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.

Ln a hypothetical
training set
x 3 window.
ped for this project
ons of feature space,
sions of interest in
ch was constructed us-
ve pixels from each of
and squares represent
present in adjacent
undary between the
sen the line joining
ing points E and F.
i the fields includes
e circles represent
Lng this boundary are
slygon circumscribed
lg a pixel in one of
classifying it in ei-
Les or the class of
the study site were
sature spaces whose
?al bands. The pan-
ial inspection of the
l are highly correlat-
.on between classes
sd. Therefore, only
ibsequent processing;
ition, consisting of
trge patch of decidu-
was used in all fea-
> locate its central
s space. As shown in
into three parts: two
.elds and a central
pixels. These two
¡t for woody vegeta-
.ture space for each
i window were then
L = 79 S = 149
Figure 3. Feature space created from window data.
classified by their locations in this space.
■ Examples of this classification method are shown in
Figures 3-5’ In these figures, squares represent woody
vegetation, triangles pointing downward and triangles
pointing upward represent the two training sets de
rived from the window, and circles represent the cen
tral pixels in the window. The circle with the dot in
dicates the pixel at the exact centre of the window.
L and S refer to the line and sample of the pixel in
the centre. In each example, the circles to the left
of or below the line are placed in the class of small
woody features. 5
Twenty subpixel woody features and two simple field
boundaries were selected from the study site for the
initial testing of the method. The woody features
greatly vary in width and length and include hedge
rows, a hedge with one large tree, single rows of
trees, and strips or patches two or more trees wide.
All of the woody features selected were found to be
incorrectly detected by a standard nearest neighbor
classification. Most of the features were also unable
to be either detected or adequately interpreted by
visual inspection of a colour-composite image.
The classification method developed in this project
correctly classified all subpixel woody features larg
er than hedgerows, properly discriminated the simple
field boundaries, and identified about 20 percent of
the hedges. Further testing is in progress.
Use of the panchromatic band will be incorporated
into future classification schemes and is expected to
improve classification accuracy for hedgerows.
Figures 3-5 illustrate the three possible ways the
training sets may be situated in feature space. The
roughly triangular shape, suggested by the training
sets in Figure 3, is the most typical arrangement.
When adjacent farm fields contain the same land cover,
the clusters of their training sets overlap as in Fig
ure 4. In both of these cases, pixels containing small
woody features are likely to be pulled toward the
woodland cluster. However, when the three training
clusters are positioned along a straight line, as in
Figure 5» pixels containing small woody features are
less likely to be pulled past the middle cluster into
the zone designating woody features. This last case
only occurred once in the 22 features tested.
Before this method can be completely operational, a
L = 134 S = 95
L = 57 S = 171
70 10 *0 ICO 110 120 130 1*0 150 140 170
Figure 5* Feature space created from window data.
few practical problems must be addressed. The most
important problem is determining an optimal size for
the window. The 13 x 3 window is frequently too big
to fit into the small comers of farm fields. This
inability to fit also contributes to the creation of
anomalies in the training sets. Four anomalous train
ing pixels are shown in Figure 6; the correct posi
tion of the division line in this feature space is
also shown. A smaller window, like a 9 x 1 window,
would fit into smaller fields and would create fewer
anomalies. A smaller window would also be more effi
cient since it would less frequently use the same
pixels in its calculations.
The automated use of this method could provide an
effective means of mapping and monitoring the pres
ence of subpixel woody features in satellite imagery.
Also, the use of this method could provide the basis
for determining specific quantitative information of
use to ecologists and resource managers.