lission
on Correct
Total %
141
26
50
27
139
73
45
69
104
74
192
74
183
64
149
82
85
69
11
6
207
89
185
48
119
79
91
87
277
85
705
93
295
41
/
+
I
V
X
/
/
nd 3
I9-0'89(u»w)
:he sixteen
From inspection of Fig 1, one would also
expect misclassification between the
spectrally-similar values of: European Larch
(1933) and (1949); Oak (1845) and (1915);
Douglas Fir (1966) and Norway Spruce/Scots
Pine (1966); and finally Corsican Pine
(1965/6) and Norway Spruce (1940). This is
borne out by reference to the confusion
matrix. The classification of Norway Spruce
(1940) would at first seem abnormal with a
classification accuracy of just 6%. In fact
93% of the training area has been
misclassified as Corsican Pine. A possible
explanation could be the result of the close
juxtaposition between the spectral values
(See Fig 1) but the higher variance in all
three bands of the Corsican Pine (See Table
2) , such that in the classification process
most pixels are classified as the latter.
At this stage it was decided to merge the
classes Oak (1845) and Oak (1915) . By
treating the two as one class, the
classification accuracy increases as shown at
the foot of Table 1. The European/Hybrid
Larch (1981) class would be more
appropriately labelled 'clearcut' as the very
young, well-spaced trees would contribute a
negligible proportion of reflectance within a
single pixel.
Table 2. Standard deviations about mean for
the 16 classes.
Band 1 Band 2 Band 3
Oak
1845
2.1
1.7
0.4
If
1915
2.2
1.4
0.5
II
1947/9
3.4
2.7
0.4
SwCh
1961
1.9
1.1
0.4
EL
1933
1.9
1.0
0.4
"
1949
2.6
1.0
0.4
EL/HL
1981
5.4
1.7
0.5
HL
1971
2.4
0.7
0.3
DF
1966
2.1
1.4
0.5
NS
1940
1.7
1.1
0.3
If
1971/2
2.8
0.9
0.4
NS/SP
1966
1.9
1.0
0.4
SP
1928
2.2
0.9
0.4
CP
1965/6
3.1
2.5
0.4
URBAN
10
8.7
0.8
AGRIC
8.2
5.3
0.6
Table 3.
Summary
of
Commis s
ion/Oruission
Errors for Evaluation Areas.
Class
Omission
Errors
Total %
Commission
Errors
Total %
Correct
Total %
Oak
1845/
1915
218
81
15
23
51
19
"
1947/9
72
61
89
65
47
39
SW CH
1961
88
100
3
100
0
0
E L
1933
170
81
11
22
40
19
**
1949
50
51
33
41
48
49
EL/HL
1981
34
55
7
20
28
45
H L
1971
67
66
3
8
34
34
D F
1966
84
47
67
41
96
53
N S
1940
101
99
0
0
1
1
”
1971/2
96
55
8
9
79
45
NS/SP
1966
150
69
17
20
67
31
S P
1928
42
51
4
9
41
49
C P
1965/6
24
21
200
68
92
79
URBAN
6
4
641
81
149
96
AGRIC
69
15
1
0
390
85
A more credible accuracy assessment is
based upon sites of known identity not used
in the training procedure. These were the
evaluation areas, and a summary of the errors
is listed in Table 3. As expected, the
overall accuracies have decreased though once
again misclassification of urban is the main
source of error. The total percentage of
pixels correctly classified is 42.9%. A
smoothing filter applied to the
classification slightly increased this to
43.6%. One would normally expect a greater
increase in accuracy than this, since the
smoothing filter eliminates stray, erronously
classified pixels within a homogenous stand
of trees. However, where certain stands have
been predominatly classified as urban, this
misclassification only becomes magnified.
The major problem to be addressed is the high
misclassification errors resulting from the
highly textured urban class. Simply omitting
this class form the training data would only
result in the misclassification of urban
areas as woodland. The drawback with the
classification algorithm is that it is based
on the spectral analysis of pixels purely on
an individual basis. Work is currently
underway in analysing the effect of applying
a textural classifier to the imagery. This
method has been shown to result in increased
classification accuracies in highly textured
regions. (Isako, 1979) .The texture band is
created by using a 'split and merge'
technique developed by Cross and Mason
(1985) . Preliminary results suggest a
significant increase in classification
accuracy. An alternative approach would be to
apply a 'layered' classification. By using
vegetation indices, vegetated and
non-vegetated regions are separated early in
the classification, preventing confusion
later in the classification process.
However, if errors are made at the first
stage they are carried on to lower levels,
regardless of the soundness of the later
decisions. If the urban areas can be
sufficiently well recognized in the enhanced
false colour composite, it is also possible
that they could be simply 'masked' from the
classification.
Given the hypothetical situation whereby
all the pixels misclassified as urban are
correctly classified, there is a drastic
increase in the proportions correctly
classified (See Table 4) . The effect of the
smoothing filter is also shown.
With reference back to Fig 1, the
distribution of the spectral curves appears
to agree with the findings of Mayer and Fox
III (1981), and later Kachhwaha (1983) .
Mayer and Fox concluded that band 5,6 and 7
Landsat MSS digital numbers have strong
correlation to the size and density of the
(coniferous) trees. High digital numbers
tended to be indicative of poor stocking in
band 5 (similar to SPOT band 2) and younger
trees in band 7 (closest to SPOT band 3) .
From Fig 1, the well-stocked stands of Norway
Spruce, Norway Spruce/Scots Pine, Corsican
Pine and Douglas Fir all have very similar
reflectance in band 2, but the younger stands
of Douglas Fir and Norway Spruce/Scots Pine
have a much higher reflectance in band 3. An
outlier is Norway Spruce (1971/2). According
to Mayer and Fox, this is typical for young
plantations of small numerous trees. One
might expect a lower overall band 2 digital
value due to the high tree density. However,
the small crown size and high percentage of
exposed bare soil probably causes the high
digital counts.