International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
In the following section we sum up shadow charecteristics used
in most dtection methods and related works in shadow detection
and information restitution from shadow. The methodologies
used to detect shadow and retrieve information under shadow
are explained in section III. The data and sites studies are
presented in section IV. Results and their analysis come in
section V and we end with conclusion and some
recommandations.
2. RELATED WORKS
The main reason of shadows presence in remote snesing images
is the obstruction of the sun light by some high objects like tall
buildings. Surfaces under shadow are poorly lighted and appear
dark on the images. The form and size of the shadow depend
geometrically on the sun rise, the height and form of the object
which generates the shadow as well as the position of the
observer or the sensor. The most characteristics used in image
analysis for shadow zone detection or its effects compensation
are:
- The low value of shadow pixels in all the visible
bands (darkness of shadow regions);
- . Shadow is like a silhouette of the object generating it.
(so, its form is function of the object form).
- One or more sides of the shadow are oriented exactly
in the sun azimuth direction;
- The shadow size depend on the sun elevation and the
object height;
- Shadow have three components: the projected shade
which represents the silhouette of the object, the self
shadow which is the part of the object under its shade
and finally the penumbra located at the shade
periphery.
- Shadow do not modify the object colour (saturation
an taint).
- Some elements of surface texture are shadow
invariant. It mean that the texture of a surface do not
greatly change when shadowed.
Many works on shadow topic are devoted to the analysis of
video graphic images to detect moving objects in video
surveillance (Prati et a/.; 2000). In remote sensing, only few
works were carried out on the phenomenon. They are related to
the detection of the shades for the recognition of the buildings
(Hertas and Nevatia, 1988), (Liow and Pavlidis, 1990), the
detection of the shades of the clouds and correction of the effect
of shade due to the relief in the zones mountainous.
The majority of the methods for detecting shadow are based on
their low value level in all the spectral bands. Thus, a simple
threshold of histogram makes it possible to discriminate the
zones of shade (Gwinner et a/., 1997). But certain dark surfaces
with low value are merged with the zones of shades, so, the
need for integrating other properties or knowledge to
differentiate the shades from these other surfaces. Certain
assumptions on the vicinity and the form (right angle,
parallelism on the sides, etc.) (Chungan and Nevatia, 1998) are
used to improve discrimination of the shades.
Over methods use the invariant properties of the color like
saturation to discriminate the zones of shade on color images.
These invariant properties make it possible to detect surface
under the shade in spite of the strong difference of the intensity
(El Salvador et a/., 2001).
Recently few works on shadow compensation or information
retrieval are published (Nakajima et al., 2002, Rau et al.; 2000).
Nakajima et al. use ALS data to simulate shadow imagery at the
same configuration as the Ikonos data acquisition. That
simulated shadow is used to extract shadow from the Ikonos.
Shadow are eliminated by using a gamma transformation to
enhance the pixel value in the shadow.
Rau et al. use an local histogram balancing to compensate the
shadow effect. Other methods of shadow compensation are
based on the physical models simulating the sun and sky
illumination (Alder-Golden et a/., 2002).
Our method for shadow detection is based on a hierarchical
analysis of an segment attributes after a segmentation of the
image. The attributes used are : radiometric (mean value and
standard deviation), geometric (form and orientation),
contextual (vicinity, relative position of objects in sun side and
shadow side) and textural. For information recovery the method
use the textural attribute and the shadow neighbouring segments
in shadow side.
Methodologies for shadow detection and for information under
shadow recovery are described with the following section.
3. METHODOLOGY
3.1 Shadow detection methodology
The method is based on segment attributes analysis, so the first
step began with the image segmentation to produce
homogeneous zones (segments), and the calculation of all
attributes: spectral (average and standard deviation), form
(length, width, surface, compactness and orientation) and
contextual (vicinity, under-segments, etc).
Sun azimuth at
acquisition time
Segmentation and
segments attributes
calculation
z / Potential ES
3 { contrast analysis and
Shadow map by \ s
interpretation a
Panchromatic
Ikonos image
atm
Fining Detection by local
À histogram and form features 4
analvsis
ei
pu i nd da
Pa Comparing segments sides "v v
l orientation and sun jede
iro azimuth ee
rm epe
4 >
Detection validating by
y
Detection results anis sing fom of segments
; d 0; segmen
T VENE A
Figure 1: Diagram for the shadow detection
The hierarchical analysis of these attributes to detect shadow is
presented on figure 1 and the most steps are:
1. First detection by contrast analysis and threshold to
get all potential shadow zones. All segments darker
than its vicinity are considered. Segments with grey
level lower than a threshold are retained as potential
shadow zones.
International Arci
bo
Local hi
used to r
the proj
The forn
and the c
3. The neig
more ge
orientati
the sun
than a fi
as havin
azimuth
shade.
4. . Each se,
by analy
the sun
merged
meaning
can be a
other ca
confirm
projecte
3.2 Methodology
The information t
invariance by sha
not significantly
surface in shadow
between the shade
in the shadow sid
object are locate
located in the shac
The principal stag
are:
1. . For eac!
shadow
shadow
attribute
2. Analysi:
segment
analysis
between
the seg
differen
same te
two seg
3. The su
segment
identifie
classific
allowed
For the shad
transformation as
OutPixel =
Where: OutPixel
