A VISIBILITY TEST ON SPOT5 IMAGES
1Vinciane Lacroix , Arnaud Hincq , Idrissa Mahamadou , 1H. Bruynseels , Olivier Swartenbroekx
{Signal and Image Center Royal Military Academy, Belgium
Vinciane.Lacroix@elec.rma.ac.be
Institut Géographique National, Belgium
KEY WORDS: Cartography, Database, Detection, Matching, SPOT, Statistics, Test.
ABSTRACT
Two visibility tests have been made on a fusion of panchromatic and multi-spectral SPOTS images. The
tests differ in the type of regions (sub-urban versus rural) and in the panchromatic image resolution (5m
versus 3m). Two independent photo-interpretors had to extract the road network and the built-up area on
part of a georeferenced image. The resulting shapes were compared with the actual database. Two methods
were used for this comparison. It appears that the visibility is highly dependent on the experience of the
photo-interpretor. The expert reaches a rate of approximately 8596 of road detection in all regions, more
than 85% of the built-up area in the sub-urban region, and approximately 65% in the rural one.
1 INTRODUCTION
The context for this study is defined by the needs of
the NGI/Belgium, which tends to set up a unique
and seamless geographical information system on
the national territory. Associated to this global pur-
pose, a "Planning" project has to work out and set
up a tool to assist the annual production scheduling
(by measuring the outdating of topo-geographical
data).
The objectives pursued for the needs of the project
Planning are to bound changing zones with certain
reliability, but not to determine single objects very
accurately. Several data sources may be used to
evaluate changes on the ground, amongst them re-
mote sensed imagery is an obvious one. A com-
plete annual cover of the territory must be acquired
at an affordable cost to make a production project
effective. The other crucial issue is to determine at
which level objects can be detected on the scenes,
hence a "visiblity test". So what level of detail can
we expect from these images?
According to (Puissant and Weber, 2002), an ur-
ban object can be detected at 10 m, identified at
5 m and analyzed at one or less meter resolution.
Therefore, SPOTS Panchromatic Data at 5 m and
at 3 m resolution and Multi-spectral data at 10 and
20 m have been purchased. A first visibility test has
been performed over a sub-urban zone in the area
of Saint-Nicolas region (600 x 800), and a second
over a rural zone around Brussels (1602 x 2004),
both in Belgium. Part of the panchromatic images
are shown in Figure 1. In the first test, two indepen-
dent operators, qualified as “expert” and “novice”
based upon their experience in image interpreta-
tion, had to extract the built-up area and the road
network from a part of the georeferenced images
which had been computed using nearest neighbor
interpolation. The multi-spectral images were used
as transparency layers put on the panchromatic im-
age, adapting their transparency coefficient depend-
ing on the context.
The operator had to mark detected objects using
polygons, according to the representation used in
the database (DB). However, due to the difference
of scale between the image and the topo-geographic
DB, a small building was marked by a dot, a thin
building or a succession of small buildings by a
line, and the other type of building by a polygon;
a road narrower than 3 pixels was noted as a line
and a larger one as a polygon.
The visibility score depends on the method used to
compare the operators’ output with the DB. The
first method refers to the confusion matrix, often
used in classification; it shows how much the de-
tected objects and the DB overlap. The second me-
thod uses more sophisticated image processing tools
to match the detected objects with the DB objects.
2 CONFUSION MATRIX
The operator output is transformed in a classified
image, which is compared with the “ground truth”
image, a rasterized version of the three classes “built-
up area” (BUA), “road network” (RN) and “noth-
ing” (NO).
