The best results however, are obtained 
with the layered classification based 
on information class structure and 
single step classifications. The 
SPOT-equivalent channels score lower 
than the full TM data set. 
The layered classifier based on in 
formation classes scores unexpectedly 
high. This classification method is 
attractive from the user's point of 
view, but is not necessarily an 
appropriate model for digital clas 
sification. However, if Table 1 is 
examined more closely, it can be 
assumed that the information classes 
also are spectrally different. This is 
more likely to be the case with ve 
getation classes. This also explains 
the usefulness of the ND as the 
discriminating parameter at the first 
node . 
The HD has also been found to be very 
useful for detailed vegetation class 
discrimination in spectrally complex 
forest types such as reported for the 
montane forests in the Himalayas by Roy 
(1987) . 
From the results in Table 3 it was 
concluded that the best thematic map 
could be obtained by interactive use of 
different classification methods. This 
involved the construction of a 
cascade-like classification procedure, 
whereby in the first step 'secondary 
forest' was classified using method 
16. The classified pixels were then 
masked, and the rest of the image was 
classified using method 18 to separate 
'secondary forest with Lithocarpus' and 
'monsoon forest' from the rest of the 
image. In a third step, 'bamboo with 
trees dominant’ was separated using 
method 20. In a fourth and fifth step, 
'bamboo with scrub dominant’ and 'rain 
forest' were separated using metods 11 
and 14 respectively. In this way, it 
was possible to reach an accuracy of 
80% for the natural vegetation classes. 
This is admittedly a time-consuming 
procedure. The alternative is to resort 
to the best single-step method, using a 
subset of TM channels, in which case 
the classification accuracy of the 
vegetation classes would be as low as 
Table 3 Classification results expressed as correctly classified pixels. The class numbers refer to Table 1, the 
classification numbers between brackets refer to Table 2. The total amount of test pixels per class is 
listed imder "IT’. The khat value refers to the complete error matrix. 
Class Layered Single step 
Spectral classes Information classes 
TM 
'SPOT' 
TM 
SPCfT’ 
TM 
'SPOT' 
TP 
(1) 
(2) 
(3) 
(4) 
(5) 
(6) 
(7) 
(8) 
(9) 
(10) 
(11) 
(12) 
(13) 
(14) 
(15) 
(16) 
(17) 
(18) 
(19) 
(20) 
LEVEL 
III CLASSIFICATION 
10 45 
32 
31 
32 
29 
18 
17 
19 
19 
23 
29 
33 
16 
0 
17 
17 
40 
0 
25 
7 
7 
11 30 
18 
19 
18 
16 
16 
12 
0 
8 
0 
10 
3 
0 
0 
0 
4 
12 
19 
22 
0 
12 
12 45 
2 
3 
2 
2 
4 
7 
0 
7 
6 
10 
3 
0 
12 
0 
8 
0 
2 
6 
5 
28 
13 124 
36 
57 
45 
62 
40 
56 
23 
32 
49 
71 
106 
23 
50 
72 
13 
90 
0 
7 
47 
2 
14 362 
219 
211 
217 
211 
219 
246 
195 
118 
220 
118 
168 
146 
177 
137 
105 
242 
104 
313 
128 
280 
15 139 
97 
57 
100 
71 
97 
37 
56 
106 
35 
106 
12 
110 
81 
115 
28 
84 
46 
52 
75 
63 
khat 
.48 
.46 
.52 
.51 
.35 
.41 
.32 
.39 
.44 
.42 
.40 
.25 
.45 
.31 
.19 
.49 
.20 
.44 
.14 
.46 
LEVEL 
II CLASSIFICATION 
4 75 
61 
58 
61 
58 
43 
30 
36 
45 
52 
60 
60 
39 
4 
42 
36 
62 
31 
55 
20 
20 
5 169 
86 
128 
97 
129 
76 
99 
29 
89 
145 
155 
156 
41 
140 
107 
50 
154 
9 
54 
80 
104 
6 501 
383 
338 
386 
383 
383 
350 
386 
343 
364 
342 
303 
455 
330 
386 
249 
409 
207 
452 
323 
423 
khat 
.60 
.61 
.66 
.69 
.47 
.55 
.47 
.56 
.66 
.61 
.59 
.41 
.62 
.46 
.37 
.61 
.34 
.65 
.33 
.63 
LEVEL I CLASSIFICATION 
1 745 
637 
641 
638 
631 
643 
625 
583 
578 
646 
659 
668 
643 
654 
659 
525 
641 
471 
690 
556 
662 
2 737 
722 
725 
697 
733 
620 
671 
641 
680 
711 
693 
671 
717 
714 
693 
652 
725 
736 
708 
434 
704 
3 90 
49 
84 
88 
88 
49 
88 
48 
88 
88 
87 
87 
42 
88 
87 
80 
88 
86 
88 
2 
88 
khat 
.81 
.86 
.83 
.86 
.70 
.79 
.66 
.75 
.86 
.85 
.83 
.80 
.87 
.85 
.66 
.87 
.67 
.90 
.39 
.87 
867