A MODIFIED, SEMI-AUTOMATED, APPROACH TO CLASSIFICATION 
As a result of the problems encountered with the automated classifi 
cation using the IDP3000 system, a modified approach was necessary. 
It would be feasible to increase the number of 'training' areas used on 
the IDP3000 in an effort to refine the classification parameters, but 
this would involve more image processing time and multiple runs of the 
'emulator', thus greatly increasing the cost. Since seven scenes had 
to be classified, this was not an economic proposition. 
It was decided to modify the method by placing greater emphasis on the 
use of software classifiers on the Honeywell 66/80 mainframe computer 
at Aberdeen University and to develop an appropriate technique for 
editing gross anomalies in the classification due to topographic and 
atmospheric effects. To facilitate storage of the classification of 
the geometrically corrected scenes, a data matrix was created which is 
effectively a mosaic of the seven scenes covering mainland Scotland. 
This matrix is 3773 by 5534 pixels in size, with joins between scenes 
along N-S or E-W grid lines. 
Scene 222/20 was then reclassified by training separately on nine sub 
scenes using a 'minimum distance to mean' classifier. This was 
generally preceded by various pre-processing transformations, especially 
band ratios and intensity normalization, to attenuate varied illumi 
nation effects particularly in upland areas with misclassification of 
heath as woodland. Experimentation had shown this method to be as good 
as any for a primary classification of land cover types. The land cover 
class of the pixels in each sub-scene was then transferred to the data 
matrix and gradually the data matrix will be filled with the classi 
fication of all seven LANDSAT scenes of mainland Scotland. By this 
method it is hoped to minimise the problem of 'distance decay' of the 
classification reliability away from 'training' areas. 
A further major problem in upland Scotland are the strong shadows cast 
by the topographic relief, especially with a satellite overpass at 
9.^0 a.m. local time. Some of this can be reduced by appropriate band 
ratioing (using 5/7 instead of 7 in multispectral classification), but 
some areas of shadow persist, resulting in heath sometimes being mis- 
classified as forest. The more glaring examples of misclassification 
due to this source have been removed by the simple but tedious process 
of manual editing of the computer file after comparing lineprinter map 
output with air photo or map 'truth'. Editing was carried out only on 
those classes which display a distinctive boundary (e.g. woodland and 
urban) and this is reflected in the results shown in Table 1. 
After classification and editing, a 'smoothing' algorithm was applied 
to remove many of the extraneous pixels causing optical 'noise' in the 
classified map. A 'minimum area' smoothing algorithm was used, moving 
a 3 x 3 grid across each pixel in turn. Water areas were not smoothed 
but for all other classes isolated pixels were removed. The effects of 
editing and smoothing for the Grantown sub-scene are given in Table 1. 
ACCURACY ASSESSMENT OF LAND COVER CLASSIFICATION 
A sampling procedure was adopted to test map accuracy after classi 
fication and editing. Results from this testing are given in Table 2. 
For this pixel block a sample size of 60 was chosen, although the 
woodland class is an amalgamation of two classes giving double sample 
348