XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3 1521
for many of the analytical techniques and is grounded in
geophysical mapping technologies and sampling. Cost is an
important consideration in the data selection process. While
purchasing a hyperspectral sensor or acquiring a service can
appear costly the data density of hyperspectral instruments
is unparalleled by other geophysical and/or geochemical
measurements. With the data density, the cost per mate-
rial spectrum (pixel) can be very low especially over time.
Moreover, once the data is acquired, it can be (re)processed
based on new geometallurgical objectives without paying
for another survey or analysis. Additionally, this data can
be processed for geotechnical objectives as well. There are
many potential approaches for the processing of hyperspec-
tral data as these data have spectral and spatial components.
HYPERSPECTRAL DATA PROCESSING
Hyperspectral imagery data has many potential approaches.
This is because these data have spectral and spatial com-
ponents for processing, analysis, and interpretation. We
will focus on the processing of these data spectroscopically
as the data are spectroscopic measurements in the scene.
Data dimensionality is an important consideration when
determining the processing methodology however, it can
be mitigated by focusing on a material’s spectroscopic sig-
nature. Generally speaking, there are two classes for pro-
cessing, data-driven and knowledge-driven approaches
(Asadzadeh and de Souza Filho 2016).
Boardman and Kruse (1994) developed an automated
spectral approach with very specific algorithms (Minimum
Noise Transform, Pixel Purity Index, and Mixture Tuned
Matched Filter (MTMF) (Figure 5). This automated spec-
tral approach developed by Boardman and Kruse is a spe-
cific use case of a generalized automated spectral (Figure 6).
The generalized automated spectral approach allows for
the incorporation of spectral mixtures analysis, expert
systems, statistical signal processing, along with other
machine learning and artificial intelligence approaches.
However, the challenge in implementing the generalized
approach is understanding the strengths and weaknesses
of the algorithms and implementing them to answer
Figure 5. Laboratory Imaging Spectrometers and the NeoSpectra point spectrometer in the University
of Arizona’s Mine Imaging INfrared Emission Raman Spectroscopy (MIINERS) Laboratory. Upper
left: SWIR imaging spectrometer is a Specim SWIR (1.0–2.5 mm). Upper Right: The FX-10 is a VNIR
imaging spectrometer from Specim (0.4–1.0 mm). Lower Left: The NeoSpectra handheld FTIR infrared
spectrometer (1.3–2.5 mm) from Si-ware. Lower Right: The FX-50 is a MWIR imaging spectrometer
(2.7–5.3 mm) from Specim. The FX10, SWIR, and FX-50 imaging spectrometers can be mounted on a
rotational stage for imaging bench faces
for many of the analytical techniques and is grounded in
geophysical mapping technologies and sampling. Cost is an
important consideration in the data selection process. While
purchasing a hyperspectral sensor or acquiring a service can
appear costly the data density of hyperspectral instruments
is unparalleled by other geophysical and/or geochemical
measurements. With the data density, the cost per mate-
rial spectrum (pixel) can be very low especially over time.
Moreover, once the data is acquired, it can be (re)processed
based on new geometallurgical objectives without paying
for another survey or analysis. Additionally, this data can
be processed for geotechnical objectives as well. There are
many potential approaches for the processing of hyperspec-
tral data as these data have spectral and spatial components.
HYPERSPECTRAL DATA PROCESSING
Hyperspectral imagery data has many potential approaches.
This is because these data have spectral and spatial com-
ponents for processing, analysis, and interpretation. We
will focus on the processing of these data spectroscopically
as the data are spectroscopic measurements in the scene.
Data dimensionality is an important consideration when
determining the processing methodology however, it can
be mitigated by focusing on a material’s spectroscopic sig-
nature. Generally speaking, there are two classes for pro-
cessing, data-driven and knowledge-driven approaches
(Asadzadeh and de Souza Filho 2016).
Boardman and Kruse (1994) developed an automated
spectral approach with very specific algorithms (Minimum
Noise Transform, Pixel Purity Index, and Mixture Tuned
Matched Filter (MTMF) (Figure 5). This automated spec-
tral approach developed by Boardman and Kruse is a spe-
cific use case of a generalized automated spectral (Figure 6).
The generalized automated spectral approach allows for
the incorporation of spectral mixtures analysis, expert
systems, statistical signal processing, along with other
machine learning and artificial intelligence approaches.
However, the challenge in implementing the generalized
approach is understanding the strengths and weaknesses
of the algorithms and implementing them to answer
Figure 5. Laboratory Imaging Spectrometers and the NeoSpectra point spectrometer in the University
of Arizona’s Mine Imaging INfrared Emission Raman Spectroscopy (MIINERS) Laboratory. Upper
left: SWIR imaging spectrometer is a Specim SWIR (1.0–2.5 mm). Upper Right: The FX-10 is a VNIR
imaging spectrometer from Specim (0.4–1.0 mm). Lower Left: The NeoSpectra handheld FTIR infrared
spectrometer (1.3–2.5 mm) from Si-ware. Lower Right: The FX-50 is a MWIR imaging spectrometer
(2.7–5.3 mm) from Specim. The FX10, SWIR, and FX-50 imaging spectrometers can be mounted on a
rotational stage for imaging bench faces