International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
2. METHODOLOGY
Earlier experiments proved that the use of laser scanner data or
multispectral images can be used to classify an urban scene and
detect buildings, but errors are expected for each data
source(Centeno et al, 2003; Centeno and Miqueles, 2003).
Therefore, the project aims at the cooperative use of both data
sources in the classification, attempting to detect buildings. À
Quickbird image and laser scanner data were used as sources.
The methodology consistes of three basic steps. Image
registration, region based segmentation and a posterior
classification of the regions, based on spectral and spatial
features. A supervised classification approach, based on fuzzy
logic, allows to describe and combine membership functions
derived for spatial and spectral features. The whole processing
was performed using the eCognition software. The quality of
the obtained result was evaluated, comparing it to the result of a
monocular restitution of the same area. In this section, we
introduce the study site and data set used. We then briefly
describe three different steps.
2.1 Study area and available data
The study area lies on the urban perimeter of the city of
Curitiba, Brazil. It is part of a residential area, characterized by
a mixture of elements such as residential and commercial
buildings, paved streets; stone covered pavement in the gardens,
roofs of different materials, and vegetation of different sizes
(trees and bushes). Two different data sets were available for
this region: a Quickbird image and a laser scanner data set.
Since both data sets don't cover exactly the same area, a region
within the overlap was chosen.
The Quickbird image was obtained on march, 2002. All four
multispectral channels and the panchromatic channel were
available. The pixel size of the multispectral image is about
2,8m while the pixel of the panchromatic image is about 0,7m.
Figure 1 shows a part of the Quickbird image of the study area.
The color composition uses the three bands of the visible.
a x LE PR
md A x
the study area
Figure 1.Quickbird image of
The laser scanner data consist of XYZI-ASCII files. The ASCII
files contain the 3-dimensional coordinates of each point and
also the intensity of the returned pulse. When the reflected
beam produced more returns, the first and the last pulse were
recorded. The laser scanner survey was obtained on may 2002,
using the Optech ALTM system by LACTEC (Instituto de
Tecnologia para o Desenvolvimento). In order to use laser
scanner data together with the satellite imagery within the
image processing software, a grid was interpolated from the
point data, with the same spatial resolution of the quickbird
image. The resulting grid, a digital surface model (DSM) was
later coded in integer values, taking a centimeter as unit.
The altitude grid is not relevant for the discrimination of
objects, because it includes the height of the terrain, the
topography. In order to avoid the effect of the topography, a
normalized DSM was computed, subtracting a DTM, already
available, from the original DSM. The so-called normalized
DSM (nDSM) is a representation of objects rising from the
terrain put on a plane (Weidner and Forstner, 1995). Figure 2
displays a perspective view of the nDSM. Darker pixels are
associated to lower elevations.
Figure 2 — 3D view of the laser scanner survey
2.2 Image Registration
Registration is a fundamental task in image processing used to
match two images. In our study, the laser data was considered
to be geometrically correct and was used to correct geometric
distortions of the satellite image. It must be taken into account
that both data were taken from different viewpoints and that the
laser data are post-processed and corrected in terms of
geometry, being represented in an orthogonal system. The
satellite image has distortions caused by the perspective view
and the inclination of the sensor in relation to the nadir. The
study area lies within a flat region. Therefore, the effect of
topography is minimal, but the effect of the height of the
buildings is a serious problem. In order to compensate the
relative displacement of the top of the buildings, a projective
model was used, based on the height of each building, derived
from the laser data. The result is a new image where the roof
are shifted closer to the base of the building.
2.3 Image Segmentation
The segmentation step consists of the division of the image in
uniform regions that correspond to the objects of interest, for
example roofs, vegetation and streets. A region growing
segmentation was used to obtain the elementary regions. This
algorithm merges adjacent pixels that have similar spectral
properties, forming regions that grow in an iterative process as
similar regions are melted together. The result is a new image,
where each pixel is labeled as member of only one region. The
algorithm called “fractal net evolution approach” (Baatz and
Schäpe 2000) avaliable in the eCognition software was used.
The algorithm allows to guide the region growing, using à
weighting function that controls the shape and spectral
uniformity of the regions. This guided segmentation allows to
improve region extraction, since the form of the elements
590
Interna
within
Antune
the pro
choose
24 C
In the «
each «
membe
interva
from d
also c
operate
differei
suitable
chosen
For the
after th
can be
topolog
consist
them i
experie
classifi
approa:
objects
segmer
objects
of the |
a mor
objects
objects
are col
stored
classifi
The da
themat
image,
differei
mainly
normal
satellite
not cor
compat
After
regions
“yards’
associa
Becaus
the the
the bu
buildin
In the
separat
present
of the «
buildin
size fil
difficul