UNCERTAINTY IN CALIBRATING FLOOD PROPAGATION MODELS WITH FLOOD
BOUNDARIES DERIVED FROM SYNTHETIC APERTURE RADAR IMAGERY
P. Matgen * *, J-B. Henry ^, F. Pappenberger ©, P. de Fraipont^, L. Hoffmann", L. Pfister *
* Centre de Recherche Public — Gabriel Lippmann (CRP-GL), L-1511 Luxembourg, Luxembourg — matgen@crpgl.lu
? Service Régional de Traitement d'Image et de Télédétection (SERTIT), Parc d'Innovation, Bld S. Brant, BP 10413,
F-67412 Illkirch Cedex, France — jb@sertit.u-strasbg.fr
* Joint Research Centre of the European Commission (JRC), Weather Driven Natural Hazards, IES, JRC, Italy —
f.pappenberger(a)lancaster.ac.uk
KEY WORDS: Remote Sensing, Floods, Classification, Calibration, SAR, Performance, Fuzzy Logic
ABSTRACT:
An instantaneous synthetic aperture radar (SAR) derived flood extent map helps retrieving the distributed conveyance parameters in
one-dimensional flood routing models. These models are generally calibrated based on the sole use of ground data. This research
aims to use earth observation (EO) data in order to establish a significant parameter retrieval strategy, providing an alternative model
calibration technique. Owing to model structural errors, parameter equifinality and the fuzziness of the available radar and ground
data used for calibration, there are some uncertainties with respect to the model predictions. In order to assess these uncertainties in a
statistical framework, Monte Carlo simulations of a well-documented flood event in the Alzette river floodplain, Luxembourg are
used to explore the parameter space of roughness coefficients. It is shown that many parameter sets perform equally well. The
subsequent generalized likelihood uncertainty estimation (GLUE) methodology is used to compare both calibration strategies. Due
to the coarse resolution of the available radar scenes and the difficulty in defining an appropriate radar backscattering threshold
value during the inundation delineation, the exact flood extent is relatively uncertain. Hence, it is recommended to use a fuzzy-rule
based calibration procedure with the available instantaneous flood boundaries derived from ERS and Envisat radar scenes. The
uncertainty bounds of the flood extension predictions are assessed for the two types of calibration procedures based on ground
survey data and earth observation data respectively. It is shown that both techniques provide similar performances. By combining
EO data with ground based data in the calibration procedure, the parameter space will be constrained providing more reliable flood
extension predictions i.e. with narrower uncertainty bounds. This study shows that earth observation data are very useful for
hydraulic model calibration and that their combined use with ground data provides more accurate inundation simulations.
1. INTRODUCTION 2003). Many different parameter values give rise to almost
equivalent model simulations in terms of performance measures
Despite the physical appeal of the river flow computations in related to the different reference data, a phenomenon commonly
hydraulic flood propagation models, calibration remains a termed equifinality (Beven, 1993). In fact, potential error
compulsory stage in inundation modelling. Elevations of high sources are numerous in hydraulic modelling (model
water marks and aerial photography are the most commonly uncertainty, parameter uncertainty, error in observed data) and
used ground data during model calibration. Due to its large no set of calibrated parameters enables an entirely successful
footprint, its all weather, day and night capabilities, the simulation of all state variables at each stage of the flood event.
synthetic aperture radar (SAR) imagery shows considerable ^ However, the uncertainties of the predicted spatially distributed
advantages over these ground data as well as over remotely flood extent can be reduced when combining objective
sensed data obtained by sensors operating at visible functions based on different types of observations available
wavelengths. Maps of flooded areas can be obtained very after the flood event. The uncertainty reduction by introducing
quickly and at low cost. Moreover, the sparseness of punctual multi-response data describing several aspects of the modelled
ground data often hampers the calibration and evaluation of system has been widely explored in recent years (Freer ct al,
distributed roughness parameters in hydraulic models. Hence, 2004).
the main objective of this study is to establish a parameter
retrieval strategy based on SAR data, thus providing an However, in the past only a few studies have investigated the
alternative model calibration technique that will be compared to potential use of earth observation data in model calibration (see
the more common one based on traditional mapping methods. for instance Horritt et al., 2001). This is why we intend in this
study to find out if the inclusion of further earth observations,
As pointed out by Aronica et al. (1998), the application of a which the model is required to replicate, can ultimately lead to
fuzzy rule based calibration technique along with a generalized a significant reduction of uncertainties related to the model
likelihood uncertainty estimation (GLUE) procedure constitutes predictions.
a new paradigm in the way we interpret the predictions of
physically based models (see also Romanowicz and Beven,
* Corresponding author.
352
Internat
2T
A flood
length,
river re
severe f
flood ev
5 years
compris
discharg
reach, st
maximu
flooding
and at t
Envisat
stages o
alluvial
(one of t
flows th
use eart
convenie
Discharge (m ?s)
Figure 1
2.1 Pre-
The SAR
(5.3 GHz
resolutior
Envisat s
acquire d
of like- a
for flood
SAR dat
combinat
polarisati
differenti:
The rada
orthorecti
geometric
horizontal
value. Th
the backs
Speckle n
moving a
