to the unsupervised classification routine in CSIRO-ORSER. Based on a 
random sample of 100 pixels, six classes were determined. For each class a 
consistent area of pixels was identified and used as a training site. 
Assuming that the pixel values for each tentative class were from a 
multivariate normal population, the mean vector and the standard deviation 
were calculated. 
naps 
the A classification was then performed of the whole subimage using a 
2 minimum Euclidean distance classifier. This particular classifier is 
à. considered to be intermediate in cost between a classifier which uses exact 
variance/covariance data and that based on a parallelepiped approach. It 
9S - is also intermediate in accuracy. In an effort to minimise the effect of 
s of the variable illumination of the ground caused by low sun angle and 
changing aspect, the original data was normalised, normalised signatures 
derived and the area classified by using the normalised Euclidean distance 
classifier. 
The classes determined were compared spatially with the photointer- 
preted map. At the same time the signatures of classes which occurred both 
within and without the nominal fire boundary were examined and compared 
£ the against signatures of classes which occurred only within the fire area. In 
.T. spite of the fact that vegetation mapping was not the object of the study 
In the various forest classes were not grouped at this stage. This enabled 
D distinctions to be maintained between them and fire affected types which 
near may be very close in signature to their unburnt original vegetation type. 
been Signatures were further refined by iteration of the process of selecting 
areas of uniformity in the broad classification map. Several more narrow 
classes were produced and appropriate statistics calculated. This process 
aimed at interpreting the data and obtaining a feeling for the distributions 
RSER rather than taking a sample and testing hypotheses about the non sample 
Or area. 
At the same time the photointerpreted map (scale 1:63 360) was digit- 
E ized by taking samples on a regular grid of points at 2 mm spacing - 
ly approximately three times the spacing of Landsat pixels. Control points 
SER were selected and digitised also from this map. 
ISIMP 
and The class maps were produced in several formats each to satisfy a 
specific requirement: 
(i) line printer character form for use as a working document and a 
the scaled map in the cheapest possible form. One of these was hand 
E coloured for ease of interpretation. 
sta (ii) colour display image on a COMTAL image display device to allow 
areas colour slide production for demonstration and to examine the 
to completeness of the classifications. 
tures (iii) 35 mm microfilm output on an III COMp 80 microfilm device (In- 
ning formation International Inc.) for production of hard copy for 
evaluation of accuracy of classification and demonstration. Some 
grouping of the unburnt forest classes was made on this product 
to simplify the grey-scale rendition of the classified image.