cover types from spectral reflectance signals. Both a 'supervised' and 
an 'unsupervised' approach have been used and automated digital classi 
fication schemes of varying complexity have been developed, from simple 
'minimum distance' and 'parallelepiped box' classifiers to the more 
complex 'Gaussian maximum likelihood' classifier. Each increase in 
sophistication of the classifier attempts to remove errors and ambi 
guities in the existing classification. Refining a classifier to mini 
mise the ambiguities and improve the classification accuracy has become 
almost an end in itself, with attention focused on the repeated classi 
fication of relatively small test areas. This has tended to shift the 
emphasis away from the study and mapping of large areas, the coverage 
of which is one of the principal advantages of LANDSAT over data from 
airborne sensors. 
In order to evaluate the suitability of LANDSAT for land cover mapping 
of a large area, a project was initiated to test the feasibility of 
producing a map of primary land cover types of mainland Scotland (about 
75,000 km 2 ) involving little or no conventional cartographic drafting. 
The intention was to use the Plessey IDP3000 image analyzer at the 
National Remote Sensing Center (Royal Aircraft Establishment, 
Farnborough, England) to effect an automated classification of land 
cover types based on a supervised approach using training data from 
aerial photography. The classified data would be output on film which 
would subsequently permit the creation of color separations for 
printing plates. A conventional color map would then be produced by 
offset-lithography, with the addition of a black plate to include 
essential cartographic elements such as names, grids and key. 
LAND COVER CLASSES 
The classes which can be included in a land cover types map are to a 
large extent constrained by the final scale of the printed map and by 
what is detectable from the data source (LANDSAT CCT). Since the aim 
is to produce a single map sheet of the whole of Scotland, the final 
scale, for practical reasons, is likely to be 1:500,000 or 1:625,000. 
The classes selected would, therefore, be a primary classification of 
land cover. Such a map of Scotland has not been produced since the 
19^*+ Land Utilisation of Scotland map (at 1:625,000) prepared by 
Professor L. Dudley Stamp from the comprehensive Land Use Survey of 
Great Britain, of which he was the prime mover from the 1930s on 
(Stamp, 1931). Stamp's map portrayed five primary categories of land 
cover: forest and woodland, arable land, medowland and permanent 
grass, heath and moorland and urban land. These classes serve as an 
indication of the level of detail which could reasonably be portrayed 
at 1:500,000 scale or smaller. 
Since it was anticipated that the final map would be of interest to 
regional and national bodies concerned with land use and resources 
planning, a questionnaire was circulated to twelve potential user 
organizations seeking their views on the classes which would be most 
useful. The results of this questionnaire formed the basis of the 
initial classification, modified subsequently by what was attainable 
from LANDSAT data. 
AN AUTOMATED APPROACH TO CLASSIFICATION 
Central to the project as initially conceived was the hire of the 
Plessey IDP3000 image analysis system (Balston and Custance, 1976) at 
the U.K. National Remote Sensing Center. The advantages of this system 
346