Table 1. Definitisn of cover classes 
2.6 Preprocessing 
514 
Cover classes Description Area-1 Area-2 
Forest 
Land covered by 
trees (woody plant 
with single stem 
and more than 5m. 
high), minimum 
area 2 ha. 
Closed forest 
Mixed broad-lea- V' 
ved forest, ca 
nopy closure>50% 
V 
Open forest 
Canopy closure, V 
20%-50% 
V 
Dense mixed 
bamboo forest 
Broad-leaved spe 
cies mixed with 
bamboos, canopy 
closure 750% 
V 
Scrub 
Vegetation type, 
the main woody 
elements of which 
are scrubs of 
more that 50cm. 
and less than 5m. 
height. They are 
woody plants with 
multiple stems of 
branching near 
the ground. 
V" 
Grassland 
Land covered by ч/ 
grass and her 
baceous plants, 
maximum height 
lm. 
Shifting 
Area which is v" 
cultivation under active 
cultivation, 
freshly burned 
or abandoned 
but regrowth 
is not more 
than two years 
old. 
Regrowth Areas left fallow 
long enough after 
shifting cultiva 
tion for vegeta 
tion to regene 
rate . 
Bare soil Land surface 
devoid of vege 
tation cover<20%. 
It includes built 
areas. 
Vindicates cover class belonging to either 
Area-1 or Area-2. 
2.5 Digital image processing 
Digital image processing was done on a 
microcomputer based interactive image proce 
ssing system housed in the Department of 
Geography, University ©f Reading,England 
with indigeneously developed software. 
Preprocessing of Landsat-2 MSS data was un 
dertaken in order to improve the image qua 
lity. 'Destriying' of the images was carried 
out by a histogram normalisation technique. 
Corrections for 'bit-slips' were also applied, 
3 CLASSIFICATION ANALYSIS 
3.1 Supervised classification 
A supervised multispectral image classifi 
cation procedure based on the minimum dis 
tance to means (Euclidean distance) algori 
thm was used in the study. In this classi 
fier a distance is computed for each pixel 
vector from the class means and the pixel 
is assigned to the class with the nearer 
means. Since this classifier is a special 
case of a more general maximum likelihood 
classifier and computationally can be pro 
grammed effecièntly it was thought ta be 
most suitable for implementing on the micro 
computer system. 
3.2 Reclassification 
In the classified image, usually, there are 
many isolated pixels whose classification 
is different from that of their neighbours. 
However, one would expect some degree of 
spatial dependence in land cover from pixel 
to pixel, if this spatial information can 
be incorporated in the classification pro 
cedure it would have the potential benefit 
of improving classification by removal of 
isolated inliers within homogeneous areas 
(Justice and Townshend 1982). One classifi 
cation smoothing algorithm is the majority 
filter which is illustrated in Fig.l. A 
spatial window of specified size (3x3, 5x5) 
is passed through the classified image and 
the center pixels classification is changed 
to the majority class of the surrounding 
pixels in the window. In this study the cla 
ssified images were reclassified using a 
3x3 majority filter. 
(a) Central pixel changed 
Original classification Reclassification 
Class Pixels 
AAA A 6 AAA 
ACB B 1 A A B 
CAA C 2 CAA 
(b) Central pixel unchanged 
Original classification Reclassification 
Class Pixels 
AAB A 3 A A B 
CCB B 3 CCB 
C B A c 3 C B A 
Figure 1. The 3-by -3 majority filter 
(Source: Schowengerdt 1983) 
4 ACCURACY ASSESSMENT 
After the classification, the results were 
evaluated to get an expression of its accu 
racy. In this study the accuracy assessment 
was conducted using 'test sets' and confu 
sion matrices. A confusion matrix is a square 
V" 
V \/