351 
size. Cochran (1977, pp.73 on) shows that a sample size of 400 will 
give an error of estimate no greater than + 5 per cent at the 95 per 
cent confidence level. In future accuracy tests of larger areas, this 
figure of 400 will he used, reduced to some extent if this exceeds 
5 per cent of class population size (Cochran, 1977, p.76). 
For the sampling process, Berry and Baker (1968, pp.91-100) advocate 
the use of a stratified systematic unaligned sampling method. This is 
followed by Rosenfield et at (1982), but is difficult to implement when 
a set sample number for each class is required. Instead, in this study 
sampling is stratified according to class, i.e. random sampling is con 
ducted on each class separately until the desired class sample size is 
reached. This method was followed by Gurney (1981) and generally gives 
an even sample distribution unless class size is very small. 
The selection of the Grantown block for accuracy testing was due largely 
to the availability of good quality 1:40,000 scale aerial photography 
of the area from the same day as the satellite overpass. Lineprinter 
output of the sampled LANDSAT data was produced in fairly accurate map 
format and was compared with the interpreted air photo information at 
the same scale. 
DISCUSSION OF RESULTS 
The 83.5 per cent lower confidence limit for map accuracy almost 
attains the 85 per cent accuracy required by Anderson et al (1976), 
mapping classes which broadly correspond to 'Level 1' of the U.S. 
Geological Survey classification system. The large area which this 
project has to cover places constraints on the complexity of and the 
time available for classification. If the classification accuracy is 
viewed in this context, then the result is encouraging and acceptable. 
The lower confidence limits for each class are also generally accept 
able with the exception of classes which are poorly represented in this 
test block. The urban class consists mainly of the town of Grantown- 
on-Spey and accuracy would be improved by post-classification smoothing 
to remove isolated 'noisy' pixels. The shadow class contains pixels 
directly shadowed from illumination and can be added to the small per 
centage of unclassified pixels. 
In the analysis of this test block, the major problem to emerge is how 
to classify land cover types which lie on a continuum (such as heath- 
land, upland scrub, grassland) rather than being separated by distinct 
boundaries (e.g. woodland, urban, water). The spatially diffuse upland 
classes account for over 50 per cent of Scotland and over 80 per cent 
of this test block. Even by accurate comparison of the air photo data 
with the sampled pixels of these classes, it is difficult to be confi 
dent of placing a pixel into the correct class. This problem will 
become more acute in the analysis of higher upland areas and is impor 
tant to note when considering any accuracy figures derived for these 
classes. 
CONCLUSIONS 
In recent years much of the interest in developing techniques of 
increased accuracy for LANDSAT digital classification has tended to 
deflect attention from the fact that LANDSAT data is pre-eminently 
suited to the simple primary classification of land cover of very large 
areas. 
When applied to land cover mapping of an area of seven LANDSAT scenes