but 
ded 
and 
and 
the 
2ast 
of 
the 
wv 
) and 
18, 
inal 
lS, 
iral 
the 
ER) 
31 
nal 
ced. 
Tt 
ve 
     
    
  
optimize the contrast in a given image so that more information can be 
extracted (Leberl et al., 1978). This involves band ratioing and canonical 
transformation of variables of data. According to Maxwell (1976) and Slater 
(1980), the ratios of corresponding pixels in two spectral bands can improve 
the signal-to-noise ratio of the classification process because the 
quantization noise in each spectral band will be reduced. Furthermore, the 
effects of some sensor radiometric errors and random changes in scene 
irradiance due to changing atmospheric conditions and topography of the 
area will be removed. The ratios of band 7/band 5 and band 5/band 4 were 
computed. Separate ratio maps were obtained from the line printer. In 
addition, the ratioed data were merged with the four channels of Landsat data 
for use in the classification. As for the canonical transformation, the 
purpose is to maximize the separation of classes by emphasizing the differences 
among the sample estimates of the means of the observations (i.e. the four 
spectral bands of the Landsat MSS data). Therefore, the training data for 
the different categories of land use were subjected to a canonical 
transformation. 
On the whole, a supervised approach involving the use of training sets 
provides the basis for the computer classification. The training data were 
derived from the 1:20 000 scale aerial photographs obtained in November 1979 
and the two sets of 1:25 000 scale aerial photographs obtained in December 1964 
and 1975 by the Government Photogrammetric Unit. The existing land use maps 
for 1966 and 1977 were also consulted. Basically, the classification program 
employed was the Euclidean distance classification, a linear classification 
rule which assigns each pixel to the class whose mean is closest to a 
limiting distance (or the threshold) set. In addition, another method of 
classification, the maximum likelihood classification, was also employed as a 
check on the accuracy. This makes the assumption that the data follow a 
multivariate normal distribution of probability. Each category is 
characterised by a characteristic mean vector and a variance-covariance matrix 
which defines the dispersion of the category population about the mean vector. 
The data are classified according to their weighted distances of separation 
from each of the categories as determined by the maximum likelihood 
criterion. 
ACCURACY EVALUATION 
The accuracy of the resultant land use maps (Figs 2 and 3) is evaluated 
with reference to the ground data obtained from the low-altitude aerial 
photographs. The selection of sample points from these land use maps for 
checking was on the stratified systematic unaligned technique which has been 
found to be the most bias-free sampling design (Berry and Baker, 1968). A 
transparent sheet of rectangles of 2.5 x 3.0 cm in dimension was placed on 
top of the land use map produced by the line printer. Each rectangular cell 
contained 100 picture elements at the scale of mapping. One picture element 
was selected in each cell according to the stratified systematic unaligned 
sampling method. The 85 per cent criterion of accuracy requires a sample size 
of at least 20 points for each category of land use before a valid assessment 
of accuracy can be accepted (Van Genderen and Lock, 1977). . In addition to 
this, the category accuracies were weighted by the percent area of each 
category in order to arrive at a more realistic evaluation (Fitzpatrick-Lins, 
1981). 
Tables 1, 2 and 3 summarize the results of this analysis for all the 
land use maps produced. The following points can be noted: 
(1) The accuracy of the resultant 1978 land use map for the whole of 
Hong Kong at 1:100 000 scale is (73.4%) lower than that for the urban area at 
1:25 000 (77.02) for the method of Euclidean classification. No improvement 
911