53 . Cross-correlation results 
The technique was applied using the same spectral library and data set as the neural network. In this case, 
clearly defined alteration zones were determined which corresponded exactly with areas previously 
defined by geologists in the area and with field studies. The technique also proved effective in identifying 
small areas of alteration previously undetected during surface mapping. Mis-classification of very similar 
minerals (kaolinite and dickite) was observed, but confined to the mixed areas at boundaries between 
major mineral distributions of alunite and kaolinite. 
6 - SUMMARY 
Three methods for the analysis of imaging spectrometer data for mineral exploration purposes were 
investigated. Parameterisation, cross-correlation and neural networks. 
The parameterisation method proved to require a large number of processing steps and a 
large user input for the expert system rules base implementation and maintenance, It was also noise and 
mixture sensitive and thus prone to mis-classifications. 
The cross-correlation method was very accurate in determining the extent of alteration 
minerals such as alunite, kaolinite and dickite. The error term produced indicating the areas showing less 
well defined distributions either due to the weakness of the spectral response in mixed pixels or due to 
higher noise levels in the normalised data.. 
The CMAC neural network offers a very fast processing speed with an accuracy of 
identification of the order of the cross-correlation method. In tests the CMAC accuracy was equivalent to 
the cross-correlation technique. However, when processing AVIRIS data the CMAC network clearly failed 
to identify the known mineral distributions in the scene. The errors are related to the presence of noise 
mixtures and artifacts of the IARR processing in the AVIRIS spectra that create addresses outside the 
range of training. These addresses randomly access the CMAC memory structure and mis-classify the data 
to the spectrally flat materials that occupy the majority of the memory locations. Further work is under 
way to determine the pre-processing steps necessary to eliminate the AVIRIS data problems causing the 
mis-classifications and to reduce the memory requirements of the network. It is expected that a 
classification accuracy equivalent to the cross-correlation method is achievable with a much reduced 
processing time. 
7 - CONCLUSIONS 
1. The cross-correlation method proved to be the most robust method for identifying the dominant 
spectral component in AVIRIS data spectra in the case with no pre-processing or training on mixtures. 
The method was relatively insensitive to both noise and low proportion mixtures, but capable of 
resolving small differences in mineralogy in most cases. 
2. The parameterisation method was rejected on the basis of the large number of processing steps, 
sensitivity to noise and mixtures and the need to develop and maintain a large expert system rules 
base for die identification step. However, further work to characterise the classification accuracy when 
compared to the other methods is necessary. 
3. The CMAC neural network was equally as capable as the cross-correlation method in identifying the 
mixed spectra and noisy spectra of the test series. However, the data problems present in the AVIRIS 
data (noise spikes, processing artifacts) prevented an accurate classification of the surface 
mineralogy. Further work is necessary to pre-process the AVIRIS data to identify and remove the non 
signal variations to enable fast, accurate classification using the CMAC neural network. 
4. An atmospheric correction is considered necessary to recover the small spectral variations present in 
the AVIRIS data. However, until the solar irradiance correction is implemented the IARR method 
provides a means of recovering the major spectral components in a scene. Future work with soils will 
need the more accurate recovery expected from an atmospheric correction. 
5. Even with the high noise levels of the 1990 AVIRIS data in the 2.0jim to 2.5 ^m wavelength region, it 
is still possible to recover accurate mineral distributions for the purposes of mineral exploration. Given 
the reduced noise levels of the more recent AVIRIS data, it should be possible to recover the small 
spectral variations associated with different lithologies and soils in a scene. 
339