OBIA Flood Delimitation Assisted by Threshold
Determination with Principal Component Analysis
Dora Roque, Nuno Afonso, Ana M. Fonseca, and Sandra Heleno
Abstract
Accurate and rapidly mapped flood boundaries are extremely
important for emergency management operations and hy-
draulic flood model calibration. This study presents a meth-
odology for automatic flood delimitation in
SAR
images. An
Object-based Image Analysis (
OBIA
) is applied to
SAR
images
and to a Digital Terrain Model (
DTM
), organizing a database
for hydraulic flood models calibration. Principal Component
Analysis (
PCA
) is proposed to automate the determination of
flood / non-flood decision thresholds. A previous classifica-
tion, with a visual threshold selection, is performed for a
small set of training images. A first
PCA
detects correlation
between the training thresholds, image, and flood parameters;
while a second
PCA
allows the automatic determination of
the threshold for the remaining dataset classification. For
the quality assessment, averages of 88 percent for properly
detected flooded area and of 10 percent for commission er-
ror are achieved. It is verified that the algorithm performs
well for images acquired during most weather conditions.
Introduction
Floods are one of the most frequent natural disasters world-
wide and are responsible for half of the human losses caused
by catastrophes (Schumann
et al
., 2010). Accurate and rapid
flood extension maps are important for local authorities (Shan
et al
., 2010), because they enable the identification of affected
zones and support response and recovery operations. Flood
maps are also useful for prevention, since they can be used
for calibration of hydraulic flood simulation models and,
therefore, inform efficient urban development.
Optical satellite images are often used for flood extension
assessment. For example, Sun
et al
. (2012) use
MODIS
images
to obtain water fraction and flood maps through a Regression
Tree, while Zhang
et al
. (2012) utilizes Brightness tempera-
tures (also from
MODIS
data) for mapping nighttime floods.
Optical data can also be used to refine the flood maps derived
from images acquired with other sensors (Shan
et al
., 2010).
However, optical images have limitations regarding flood
identification, such as the impossibility to observe the surface
in cloudy conditions. Earth observation data obtained through
active microwave sensors with synthetic aperture, called
SAR
,
can overcome those limitations, since they can be acquired in
all-weather conditions and during both day and night. Flood
extraction from
SAR
images may seem intuitive, since smooth
water surfaces present low backscatter coefficient values. Nev-
ertheless, strong wind or rain and the presence of vegetation
roughen the water surface, leading to radiometric similarity
to other land cover classes, which increases the complexity of
accurate flood extraction.
Many recent flood assessment studies have been performed
using
SAR
images.
RGB
color composite and thresholding are
among the most used techniques (Angiati and Dellepiane,
2009; Chini
et al
., 2012; Martinis
et al
., 2009; Matgen
et al
.,
2007 and 2011). Matgen
et al
. (2011), in particular, perform
a pixel-based thresholding with threshold selection based
on the statistical modeling of the image histogram. A region
growing approach, possibly followed by an optional change
detection analysis, enables the automatic flood delimitation.
Multi-seed growing segmentation procedures are also found
in literature, in which selected water pixels grow to their
neighborhood, on a pixel-by-pixel basis (Angiati and Del-
lepiane, 2009). Active Contour Models are frequently used
for flood delineation (Horritt
et al
., 2001; Matgen
et al
., 2007;
Silveira and Heleno, 2008), on which an initial water region
is selected and its neighborhood is evaluated in order to mini-
mize an energy function. Most research projects consider only
SAR
images to perform this analysis, but Mason
et al
., (2007
and 2010) also use a Digital Terrain Model (
DTM
), in which
a constraint in the vertical component avoids the growth of
the contour to regions with higher altitude. Other studies also
consider altitude as an important element for flood assess-
ment, such as Martinis
et al
. (2009) who use a Digital Eleva-
tion Model (
DEM
) to identify areas surrounded by water and
with lower altitude than the radiometry-classified flooded
regions. Besides
SAR
imagery and
DEM
, Pulvirenti
et al
.
(2011) and Pierdicca
et al
. (2013) use land cover information,
combining radiometry, distance to water bodies and altitude
difference in relation to water bodies through fuzzy logic. The
increase in the number of classification algorithms using
DEMs
is expectable, since the availability of high resolution global
DEMs
is growing, for example the
TanDEM
-
X
global
DEM
.
The accuracy of most flood detection algorithms using
SAR
amplitude data decreases for water surfaces disturbed by
wind, rain or emerged elements (Mason
et al
., 2007 and 2010;
Zwenzner and Voigt, 2009). Besides, the classification method
must be applied automatically, in order to minimize the
processing time and the subjectivity of results. In this study,
an Object-based Image Analysis (
OBIA
) leads to an automatic
delineation of flooded areas. A radiometry-based thresholding
enables the identification of undisturbed flood corresponding
objects. Those seed objects are used to start a region growing
procedure based on both radiometry and altimetry informa-
tion. This method enables the classification of flooded areas
with disturbed (rough) surfaces. The data gathered during
a previous classification of a small training dataset of im-
ages with visual threshold selection is used to calculate the
thresholds to classify a validation dataset, through a Principal
Dora Roque, Nuno Afonso, and Ana M. Fonseca are at Labo-
ratório Nacional de Engenharia Civil, Av. do Brasil 101, 1700-
066 Lisboa, Portugal (
).
Sandra Heleno is at CERENA/Instituto Superior Técnico, Av.
Rovisco Pais 1, 1049-001 Lisboa, Portugal.
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 5, June 2014, pp. 551–557.
0099-1112/14/8006–551
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.6.551
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2014
551
