331 
Neural networks, Cross-correlation and Parameterisation : A Comparison of 
approaches to mineralogical mapping using imaging spectrometer data. 
S. MACKIN and S. A. BRIGGS 
British National Space Centre, RSADU, ITE Monkswood, Abbots Ripton, Cambridgeshire, PE17 2LS. 
G. SIMPSON 
Earth Observation Sciences, Broadmede, Famham Business Park, Famham, Surrey, GU9 8 QL. 
ABSTRACT: 
This paper presents the preliminary results of the application of techniques for the recognition of 
characteristic features of minerals present in high spectral resolution imaging spectrometer data. A number 
of techniques have been developed to extract important mineral compositional information, including 
parameterisation techniques, cross-correlation and the more novel approach using an artificial neural 
network. All three techniques have been evaluated and the neural network and cross-correlation methods 
compared using test data sets of known mixture pairs with Gaussian noise and with AVIRIS data sets 
collected in Nevada, USA in the summer of 1990. The neural network and cross-correlation methods 
accurately recovered the mineral identities of the test series, the network classification being an order of 
magnitude faster than the cross-correlation technique. However, the network performed badly compared to 
the cross-correlation method when classifying AVIRIS data. The problems seem to be associated with the 
variability within the AVIRIS data itself (noise, mixtures, processing artifacts). Further work is in progress 
to improve the network operation to produce a more accurate classification. 
KEY WORDS : Neural networks, Parameterisation, Cross-correlation, Spectrometry, Mapping 
1 - INTRODUCTION 
Imaging spectrometers collect a vast amount of detailed information on the spectral characteristics of the 
surface composition of the earth. In principle they operate much like the LANDSAT Thematic Mapper 
(TM) series of satellites collecting the reflected solar radiation from the earths surface in discrete bands. 
These bands cover visible and infrared wavelengths of the solar spectrum. In the case of the LANDSAT 
satellites, six bands are collected in the 0.4pm to 2.5pm wavelength range, while the new imaging 
spectrometers such as the Airborne Visible / InfraRed Imaging Spectrometer (AVIRIS) collect 224 partly 
overlapping bands in the same wavelength region. 
The detailed spectral information reveals characteristic narrow absorption features due to 
the chemical composition and structure of the material reflecting the solar radiation. In the case of 
economically significant minerals, the differences are great enough to provide a means of direct 
identification of the spectrally dominant surface mineralogy from the spectra alone. 
2 - IMAGING SPECTROMETERS - DATA PROCESSING 
To identify the composition of the surface mineralogy the unknown spectra are compared with a reference 
set of library minerals. The primary aim of the data processing is to convert the imaging spectrometry data 
to the same form as the reference library of minerals. In this case, reflectance. 
2.1. AVIRIS data 
The AVIRIS sensor used in this study is a scanner which feeds the reflected solar radiation from the 
surface through four separate diffraction gratings onto four separate line arrays of detectors covering the 
0.4pm to 2.5pm wavelength range. Of the collected data a number of bands have no significant signal due 
to absorption of the solar radiation during its passage through the atmosphere to the surface. The primary 
absorbers are H 2 O, 02 and CO 2 . The affected bands were not used during processing, Overlapping bands 
between the four spectrometers of the instrument were also excluded during processing, along with the 
longer wavelength bands of the fourth spectrometer which were dominated by noise.