F the 
sugar- 
] in- 
infra- 
) 
and 
cla- 
y pi- 
raine 
re.As 
nter- 
iple 
simul- 
esults 
natic 
ual 
nfra- 
des- 
No.l 
s used 
llows: 
ers 
ide 
mi - 
ocus- 
  
= 1693 - 
ces (which give an electrical signal output). The readings of 
the corresponding meters on the three channels give the spec- 
tral features of the pixel on which the input light is focus- 
sed. 
For various technical reasons, the digitization was done 
manually using a grid accurately drawn to scale. Each pixel of 
the enlarged color print corresponds to & ground element of 
size Tm.x Tm. Fig.2 gives & schematic of the pattern recogni- 
tion system. 
(b) Classification of Categories, : For each of the categories 
listed in Sec.1, training samples were selected on the basis of 
the known interpretation of the scene {rie.3). 
Fig.l gives the computer plot of the relative spectral 
responses of the categories as obtained from the training set, 
and Fig.5, the histogram which throws some light on the sepa- 
rability or otherwise of the various categories. It is observ- 
ed from Fig.lh that the mean values of the spectral features are 
different for most of the categories. 
The classification algorithm is derived from the maximum 
likelihood criterion, on the assumption that the measured (vec- 
tor)features for all the categories are normally distributed. 
For each category, therefore, the parameters to be estimated 
are the mean vector and the covariance matrix. 
Out of about 5100 pixels of the color infrared print, the 
following constituted the training sets 
a) Paddy freshly planted (A1.1) 120 
well grown (A1.2) 60 
mature (A1.3) 120 
harvested (A1. ) 30 
b) Sugarcane freshly planted (A2.1) 100 
young ratoon (42.2) 80 
well grown (A2.3) 60 
c) Fallow land (A19 ) 60 
"Total samples 630 
Mean vectors and covariance matrices for these training 
sets have been calculated. These parameters correspond to the 
frame under study. However, it is believed that they could be 
used for the classification of other frames in the same reel. 
The classification results obtained from an application of 
the maximum likelihood criterion are compared with the exact 
interpretation (Fig.3) of the area under study , in Tables 1-5. 
Figures 6-11 are the computer print-outs of the classification 
results. As regards the error analysis results, the correct in- 
terpretation (Fig.3) was overlaid on the computer classification 
and the "misfits" were treated as contributions to the error.