347 
include: the rapid application of certain pre-processing functions; 
interactive manipulation of LANDSAT data for a selected area on a color 
TV monitor which, when combined with training data, can establish land 
cover class parameters in single or multispectral mode and the computer 
extrapolation of this classification, via an 'emulator', to the 
remainder of the scene. 
Since mainland Scotland consists of seven complete LANDSAT scenes, it 
seemed that some sort of automated classification approach was essential. 
From the outset it was decided to use the CCTs which the National 
Remote Sensing Center had made for the production of a U.K. mosaic, 
since they were already geometrically corrected to fit the National 
Grid map reference system. The scenes used had also been resampled to 
give 100 m square pixels. The scene selected for the initial classi 
fication attempt was of the Central Highlands (220/22) since it covered 
a wide range of different landscapes, from coastal lowlands to the 
highest mountain plateau in Scotland. This scene was of 2h August 1976 
and, fortuitously, this was also the date of a large aerial photo 
graphic sortie in Scotland. The resultant photography, at scales from 
1:25,000 to 1:50,000, was invaluable in preparing land cover maps of 
'training' areas. These areas were selected to be representative of 
the lowland (Black Isle/Inverness) and the upland (Grantown area) land 
scapes in the scene. 
During the first experiment procedures tested on the IDP3000 for the 
'training' areas included: contrast stretching, principal components 
analysis, density slicing, band ratioing and multispectral classifi 
cation. At each stage the setting parameters for the IDP3000 were 
recorded and slides were taken of the color monitor to facilitate later 
decisions on the optimal choice for classification. On subsequent 
visits to Farnborough, the classification technique adopted was 
essentially a contrast stretch of the scene, to optimise the data, 
followed by a multispectral classification of the 'training' areas. 
This established a four-dimensional signature for each class, the para 
meters of which were used later by the 'emulator' to extend the classi 
fication to the rest of the scene. The classified tape was then used 
to drive a film writer. (LIN0SCAN 204d) at R.A.E., producing three film 
negatives of the blue, green and red components of the appropriate 
class color for each pixel. These negatives, at 1:1.2 million scale, 
were psed subsequently in the Geography Department at Aberdeen Uni 
versity to produce screened printing plates of yellow, magenta and cyan 
to allow printing of a preliminary color map of the classification. 
Examination of the resultant color map revealed significant areas which 
were misclassified or unclassified. It became apparent that extra 
polation of the classification beyond the 'training' areas gave poor 
results except where: 
(i) Land cover and topography closely resembled the training 
areas. 
(ii) Atmospheric haze was constant. Increased haze alters the 
reflectance values and consequently the classification 
parameters. There was variable haze in the scene 222/20. 
(iii) Areas were geographically close to 'training' areas, in 
which case factors such as field size and density of settle 
ment are likely to be fairly similar to the 'training' 
areas.