narrow or absent. Real SPOT imagery of the United
Kingdom would be recorded in the morning and would
contain larger shadows, which are expected to be use
ful in detecting hedgerows.
2.2 Preprocessing of simulated imagery
The preprocessing of the simulated imagery included
two image rectification techniques applied to the raw
data recorded by the Daedalus scanner (Hunting Geology
and Geophysics Ltd. 1984). One was a linear scaling
technique by which the swath was scaled to fit a map.
The second technique was an S-bend correction which
rectified distortions caused by the radial rotation of
the scan mirror. This latter correction is unnecessary
in data from push-broom scanners like those on board
the SPOT-1 satellite.
During the course of the project, it was discovered
that the three simulated multispectral bands were not
properly registered to one another. Whether this prob
lem is due to preprocessing or subsequent processing
is presently unknown. The misregistration was verified
by two methods!
1. By locating the edges of obvious features in
printed arrays of DN values and comparing the loca
tions in each spectral band.
2. By visually comparing registered and unregistered
colour-composite images on adjacent display terminals.
This misregistration was corrected by shifting band
1 one pixel to the west and by shifting band 3 one
pixel to the east.
3 PREVIOUS METHODS FOR CLASSIFYING MIXED PIXELS
Figure 1. Classification zones in a hypothetical
feature space.
training set
mixed pixel
region
training set
In the literature, there are two types of methods for
classifying mixed pixels. In the first type, mixed
pixels are treated as whole entities and are assigned
to a single class. It is assumed that the class as
signment corresponds to the dominant constituent
class. Textural (Haralick 1979) or contextual informa
tion (Gurney & Townshend 1983) may be incorporated in
to the classification procedure.
The second type of classification method for han
dling mixed pixels involves pixel splitting, the proc
ess of breaking a pixel into its component parts and
classifying the fractions. Pixel splitting methods are
based on the premise that the gray tone represented by
a digital number of a pixel is proportional to the
gray tones of the constituent classes. Four statisti
cal procedures for pixel splitting have been found in
the literature! weighted averaging (Marsh et al. 1980),
linear regression (Nalepka et al. 1972, Richardson &
Weigand 1977), maximum likelihood classification
(Chittineni 1981, Horwitz et al. 1971), and linear
discriminant analysis (Marsh et al. 1980). The first
three of these procedures have been or could easily be
used in the three-class case, the usual condition for
pixels containing narrow woody features.
For a number of reasons, the methods already devel
oped seemed inappropriate for this project. Treating
mixed pixels as whole entities was not appropriate be
cause linear woody features are seldom the dominant
constituent class. The variances of the major classes
in the subscene are significantly different; the var
iance of woody vegetation is significantly larger than
the variance of any other class. Therefore, it seemed
unsuitable to use either weighted averaging which does
not take variance into consideration or linear dis
criminant analysis which requires the variances of all
classes to be equal. The linear regression method was
eliminated since development of the model requires
specific information regarding the proportions of com
ponent parts within mixed pixels. The small sizes of
the features of interest and the geometric distortions
in the image make collection of this specific informa
tion difficult. Finally, the maximum likelihood method
seemed inappropriate because it is the most computa
tionally demanding of the known methods. Since various
postprocessing algorithms are planned, it seemed de
sirable to conserve computer time.
Figure 2. Three parts of the 13 x 3 window.
4 NEW METHOD FOR CLASSIFYING SUBPIXEL WOODY FEATURES
The classification method developed for this project
discriminates selected subdivisions of feature space.
Figure 1 illustrates the subdivisions of interest in
a hypothetical feature space which was constructed us
ing two spectral channels and five pixels from each of
three classes. If the triangles and squares represent
two crops or any two land covers present in adjacent
fields, pixels containing the boundary between the
fields are expected to fall between the line joining
points B and C and the line joining points E and F.
If, however, the boundary between the fields includes
a linear woody feature and if the circles represent
woody vegetation, pixels containing this boundary are
more likely to fall within the polygon circumscribed
by points A, B, G, and D. Locating a pixel in one of
these two zones is the basis for classifying it in ei
ther the class of normal boundaries or the class of
small woody features.
Nine major classes present in the study site were
plotted in the two-dimensional feature spaces whose
axes are pairs of the multispectral bands. The pan
chromatic data was not used. Visual inspection of the
plots revealed that bands 1 and 2 are highly correlat
ed and that the greatest separation between classes
occurs when bands 2 and 3 are used. Therefore, only
bands 2 and 3 were included in subsequent processing;
band 1 was eliminated due to its redundancy.
A training set for woody vegetation, consisting of
80 pixels, was selected from a large patch of decidu
ous woodland in the subscene and was used in all fea
ture-space calculations.
A 13 x 3 window was employed to locate its central
pixels in the two-channel feature space. As shown in
Figure 2, the window was divided into three parts: two
training sets for the adjacent fields and a central
region expected to contain mixed pixels. These two
training sets and the training set for woody vegeta
tion were used to construct a feature space for each
window. The central pixels in the window were then
Figure 3* Fea-t
classified by
■ Examples of
Figures 3-5. I
vegetation, tr
pointing upwar
rived from the
tral pixels in
dicates the pi
L and S refer
the centre. In
of or below th
woody features
5 RESULTS AND :
Twenty subpixe!
boundaries werf
initial testing
greatly vary ir
rows, a hedge \
trees, and stri
All of the wooc
incorrectly del
classification.
to be either de
visual inspects
The classifie
correctly class
er than hedgero
field boundarie
the hedges. Fur
Use of the pa
into future cla
improve classif
Figures 3-5 1
training sets m
roughly triangu
sets in Figure ;
When adjacent f;
the clusters of
ure 4. In both <
woody features ;
woodland cluster
clusters are po:
Figure 5» pixel:
less likely to 1
the zone design:
only occurred or
Before this me
