865 
verage of Xishuangbanna is illustrated 
in Figure 1. 
All processing has been executed on a 
I 2 S model 75 with S600 software, linked 
to a MicroVax as a host computer. 
Band 1 (blue wavelength) was found to 
be totally corrupted upon receipt and 
was dropped from further analysis. 
The TM image was subjected to a 
radiometric standardisation whereby 
digital numbers were converted to 
radiances using the method proposed by 
Price (1987). This standardisation is 
in fact not necessary for the type of 
exercise described here, but was 
conducted anyway since the TM image was 
also part of another exercise 
involving multitemporal analysis in 
conjunction with a Landsat MSS image. 
The geometric correction involved a 
projection to the UTM coordinate system 
using a cubic convolution procedure. 
From the complete frame , a subimage of 
805 x 515 pixels, containing the 
Mengyang Nature Reserve was extracted. 
3.3. Experimental set-up 
3.3.1. Vegetation classes 
The total amount of pixels that were 
identified in the field was divided at 
random in a subset for supervised clas 
sification training and a subset for 
classification accuracy testing. 
The list of training classes, and their 
respective amount of pixels is shown in 
Table 1. 
The individual classes were further 
aggregated to form a multilevel land 
cover classification system (Sabins 
1987) . 
3.3.2. Classification strategies 
Since the characteristics of the images 
and the areas to be classified vary 
greatly, it is not possible to put 
forward a certain classifier that will 
yield the best results (Campbell 1987). 
Hence alternative strategies have to be 
tested. 
The experimental set-up of this 
exercise contained several variables. 
The single-step classification was 
tested against a 'decision-tree' 
(Mather 1987) or 'layered clas 
sification'. This classification method 
has been found to be cost-effective in 
terms of CPU time, while yielding re 
sults similar to the single-step clas 
sifier (Lee and Richards 1985). 
The layered classifiers were divided 
into two types. In the first type the 
hierarchical structure was based on 
spectral subsets. In this case the 
tree structure was determined by 
visual inspection of graphs depicting 
mean radiance ± 1 standard deviation 
for every channel and every class . 
The second type had a fixed structure, 
based on the multilevel structure pre 
sented in Table 1. This corresponds 
with an information class hierarchical 
structure. 
Another experimental variable consisted 
of using all available TM channels vs. 
using only the SPOT-l-equivalent chan 
nels 2, 3 and 4. 
Furthermore, the use of single channels 
was tested against the use of the best 
subset of channels determined from the 
use of the average pairwise divergence 
as a separability measure (Swain 1978). 
Since almost all classes involved ve 
getation types, the use of the nor 
malized difference (ND) (Rouse et al. , 
cit. Bariou et al. 1985) at the first 
Table 1 List of training classes (with amount of training pixels between brackets) 
Class Level III 
Class Level II 
Class Level I 
1. tea (266) 
2. rubber (233) 
3. rubber/tea mixture (126) 
4. citrus (7) 
5. rice (155) 
6. mixed agriculture (109) 
7. recent shifting cultivation (43) 
8. weeds (Eupatoriinn sp.) (23) 
9. alpine meadows (140) 
1. plantation agriculture 
(632) 
2. early crops (330) 
1. agriculture (1102) 
3. grasslands (140) 
10. secondary forest (40) 
11. secondary forest 
(with Lithocarpus sp.) (10) 
4. secondary forest (146) 
12. forest with dominant bamboo (122) 
13. scrub with dominant bamboo (220) 
5. vegetation with 
bamboo (342) 
2. forest (1800) 
14. monsoon forest (1178) I 6. primary forest (1352) 
15. tropical rain forest (174) 
16. fishing pond (14) 
17. river (33) 
7. water (47) | 3. water (47)