See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/356647026

Validation and Improvement of Data Assimilation for Flood Hydrodynamic

Modelling Using SAR Imagery Data

Conference Paper · October 2021

CITATIONS READS

11 113

6 authors, including:

Thanh Huy Nguyen Christophe Fatras

Luxembourg Institute of Science and Technology (LIST) Collecte Localisation Satellites
41 PUBLICATIONS 153 CITATIONS 35 PUBLICATIONS 370 CITATIONS

SEE PROFILE SEE PROFILE

Peter Kettig A. Piacentini

Centre National d’Etudes Spatiales
103 PUBLICATIONS 1,945 CITATIONS
18 PUBLICATIONS 97 CITATIONS
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Thanh Huy Nguyen on 01 December 2021.

The user has requested enhancement of the downloaded file.

 Validation and Improvement of Data Assimilation for
Flood Hydrodynamic Modelling Using SAR Imagery
Data
Thanh Huy Nguyen1, Anthéa Delmotte1, Christophe Fatras2, Peter Kettig3, Andrea Piacentini1, and Sophie Ricci1
1
CECI, CERFACS/CNRS UMR 5318, Toulouse, France
2
Collecte Localisation Satellites (CLS), Toulouse, France
3
Centre National d’Etudes Spatiales (CNES), Toulouse, France

Corresponding author: [email protected]

Abstract—Relevant comprehension of flood hazards has Indeed, flood simulation and forecast capability have been
emerged as a crucial necessity, especially as the severity and greatly improved thanks to the advances in data assimilation
the occurrence of flood events may intensify with climate (DA). Such methods, notably Ensemble Kalman Filter
changes. Flood simulation and forecast capability have been (EnKF), aim at combining in-situ gauge measurements with
greatly improved thanks to advances in data assimilation. numerical models to correct the hydraulic states and reduce the
This approach combines in-situ gauge measurements with uncertainties in the model parameters (e.g., friction coeffi-
hydrodynamic models, which aims at correcting the cients, upstream inflow). These filters rely on the stochastic
hydraulic states and reducing the uncertainties in the model computation of the forecast error covariance matrix, within a
parameters, e.g., friction coefficients, inflow discharge. These limited number of simulations. The sources of uncertainty,
methods depend strongly on the availability and quality of represented by the control vector, is updated over each assim-
observations, thus requiring other data sources to improve ilation window. Nevertheless, this approach depends strongly
the flood simulation and forecast quality. Sentinel-1 images on the availability and quality of observations, as its perfor-
collected during a flood event were used to classify an mance relies on the spatial and temporal density of the observ-
observed scene into dry and wet areas. The study area ing network [2]. As a matter of fact, limnimetric in-situ obser-
concerns the Garonne Marmandaise catchment, and focuses vations providing water levels are only available at a few
on the recent flood event in January-February 2021. In this sparse locations along a river catchment, due to installation
paper, seven experiments are carried out, two in free run and maintenance costs [3]. This is a limiting factor for numer-
modes (FR1 and FR2) and five in data assimilation modes ical model precision in simulation and forecast, especially in
(DA1 to DA5). A model-observation bias was diagnosed and the floodplains. Such a situation requires efforts to leverage
corrected over the beginning of the flood event. Quantitative other sources of data such as remote sensing-derived flood
assessments are carried out involving 1D metrics at Vigicrue maps to validate and improve the flood simulation and fore-
observing stations and 2D metrics with respect to the casting performance. In this work, we carry out the flood ex-
Sentinel-1 derived flood extent maps. They demonstrate tent mapping by applying a Random Forest (RF) segmentation
improvements on flood extent representation thanks to the on Synthetic Aperture Radar (SAR) images such as Sentinel-
data assimilation and bias correction. 1 (S1) [4]. The inferred flood extent maps are then compared
with the flood extents simulated by TELEMAC-2D with
I. INTRODUCTION EnKF assimilation.
Simulations based on hydrodynamic numerical models in This work highlights the merits of using SAR-derived
analysis and forecast modes are crucial to mitigate flood flood extent maps to validate and improve the simulation
impacts. The analysis mode could be carried out to obtain results based on hydrodynamic numerical models with EnKF
better estimates of the dynamic footprints of past flood events, DA. It illustrates how SAR imagery data could be used to
as well as to assess flood damages and design future flood overcome the limits of the calibration and validation process
defense systems, whereas the forecast mode is used by civil which was done using river-gauge data only. For instance, a
security services and industry. However, these numerical bias between the models and in-situ observations has been
models remain imperfect because the uncertainties inherently identified and corrected, yielding better flood extent
existing within the models and the inputs, e.g., friction and representation. Quantitative performance assessments are
boundary conditions, seemingly translate into uncertainties in carried out by comparing the simulated and observed water
the model outputs. A well-established method for reducing level time-series at several in-situ gauge locations, as well as
uncertainties and generating more reliable predictions is to involving Critical Success Index measured between the
periodically adjust these models, for instance, by assimilating simulated flood extent maps and the SAR-derived maps. They
various observations as they become available [1]. underline the benefits of using spatially distributed remote
sensing data that inform on the floodplain dynamics.
 2020 TELEMAC User Conference Antwerp, BE, 14-15 October, 2021

II. STUDY AREA, DATA, MODEL and lastly, B% [m1/3.s-1] is the river bed and floodplain friction
Hydrodynamic numerical models, such as TELEMAC-2D coefficient using the Strickler formulation [7]. In order to
(www.opentelemac.org), are used to simulate and predict solve Eq. (1)-(3), initial conditions {L(', *, % = 0) =
water surface elevation and velocity from which the flood risk L2 (', *); "(', *, % = 0) = "2 (', *); #(', *, % = 0) =
can be assessed for lead times ranging from a couple of hours #2 (', *)} are provided, and boundary conditions (BC) are de-
to several days. TELEMAC-2D solves the Shallow Water scribed with a time-dependent hydrograph upstream and a rat-
Equations (SWE) with an explicit first-order time integration ing curve downstream. The Strickler coefficient is prescribed
scheme, a finite element scheme and an iterative conjugate as uniform over subdomains, and calibrated according to the
gradient method [5]. At each point within the mesh observing network.
representing the model topography and bathymetry (for mesh B. Study area
nodes in the river channel), the results of the simulation are The study area concerns the Garonne Marmandaise
water depth and velocity averaged over the azimuth axis. catchment (Southwest France) which extends over a 50-km
A. Shallow Water Equations (SWEs) in TELEMAC-2D reach of the Garonne River, between Tonneins, downstream of
the confluence with the river Lot, and La Réole. This part of
The non-conservative form of SWEs is written in terms of the valley is identified as an area at high flood risk. Since the
water depth (ℎ [m], also called water level) and horizontal 19th century, it has been equipped with infrastructures to
components of velocity (" and # [m.s-1]). They express mass protect the Garonne floodplain from flooding events such as
and momentum conservation averaged in the vertical the historic flood of 1875. A system of longitudinal dykes and
dimension while assuming that: weirs was progressively constructed to protect floodplains and
• The horizontal length scale is significantly greater than manage submersion and flood retention areas.
the vertical scale;
• Vertical pressure gradients are hydrostatic;
• Horizontal pressure gradients are due to the displacement
of the free surface.
As such, using similar mathematic notations to [6], the SWEs
read:
$ℎ $ $
+ (ℎ") + (ℎ#) = 0 (1)
$% $' $*
$" $" $" $0 1
+" +# = −/ + 1! + 23# 4ℎ5" 9999999999⃗
/782("); (2)
$% $' $* $' ℎ
$# $# $# $0 1
+" +# = −/ 9999999999⃗ (#); (3)
+ 1# + 23# 4ℎ5" /782
$% $' $* $* ℎ
where 0 [m NGF69] is the water surface elevation (ℎ = 0 −
0$ with 0$ [m NGF69] being the bottom elevation) and 5"
[m2.s-1] is the water diffusion coefficient. / [m.s-2] is the grav- (a) Mesh and friction zoning
9999999999⃗ are respectively
itational acceleration constant. 23# and /782
the divergence and gradient operators. 1! and 1# [m.s-2] are
the horizontal components of external forces (friction, wind
and atmospheric forces), defined as follows:
& &
⎧ 1! = − / "√" + # − 1 $E+,-
⎪ B%& ℎ '/) D* $'
⎪ 1 D+./
⎪ + F G HG & + G*,# &
ℎ D* 0 *,! *,!
(b) Topography and bathymetry (c) Hillshade representation
⎨ / #√"& + # & 1 $E+,- Figure 1: Garonne Marmandaise TELEMAC-2D model.
⎪1# = − B%& ℎ '/)
−
D* $*
⎪ A TELEMAC-2D model (Figure 1) was developed and
⎪ 1 D +./ calibrated over this catchment, which was built on a mesh of
+ F G HG & + G*,# &
⎩ ℎ D* 0 *,# *,! 41,000 nodes using bathymetric cross-sectional profiles and
topographic data [8]. It involves a triangular unstructured
where D* /D+./ [kg.m-3] is the water/air density, E+,- [Pa] mesh, with an increased mesh resolution around the dykes and
is the atmospheric pressure, G*,! and G*,# [m.s-1] are the hor- in the river bed. The local rating curve at Tonneins (established
izontal wind velocity components, F0 [-] is the wind drag co- from a number of water level-discharge measurements)
efficient that relates the free surface wind to the shear stress, translate the observed water levels into a discharge that is then
applied over the entire upstream interface, both river bed and

2
 2020 TELEMAC User Conference Antwerp, BE, 14-15 October, 2021

floodplain boundary cells. This modeling strategy was peak on February 4. In this work, we examine an extended
implemented by Electricité de France R&D to allow a cold length of this event, i.e., between January 16 and February 15.
start of the model with any inflow value. However, it prompts
an over-flooding of the upstream first meander (near
Tonneins), until the water returns to the river bed. The
downstream BC at La Réole is described with a local rating
curve. Over the simulation domain, the friction coefficient B%
is defined over four areas. Their values resulted from a
calibration procedure over a set of non-overflowing events and
are set respectively equal to: B%! = 45, B%" = 38 and B%# =
40 [m1/3s-1] for the upstream, middle and downstream part of
the river bed and B%$ = 17 [m1/3s-1] for the floodplain. They
are characterized by a discrete zoning of uniform B% values
into subdomains within the catchment, restricted by the
limited number of in-situ measurements. Such a friction
Figure 2: Water level time series at Vigicrue observing stations, and
coefficient setting is indeed prone to uncertainty related to the Sentinel-1 overpass times.
zoning assumption, the calibration procedure and the set of
calibration events. This uncertainty is more significant in the Figure 2 depicts the in-situ water level (15-minute time
floodplain area where no observing station is available. step) observed during the flood event at Vigicrue observing
The probability density function (PDF) for the Strickler stations: Tonneins (blue curve), Marmande (orange curve) and
coefficients is assumed to follow a gaussian distribution with La Réole (green curve). The event is observed by nine S1
mean and standard deviation set accordingly to the calibration images, indicated by the vertical black dashed lines in Figure
process and expert knowledge. The limited number of in-situ 2. S1 works as a constellation of two satellites in a phased
observations also yields errors in upstream inflow as the orbit, S1A and S1B, each with a 6-day revisit frequency. They
expression of the inflow relies on the use of the local rating are part of the Copernicus program launched by ESA with
curve, usually involves extrapolation for high flows. In order contributions from CNES. The flood peak was covered by the
to account for uncertainties in the upstream BC (i.e., time- ascending orbit 30 on February 2 18:55 and by the ascending
dependent discharge T34 (%)) while limiting the dimension of orbit 132 the next day on February 3 18:48.
the uncertain input space, the perturbation added to BC is
applied via a parametric formulation that allows for a III. METHODS
multiplicative, an additive and a time-shift error, as proposed
by [9]:
TU34 (%) = 8 × T34 (% − W) + X (4)
where (a, b, c) ∈ ℝ) , and their PDF follows gaussian distri-
bution, centered at their default values. The characteristics of
the friction- and inflow-related uncertainty PDFs are summa-
rized in Table 1. Other works dealing with this uncertainty
have been put forth using EnKF [10] or Extended Kalman fil-
ter [11]. Figure 3: Proposed methodology.

The merits of the complementary use of in-situ data for

Table 1: Gaussian PDF of uncertain input variables related to friction and assimilation and remote sensing (RS) data for validation are
inflow discharge coefficients.
assessed with different deterministic and ensemblist numerical
Vari-
Unit
Calibrated/de- Standard 95% confi- experiments. Four main steps to sum up the dynamic and
ables fault value ! ! deviation "" dence interval novelty of the proposed methodology (Figure 3):
5!! 6"/$ 7 %" 17 0.85 17 ± 1.67
##! 6"/$ 7 %" 45 2.25 45 ± 4.41 (1) DA combines in-situ data with this TELEMAC2D model
##" 6"/$ 7 %" 38 1.9 38 ± 3.72 to reduce the uncertainties within the model and inputs.
### 6"/$ 7 %" 40 2.0 40 ± 3.92 The control variables are the friction coefficients and
$ - 1 0.06 1 ± 0.118 inflow parametric coefficients;
% 6$ 7 %" 0 100 0 ± 196 (2) S1 imagery data is used to generate flood extent maps,
& s 0 900 0 ± 1760 represented by binary flood maps (black pixels: dry,
C. 2021 flood event and observations white pixels: wet);
(3) Several virtual observation regions are defined with
A substantial flood event occurred in late January and expertise at important areas within the simulation
February 2021 as it exceeded the yellow risk level alert set out domain. They measure the number of wet pixels within
by the French national flood forecasting center (SCHAPI) in defined boxes and allow to validate local performance of
collaboration with the departmental prefect, and reached its the flood simulations;

3
 2020 TELEMAC User Conference Antwerp, BE, 14-15 October, 2021

(4) Based on the performed 1D and 2D validations, technique is an alternative to anomalies inflation for avoiding
improvements are made upon the model and simulation the ensemble collapse, while preserving part of the
setting, allowing to better configure the control space, information from the background statistical characteristics. In
correct the existing bias between the model and the following implementation, g is respectively set equal to
observation, and so on. D,.
0.3. The background hydraulic state, denoted by t9 ,
A. Ensemble-based data assimilation algorithm (EnKF) associated with each member of the ensemble results from the
integration of the hydrodynamic model ℳ9 : ℝJ → ℝ- from
Continuous time-series of measured water levels and/or the control space to the model state (of dimension x) over b9 :
discharges recorded at discrete locations are traditionally used
for model calibration and validation of DA algorithms for real- t9D,. = ℳ9 lt9GF
+,.
, ^9D,. n (6)
time constraints of hydraulic flood prediction models [1] [12].
In this work, the measured water levels at the 3 Vigicrue The initial condition for ℳ9 at %%,+/, is provided by a user-
stations (Tonneins, Marmande and La Réole) are assimilated defined restart file for the first cycle. For the following cycles,
+,.
with the EnKF algorithm in the TELEMAC-2D Garonne it stems from the analyzed model state t9GF , saved from the
model presented previously to sequentially correct the friction previous cycle. The control vector equivalent in the observa-
and inflow discharge. tion space for each member, noted y9D,. , stems from:
1) Description of the control vector:
y9D,. = ℋ9 lt9D,. n (7)
The DA algorithm consists in a cycled stochastic EnKF,
where the control vector ^ is composed of the friction -
where ℋ9 : ℝ → ℝ H'()
is the observation operator from the
coefficients (four scalars B%% , 3 ∈ 0, . . ,3) and parameters that model state space to the observation space (of dimension
modify the time-dependent upstream BC (three scalars 8, X, W). dL$% ) that selects, extracts and eventually interpolates model
` denotes the size of the control vector. These seven outputs at times and locations of the observation vector y9L
parameters are assumed to be constant over a DA cycle, yet over b9 . It should be noted that, in the following, the obser-
their evolution in time is made possible by DA between cycles. vation operator may also include a bias correction to take into
The DA cycle a covers a time window, noted b9 = account a systematic model error. Eq. (7) thus reads
[%7;<=; , %>?@] of 12-hour length over which dAB7 in-situ
observations are assimilated. The cycling of the DA y9D,. = ℋ9 lt9D,. n − y$.+% (8)
algorithms consists in sliding the window by a period e%C.D, = where y$.+% is an a priori knowledge of the model-observation
6 hours so that the cycles b9 and b9EF overlap. The EnKF bias.
algorithm relies on the propagation of d" members with 3) EnKF analysis step:
perturbed values of ^ (denoted by ^ . ) into the forecast values The EnKF analysis step stands in the update of the control
denoted by ^9D,. with 3 ∈ [1, d" ] represents the ensemble and model state vectors. When applying a stochastic EnKF
member index. [13], the observation vector y L,. is perturbed, thus an ensemble
2) EnKF forecast step: of observations y9L,. (3 ∈ [1, d" ]) is generated:
The EnKF forecast step stands in the propagation in time y9L,. = y9L + {9 with {9 ∼ q(0, Ä 9 ) (9)
of the control and model state vectors over the assimilation
window b9 that gathers dAB7 observations. The EnKF is here where Ä 9 = sL$% & ÉH'() is the observation error covariance
applied to model parameters that, by definition, do not evolve matrix, here assumed to be diagonal, of standard deviation
in time over the b9 . In order to avoid ensemble collapse, sL$% (and ÉH'() is the dL$% × dL$% identity matrix), as the ob-
artificial dispersion is introduced within the sampling with the servation errors are assumed to be uncorrelated, Gaussian and
addition of perturbations θ to the difference between the mean with a standard deviation proportional to the observations
of the analysis from the previous cycle (a − 1) and the sL$%,9 = Ñy9L . The innovation vector over b9 is the difference
previous cycle analysis. The two terms are weighted by the
between the perturbed observation vector y9L,. and the model
hyperparameter g. The forecast step thus reads:
equivalent y9D,. from Eq. (7) (or Eq. (8)) and Eq. (9). It is
^2 + i.9 3j a = 1 weighted by the Kalman gain matrix Ü 9 and then added as a
^9D,. =h (5)
+
^9GF + +,.
gF 4^9GF +
− ^9GF ; + (1 − gF )i.9 3j a > 1 correction to the background control vector ^9D,. , so that the
analysis control vector ^9+,. is computed in Eq. (10),
+
with ^9GF = l∑.IF
& +,.
^9GF
H
nod" ∈ ℝJ and i.9 ∼ q(r, sK& ). ^9+,. = ^9D,. + Ü 9 ly9L,. − y9D,. n (10)
For the first cycle, the perturbed friction and upstream The Kalman gain reads:
GF
forcing coefficient values are drawn within the PDFs Ü 9 = á9!,# àá9#,# + Ä 9 â (11)
described in Table 1. For the next cycles, the set of coefficients
issued from the analysis at the previous cycle is further with á9#,#being the covariance matrix of the error in the back-
dispersed by combining the analysis anomalies with ground state equivalent in the observation space y9D and á9!,#
perturbations θ drawn from the gaussian distribution centered the covariance matrix between the error in the control vector
at 0 and with the standard deviation described in Table 1. This

4
 2020 TELEMAC User Conference Antwerp, BE, 14-15 October, 2021

and the error in y9D , stochastically estimated within the ensem- detected binary flood maps, a majority filter (with size of 3)
ble: was applied on the resulting flood binary map. The ground
1 M sampling distance of the S1 images and the derived flood
á9!,# = ä ã ∈ ℝJ×H'() binary map is 10 x 10 meters. The generated flood extent maps
d" 9 9 are used to improve flood visualization and reduce modeling
1 M
á9#,# = ã ã ∈ ℝH'()×H'() uncertainty.
d" 9 9
with C. Experimental setup
D,H Two free run simulations and five DA simulations were
ä9 = å^9D,F − ^9D , ⋯ , ^9 & − ^9D é ∈ ℝJ×H& implemented, summarized by Table 2. Their variations
ã9 = åy9D,F − y9D , ⋯ , y9
D,H&
− y9D é ∈ ℝH'()×H& concerns whether the experiment consists in a Free Run or
involves a DA approach, as well as whether or not the model-
where observation bias y$.+% is taken into account, and the value of
F F
^9D = ∑H D,. J D
.IF ^ 9 ∈ ℝ and y9 =
&
∑H D,.
.IF y9 ∈ ℝ
& H'() Ñ (for sL$% ). The diagnosed bias was estimated during the 24
H& H&
hours of January 15 which are composed of quasi-stationary
The analyzed hydrodynamic state, associated with each ana- non-overflowing discharge, result in y$.+%,MLJJ".J% = 0.72,
lyzed control vector ^9+,. is denoted by t9+,. . It results from the y$.+%,O+/-+JP" = 0.4, and y$.+%,QR = −0.24 meters.
integration of the hydrodynamic model ℳ9 with updated fric- Table 2: Summary of the realized experiments.
tion and upstream forcing over b9 , starting from the same Exp. Bias cor-
initial condition (for the first cycle), then from each back- name rection
DA !' Ñ (%) Control variables
ground simulation within the ensemble: FR1 No No 1 - -
t9+,. = ℳ9 lt9GF
+,.
, ^9+,. n (12) FR2 Yes No 1 - -
DA1 No Yes 24 15 ##[!:&] , $, %, &
B. Flood extent mapping using S1 images
DA2 Yes Yes 24 15 ##[!:&] , $, %, &
In recent years, SAR image data has been widely used in DA3 Yes Yes 24 1 ##[!:&] , $, %, &
flood management due to its ability to collect day and night
DA4 Yes Yes 24 99 ##[!:&] , $, %, &
images in all weather and to map flood extents in large areas
in near real-time. Water bodies and flooded areas usually DA5 Yes Yes 24 15 ##[!:&]
appear on SAR images with low backscatter intensity because
most of the incident radar signals are reflected away from the IV. RESULTS AND DISCUSSIONS
SAR antenna. Therefore, with a few exceptions such as the In terms of validation, two assessments are carried out.
built environment and vegetation areas, the detection of such First, a 1D validation between the simulated and observed
areas is straightforward on SAR images. time-series water levels at Vigicrue observing stations was
achieved. Second, 2D assessments between the flood extent
In this work, a Random Forest (RF) segmentation [14] was
maps simulated by TELEMAC-2D and those derived from
trained over a training dataset with 223 S1 images from 12
Sentinel-1 images by RF algorithm was carried out in order to
non-coastal Copernicus EMS Rapid mapping flood cases from
multiple regions of the world. The training dataset consists of validate the overall performance of simulated flood extents,
and to analyze the local behavior at individual virtual
permanent water pixel samples selected according to the
observation locations.
Global Surface Water Occurrence products [15].
A. Simulated water levels
Figure 5 depicts the water levels simulated by the
performed experiments (from FR1 to DA5) represented by
solid curve with respect to the observed water levels (black
dashed curve) at the 3 Vigicrue observing stations. These plots
focus on the period near the flood peak, i.e., between January
31 and February 9. The bottom panel on each sub-figure shows
the differences between the simulated and observed water
Figure 4: SAR S1 image (left) acquired on 2021-02-02 and the flood extent levels.
map inferred by RF (right).

Both VV and VH polarizations of S1 were used, in

conjunction with the use of the local slope derived from the
MERIT DEM [16]. The S1 images are calibrated and
orthorectified, then inferred by the RF algorithm to produce
the binary flood map, indicating flood (white pixels) and non-
flood areas (black pixels), as shown in Figure 4. Using cuML,
an open-source GPU-accelerated machine learning library, the
RF algorithm is able to generate a flood extent map in a couple
of minutes. To remove noises and artifacts in the resulting

5
 2020 TELEMAC User Conference Antwerp, BE, 14-15 October, 2021

observed water levels. For each metrics (each column) the best
score (i.e., lowest RMSE and lowest maximum absolute error)
is highlighted in boldface, whereas the second best is
underlined.
Considering the free run experiments (FR1 and FR2),
Table 3 shows that the bias correction leads to a better
agreement between the model and observations at Tonneins
and Marmande. This is however not the case for La Réole.
Next, with the involvement of DA, the resulting water level
errors have been reduced significantly. For example, from FR1
to DA1, the RMSE are reduced by 88.1%, 90.6% and 63.8%
at Tonneins, Marmande and La Réole, respectively. Overall,
DA3 yields the water levels the closest to the observations at
every Vigicrue stations (as highlighted by the boldfaced
values), both the RMSE and maximum absolute error, as a
result of the σSTU set equal to 0.01 in which the observation
vector was considered with very little perturbation (Eq. (9)).
On the other hand, high uncertainty of observation vector is
(a) Tonneins assumed for the DA4 experiment (σL$% = 0.99), which leads
to the resulting errors greater than DA2 and DA3.
In practice, an inflow discharge used as input for a
hydraulic model like TELEMAC-2D may originate from a
hydrologic model in which statistical corrections were also
carried out [17]. This leads to the assumption that no
uncertainty exists in the inflow discharge, based on which the
DA5 was realized. However, the comparison between DA2
(with 7 parameters in the control vector) and DA5 (with only
4 friction coefficients in the control vector) demonstrates the
improvements from DA5 to DA2 emphasized by the water
level errors, which are reduced both in terms of RMSE and
maximum absolute errors. This advocates for the
consideration of uncertainty within T34 (%).
Table 3: 1D assessment metrics w.r.t. in-situ data measured at Vigicrue
observing stations, computed over the whole simulation duration.
(b) Marmande Root-Mean-Square Error (m) Max Absolute Error (m)
Exp. La La
Tonneins Marmande Tonneins Marmande
name Réole Réole
FR1 0.756 0.625 0.409 1.062 1.870 1.721
FR2 0.102 0.338 0.505 0.404 1.472 1.956
DA1 0.090 0.059 0.148 0.456 0.352 0.690
DA2 0.084 0.104 0.138 0.330 0.590 0.676
DA3 0.034 0.034 0.043 0.238 0.258 0.212
DA4 0.171 0.260 0.298 0.756 1.319 1.295
DA5 0.093 0.226 0.480 0.299 1.036 1.796

C. Control vector analysis

The analyzed values from DA experiments for friction and
inflow parameters are shown in the first seven panels on
Figure 6. The inflow discharge T34 (%) is also depicted on the
eighth panel, and the two bottom panels reveal the water level
errors from FR1 and FR2 at the Vigicrue stations. Since the
(c) La Réole
model-observation bias is not taken into account in DA1, a
Figure 5: Simulated and observed time-series water levels around the larger increment is required in the control vector to make up
flood peak, between January 31 and February 9. for the difference between the model and observation. In order
B. 1D Assessment Metrics to make the model as close to the observation as possible, DA3
(with sL$% = 0.01) commits very large correction upon the
Table 3 summarizes the 1D assessment metrics (RMSE control vector, as shown by the large margin from the violet
and maximum absolute error) computed over the whole curves in all 7 control variable panels. Due to this constraint,
simulation duration, from January 16 to February 15, between
the friction coefficient for the floodplain B%$ is even
the water levels simulated by TELEMAC-2D and the

6
 2020 TELEMAC User Conference Antwerp, BE, 14-15 October, 2021

diminished below 10 m1/3s-1 near the flood peak which is not prediction is the number of non-flooded pixels incorrectly
a realistic value for the floodplain in this catchment. On the predicted as flooded, True Negatives (ed) is the number of
other hand, DA4 with a sL$% set equal to 0.99, i.e., full pixels correctly identified as non-flooded, and False Negatives
uncertainty on the observation vector, presents very small (1d) or under-prediction is the number of missed flooded
corrections upon the control vector. As such, the brown curves pixels. The RF-inferred flood extent maps are considered the
remain quite stable from the calibrated and default values. reference flood map based on which the TELEMAC-2D flood
Even though DA2 and DA5 experiments were configured with extent map will be evaluated.
the same sL$% , the respective increments on friction Table 4 summarizes the CSI measured from the
coefficients of DA5 (magenta curves) are much higher than experiments with respect to the RF-inferred flood extent maps.
those of DA2 (crimson curves). This is due to the fact that the On each day, the best score (i.e., highest CSI) is highlighted in
control vector in DA5 is only composed of the friction boldface, whereas the second best is underlined. At the
coefficients, whereas DA2 involves all 7 variables, hence the beginning of the event, the water prevails mainly in the river
contribution to compensate the water level error was shared bed and only occupies a small portion of the simulation
among them. domain. However, the CSI scores during these non-flooding
times were degraded because of several over-predicted regions
such as the numerical artificial flooding of the upstream first
meander (previously mentioned in subsection II.B). Let us
focus on the 2021-02-03 which is the S1 overpass time the
nearest to the flood peak. DA4 with the most uncertainty
hypothesized on the observation vector allows the highest CSI
score. However, reducing the sL$% from 0.99 (DA4) to 0.15
(DA2) only decreases the resulting CSI 0.13 points (i.e., from
63.97% to 63.84%). DA3, by forcing the simulated water
levels to be as close to the observation as possible, yields
significantly lower CSI, even smaller than the free runs. After
the flood peak, the interpretability of the CSI may also be
limited due to a number of water puddles that remain in the
floodplain as the model struggles to simulate the water
recession.
Table 4: CSI metric with respect to S1-derived flood extent maps.
CSI (%)
Exp. Jan 26 Jan 27 Jan 28 Feb 2 Feb 3 Feb 7
name 07:00 19:00 19:00 19:00 19:00 7:00
FR1
28.13 27.81 25.09 44.39 55.86 22.10
FR2
DA1 25.82 25.92 23.58 41.91 61.97 20.00
DA2 27.66 27.49 24.95 43.98 63.84 20.87
DA3 27.65 27.49 24.93 44.22 50.82 23.75
DA4 27.52 27.44 24.84 47.08 63.97 20.92
DA5 27.74 27.59 24.92 44.60 49.31 25.41

E. Analyzing local behavior at virtual observation regions

Several locations within the simulation domain were
Figure 6: Evolution of controlled variables. selected to analyze local behavior of the flood simulation,
D. Assessment of simulated flood extents especially in the floodplains, as shown by the rectangular
boxes in Figure 7. From these boxes, the number of wet pixels
The flood extent maps simulated by TELEMAC-2D from is counted from the flood extent maps (there are 9 in total, cf.
the experiments are generated by applying a threshold of 5cm Figure 2). These counts from the experiments (represented by
above which the node is considered wet (and dry otherwise).
the dashed curves) and the RF-inferred flood extent maps
A rasterization of TELEMAC-2D water level output field onto (black solid curve) are depicted in Figure 8.
a regular grid is also carried out. They are then compared with
the flood extent maps generated from the RF algorithm on S1
images. Critical Success Index (CSI) is widely used to assess
the performance of flood extent mapping, it is defined as
follows:
eE
CSI = (13)
eE + 1E + 1d
where True Positives (eE) is the number of pixels
correctly predicted as flooded, False Positives (1E) or over-

7
 2020 TELEMAC User Conference Antwerp, BE, 14-15 October, 2021

V. CONCLUSIONS
In this paper, flood extents observed from Sentinel-1
images were extracted using Random Forest and compared
with the flood extent maps simulated by TELEMAC-2D. The
study was carried out over the Garonne Marmandaise
catchment, focusing on the flood event occurred in January
and February 2021. Seven experiments were realized, namely
two in free run mode and five in DA mode. The DA was
performed using only in-situ observations, and it was
implemented by an EnKF with a 12-hour assimilation window
sliding with 6-hour overlapping between windows.
Figure 7: Virtual observation regions overlapped on a RF-inferred flood Several key remarks can be drawn from this work. First,
extent map, generated from S1 image on February 3.
ensemble-based data assimilation allows time-varying
All of virtual observation regions are filled very quickly correction of friction and inflow leading to improved
with respect to the S1 observations. One particular region is simulation and forecast in the river bed and the flood plain.
box number 4 in which the number of flooded pixels is very Second, it was shown that the bias correction leads to properly
low according to the observation, but it is over-flooded in the corrected water levels in the river bed and floodplains. The
experiments, with varying degree (FR2 and DA4 by a small uncertainty assumed on the observation vector through
margin, and the others by a lot). Such different behaviors parametrization of sL$% was also demonstrated, for which the
between boxes number 3, 4, and 5, despite they are all on the compromise between a good fit on simulated water level (i.e.,
floodplain and near to each other, advocate for an adjustment, DA3) and a high CSI/good agreement between flood extents
e.g., DA, which corrects the friction of the area around box (i.e., DA4). Lastly, several limitations concerning local
number 4. DA1, without taking into account the bias behaviors of the flood simulation have been revealed with the
correction, tends to overflood the whole catchment (i.e., for all virtual observation regions.
boxes). Lastly, the flood recession period associated with the Moving forward, a particular perspective for this study
last two S1 overpass times is poorly modelled, particularly for concerns resolving the limitation of the control vector size and
box number 2. This stems from the fact that evaporation and refining the spatial friction zoning definition, assimilating
ground infiltration physical processes are not accounted for in information from RS data such as number of wet pixels in the
the Garonne model. flood plain (which were used for validation in this work)
and/or flood extent information to improve the description of
the refined friction subdomains in the floodplains and to
calibrate them. This study paves the way toward a cost-
effective and reliable solution for flood forecasting and flood
risk assessment over poorly gauged or ungauged catchments,
thanks to the use of remote sensing data. Such developments,
once generalized, could potentially lead to hydrology-related
disaster risk mitigation in other regions. Future progresses
built upon this work will involve a more refined approach for
friction zoning and calibration, especially for the floodplains,
as well as improving the current models by assimilating flood
extent maps.
ACKNOWLEDGEMENT
This work was supported by CNES, CERFACS and SCO-
France. The authors gratefully thank the Electricité de France
R&D for providing the TELEMAC-2D model for the Garonne
River, SCHAPI, SPC Garonne-Tarn-Lot and SPC Gironde-
Adour-Dordogne for providing the in-situ data. Lastly, the
authors would like to thank the anonymous reviewer whose
comments and suggestions helped improve this manuscript.

Figure 8: Wet pixel counts at virtual observation regions.

8
 2020 TELEMAC User Conference Antwerp, BE, 14-15 October, 2021

REFERENCES
[1] J. Neal, P. Atkinson, and C. Hutton, “Flood inundation model updating
using an ensemble Kalman filter and spatially distributed measure-
ments,” J. Hydrol., 336(3-4), 2007, pp. 401-415.
[2] I. Mirouze, S. Ricci, and N. Goutal, “The impact of observation spatial
and temporal densification in an ensemble Kalman Filter,” in 26th TE-
LEMAC-MASCARET User Conference, 2019, Toulouse.
[3] D. Mason, G. Schumann, J. Neal, J. Garcia-Pintado, and P. Bates, “Au-
tomatic near real-time selection of flood water levels from high resolu-
tion Synthetic Aperture Radar images for assimilation into hydraulic
models: A case study,” Remote Sens. Environ., 124, 2012, pp. 705-716.
[4] P. Kettig et al., “The SCO-FloodDAM Project: New Observing Strate-
gies for Flood Detection, Alert and Rapid Mapping,” in 2021 IEEE In-
ternational Geoscience and Remote Sensing Symposium, 2021, in
press.
[5] J. M. Hervouet, “Hydrodynamics of free surface flows: modelling with
the finite element method,” John Wiley & Sons, 2007.
[6] Laborie, V., Ricci, S., De Lozzo, M., Goutal, N., Audouin, Y. and Ser-
gent, P., 2020. Quantifying forcing uncertainties in the hydrodynamics
of the Gironde estuary. Computational Geosciences, 24(1), pp.181-202.
[7] P. Gauckler, “Etudes Théoriques et Pratiques sur l'Ecoulement et le
Mouvement des Eaux,” Gauthier-Villars, 1867.
[8] A. Besnard and N. Goutal, “Comparison between 1D and 2D models
for hydraulic modeling of a floodplain: case of Garonne River,” Houille
Blanche-Revue Internationale De L‘Eau, vol. 3, 2011, pp. 42-47.
[9] Ricci, S., Piacentini, A., Thual, O., Pape, E. L., & Jonville, G. (2011).
Correction of upstream flow and hydraulic state with data assimilation
in the context of flood forecasting. Hydrology and Earth System Sci-
ences, 15(11), 3555-3575.
[10] Laborie, V., 2020. Quantification d'incertitudes et assimilation de
données pour la modélisation hydrodynamique bidimensionnelle:
application au modèle de prévision des hautes eaux de l’estuaire de la
Gironde (Doctoral dissertation, Université Paris-Est).
[11] Habert, J., Ricci, S., Le Pape, E., Thual, O., Piacentini, A., Goutal, N.,
Jonville, G. and Rochoux, M., 2016. Reduction of the uncertainties in
the water level-discharge relation of a 1D hydraulic model in the
context of operational flood forecasting. Journal of Hydrology, 532,
pp.52-64.
[12] Zaoui, F., Goeury, C., & Audouin, Y. (2018). Ensemble Integrations of
Telemac-Mascaret for the optimal model calibration. In Proceedings of
the XXVth TELEMAC-MASCARET User Conference, 9th to 11th Octo-
ber 2018, Norwich (pp. 169-175).
[13] M. Asch, M. Bocquet, and M. Nodet, “Data assimilation: methods, al-
gorithms, and applications,” Society for Industrial and Applied Mathe-
matics, 2016.
[14] M. Pal, “Random Forest classifier for remote sensing classifica-
tion,” Int. J. Remote Sens., 26(1), 2005, pp. 217-222.
[15] J. Pekel, A. Cottam, N. Gorelick, and A. Belward, “High-resolution
mapping of global surface water and its long-term changes,” Na-
ture, 540(7633), 2016, pp. 418-422.
[16] D. Yamazaki et al., “MERIT Hydro: a high-resolution global hydrogra-
phy map based on latest topography dataset,” Water Resources Re-
search, 55, 2019, pp. 5053– 5073.
[17] R. Hostache et al., “Near-real-time assimilation of SAR-derived flood
maps for improving flood forecasts,” Water Resour. Res., 54(8), 2018,
pp. 5516-5535.

View publication stats