4.1. Cross-correlation 
The results are summarised in figure 2. Clearly the mixture proportions are not significant until the 60-40 
level for the mixture combinations used, while noise has a limited impact on identification until the 20-1 
signal to noise level is reached. Given the expected signal to noise level of the 1990 AVIRIS data to be 
between 25-1 and 50-1 in the 2.03(xm to 2.37 pm wavelength region the method should be effective for 
mapping areas dominated spectrally by a single mineral. 
4.2. CM AC neural network 
The results from the same test sequence using the neural network are shown in figure 3. They show a 
strong similarity to the results obtained using the cross-correlation method. The neural net results were 
poorer in dealing with mineral mixtures showing a drop in classification numbers at the 70-30 level. Note 
that if the CMAC had been trained on mixtures of the two minerals there would be no reduction in 
classification accuracy. The results with varying the noise level were very similar to that of the cross- 
correlation method which suggests a lower limit of separation of these two minerals at the 20-1 signal to 
noise level for a 50% reflector. The major advantage of the CMAC was the much faster processing speed 
exceeding that of the cross-correlation technique by a factor of 10. 
5. AVIRIS DATA ANALYSIS 
The AVIRIS data described were collected over the Paradise Peak Mine area of Nevada, USA during the 
summer of 1990. The area is semi-arid with very limited dry vegetation cover, good outcrop and showing 
extensive mineralisation at the surface. 
5.1. Geological background 
The mine area shows a distinct alteration sequence related to high sulphur epithermal gold mineralisation 
(John et al., 1989) producing an alunite rich rock above the main gold bearing horizon and an alteration 
suite of minerals including dickite and kaolinite. 
The main target areas for further exploration include alunite rich rocks and major zones 
of kaolinite and dickite. All the altered areas coincided with extensive introduction of silica. 
5.2. CMAC neural network results 
The network was trained using primarily the JPL reference mineral library. A noise element was added to 
the mineral spectra prior to normalisation and training simulating the expected noise in the AVIRIS data 
in the 2.0pm to 2.37pm wavelength range. This is a necessary part of the CMAC network training. 
The results of the classification were compared with both the alteration map produced by 
the mine geologists and fieldwork studies in the area. The results were disappointing, the network failed to 
identify any of the known alteration areas in the scene, classifying most pixels as one of the three 
spectrally flat materials in the wavelength sub-section of the library used (haematite, goethite, green 
vegetation). 
The poor classification results are due to a combination of problems explained previously. 
The AVIRIS data is generally noisy, but also contains noise spikes which produce addresses outside the 
normal range of training, this combined with the obviously mixed surfaces in the area and the artifacts 
introduced during IARR processing creates virtual addresses that randomly access memory. 
The mis-classifications to the spectrally flat materials are due to the state of the network 
after paining. A flat material with noise presented several thousand times to the network will produce 
several thousand different addresses and hence occupy several thousand memory locations due to the 
influence of the noise, while a material with a large specPal variation will be less affected by noise and 
occupy a very small number of memory locations. Therefore, if the memory is accessed randomly, it is 
much more likely that the specPally flat material is identified. Examination of the classification maps 
from the network matching of the AVIRIS data sPongly confirm this interpretation. Further work is in 
progress to match the spectra without normalisation. This should increase classification accuracy when 
combined with noise spike suppression and the inclusion of known mixtures during training. 
337