absorption bands and absorption band characteristics 
for identification of specific materials and to define 
the facts and rules utilized in the expert system. A 
tree hierarchy was chosen to minimize analysis time 
(Figure 3). The tree was designed to model the 
spectral analysis procedures and decision processes 
followed by an experienced analyst. Facts and rules 
were written for each material or group of materials 
in the database based on prior analysis of the spectral 
library. The spectral library itself was never accessed 
during the analysis of a spectrum. The strongest 
absorption feature for a given spectrum was 
determined, and the spectrum broadly classified (eg. 
clay, carbonate, iron oxide). Primary band 
characteristics (eg. doublet, triplet) and 
secondary/tertiary absorption bands were used to 
progress through the tree structure until an 
identification was made. If the decision process failed 
at any point because there was insufficient 
information to identify a specific material, then the 
last classification was used to give the best answer 
possible. 
SPECTROMETER DATA 
Figure 3. Tree Hierarchy for analysis of 
reflectance spectra 
LABORATORY EXAMPLE 
The prototype expert system was tested on selected 
mineral spectra using the continuum-removal and 
feature extraction algorithms and a basic set of rules 
derived from the analysis. The laboratory reflectance 
spectra were convolved with random noise to test 
the possibilities of success on real field and aircraft 
spectra. Because identification was very adversely 
affected by low signal-to-noise ratios (similar to those 
expected with aircraft data), a binary encoding 
algorithm (Mazer et al., 1988) was included to reduce 
noise sensitivity. Figure 4 shows a flow diagram of 
the analysis procedures. Note that the feature 
extraction and binary encoding procedures operate 
independently to select material candidates. Features 
matching specific rules are given greater weight than 
binary matches, however, in noisy data, the binary 
encoding has a significant effect on the decision 
process. 
Figure 5 shows an example of rules for 
identification of calcite and dolomite. The decisions 
follow the hierarchical tree from broad to specific 
classifications. If the process fails at some level, then 
Figure 4. Flow diagram showing procedures for 
analysis of reflectance spectra using the 
expert system (From Kruse et al., 1990a). 
the identification at the previous level is returned as 
the best possible answer. If the expert system is 
unable to identify the material, then the spectrum is 
flagged as an unknown material. 
1st Decision ( Surface class ) 
if not broad spectral bands near 1.4 and 
1.9 pm (vegetation) 
then look for in rock class 
2nd Decision ( rock level ) 
if it has a deep band in 2.30 - 2.35 pm region 
then look for in "carbonate" species. 
3rd Decision ( "carbonate" species ) 
if the strongest absorption feature is a single 
band near 2.34 pm 
then it is calcite 
else if the strongest absorption feature is a 
single band near 2.32 pm 
then it is dolomite. 
Figure 5. Example of rules: calcite vs dolomite 
(From Kruse et al., 1990a). 
Figure 6 shows the results of analysis of the 
calcite spectrum shown in Figure 2 using the expert 
system. The results of the feature extraction 
procedure are shown in Figure 6a. The interaction of 
the absorption features and the rules result in the 
broad classification shown in Figure 6b. Figure 6c 
shows the percentage match of the spectrum with the 
binary library. Note the 100% match between the 
library calcite and the unknown spectrum. This is 
because the calcite spectrum comes from the library. 
Figure 6d shows the final result of the decision 
process when the binary encoding is used to weight 
the feature extraction process. Figures 6e and 6f show 
the justification for the identification of calcite versus 
dolomite.