×

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

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

426
2.2 Image acquisition
The study site was flown on 13th May 1984 as
part of the UK SPOT Simulation Campaign, and
the data was received some months later in
the form of a computer-compatible tape. The
20 m resolution multispectral digital data
was acquired, but not the 10 m resolution
panchromatic data.
2.3 Analysis techniques
Computer-assisted interpretation of the image
data was carried out using the GEMS image
processing system at the National Remote
Sensing Centre, Farnborough, and at Silsoe
College, Bedford. A manual contrast stretch
was interactively selected to give the best
visual definition and separation of tree
types within the forest. A maximum
likelihood classification was then applied to
the resulting enhanced image. The training
areas were delineated using 1:10,000
Forestry Commission stock maps which provide
information on species composition, exact
location and date of planting.
At least two sets of training areas were
compiled for each class within the
classification. The first set, called
'training' data, were used to develop the
statistics for the classification algorithm,
and were also used for a preliminary accuracy
assessment. The second set of training
areas, called the 'evaluation' data, were
used as an independent assessment of the
classification accuracy.
Based upon a visual inspection of the
original enhanced image, it was decided to
proceed with the selection of training areas
at Anderson level III. No attempt was made
at this stage to use ratio, principal
component or texture algorithms in the
classification.
3 RESULTS AND DISCUSSION
The final supervised classification used 16
interpreter-defined classes. The simplest
form of accuracy measurement is to compare
the classified data with the training data
used to generate the classification. This is
presented in the form of a confusion matrix
(see Dury et al 1986) . Inspection of the
confusion matrix reveals the kinds of errors
generated by the classification process
(Table 1). The spectral curve patterns (Fig
1) and standard deviations (Table 2) for each
class also help to explain errors.
The most: common errors of omission are the
assignment of areas of forest to the urban
category. All but one species, Norway
Spruce (planted 1971) have been misclassified
as urban. Both Oak (1845) and Oak (1915)
have a massive 60% of the training area
misclassified as urban. However the younger
stands of Oak (1947/9) were not affected to
the same extent. This over-classification of
the urban class is attributable to the high
degree of spectral variance within the class
(compare the standard deviations for the
urban class in Table 2 with the other
classes) . It is interesting to note that
although the agricultural class has high
spectral variance, it has a very high mean in
Band 3, the near infra-red, causing it to
remain spectrally distinct from other
classes.
Table 1. Summary of Commission/Omission
Errors
Class
Omission
Errors
Total %
Commission
Errors
Total %
Correct
Total
%
Oak
1845
399
74
29
17
141
26
If
1915
132
73
84
63
50
27
If
1947/9
52
27
63
31
139
73
SW CH
1961
20
31
10
18
45
69
E L
1933
37
26
24
19
104
74
ft
1949
67
26
33
15
192
74
EL/HL
1981
104
36
9
5
183
64
H L
1971
33
18
5
3
149
82
D F
1966
38
31
134
61
85
69
N S
1940
187
94
2
18
11
6
ft
1971/2
26
11
0
0
207
89
NS/SP
1966
197
52
23
11
185
48
S P
1928
32
21
11
8
119
79
C P
1965/6
13
13
249
73
91
87
URBAN
48
15
641
70
277
85
AGRIC
42
7
33
4
705
93
Oak
1845 &
1915
427
59
9
3
295
41
Fig 1. Spectral curves for the sixteen
cover-types.