336 
Vnm = BASE(n) + 1 (4) 
Where n is the number of bands and m is the number of the address created (up to NUMCEL). Note that if 
Vnm is less than STATE(n) then Vnm has NUMCEL added to its value. 
The system can respond to small variations in the expected values. If for example the 
first value of our three input tests case (figure 1) is increased by two, the BASE values stay the same, but 
the first two virtual addresses will be changed. In other words a small perturbation in the curve shape 
produces a reduced accuracy of match as only three of the original five virtual addresses are the same. 
When processing a scene, each unknown spectrum consists (using the 38 band AVIRIS 
data set) of a 38 value STATE vector converted to a 38 value BASE vector. From this NUMCEL (in this 
case equal to five) virtual addresses are created which point to NUMCEL rows in real memory. Each row 
has a memory cell for each of the library minerals used in training which contains a value. These cells for 
each mineral are summed across the five memory rows accessed, the mineral with the highest total is 
chosen and the pixel classified with the chosen mineral identity. 
3.3.2. Problems of use. 
1. Sensitivity to noise spikes. If the magnitude of a noise spike exceeds NUMCEL, it will change all the 
virtual addresses and therefore randomly select locations in real memory producing a mis-match. The 
problem is independent of the number of bands in the data input. 
2. Sensitivity to mixtures. Again mixtures would produce offsets in all the values of the BASE address, 
producing virtual addresses outside the range of normal training producing mis-matches. However, if 
known mixtures are present in a scene, they can be given as training examples to avoid this problem. 
3. The memory requirements for a relatively small library (25 materials) with a 38 band input (STATE) 
vector exceed 20Mb if identification performance is not to be degraded. This exceeds the memory 
requirements of some workstations. This problem is currently being addressed with some success. It is 
expected that a much larger number of library minerals will be used in the updated version with much 
reduced memory requirements. 
4. Sensitivity to artifacts of processing such as those found after using IARR processing. These artifacts 
produce effects similar to mixing and noise and can be treated as an additional noise element. 
3.3.3. Advantages of the CMAC neural network. 
1. The training changes the values in memory to maximise the separation between the chosen minerals. 
For example if two very similar minerals have a large number of virtual addresses in common, the 
values of the common addresses are reduced while those not shared are increased in value thus 
weighting the features that separate the two minerals. 
2. An important factor is speed of identification. Once successfully trained the system takes a few 
minutes to perform a classification that takes 30 minutes using the cross-correlation algorithm. This is 
due to the very fast Look-Up-Table (LUT) nature of the identification process. 
3. The method of generalisation of the CMAC allows the user to flag inputs outside the range of the 
training of the network rather than attempt to mis-classify the unknown as happens with other network 
architectures. 4 
4 ■ TEST SERIES 
A test series was generated on the system in which two component mixtures of muscovite and dolomite 
were created with Gaussian noise at various signal to noise ratios for a 50% reflector. Prior to testing, the 
parameterisation technique was rejected based on the evaluation of the problems associated with its use. 
The results from the cross-correlation and CMAC neural network are presented below. Note that a more 
thorough evaluation of the parameterisation method is planned to accurately determine the classification 
accuracy when compared to the other techniques discussed.