potential competing interests:
on network models considering hydrological-hydraulic couplings. The Pierre-Olivier Malaterre, INRAE, UMR G-Eau, Montpellier; Igor
introduction of pertinent signatures, scales and constraints in the for­ Gejadze, INRAE, UMR G-Eau, Montpellier; Hind Oubanas, INRAE, UMR
ward-inverse models (e.g. forward operators, covariance matrices, cost G-Eau, Montpellier; Sophie Ricci, INRAE, UMR G-Eau, Montpellier.

Appendix A. The computational inverse method

The computational inverse method is based on Variational Data Assimilation (VDA) applied to the 1D Saint–Venant model (Eq. (1)). The
computational inverse method is the one presented in Brisset et al. (2018) and Larnier et al. (2020) with an augmented composite control vector c
(Eq. (4)): c contains a spatially distributed friction coefficient enabling to model complex flow zones (while it is an uniform friction law K (h) in
Larnier et al. (2020)). This definition of K (x , h) enables to consider more heterogeneous bathymetry controls. c also contains lateral flow hydro­
graphs Qlat , i (t ) to deal with in/offtakes due to tributaries or underground flows. It is important to point out that the imposed downstream BC is an
unknown of the inverse problem. It is constrained with the observed water elevations and inferred river bottom slope using a locally uniform flow
hypothesis (i.e. Manning equation, see subSection 2.1) – except in the last real case above. The cost function j (c ) is defined as:
j (c ) = jobs (c ) + jreg (c ) (A.1)
where > 0 is a weighting coefficient of the so-called “regularization term” jreg (c ) . The term jobs (c ) measures the misfit between observed and
modeled WS elevations such that:
1 2
jobs (c ) = (Z (c ) Zobs )
2
O
(A.2)
The norm · O = 2 is defined from an a priori positive definite covariance matrix O . Assuming uncorrelated observations O = diag ( Z ) . The
O 1/2·
modeled WS elevations Z depend on c through the hydrodynamic model (Eq. (1)) and the inverse problem reads as
c = argminc j(c ) (A.3)
This optimal control problem is solved using a Quasi-Newton descent algorithm: the L-BFGS algorithm version presented in Gilbert and
Lemaréchal (1989). The cost gradient j (c ) is computed by solving the adjoint model; the latter is obtained by automatic differentiation using
Tapenade software (Hascoët and Pascual, 2013). Detailed know-hows on VDA may be found e.g. in the online courses (Bouttier and Courtier, 1999;
Monnier, 2014). see Fig. 5.
To be solved efficiently this optimization problem needs to be “regularized”. Indeed the friction and the bathymetry may trigger indiscernible
surface signatures therefore leading to an ill-posed inverse problem; we refer e.g. to Kaltenbacher et al. (2008) for the theory of regularization of such
inverse problems and to Larnier et al. (2020) for a discussion focused on the present inverse flow problem.
Following (Larnier et al., 2020), the optimization problem (Eq. (A.3)) is regularized as follows. First the regularization term jreg is added to the
cost function (Eq. (A.1)). We simply set: jreg (c ) = 2 b (x ) 22 . Therefore this term imposes (as weak constraints) the inferred bathymetry profile b (x )
1

to be an elastic interpolating the values of b at the control points (i.e. a cubic spline).
A specificity of the present context is the large inconsistency between the large observation grid (altimetry points) and the finer model grid.

18
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Between the sparse observations points (equivalently the control points), the bathymetry profile b (x ) is reconstructed as a piecewise linear function.
It is worth to point out that the resulting reconstruction is consistent with the physical analysis presented in Montazem et al. (2017) and Montazem
(2018). (This study analyses the adequation between the SW model (Eq. (1)) behavior and the WS signature).
Next and following (Lorenc et al., 2000; Weaver and Courtier, 2001; Larnier et al., 2020), the following change of control variable is made:
k=B 1/2 (c cprior ) (A.4)
where c is the original control vector, cprior is a prior value of c and B is a covariance matrix. The choice of B is crucial in the VDA formulation; its
expression is detailed below. After this change of variable the new optimization problem reads:
min J(k ) with J(k ) = j (c )
k (A.5)
It is easy to show that this leads to the following new optimality condition: j (c ) = 0 ; somehow a preconditioned optimality condition. For
B1/2
more details and explanations we refer to Haben et al. (2011a,b) and Larnier et al., 2020 in the present inversion context.
Assuming uncorrelated controls the matrix B is defined as block diagonal such that B = diag (BQ , BQlat,1 ,…, BQlat,L , Bb , B , B ) . Still following
(Larnier et al., 2020), the covariance matrices BQ , BQlat and Bb are set as the classical second order auto-regressive correlation matrices:

2 exp
|t j ti | 2 exp
|t j t i| |xj xi |
(BQ )i, j = ( Q) , (BQlat,l ) i,j = ( Q lat ) and (Bb )i, j = ( b ) 2 exp
tQ tQ Lb (A.6)
The VDA parameters tQ and Lb represent prior hydraulic scales and act as correlation lengths. We refer to Brisset et al. (2018) for a thorough
analysis of the discharge inference in terms of frequencies and wave lengths and (Larnier et al., 2020) in the present river-observation context. In the
present study, the friction parameters applied to deca-kilometric patches are assumed to be uncorrelated thus the matrices B and B are diagonal:
(B )i, i = ( )2 , (B )i, i = ( )2 (A.7)
The scalar values may be viewed as variances; their values are given in the numerical results section.
Finally, in a noised observation context and to avoid overfitting noisy data, we denote by the noise level such that Zobs Ztrue 2 with Zobs
the observed and Ztrue the true WS elevation profiles. A common technique to avoid overfitting noisy data, in the context of Tykhonov’s regular­
ization of ill-posed problems, is Morozov’s discrepancy principle, (see e.g. (Kaltenbacher et al., 2008) and references therein): the regularization
parameter (see Eq. (A.1)) is chosen a posteriori such that j does not decrease below the noise level.

Appendix B. Extended friction calibration results

Fig. B.16. ENVISAT WS elevation misfit after friction calibration at all stations for M 2a .

19
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Fig. B.17. ENVISAT WS elevation misfit after friction calibration at all stations for M 2b .

20
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Appendix C. Additional graphs and RMSE for lateral hydrograph inferences on the Negro River with DenseSet and SWOTNoiseSet
observation patterns

Fig. C.18. Lateral hydrograph inferences from SWOTNoiseSet and DenseSet, using the QFG
+30
inflow prior.

21
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Table C.7
Inferred lateral inflows parameter weights and RMSE.

22
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

Appendix D. Technical specifications

Host bridge: Intel Corporation 8th Gen Core Processor Host Bridge/DRAM Registers.
PCI bridge: Intel Corporation Xeon E3-1200 v5/E3-1500 v5/6th Gen Core Processor PCIe Controller.
Memory: 2x16Gb SODIMM DDR4 Synchronous 2667 MHz (0.4 ns)
Resolution mode: sequential.
Resolution method: implicit-explicit preissmann scheme.
Sample run: inference of the full triplet on the Negro model (inferred control vector cfilter in subSection 5.2).

• Control vector components: 4 × 740 flow points, 436 bathymetry points, 17 friction patches (3413 total sought values)
• Total run time (direct): under 15 min
• Total run time (inverse): 20.8 h
• Number of iterations: 35
• Average iteration time length: 35.8 min

References doi/abs/10.1002/2017WR021543.
Cunge, J.A., Holly, M., F., Verwey, A., 1980. Practical Aspects of Computational River
Hydraulics. Pitam Publishing.
Agência Nacional de Âguas (ANA). Brasil hydrografia. http://www.portalbrasil.net/ Da Silva, J.S., Seyler, F., Calmant, S., Rotunno Filho, O.C., Roux, E., Aranjo, A.A.M.,
brasil_hidrografia.htm. Accessed: 2020-02-06. Guyot, J.-L., 2012. Water level dynamics of amazon wetlands at the watershed scale
Allen, G.H., Pavelsky, T.M. 2018. Global extent of rivers and streams. Science. ISSN 0036- by satellite altimetry. International Journal of Remote Sensing 33 (11), 3323–3353.
8075. doi: 10.1126/science.aat0636. http://science.sciencemag.org/content/early/ https://doi.org/10.1080/01431161.2010.531914.
2018/06/27/science.aat0636. Data assimilation for free surface flows. Technical report, Mathematics Institute of
Allen, G.H., Pavelsky, T.M., 2015. Patterns of river width and surface area revealed by the Toulouse-INSA group-C&S corp.CNES-CNRS, 2019. URL http://www.math.univ-
satellite-derived north american river width data set. Geophysical Research Letters 42 toulouse.fr/DassFlow.
(2), 395–402. ISSN 1944–8007. doi: 10.1002/2014GL062764. de Saint-Venant, Adhemar J-C., 1871. Théorie du mouvement non-permanent des eaux,
Alsdorf, D., Bates, P., Melack, J., Wilson, M., Dunne, T., 2007. Spatial and temporal avec application aux crues des rivières et à l’introduction des marées dans leur lit. CR
complexity of the Amazon flood measured from space. Geophysical Research Letters Académie des Sciences Paris 73 (147–154), 5.
34 (8). doi: 10.1029/2007GL029447. https://agupubs.onlinelibrary.wiley.com/doi/ Durand, M., Andreadis, K.M., Alsdorf, D.E., Lettenmaier, D.P., Moller, D., Wilson, M.,
abs/10.1029/2007GL029447. 2008. Estimation of bathymetric depth and slope from data assimilation of swath
Ancey, C., 2018. Hydraulique à surface libre. Ecole Polytechnique Fédérale de Lausanne altimetry into a hydrodynamic model. Geophysical Research Letters 35 (20). doi: 10.
Ecublens. 1029/2008GL034150. https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/
Andreadis, K.M., Clark, E.A., Lettenmaier, D.P., Alsdorf, D.E., 2007. Prospects for river 2008GL034150.
discharge and depth estimation through assimilation of swath-altimetry into a raster- Durand, M., Fu, L., Lettenmaier, D.P., Alsdorf, D.E., Rodríguez, E., Esteban-Fernandez, D.,
based hydrodynamics model. Geophysical Research Letters 34 (10). doi: 10.1029/ May 2010. The surface water and ocean topography mission: observing terrestrial
2007GL029721. https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/ surface water and oceanic submesoscale eddies. Proceedings of the IEEE 98 (5),
2007GL029721. 766–779. https://doi.org/10.1109/JPROC.2010.2043031.
Andreadis, Konstantinos M., Schumann, Guy J-P., Pavelsky, Tamlin, 2013. A simple Durand, M., Neal, J., Rodríguez, E., Andreadis, K.M., Smith, L.C., Yoon, Y., 2014.
global river bankfull width and depth database. Water Resources Research, 49, Estimating reach-averaged discharge for the river Severn from measurements of river
7164–7168. ISSN 1944–7973. doi: 10.1002/wrcr.20440. water surface elevation and slope. Journal of Hydrology 511, 92–104. https://doi.
Biancamaria, S., Durand, M., K.M. Andreadis, K.M., Bates, P.D., Boone, A., Mognard, N. org/10.1016/j.jhydrol.2013.12.050.
M., Rodríguez, E., Alsdorf, D.E., Lettenmaier, D.P., Clark, E.A., 2011. Assimilation of Durand, M., Gleason, C.J., Garambois, P-A., Bjerklie, D., Smith, L.C., Roux, H., Rodriguez,
virtual wide swath altimetry to improve arctic river modeling. Remote Sensing of E., Bates, P.D., Pavelsky, T.M., Monnier, J., Chen, X., Di Baldassarre, G., Fiset, J-M.,
Environment 115 (2), 373–381. ISSN 0034–4257. doi: 10.1016/j.rse.2010.09.008. Flipo, N., Frasson, R.P.d.M., Fulton, J., Goutal, N., Hossain, F., Humphries, E.,
http://www.sciencedirect.com/science/article/pii/S0034425710002816. Minear, J.T., Mukolwe, M.M., Neal, J.C., Ricci, S., Sanders, B.F., Schumann, G.,
Biancamaria, S., Lettenmaier, D.P., Pavelsky, T.M., 2016. The SWOT mission and its Schubert, J.E., Vilmin, L., 2016. An intercomparison of remote sensing river dis­
capabilities for land hydrology. Surveys in Geophysics 37 (2), 307–337. ISSN charge estimation algorithms from measurements of river height, width, and slope.
1573–0956. doi: 10.1007/s10712-015-9346-y. Water Resources Research 52 (6), 4527–4549. https://doi.org/10.1002/
Bjerklie, D.M., Birkett, C.M., Jones, W., Carabajal, C., Rover, J.A., Fulton, J.W., 2015WR018434. URL https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1002/
Garambois, P-A., 2018. Satellite remote sensing estimation of river discharge: 2015WR018434.
Application to the Yukon river Alaska. Journal of Hydrology 561, 1000–1018. ISSN Fassoni-Andrade, A.C., de Paiva, R.C.D., Fleischmann, A.S., 2020. Lake topography and
0022–1694. doi: 10.1016/j.jhydrol.2018.04.005. http://www.sciencedirect.com/ active storage from satellite observations of flood frequency. Water Resources
science/article/pii/S0022169418302464. Research 56 (7). doi: 10.1029/2019WR026362. https://agupubs.onlinelibrary.wiley.
Bouttier, F., Courtier, P., 2002. Data assimilation concepts and methods March 1999. com/doi/abs/10.1029/2019WR026362.
Meteorological Training Course Lecture Series. ECMWF, p. 59. http://msi.ttu.ee/ Fekete, B.M., Vorosmarty, C.J., 2002. The current status of global river discharge mon­
elken/Assim_concepts.pdf. itoring and potential new technologies complementing traditional discharge mea­
Brisset, P., Monnier, J., Garambois, P.-A., Roux, H., 2018. On the assimilation of alti­ surements. IAHS - PUB.
metric data in 1D Saint-Venant river flow models. Advances in Water Resources 119, Filizola, N., Spínola, N., Arruda, W., Seyler, Fr., Calmant, S., Silva, J., 2009. The Rio
41–59. Negro and Rio Solimões confluence point-hydrometric observations during the 2006/
Cacuci, D.G., Navon, I.M., Ionescu-Bugor, M., 2013. Computational Methods for Data 2007 cycle. River, Coastal and Estuarine Morphodynamics-RCEM 2009, 1003–1006.
Evaluation and Assimilation. Taylor and Francis CRC Press, Boca Raton. Fleischmann, A.S., Paiva, R.C.D., Collischonn, W., Sorribas, M.V., Pontes, P.R.M., 2016.
Callède, J., Moreira, D.M., Calmant, S., 2013. Détermination de l’altitude du zéro des On river-floodplain interaction and hydrograph skewness. Water Resources Research
stations hydrométriques en Amazonie brésilienne. Application aux lignes d’eau des 52, 7615–7630. https://doi.org/10.1002/2016WR019233. https://agupubs.
Rios Negro, Solimões et Amazone. Journal of Water Science 26 (2), 153–171. https:// onlinelibrary.wiley.com/doi/abs/10.1002/2016WR019233.
doi.org/10.7202/1016065ar. Fleischmann, A.S., Collischonn, W., Paiva, R., Tucci, C.E., 2019. Modeling the role of
Chow, V.T., 1964. Handbook of Applied Hydrology. McGraw-Hill Book Co., New-York, p. reservoirs versus floodplains on large-scale river hydrodynamics. Natural Hazards 99
1964. (2), 1075–1104. ISSN 1573-0840. doi: 10.1007/s11069-019-03797-9.
Chow, V.T., 1959. Open-Channel Hydraulics. Mc Graw-Hill, New-York, USA. Garambois, P-A., Monnier, J. 2015. Inference of effective river properties from remotely
Cochonneau, G., Sondag, F., Guyot, J-L., Geraldo, B., Filizola, N., Fraizy, P., Laraque, A., sensed observations of water surface. Advances in Water Resources 79, 103–120. URL
Magat, P., Martinez, J-M., Noriega, L., Oliveira, E., Ordonez, J., Pombosa, R., Seyler https://www.sciencedirect.com/science/article/abs/pii/S0309170815000330.
F., Sidgwick, J., Vauchel, P., 2006. L’Observatoire de Recherche en Environnement, Garambois, P-A., Calmant, S., Roux, H., Paris, A., Monnier, J., Finaud-Guyot, P.,
ORE HYBAM sur les grands fleuves amazoniens. ISBN 978-1-901502-78-7. Montazem, A.S., da Silva, J.S., 2017. Hydraulic visibility: Using satellite altimetry to
CNES. SWOT hydrology toolbox. URL https://github.com/CNES/swot-hydrology- parameterize a hydraulic model of an ungauged reach of a braided river.
toolbox. Hydrological Processes, 31 756–767. ISSN 1099–1085. doi: 10.1002/hyp.11033.
Collischonn, W., Allasia, D., Da Silva, B.C., Tucci, E.C.M., 2007. The MGB-IPH model for Garambois, P-A., Larnier, K., Monnier, J., Finaud-Guyot, P., Verley, J., Montazem, A.S.,
large-scale rainfall-runoff modelling. Hydrological Sciences Journal 52 (5), 878–895. Calmant, S., 2020. Variational estimation of effective channel and ungauged ana­
https://doi.org/10.1623/hysj.52.5.878. branching river discharge from multi-satellite water heights of different spatial
Collischonn, W., Fleischmann, A.S., Paiva, R.C.D., Mejia, A., 2017. Hydraulic causes for sparsity. Journal of Hydrology 581, 124409. ISSN 0022–1694. doi: 10.1016/j.jhy­
basin hydrograph skewness. Water Resources Research 53 (12), 10603–10618. drol.2019.124409.
https://doi.org/10.1002/2017WR021543. https://agupubs.onlinelibrary.wiley.com/ Gejadze, I., Malaterre, P-O., 2017. Discharge estimation under uncertainty using

23
L. Pujol, et al. Journal of Hydrology 591 (2020) 125331

variational methods with application to the full Saint-Venant hydraulic network agupubs.onlinelibrary.wiley.com/doi/abs/10.1002/2017WR021735.
model. International Journal for Numerical Methods in Fluids 83 (5), 405–430. ISSN Paiva, R.C.D., Buarque, D.C., Collischonn, W., Bonnet, M.-P., Frappart, F., Calmant, S.,
1097–0363. doi: 10.1002/fld.4273. doi: 10.1002/fld.4273. fld.4273. Carlos, E.M., 2013a. Large-scale hydrologic and hydrodynamic modeling of the
Gilbert, J.C., Lemaréchal, C., 1989. Some numerical experiments with variable-storage amazon river basin. Water Resources Research 49 (3), 1226–1243. https://doi.org/
quasi-newton algorithms. Mathematical Programming 45 (1–3), 407–435. 10.1002/wrcr.20067. https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1002/
Gleason, C.J., Smith, L.C., 2014. Toward global mapping of river discharge using satellite wrcr.20067.
images and at-many-stations hydraulic geometry. Proceedings of the National Paiva, R.C.D., Buarque, D.C., Collischonn, W., Bonnet, M.-P., Frappart, F., Calmant, S.,
Academy of Sciences 111 (13), 4788–4791. ISSN 0027–8424. doi: 10.1073/pnas. Carlos, E.M., 2013b. Large-scale hydrologic and hydrodynamic modeling of the
1317606111. http://www.pnas.org/content/111/13/4788. amazon river basin. Water Resources Research 49 (3), 1226–1243. https://doi.org/
Haben, S.A., Lawless, A.S., Nichols, N.K., 2011a. Conditioning and preconditioning of the 10.1002/wrcr.20067. https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1002/
variational data assimilation problem. Computers & Fluids 46 (1), 252–256. wrcr.20067.
Haben, S.A., Lawless, A.S., Nichols, N.K., 2011b. Conditioning of incremental variational Paris, A., de Paiva, R.C.D., da Silva, J.S., Moreira, D.M., Calmant, S., Garambois, P.-A.,
data assimilation, with application to the met office system. Tellus A 63, 782–792. Collischonn, W., Bonnet, M.-P., Seyler, F., 2016. Stage-discharge rating curves based
Hascoët, L., Pascual, V., 2013. The Tapenade Automatic Differentiation tool: Principles, on satellite altimetry and modeled discharge in the amazon basin. Water Resources
Model, and Specification. ACM Transactions On Mathematical Software 39 (3). doi: Research 52 (5), 3787–3814. https://doi.org/10.1002/2014WR016618. https://
10.1145/2450153.2450158. agupubs.onlinelibrary.wiley.com/doi/abs/10.1002/2014WR016618.
Honnorat, M., Lai, X., le Dimet, F-X., Monnier, J., 2006. Variational data assimilation for Park, E., Latrubesse, E.M., 2017. The hydro-geomorphologic complexity of the lower
2D fluvial hydraulics simulation. CMWR XVI-Computational Methods for Water amazon river floodplain and hydrological connectivity assessed by remote sensing
Ressources. Copenhagen. and field control. Remote Sensing of Environment 198, 321–332. ISSN 0034-4257.
Hostache, R., Lai, X., Monnier, J., Puech, C., 2010. Assimilation of spatially distributed doi: 10.1016/j.rse.2017.06.021. http://www.sciencedirect.com/science/article/pii/
water levels into a shallow-water flood model. Part II: Use of a remote sensing image S0034425717302808.
of Mosel River. Journal of Hydrology 390, 257–268. Pekel, J.-F., Cottam, A., Gorelick, N., 2016. High-resolution mapping of global surface
Kaltenbacher, B., Neubauer, A., Scherzer, O., 2008. Iterative regularization methods for water and its long-term changes. Nature 540 (418), 19. https://doi.org/10.1038/
nonlinear ill-posed problems, vol. 6 Walter de Gruyter. nature20584. https://www.nature.com/articles/nature20584#supplementary-
Lai, X., Monnier, J., 2009. Assimilation of spatially distributed water levels into a shallow- information.
water flood model. Part I: mathematical method and test case. Journal of Hydrology Pontes, P.R.M., Fan, F.M., Fleischmann, A.S., Paiva, R.C.D., Buarque, D.C., Siqueira, V.A.,
377, 1–11. http://www.sciencedirect.com/science/article/pii/S0022169409004508. Jardim, P.F., Sorribas, M.V., Collischonn, W., 2017. MGB-IPH model for hydrological
1-2. and hydraulic simulation of large floodplain river systems coupled with open source
Larnier, K., Monnier, J., 2020. Hybrid neural network – data assimilation algorithm to GIS. Environmental Modelling & Software 94, 1–20. ISSN 1364-8152. doi: 10.1016/j.
infer river features. submitted. envsoft.2017.03.029. http://www.sciencedirect.com/science/article/pii/
Larnier, K., Monnier, J., Garambois, P.-A., Verley, J., 2020. On the estimation of river S136481521630189X.
discharges from altimetry. Inverse Problems Science Engineering (IPSE) Accepted, to Rodríguez, E., 2012. SWOT Science requirements document. JPL document, JPL, 2012.
appear. URL https://hal.archives-ouvertes.fr/hal-01811683. URL https://swot.jpl.nasa.gov/system/documents/files/2176_2176_D-61923_SRD_
Latrubesse, E.M., Franzinelli, E., 2005. The late quaternary evolution of the negro river, Rev_B_20181113.pdf.
amazon, brazil: Implications for island and floodplain formation in large ana­ Rodríguez, E., Esteban-Fernandez, D., Peral, E., Chen, C.W., Blesser, J.-W., Williams, B.,
branching tropical systems. Geomorphology 70 (3), 372–397. ISSN 0169–555X. doi: 2018. Wide-swath altimetry: A review. In: Stammer, D., Cazenave, A. (Eds.), Satellite
10.1016/j.geomorph.2005.02.014. http://www.sciencedirect.com/science/article/ Altimetry over Oceans and Land Surfaces (Chapter 2). CRC Press.
pii/S0169555X05000826. Tropical Rivers. Rodríguez, E., Durand, M., de, R.P., Frasson, M., 2020. Observing rivers with varying
Lorenc, A.C., Ballard, S.P., Bell, R.S., Ingleby, N.B., Andrews, P.L.F., Barker, D.M., Bray, spatial scales. Water Resources Research. doi: 10.1029/2019WR026476. URL
J.R., Clayton, A.M., Dalby, T., Li, D., et al., 2000. The met. office global three-di­ https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2019WR026476.
mensional variational data assimilation scheme. Quarterly Journal of the Royal Roux, H., Dartus, D., 2006. Use of parameter optimization to estimate a flood wave:
Meteorological Society 126 (570), 2991–3012. Potential applications to remote sensing of rivers. Journal of Hydrology 328,
Guinot, V., 2010. Wave propagation in fluids: models and numerical techniques, second 258–266.
ed., vol. 49. Samuels, P.G., 1989. Backwater lengths in rivers. Proceedings of the Institution of Civil
Monnier, J., 2014. Variational data assimilation: from optimal control to large scale data Engineers 87, 571–582. https://doi.org/10.1680/iicep.1989.3779.
assimilation. Open Online Course, INSA Toulouse. URL https://www.math.univ- Schneider, R., Godiksen, P.N., Villadsen, H., Madsen, H., Bauer-Gottwein, P., 2017.
toulouse.fr/%7Ejmonnie/Enseignement/CourseVDA.pdf. Application of CryoSat-2 altimetry data for river analysis and modelling. Hydrology
Monnier, J., Couderc, F., Dartus, D., Larnier, K., Madec, R., Vila, J.-P., 2016. Inverse and Earth System Sciences 21 (2), 751–764. https://doi.org/10.5194/hess-21-751-
algorithms for 2D shallow water equations in presence of wet dry fronts. application 2017. https://www.hydrol-earth-syst-sci.net/21/751/2017/.
to flood plain dynamics. Advances in Water Resources 97, 11–24. Schuite, J., Flipo, N., Massei, N., Rivière, A., Baratelli, F., 2019. Improving the spectral
Montazem, A., 2018. Représentation et segmentation hydraulique effective de rivières analysis of hydrological signals to efficiently constrain watershed properties. Water
pour le calcul de debit par altimetrie SWOT à l’échelle globale. Universite de Resources Research 55 (5), 4043–4065. https://doi.org/10.1029/2018WR024579.
Toulouse III Paul Sabatier, Toulouse, France, Thèse de doctorat. https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2018WR024579.
Montazem, A., Garambois, P.-A., Finaud-Guyot, P., Calmant, S., Monnier, J., Moreira, D., Thual, O., 2010. Hydrodynamique de l’environnement. Editions de l’Ecole Polytechnique.
2017. Physical basis for river segmentation from water surface observables. In: AGU Tuozzolo, S., Lind, G., Overstreet, B., Mangano, J., Fonstad, M., Hagemann, M., Frasson,
Fall Meeting Abstracts, vol. 2017, p. H44H-06. URL https://ui.adsabs.harvard.edu/ R.P.M., Larnier, K., Garambois, P.-A., Monnier, J., Durand, M., 2019. Estimating river
abs/2017AGUFM.H44H..06S. discharge with swath altimetry: A proof of concept using AirSWOT observations.
Montazem, A.S., Garambois, P-A., Calmant, S., Finaud-Guyot, P., Monnier, J., Moreira, D. Geophysical Research Letters 46 (3), 1459–1466. https://doi.org/10.1029/
M., Minear, J.T., Biancamaria, S., 2019. Wavelet-based river segmentation using 2018GL080771. https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/
hydraulic control-preserving water surface elevation profile properties. Geophysical 2018GL080771.
Research Letters 46 (12), 6534–6543. ISSN 0094–8276. doi: 10.1029/ Weaver, A., Courtier, P., 2001. Correlation modelling on the sphere using a generalized
2019GL082986. diffusion equation. Quarterly Journal of the Royal Meteorological Society 127 (575),
Moreira, D.M., 2016. Apport des données de géodésie spatial pour l’étude du bassin hy­ 1815–1846.
drologique amazonien. PhD thesis, LEGOS – Laboratoire d’Etudes en Géophysique et World Meteorological Organization, 1997. GCOS/GTOS plan for terrestrial climate-re­
Océanographie Spatiale. http://www.theses.fr/2016TOU30372. lated observations, version 2.0. URL.
O’Loughlin, F.E., Neal, J., Schumann, G.J.P., Beighley, E., Bates, P.D., 2019. A LISFLOOD- World Meteorological Organization. GCOS systematic observation requirements for sa­
FP hydraulic model of the middle reach of the Congo. Journal of Hydrology, p. tellite-based products for climate 2011 update: Supplemental details to the satellite-
124203. ISSN 0022-1694. doi: 10.1016/j.jhydrol.2019.124203. http://www.scien­ based component of the implementation plan for the global observing system for
cedirect.com/science/article/pii/S0022169419309382. climate in support of the UNFCCC (2010 update). https://library.wmo.int/doc_num.
Oubanas, H., Gejadze, I., Malaterre, P-O., Mercier, F., 2018. River discharge estimation php?explnum_id=3710.
from synthetic SWOT-type observations using variational data assimilation and the Yoon, Y., Durand, M., Merry, C.J., Clark, E.A., Andreadis, K.M., Alsdorf D.E., 2012.
full Saint-Venant hydraulic model. Journal of Hydrology 559, 638–647. ISSN 0022- Estimating river bathymetry from data assimilation of synthetic SWOT measure­
1694. doi: 10.1016/j.jhydrol.2018.02.004. http://www.sciencedirect.com/science/ ments. Journal of Hydrology 464–465 (0), 363–375. ISSN 0022–1694. doi: 10.1016/
article/pii/S0022169418300805. j.jhydrol.2012.07.028. http://www.sciencedirect.com/science/article/pii/
Oubanas, H., Gejadze, I., Malaterre, P.-O., Durand, M., Wei, R., Frasson, R.P.M., S0022169412006294.
Domeneghetti, A., 2018b. Discharge estimation in ungauged basins through varia­ Yoon, Y., Garambois, P.-A., Paiva, R., Durand, M., Roux, H., Beighley, E., 2016. Improved
tional data assimilation: The potential of the SWOT mission. Water Resources error estimates of a discharge algorithm for remotely sensed river measurements: Test
Research 54 (3), 2405–2423. https://doi.org/10.1002/2017WR021735. https:// cases on Sacramento and Garonne rivers. Water Resources Research 52 (1), 278–294.

24