868 
55%. 
Bamboo forest is difficult to classify, 
which is probably due to the 
heterogeneity of the canopy. It is 
frequently misclassified as bamboo 
scrub. Difficulties in classifying open 
tropical forests have also been re 
ported by Singh (1987). The separation 
between the two classes of secondary 
forest is also rarely achieved. Re 
lative dominance of one tree species 
(Lithocarpus sp.) does apparently not 
alter the spectral response pattern of 
secondary forest. As can be seen for 
the Level II results, when merged, the 
secondary forest is rather accurately 
classified. 
Without too much loss of accuracy, a 
subset of TM channels is the best 
single method to classify natural 
vegetation into 'secondary forest’, 
'bamboo vegetation' and 'primary 
forest' . An accuracy of 85 % can be 
obtained, as opposed to 88% for the 
optimal combination (cascade method) of 
classification methods. 
For level I classification several 
methods attain comparable accuracies 
ranging between 90% and 95%. The 
combination of ND + TM5 allows for a 
93% correct classification of natural 
vegetation ('forest') 
From the complete error matrices, 
which, for conciseness are not 
displayed, it was also concluded that 
there is little confusion between agri 
culture classes and natural vegetation 
classes. Only agriculture class 6 
(mixed agriculture) overlaps spectrally 
with forest classes 11, 12 and 14. 
In almost all cases the use of the full 
TM channel set yields results that are 
superior to the SPOT-1-equivalent chan 
nels. 
5. CONCLUSIONS 
Despite the heterogeneity of cover 
types and the spectrally complex nature 
of the tropical forest in Mengyang, 
reasonable good classifications results 
could be obtained. It is shown that 
secondary forest can be separated from 
primary forest, which is an important 
asset to evaluate forest degradation in 
Mengyanq. Agriculture can be separated 
from natural vegetation (forest) clas 
ses with an accuracy of about 95%. 
However, no single classification 
method that yields the highest accuracy 
for all Level III classes, can be re 
commended. In order to obtain the best 
result, the classification methods have 
to be combined into a cascade-like 
structure, using a masking technique at 
every step. In this way it is possible 
to obtain an accuracy of 80 %. 
The obtained results are intended to be 
used in the framework of a nature 
conservation management database. Other 
data, such as habitat of endangered 
animals, location of illegal villages 
and cash crop plantations etc. should 
be digitally overlaid with the thematic 
vegetation map in a GIS concept. In 
fact, this would amount to a digital 
version of the type of work already 
done for the giant panda habitat in 
China (De Wulf et al. 1988) . 
ACKNOWLEDGEMENTS 
This paper presents research results of 
the Belgian Scientific Research Pro 
gramme for Remote Sensing from Space 
(TELSAT-II/04). The scientific res 
ponsibility is assumed by the authors. 
The international scientific 
cooperation, of which this paper 
contains the partial results, involved 
the gratefully acknowledged assistance 
of several persons and organisations. 
Logistic support for the field survey 
preparations was provided by Mr. Li 
Lukang (Ministry of Forestry, Beijing), 
the staff of WWF Hongkong, and Mrs. 
Pascale Moehrle and Mr. Christopher 
Elliott (WWF International, Gland). In 
the field, Mr. Bruno Verbist provided 
enjoyable company and invaluable help. 
Further assistance in China was pro 
vided by Prof. Yang Chun Jian and Mr. 
Wang Hong. 
Image preprocessing was taken care of 
by Mr. Dirk Tietens, while Mr. Peter 
Haelvoet helped in the paper editing 
phase. 
LITERATURE 
Bariou, R., Lecamus, D. and Le Henaff, 
F., 1985. Indices de végétation. 
Dossiers de Télédétection, Centre 
Régional de Télédétection, Rennes- 
France . 
Campbell, J.B., 1987. Introduction to 
remote sensing. The Guilford Press, 
London. 
De Wulf, R.R., Goossens, R.E., 
MacKinnon, J.R. and Wu Shencai , 1988. 
Remote sensing for wildlife management: 
Giant panda habitat mapping from 
Landsat MSS images. Geocarto 
International, 3 (1): 41-50. 
De Wu]f, R.R. and Goossens, R.E., 1989. 
Classification of small-scale forests 
in Flanders using TM digital data. In: 
Monitoring the Earth's environment, 
Guyenne, T.D. and Calabresi, G. 
(editors), ESA SP-1102, ESTEC, 
Moordwijk-The Netherlands. 
Lee, T. and Richards, J.A., 1985. A 
low-cost classifier for multitemporal 
applications. International Journal of 
Remote Sensing, 6 (8): 1405-1418.