Although the identity of these grouping are not known, comparison of the spectral res- 
ponse curve with those generated in the supervised classification, led to the identificationn of 
some of the classes. The classification was confirmed bv comparison spatially with ground truth 
information. Similar ^ clusters were aggregated into one cluster on the basis of similarity 
of spectral response curves and also the extreme smallness of the areal distribution or some 
clusters. This suggests that some supervision of the computer classification is still necessary in 
what Lillesand and Kieffer (1979) called “hybrid” classification. 
Enhancements The first enhancement process is the display of the data in FCC of 4 in 
blue, 5 in green and 7 in red, a process that helped in the identification of soil physiographic 
units. 
An edge enhancement was also run by using a 3 x 3 pixel window. This performs a 3 x 3 
convolution of the screen image. The value of the centre pixel is the average of the values con 
tained in the 3x 3 window surroundiag the pixel. 
Although the distinction between the classes was clearer with this enhancement, this was 
not translated into the area composition statistics or to the hard copy printout. This is a de 
finite limitation of the software, Perhaps if the enhancement was done before the selection of 
the training sites, the results of the digital classification could have been improved. 
Rectification and registration of GT map and the LANDSAT generated maps 
Up to this point 4 separate files have been created: the file of the ground truth (GT) or 
the reference soil map, maximum likelihood, minimum distance and cluster anlysis from the 
LANDSAT data. 
The GT maps were set up using arbitrary UTM coordinates, prior to digitizing. A 25 x 25 
m cell size was selected for the grid conversion so that the small flood plain units and valley 
bottom units could be resolved. These cells were aggregated into 50 m x 50 m cells in the GIS 
file using a 2 x 2 pixel template. The most frequently occuring reflectance value is used as the 
output value while a tie is broken by a user—specified class of prioritiesin favour of the larger 
or the smaller value. 
The LANDSAT maps were rectified to this map. Apart from rotation, the pixel were re- 
sampled to 50 m x 50 m pixels to match the GT map. 
RESULTS 
(1) Qualitative comparison:  1he three LANDSAT classification maps and GT soil map were 
displayed separately for examination on the colour monitor. Since it was not possible to do a 
satistactory comparison this way, hard copy continuous greytone printouts of the images were 
generated at a suitable scale by the computer (Scale: 1 : 36, 750). 
Although the LANDSAT generated imageries and the GT map were accurately registered, 
it was not possible to accurately and satisfactorily trace the coincidence of the classes in each 
of comparison as Weismiller et al (1977) found out. The computer method of overlaying 
pairs of maps was used. The GT map was first displayed and then, one after the other the LAND- 
SAT maps, with a green image plane — the best of the three primary colours. 
Each of the classes was highlighted to see its coincidence or distribution on the GT mao. 
The following are the general and specific statements about the comparison of the digita! Z,a:3- 
ification of the LANDSAT data with the GT map: 
     
  
  
   
  
  
   
  
  
   
  
  
   
  
  
  
  
  
  
  
   
  
  
  
  
    
  
   
  
  
   
  
  
   
  
  
  
  
   
  
     
ant 
me 
alc