Spatial Co-Registration and Spectral
Concatenation of Panoramic Ground-Based
Hyperspectral Images
Unal Okyay and Shuhab D. Khan
Abstract
Most of the studies employing ground-based hyperspectral
cameras have been restricted to either visible-near infrared
(
VNIR
) or shortwave infrared (
SWIR
) spectrum. In this study,
continuous
VNIR+SWIR
image spectra were obtained from
two separate hyperspectral cameras. Spatial co-registration
of panoramic hyperspectral images was performed us-
ing homologous points automatically extracted by the
scale-invariant feature transform (
SIFT
) algorithm. The
effect of input image selection on identifying homologous
points and different transformation techniques on image
co-registration were evaluated. Although proper spatial
image co-registration was expected to provide persistent
spectral data, spectral concatenation of the hyperspectral
images required characterizing and correcting the errors
and inconsistencies due to camera sensor properties and
limitations. Bringing spectral data from two hyperspectral
cameras together provided the opportunity to identify and
differentiate lithological units, which would not have been
possible using a single sensor, and thus allow more detailed
and extended spectral analysis of near-vertical outcrops.
Introduction
Imaging spectroscopy, i.e., hyperspectral imaging, has been
successfully utilized in geology to obtain chemical-mineralog-
ical information about the Earth’s surface for several decades
(Goetz, 2009). Hyperspectral imaging instruments collect
reflectance data from hundreds of very narrow bands within
their operational spectral range, which allows a detailed analy-
sis of diagnostic absorption features (Goetz
et al
., 1985). Until
recently, hyperspectral imaging data have been acquired most-
ly from airborne platforms along with some from spaceborne
platforms (Bishop
et al
. 2018). With the advent of compact
hyperspectral cameras, there has been an increasing interest in
the use of these instruments for close-range imaging spectros-
copy applications in geological studies. These applications
may include ground- and/or laboratory-based hyperspectral
data collection (e.g., Kurz
et al
., 2012; Okyay
et al
., 2016; Zaini
et al
., 2014). While laboratory-based non-imaging instruments
have been utilized for spectrometry (Goetz
et al
., 1985) and
more recently for drill core scanning (Huntington
et al
., 2006;
Mauger, 2007), field-deployable imaging systems provided an
unprecedented way of scanning near-vertical geologic outcrops
(e.g., cliffs, roadcuts, and quarry walls) with high resolution.
Such outcrops can provide excellent rock exposures for de-
tailed geological observations; however, they are often inac-
cessible for direct measurements and airborne hyperspectral
platforms fail to provide an adequate viewing angle.
Most of the studies employing ground-based hyperspectral
cameras have been restricted to either visible near-infrared
(
VNIR
: 0.4 – 1.0 µm) spectrum (Okyay and Khan, 2016; Sun and
Khan, 2016) or shortwave infrared (
SWIR
: 1.0 – 2.5 µm) spec-
trum (Alonso de Linaje
et al
., 2018; Alonso de Linaje and Khan,
2017; Krupnik
et al
., 2016; Kurz
et al
., 2012; Okyay
et al
., 2016;
Sun
et al
., 2018 and 2017; Zaini
et al
., 2014). This is mainly
because
VNIR
and
SWIR
spectra require separate detectors of dif-
ferent types and sensitivity while most of the existing cameras
have only one type of detector, and the use of multiple cameras
is not widespread due to cost and operational complexity. Only
a few studies (Khan
et al
., 2018; Murphy
et al
., 2014 and 2012)
have employed
VNIR
and
SWIR
hyperspectral cameras together
for spectral analysis. Depending on the chemical-mineralogical
properties and conditions of the outcrops in question, using a
single camera, commonly in
SWIR
range for geological applica-
tions, may be sufficient to identify the surface geochemistry. A
number of rock-forming minerals (and thus, rocks), however,
have absorption features both in
VNIR
due to electronic pro-
cesses and
SWIR
due to vibrational processes (Clark, 1999; Clark
et al
., 1990). Therefore, a continuous
VNIR
+
SWIR
spectrum is
desired for ground-based hyperspectral imaging in geological
applications to capture diagnostic absorption features over the
wavelength range from 0.4 to 2.5 µm.
Combining spectral data from different hyperspectral imag-
ing instruments requires spatial image co-registration. Con-
ventionally, this is performed using homologous (correspond-
ing) points commonly extracted from input images manually.
Considering the dissimilar spatial resolutions, spectral ranges,
and slightly different viewing angles of the cameras, extract-
ing multiple matching points from natural surfaces in pan-
oramic ground-based hyperspectral images manually is not
straightforward and often time consuming. Thus, an auto-
mated method for point extraction and matching is desired.
There are two main approaches to automatically extract paired
homologous points from input images: intensity-based and
feature-based image matching (Zitová and Flusser, 2003).
Unlike intensity-based algorithms, feature-based algorithms
are generally insensitive to changes in rotation, scale, and
viewpoint, which makes them preferable in remote sensing
applications. There are several well-known feature-based
algorithms, such as Speeded-Up Robust Features (
SURF
) (Bay
Unal Okyay is with the National Center for Airborne Laser
Mapping, Department of Civil and Environmental Engineering,
University of Houston, 5000 Gulf Freeway, Bldg.4, Ste.216,
Houston, TX, 77204; and formerly with the Department of Earth
and Atmospheric Sciences, University of Houston, 3507 Cullen
Blvd. SR Bldg.1, Rm.312, Houston, TX, 77204 (
)
Shuhab D. Khan is with the Department of Earth and
Atmospheric Sciences, University of Houston, 3507 Cullen
Blvd. SR Bldg.1 Rm.312, Houston, TX, 77204
Photogrammetric Engineering & Remote Sensing
Vol. 84, No. 12, December 2018, pp. 781–790.
0099-1112/18/781–790
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.12.781
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
December 2018
781
