See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/284479221

Algorithms and software for the automated identification of minerals using field
spectra or hyperspectral imagery

Article · January 1999

CITATIONS READS

42 1,017

3 authors, including:

Mark Berman Jon Huntington

The Commonwealth Scientific and Industrial Research Organisation Mineral Resources, The Commonwealth Scientific and Industrial Research Organisat…
72 PUBLICATIONS 5,691 CITATIONS 80 PUBLICATIONS 2,898 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Mark Berman on 06 July 2017.

The user has requested enhancement of the downloaded file.

 ALGORITHMS AND SOFTWARE FOR THE AUTOMATED IDENTIFICATION
OF MINERALS USING FIELD SPECTRA OR HYPERSPECTRAL IMAGERY




M. Berman , L. Bischof and J. Huntington

CSIRO Mathematical & Information Sciences, North Ryde, NSW, Australia
CSIRO Exploration & Mining, North Ryde, NSW, Australia

ABSTRACT

Sophisticated algorithms and software for identifying pure minerals and mixtures of min-
erals from short-wave infrared reflectance spectra are described. The software, called
The Spectral Assistant (TSA), has been trained using a spectral reference library of 551
samples, representing 49 “pure” minerals and water, measured with a PIMA-II spec-
trometer. It reliably identifies mixtures of up to two minerals in the library, and does
so in a fraction of a second on a Pentium. This means that hundreds of spectra can
be analysed in a matter of minutes. TSA Version 4 has also been incorporated in The
Spectral Geologist, a larger package designed by the CSIRO to analyse spectra specifi-
cally with exploration and mining applications in mind. Recent work, which has adapted
TSA so that it can be applied to atmospherically-corrected hyperspectral imagery is also
described. The algorithms and software are illustrated using field and airborne spectro-
scopic examples.

1. INTRODUCTION

Field and airborne spectroscopy are being used increasingly for mineral identification, in both exploration
and mining contexts. There are two primary forces encouraging the development of automated spectral
identification systems for minerals. The first is the fact that most geologists are not experts at interpreting
spectra. This is especially so for mineral mixtures. The second is that the increasingly large volumes
of spectra being measured (especially at the airborne level) make human intepretation of spectra an
impractical proposition in many cases.

A variety of spectral identification systems have been developed over recent years. Many try to mimic
human experts directly by first “extracting” the major absorption features in the spectrum of interest, then
measuring relevant quantities representing the size, shape and wavelength positions of these features,
and finally devising “expert-system” rules based on these quantities for identifying the (assumed pure)
mineral represented by the spectrum (Ali et al, 1989; Kruse et al, 1993; Lister, 1990). Apart from the fact
that new (ad hoc) rules need to be added to the system whenever a new mineral is added to the database
on which it is based, such a system has great difficulty in dealing with mixtures.

Presented at the Thirteenth International Conference on Applied Geologic Remote Sensing,

Vancouver, British Columbia, Canada, 1-3 March 1999.
Our approach has been to build a relatively simple statistical model to represent typical short-wave in-
frared (SWIR) reflectance spectra. This model deals naturally with mixtures. Associated with the model
is an identification algorithm, which has been optimised for speed. The basic model and identification
algorithm are described in Section 3. Beforehand, in Section 2, we discuss the data on which the model
and algorithm are based. In Section 4, we describe and illustrate how we have adapted them to deal with
atmospherically-corrected hyperspectral imagery. Future directions are outlined in Section 5.

2. A LIBRARY FOR TRAINING AND TESTING FIELD SPECTRA

All the field reflectance spectra that we shall be discussing in this and the next Section were measured us-
ing a PIMA-II spectrometer, developed and marketed by Integrated Spectronics Pty. Ltd. This spectrom-
eter records a reflectance measurement every 2nm in the range [1300,2500] nanometres (nm), thereby
producing 601 numbers per spectrum. In early versions of the instrument, we sometimes observed spu-
rious measurements at either end of some of our spectra. Therefore, we only consider the range [1304,
2496]nm in our analyses. For statistical reasons that will become clear later, we also only consider ob-
servations every 4nm (starting at 1304nm). This reduces the number of measurements per spectrum used
in the current version of our software to 299.

Our training library consists of 551 spectra representing 49 (nominally pure) mineral classes, together
with a class consisting of 10 Water spectra. The class sizes range between 3 and 37 samples. Some
of the geologically more important minerals in the library are: Kandites (e.g. Kaolinite, Dickite and
Halloysite); White Micas (e.g. Illite, Muscovite, Phengite and Paragonite); other AlOH minerals (e.g.
Montmorillonite, Pyrophillite); three varieties of Chlorite (Magnesium, Iron and Intermediate); a variety
of other MgOH minerals (e.g. Biotite, Phlogopite, Nontronite, Talc, Tremolite and Actinolite); Sulphates
(e.g. Alunite, Jarosite and Gypsum); Carbonates (e.g. Calcite, Dolomite and Magnesite); Epidote and
Opaline Silica. A complete list of the pure minerals in the library can be found in Tables 1a, 1b and 1c
in Berman and Bischof (1997).

The algorithms we shall describe have also been tested on a variety of data sets containing real mixtures
(and pure minerals), as well as powder mixtures of known minerals with known weight proportions.
Details of the test sets are given in Berman and Bischof (1997, Section 2.2).

3. THE UNDERLYING MODEL AND IDENTIFICATION ALGORITHM

Figures 1a and 1b show 5 Kaolinite spectra and 5 Illite spectra respectively from our training library.
What visually distinguishes the Kaolinites from the Illites are the shapes of the absorption features, es-
pecially around 1400 and 2200nm. Indeed, experts distinguish between most mineral classes using these
intermediate-frequency absorption features, rather than the low-frequency background, often called the
“continuum”. Typically, this background is caused by the presence of other materials in the sample,
which do not have any distinctive absorption features of moderately high frequency in the SWIR. Com-
mon minerals without such absorption features in this region include Quartz, Feldspar and Iron Oxide-
bearing minerals. We shall henceforth refer to such minerals as “inactive” (at least as they pertain to the
SWIR). How do the spectra of minerals mix in a single sample? According to Beer’s Law (Skoog and

  
 
West, 1980, eqn. (6-11)), this mixing is multiplicative providing there is no interaction among the vari-
ous species. Hence, if we take logarithms of a spectrum, the mixing is additive. Letting
denote the reflectance spectrum of a mixture of 
minerals, we therefore have

" $&'$
 ! #$ 
%
(3.1)

&'$ $
(
where is the typical spectrum of the th mineral in the mixture and  is (in some sense) proportional
to the amount of that mineral in the mixture.
0.8
Reflectance

0.6

)
0.4

)
0.2

1400 1600 1800 2000 2200 2400

Wavelength (in nanometres)

(a)

)
)
0.7

)
0.6
0.5
Reflectance

0.4
0.3
0.2

1400 1600 1800 2000 2200 2400

Wavelength (in nanometres)

(b)

Figure 1. (a) 5 Kaolinite Spectra, (b) 5 Illite Spectra

Some of the minerals in a mixture will be inactive. The conventional way of removing their effect is first
to estimate them by finding a low-frequency curve (such as the upper convex hull, Green and Craig, 1985)
that lies above the spectrum while being as close as possible to it. Background correction is then achieved
by taking the ratio or difference of the two curves. This was the approach we followed in early versions
of our algorithm and their software implementations, which we call The Spectral Assistant (TSA). This
was quite adequate for versions 1 to 3 which ignore the spectral regions where Water has significant
features, specifically [1400,1600]nm and [1850,2100)nm. We shall call these the water regions.
However, some computer experiments that we conducted, which attempted to extend the algorithms to
include the water regions, indicated to us that methods, which first removed the background and then tried
to identify the mineral composition represented by the spectrum, were often fitting the background too
closesly to the observed spectra; see Berman and Bischof (1997, Section 4.1) for details. We concluded
that this problem could only be overcome by carrying out the background removal and identification
simultaneously. The algorithm which does this is described in this paper. It has been implemented in
TSA Version 4, which itself has been incorporated in The Spectral Geologist (TSG), a larger package
designed by the CSIRO to analyse spectra specifically with exploration and mining applications in mind
(AusSpec, 1998).

It is easiest to explain the algorithm in two stages. First, in subsection 3.1, we assume that the spectrum
contains only one “active” mineral, which we shall henceforth call “singletons”. In subsection 3.2, we
deal with mixtures of active minerals.

3.1. A MODEL AND IDENTIFICATION ALGORITHM FOR SINGLETONS

The approach we have adopted is to model the background as a low-frequency cubic spline (Wahba,
1990) with “knots” at 1700nm and 2100nm. This means that in each of the ranges [1304,1700], [1700,
2100] and [2100,2496]nm, the background is modelled as a cubic polynomial. However, the three cubics
are constrained to join continuously and smoothly. It has the particular advantage that it can be repre-
sented as the weighted sum of six simple basis functions; see Berman and Bischof (1997, Section 4.2) for
details. In addition, we have found it convenient to always fit Water in the model. This can be thought
of as a seventh basis function. We have estimated a typical water curve from the samples in our library.
This leads to the following model for singletons:

&+* $/0$
  $# , 21
%
.- (3.2)
&+* / $
where
$
is the typical dry spectrum of interest, (354 687 are the basis functions representing the
background and water,  and are the weights given to these spectra, and 1 accounts for errors in the
-
model. Compare (3.2) with (3.1).
/9$
&+*
The only quantities known a priori are the ’s. For each class : in our training library ( :;<4 =?> ),
: . This does not have a unique solution! We can make it unique by imposing/ $appropriate constraints
we estimate their typical dry spectrum, , via a least squares fit of (3.2) to all the sample spectra in class

such as that the background curve (as represented by the weighted sum of the ’s in (3.2)) always lies
*
above the observed spectrum; see Berman and Bischof (1997, Section 4.3) for details. However, from
*
a fitting perspective, any of the least squares solutions will do, as they all give rise to the same fits. For
later use, it will be convenient to let1B@ A C
denote the residual from the fit for the th spectrum in class , :
and D@ A 
the estimate of for that fit.

For a spectrum representing a singleton whose class is unknown, we could just find that class (among
the 49) with the best least squares fit. However, this is statistically inefficient, because the fact that the
data are in the form of curves implies that observations at nearby wavelengths will be highly correlated.
1
Let us assume that, for each spectrum, in (3.2) is Normally distributed with zero mean and unknown
E
within-class covariance matrix . It will be computationally convenient (but almost certainly untrue) to
F
assume that is the same for all 49 classes. Then, under the assumption that each of the mineral classes
is equally likely, the rule with the minimum error rate chooses the class (and associated parameters) to :
minimise the (generalised) Mahalanobis distance, which in matrix notation is

* &+* , $/9$ &+* , $L/9$

   HG  G $# BI KE J
 HG  G #$ 8
%
.- %
.- (3.3)

I
where
&+*
represents the transpose of a matrix.

M >N>PO M >N>
QM >N>ROTSVUNU?W M
In (3.3), we can use the estimate of described above. However, how do we estimate the
F
matrix which consists of 44850 ( ) different elements! We have adopted a modification
of an approach called penalised discriminant analysis (Hastie et al, 1995) , which is designed to take
advantage of the fact that the data are curves. In this approach, is estimated by E
X X Y^]R_a`  X W _a`  ] 8
E5  4 GZY  E\Z
[ bE [ (3.4)
X * *
where Eb[ Y
is the average within-class covariance matrix of normalised residuals ; is a “rough- 1a@ A W D@ A ]
_a`
ness”-type penalty matrix; is a regularisation parameter which trades off fidelity to the data against the
smoothness of the solutions; and “ ” represents the trace operator. usually corresponds to a squared ]
Y
derivative (or difference operator) of a certain order; see Berman and Bischof (1997, Section 3.3) for
details of our choice. We have chosen using our training library of 551 singleton spectra only and
applying the method of cross validation to them.

3.2. EXTENDING THE MODEL AND ALGORITHM TO DEAL WITH MIXTURES

Suppose that we have a spectrum representing a mixture of minerals in our training library. For simplicity,
we shall assume that there are 2 such minerals in the mixture (classes and of the database, say), : 
although the approach we will outline generalises to more than 2 minerals in an obvious way. The
obvious extension to (3.3) is to minimise

*dc &+* &+c , $/0$ &+* &+c , $L/9$

 #

  HG 
G   G %
.-  I E J G 
G   G #$ %
.- 8
$  (3.5)
$
with respect to :  
6   , and - (be4 687 . Given : and  , estimation of the other *dc
straightforward using generalised least squares. However, finding the classes : and  minimizing 
parameters is
quickly is less straightforward, because on the face of it, there are f 6g  h4N4 7Vi possible combinations
to examine.

The way we deal with this issue is to apply a linear transform to the data called the Canonical Variate

E@
transformation (McLachlan, 1992, Section 3.9.2). Apart from being the optimal linear transformation
for class discrimination, it has several other advantages. First, it converts the matrix into the identity
matrix, so that minimisation of (3.3) or (3.5) reduces to ordinary least squares. Second, in this new
*
space, the intrinsic dimensionality of the data is reduced from 299 to the number of possible explanatory
*dc
&+ * &+ c
variables minus 1. In this case, there are 49 classes plus 7 basis functions, so that the dimensionality is
reduced to 55. This means that, although either
299 terms, 244 of them will be independent of
or
* *dc
can be represented as the sum of squares of
and , and so only 55 of these terms need to be used * *jc
for classification. We will label this part of or
*jc as or  
respectively.  
: 
Finding the and minimizing 
(or indeed analogues of this statistic for mixtures of 3, 4 or more
minerals) is a classic example of what statisticians call the “subset selection problem”. In such a prob-
lem, there are a number of possible (multivariate) explanatory variables (in our case the 49 typical dry
spectra plus 7 basis functions) and a (multivariate) response variable (in our case a measured spectrum).
However, typically a linear combination of a small, but unknown, subset of the explanatory variables (in
our case 1 or 2 typical dry spectra plus the 7 basis functions) is believed to explain the variation in the
response variable. The problem is to find the best-fitting subset(s) of a given size. Efficient algorithms
to solve this problem are given in Miller (1990, Chapter 2 and Section 3.6). The author has provided
* *dc
us with code implementing these algorithms. This code (modified a little by us) is at the heart of all
versions of TSA. For any number of components in a mixture, the code produces the best fits to the
data (as measured by  ,  or their extensions to more than 2 components). The code is very fast; a
typical spectrum is analysed in a fraction of a second on a standard Pentium, thus enabling the analysis
of hundreds of spectra in just a few minutes.
* *dc
In reporting the results, we usually report the 3 best mineral fits minimizing each of and , often  

 
called the best fitting singletons and the best fitting mixtures, for short. We do not report for the best
fitting singletons. For mixtures, we only report the 3 best fits for which both are are positive. We
do not report these values, but instead we report the proportion estimates

k @ Q
W  
l  8 4 G k @ (3.6)
*
*dc
For each of the reported fits, we also give a standardised version of the corresponding value of or 
 , which we call the Standardised Residual Sum of Squares (SRSS). First, the spectra are normalised
so that the amplitudes of the absorption features in the same mineral class are comparable; see Figures

* *dc
1a and 1b. This is a moderately complicated process, designed to ensure reasonable consistency across
the different versions of TSA; see Berman and Bischof (1997, Sections 3.1, 3.2 and Appendix A) for
details. Second, the resulting or  
are multiplied by 100 (for reporting purposes) and divided
by the intrinsic dimensionality of the data (55). Under certain idealised assumptions, it can be shown
that, when the correct mineral components have been found, a typical SRSS score should be about 100.
 0.70
0.65
0.60
Reflectance

0.55
0.50
0.45
1400 1600 1800 2000 2200 2400

Wavelength (in nanometres)

Figure 2. A Spectrum Representing a Mixture of Kaolinite and Muscovite

Empirically, we have in fact found that an SRSS less than about 250 is usually indicative of a good fit.

As an example, we consider the spectrum in Figure 2, which has been identified by an expert as a mixture
of Kaolinite and Muscovite (or possibly the very similar Illite). Figure 3 shows the background-corrected
version of this spectrum together with the 3 best fitting singletons and the 3 best fitting mixtures. We
note that the best fitting singleton is none of Kaolinite, Muscovite or Illite, but is Halloysite, which,
like Kaolinite, belongs to the Kandite family. However, the best fitting mixture is Kaolinite and Illite,
followed by Kaolinite and Muscovite, and thirdly Halloysite and Illite.

The next obvious question is: What are sensible rules for deciding whether a spectrum represents a sin-

monbm!m

W monbmpPm q  rS
gleton or a mixture? Although there are theoretical answers to this question, we have found empirically

st(vu  k @ 4 G k @  q .U 4 w
that the following rule usually works well. Let and be the SRSS for the best fitting
singleton and best fitting mixture respectively. If monbm!m monbm!m 4and ,
then choose the mixture; otherwise choose the singleton. On that basis, we would conclude from the
results in Figure 3 that the spectrum probably represents a mixture of Kaolinite and Illite.

3.3. SOME LIMITATIONS OF THE ALGORITHM

Although extension of the above algorithm to deal with mixtures of 3 or more spectra is, in principle,
straightforward (and indeed has been implemented in the software), we have not yet made these exten-
sions publicly available because we have not had enough data to develop thresholds for reliably identify-
ing 3 or more spectra. We expect to remedy this situation in the near future. Such mixtures will typically
give a poor fit to the best fitting singletons or mixtures of two minerals (i.e. have an SRSS greater than
250). Other reasons for a poor fit include: (i) some spectra, especially dark ones, are noisy; (ii) the
sample contains a mineral not in the training library; or (iii) the instrument measuring the spectrum has
not been adequately calibrated, which leads to wavelength shifts of absorption features.
 1.0
Corrected Spectrum

0.9
0.8
Spectrum
Hall ( SRSS = 181.18 )
Kaol ( SRSS = 185.82 )

0.7
Illi ( SRSS = 256.24 )

1400 1600 1800 2000 2200 2400

Wavelength (in nanometres)

1.0 (a)
Corrected Spectrum

0.9
0.8

Spectrum
0.51 * Kaol + 0.49 * Illi ( SRSS = 33.816 )
0.51 * Kaol + 0.49 * Musc ( SRSS = 51.325 )
0.7

0.46 * Hall + 0.54 * Illi ( SRSS = 79.431 )

1400 1600 1800 2000 2200 2400

Wavelength (in nanometres)

(b)

Figure 3. Background-Corrected Spectrum with (a) 3 Best Fitting Singletons, and (b) 3 Best Fitting
Mixtures of 2 Minerals

3.4. OPERATIONAL ISSUES AND ACCURACIES

TSA has been found to be a satisfactory method for quickly identifying the most likely components of
mineral mixtures in PIMA spectra for which it has so far been trained. Output from other spectrometers
could, in theory, be accommodated with appropriate training and the results would be commensurate
with the quality of those spectra. TSA works particularly well on large sample suites where reproducible
results and patterns can be used to increase confidence and exclude the odd “outlier” interpretation. As
a result of tests on physical mixtures, it was found that the abundance thresholds at which it recognises
the most common minerals are typically between 10% and 20% (which is one of the reasons we usually
recommend the 15% threshold mentioned at the end of Section 3.2). However, several cases of even
better detection thresholds have been found for some minerals. Predicted abundances on these physical
mixture sets are typically within 5 to 10% of actual abundances.

4. EXTENDING THE ALGORITHM AND SOFTWARE TO DEAL WITH

HYPERSPECTRAL IMAGERY

It is straightforward to extend the above algorithm to atmospherically-corrected hyperspectral imagery,

provided there exists an appropriate library of spectra (specific to the instrument in question), which has
been atmospherically corrected by the same procedure. The working name of the software is The Spatial
Spectral Assistant (TSSA). It produces numerical output analogous to that in Figure 3, but in the form
of images: images of best-fitting singletons, and mixtures of 2 or 3 or more minerals; images of mixture
proportions; and images of SRSS scores. These images, which contain the basic information, can then
be further processed to produce summary information in image or other form. For instance, the SRSS
scores and mixing proportions can be used to produce images showing the likely composition at each
pixel; or false colour composites of three minerals of interest can be displayed.

As an example, we have analysed part of a 4 S M =POi 4 =

AVIRIS scene of Oatman, Arizona (Huntington
and Boardman, 1995). Our collaborator, Joe Boardman, has taken the last 56 bands (covering the region
[1961, 2509]nm), atmospherically corrected the scene using the ATREM 2.0 package (Gao et al., 1990,
1996)) and EFFORT post-processing (Boardman, 1998). He has then applied his endmember extraction
methods based on convex geometry ideas embodied in the ENVI software package (ENVI, 1997) and
identified 15 clusters of endmembers in the scene. The most important endmembers identified were:
Alunite, Dickite, Kaolinite, Long and Short Muscovites, Opaline Silica, Dolomite, Calcite, Chlorite, Dry
Vegetation and several other classes with less specific names, such as “Unknown MgOH”, “Unknown
Smooth 1” and “Unknown Smooth 2”.

We have taken 106 atmospherically-corrected spectra from 14 of the 15 endmember clusters identified
by Joe (a class called “Dark” was excluded and one very large cluster was sub-sampled to make its
size comparable to the others) and used it to train TSSA. Unlike TSA, we have only used one basis
function - a constant, because we wish to include smooth spectra as endmembers (see above), and Water
is inactive in the spectral region in which we are working. Also, we have used an additive rather than

w.4 M Z
O i 4 =O w i
a multiplicative mixture model (see (3.1)), because this is likely to be a better model for mixed pixels.

4 UNUO 4 UNU
TSSA was applied to a subset of the Oatman image. It took about 90 minutes on a
single processor Sun Sparc Ultra to process this image. Figure 4a shows a pixel subscene
of the 2091nm band, while Figure 4b shows most of the best fitting singletons found by TSSA in this
region. A colour display of the regions is more informative, but most of the information is contained
in this greyscale image. The dominant mid-grey represents Dry Vegetation and the dark region on the
centre right is Dickite. It partially surrounds a dark grey region of Alunite. The very light region to the
left of the Dickite, and elsewhere in the image, is Muscovite, and the light grey to its left and elsewhere
is “Unknown MgOH”. Along the dry stream bed (the darkest feature in Figure 4a), the dark pixels are
Dolomite, while the grey ones at its northern end are Chlorite. Obviously, an analysis of the mixture
images would further refine this classification.

As with all automatic methods of classification, it is important to be able to convince the reader that such
results are essentially correct. Validation of mineral mapping results is essential to its acceptance, and
may be done using field verification, repeat classifications with different sensors and algorithms, together
with models of assemblages predicted from a priori knowledge. For the subscene shown in Figure 4a, we
believe the results (the first application of the original TSA to spatial data) to be about 80% reliable for
the dominant material, judged against detailed field work, field spectral sampling and comparison with
other methods (Huntington and Boardman, 1995; Boardman and Huntington, 1997). Further work needs
to be carried out to reliably identify secondary minerals, particularly because of ambiguities caused by
 (a) (b)

Dickite Dolomite Alunite Opal Dry Veg. Chlorite MgOH Muscovite

Figure 4. (a) 4 UNU;O 4 UNU pixel subscene of Oatman image (2091nm band), (b) Best Fitting Singletons

the coarser spectral sampling of AVIRIS (10nm) than PIMA (2nm).

5. FUTURE DIRECTIONS

One of our principal aims is to develop rules for reliably identifying 3 or more mixtures from spectra.
This will require the collection of suitable test spectra. We also intend to extend the TSA training library
to include other important minerals, as well as at least two varieties of water, bound and unbound; in
some situations, TSA tries to fit an additional mineral (such as Opaline Silica) to accomodate water
features which do not quite match those in our current library.

Regarding TSSA, we intend to build up an analogous library of atmospherically-corrected endmembers

obtained from a variety of scenes. In addition, we intend to use spatial context both to improve the
classification accuracy and to increase its speed. This may also be achieved using multi-processors. A
longer-term desire is to incorporate atmospheric correction into the fitting procedure, much as we have
done in incorporating the background and water into the algorithm used by TSA.

6. ACKNOWLEDGEMENTS

Between October, 1994 and October, 1997, the development of TSA and its underlying algorithms was
supported by the Australian Mineral Industries Research Association. We thank it and the sponsor-
 x
ing exploration companies who supported the research. TSA development has also benefitted from the
PIMA-II measurement of samples from the USGS spectral library, and we thank Dr. Roger Clark for
making this possible. Thanks also go to Peter Mason for the development of an interface between TSA
and the TSG package, to NASA/JPL for acquiring the AVIRIS scene of Oatman, and to Joe Boardman
for providing us with his endmember information for this scene.

REFERENCES

1. K. Ali, C.J. Horsfall, and R. Lister. Towards knowledge-based identification of mineral mixtures from re-
flectance spectra. Knowledge Based Systems, 2:5–13, 1989.
2. AusSpec International, P.O. Box 2235, Kew MDC, Victoria 3101, Australia. The Spectral Geologist Users
Manual, Version 1.1, 1998. 92 pages.
3. M. Berman and L. Bischof. The algorithms underlying The Spectral Assistant software. Technical Report
97/87, CSIRO Mathematical and Information Sciences, 1997.
4. J.W. Boardman. Post-ATREM polishing of AVIRIS apparent reflectance data using EFFORT: a lesson in
accuracy versus precision. In R.O. Green, editor, Summaries of the Seventh Annual JPL Airborne Earth
Science Workshop, volume 1, AVIRIS Workshop, pages 53–54. JPL Publ. 98-1, NASA, Pasadena, California,
1998.
5. J.W. Boardman and J.F. Huntington. Data analysis and product simulation workshop manual for the Aus-
tralian Resource Information and Environmental Satellite (ARIES-1). Technical Report 369R, CSIRO Ex-
ploration and Mining, 1997. 93 pages.
6. ENVI. The Environment for Visualising Images Users Guide, Version 3.0. Better Solutions Consulting LLC,
La Fayette, Colorado, U.S.A., 1997. 614 pages.
7. B.C. Gao and A.F.H. Goetz. Column atmospheric vapour and vegetation liquid water retrievals from airborne
imaging spectrometer data. Journal of Geophysical Research, 95:3549–3564, 1990.
8. B.C. Gao, K.B. Heidebrecht, and A.F.H. Goetz. Atmospheric Removal Program (ATREM) User’s Guide,
Version 2.0. Centre for the Study of Earth from Space (CSES), CIRES, University of Colorado, Boulder,
Colorado, 1996.
9. A.A. Green and M.D. Craig. Analysis of aircraft spectrometer data with logarithmic residuals. In G. Vane
and A. Goetz, editors, Proceedings of the Airborne Imaging Spectrometer Data Analysis Workshop, pages
111–119. JPL Publ. 85-41, NASA, Pasadena, California, 1985.
10. T. Hastie, A. Buja, and R. Tibshirani. Penalized discriminant analysis. Annals of Statistics, 23:73–102, 1995.
11. J.F. Huntington and J.W. Boardman. Semi-quantitative mineralogical and geological mapping with 1995
AVIRIS data. In Proceedings of Spectral Sensing Research ’95, ISSSR. AGPS, 1995.
12. F.A. Kruse, A.B. Lefkoff, and J.B. Dietz. Expert system-based mineral mapping in northern Death Val-
ley, California, using the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS). Remote Sensing of
Environment, 44:309–336, 1993.
13. R. Lister. Toward context dependent classification of infrared spectra by energy minimisation. In Proceedings
of the International Joint Conference on Neural Networks, San Diego, pages 1–6, 1990.
14. G.J. McLachlan. Discriminant Analysis and Statistical Pattern Recognition. John Wiley & Sons, New York,
1992.
15. A.J. Miller. Subset Selection in Regression. Chapman and Hall, London, 1990.
16. D.A. Skoog and D.M. West. Principles of Instrumental Analysis. Saunders College, Philadelphia, 2nd
edition, 1980.
17. G. Wahba. Spline Models for Observational Data. SIAM, Philadelphia, 1990.

View publication stats