You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Remote sensing for resources development and environmental management
Damen, M. C. J.

Symposium on Remote Sensing for Resources Development and Environmental Management / Enschede / August 1986
Per-field classification of a segmented SPOT simulated image
Commission of the European Communities, Joint Research Centre, Ispra Establishment, Italy
ABSTRACT: A per-field classification has been applied to a SPOT simulated image of an area in southern France
with mainly agricultural landuse. Using edge preserving smoothing, edge extraction, and edge tracking
techniques the field boundaries are derived from the panchromatic channel. The segmented image is replaced by
a property table in which the spectral signatures, position, and simple form describing factors of all
segments are stored.
RESUME: Une méthode de classification 'par champs' a été appliquée à une image de simulation SPOT dans une
région à prédominance agricole du sud de la France. Les limites de champs ont été déterminées sur le canal
panchromatique par des techniques de lissage conservant les contours, puis l'extraction et le suivi de ces
contours. On substitue à l'image ainsi segmentée une table d'attributs où sont mémorisées, pour chaque
segment, les signatures spectrales, la position et de simples facteurs de forme.
New satellite platforms provide us with more
detailed information due to the higher spectral and
spatial resolution. For research of the spectral
properties of objects methods will be used which have
proven their use in the past, only shrinked to a
space with less dimensions. However higher spatial
resolution will lead to detection of more objects in
a scene. Analysing a satellite image becomes more and
more a matter of spatial analysis. Research of the
spatial relationship of objects to for instance
landuse can not refer to established methods. Not
many attempts have been made partly due to the fact
that much effort has to be put in co-operation
between computer scientists and experts.
Only recent commercial software packages are
distributed on the common market which contain the
required basic processes needed for spatial analysis,
therefore opening possibilities for experts to get
involved in spatial analysis. When the development of
methods for spatial analyses and "reasoning" gets a
higher priority in research programs, better results
in this field are to be expected.
What I have tried in this direction is to design an
'automatic' per-field classification system to
analyse the spectral and spatial characteristics in a
complex simulated satellite scene and learn from it.
Automatic stands for the field boundaries detection
by segmentation and is placed in quotation marks
because parameters have to be fed to the system which
can influence the result considerably. Image
classification in this paper must be seen as a
display of intermediate results. To learn from the
analysis is more important.
The image under investigation is a subset from a
SPOT simulated image consisting of 231 x 216 pixels
in the panchromatic band of the "Les Vans" area of
the department of Ardeche in southern France(fig.
la). It is part of a least favoured area as defined
by the Commission of the European Communities. It is
a rather flat area with mainly agricultural landuse
and some outcrops of hard rocks surrounded by
woodlands. It was previously investigated by Megier
(1984) who with the help of digitized field borders
tested some parameters for the per-field
classification. These fields are used in the
following to measure the performance of a
classification and to analyse form parameters i.e. a
measure for elongatedness, irregularity, and rotation
of known fields. The SPOT satellite works well at
this time and will produce images from the 'Les Vans'
area soon. Probably this article is the last one
dealing with simulated images.
Segmentation of the image is required to obtain a
list of polygons describing the fields of landuse.
There are many ways to segment an image.
The procedure to prepare an image for segmentation
proposed here includes an edge preserving smoothing
technique. Edge extraction produces an edge-map. This
edge-map has to be cleaned before tracking. Edge
tracking and labeling is needed for using the
segments as individual and separated regions.
Fully aware of the existence of better segmentation
procedures I would like to propose a filtering method
based on the detection of local features. Other
methods have been tested. For the near future
segmentation by contouring will be developed but at
the time it is not ready.
2.1 Edge preserving smoothing
The edge preserving smoothing used is described in
Stakenborg(1985). A short explanation will be given.
Smoothing takes place with a 3 by 3 mask.
Thresholding of the differences from the central
pixel divides the kernel into a homogeneous and a
non-homogeneous part consequently creating a bit
pattern. If this bit pattern is found in a set of
pre-defined edge patterns the pixel is left alone
otherwise only the mean of the pixels under the
pattern is taken. This is a slight improvement of the
method proposed by Lev et.al.( 1977) . Threshold
selection is critical. By defining a threshold range
and counting the number of selected patterns from the
set for each pixel and each threshold value, the
'ideal' threshold can be determined and lies around
the peak of the threshold histogram.
Tests have shown that a threshold from 4 to 6 for
the Landsat Thematic Mapper channel 3 and threshold
10 for the SPOT simulated data channel 4 work well.
Blockdiagrams in figure 2 show the effect of the edge
preserving smoothing of the subset in figure la seen
from the N-E.
Other methods for edge preserving smoothing work