ArticlePDF Available

Abstract and Figures

While uncrewed aerial vehicles are routinely used as platforms for hyperspectral sensors, their application is mostly confined to nadir imaging orientations. Oblique hyperspectral imaging has been impeded by the absence of robust registration and correction protocols, which are essential to extract accurate information. These corrections are especially important for detecting the typically small spectral features produced by minerals, and for infrared data acquired using pushbroom sensors. The complex movements of unstable platforms (such as UAVs) require rigorous geometric and radiometric corrections, especially in the rugged terrain often encountered for geological applications. In this contribution we propose a novel correction methodology, and associated toolbox, dedicated to the accurate production of hyperspectral data acquired by UAVs, without any restriction concerning view angles or target geometry. We make these codes freely available to the community, and thus hope to trigger an increasing usage of hyperspectral data in Earth sciences, and demonstrate them with the production of, to our knowledge, the first fully corrected oblique SWIR drone-survey. This covers a vertical cliff in the Dolomites (Italy), and allowed us to distinguish distinct calcitic and dolomitic carbonate units, map the qualitative abundance of clay/mica minerals, and thus characterise seismic scale facies architecture.
This content is subject to copyright.


Citation: Thiele, S.T.; Bnoulkacem, Z.;
Lorenz, S.; Bordenave, A.; Menegoni,
N.; Madriz, Y.; Dujoncquoy, E.;
Gloaguen, R.; Kenter, J. Mineralogical
Mapping with Accurately Corrected
Shortwave Infrared Hyperspectral
Data Acquired Obliquely from UAVs.
Remote Sens. 2022,14, 5. https://
doi.org/10.3390/rs14010005
Academic Editors: Prasad
S. Thenkabail and Norman Kerle
Received: 20 October 2021
Accepted: 17 December 2021
Published: 21 December 2021
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2021 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
remote sensing
Article
Mineralogical Mapping with Accurately Corrected Shortwave
Infrared Hyperspectral Data Acquired Obliquely from UAVs
Samuel T. Thiele 1, * , Zakaria Bnoulkacem 1,2, Sandra Lorenz 1, Aurélien Bordenave 3, NiccolòMenegoni 4,
Yuleika Madriz 1, Emmanuel Dujoncquoy 5, Richard Gloaguen 1and Jeroen Kenter 5
1Helmholtz-Zentrum Dresden-Rossendorf, Helmholtz Institute Freiberg for Resource Technology,
09599 Freiberg, Germany; z.bnoulkacem@hzdr.de (Z.B.); s.lorenz@hzdr.de (S.L.);
y.madriz-diaz@hzdr.de (Y.M.); r.gloaguen@hzdr.de (R.G.)
2Department of Photogrammetry and Cartography ESGIT, IAV Hassan II University, Rabat 10101, Morocco
3Applied Geology and Sedimentology, Technopole Hélioparc, 64000 Pau, France;
aurelien.bordenave@ageos-science.com
4Department of Earth and Environmental Sciences, University of Pavia, 27100 Pavia, Italy;
niccolo.menegoni01@universitadipavia.it
5TotalEnergies, 64000 Pau, France; emmanuel.dujoncquoy@total.com (E.D.);
jeroen.kenter@totalenergies.com (J.K.)
*Correspondence: s.thiele@hzdr.de
Abstract:
While uncrewed aerial vehicles are routinely used as platforms for hyperspectral sensors,
their application is mostly confined to nadir imaging orientations. Oblique hyperspectral imaging
has been impeded by the absence of robust registration and correction protocols, which are essential
to extract accurate information. These corrections are especially important for detecting the typically
small spectral features produced by minerals, and for infrared data acquired using pushbroom sensors.
The complex movements of unstable platforms (such as UAVs) require rigorous geometric and
radiometric corrections, especially in the rugged terrain often encountered for geological applications.
In this contribution we propose a novel correction methodology, and associated toolbox, dedicated to
the accurate production of hyperspectral data acquired by UAVs, without any restriction concerning
view angles or target geometry. We make these codes freely available to the community, and thus hope
to trigger an increasing usage of hyperspectral data in Earth sciences, and demonstrate them with the
production of, to our knowledge, the first fully corrected oblique SWIR drone-survey. This covers
a vertical cliff in the Dolomites (Italy), and allowed us to distinguish distinct calcitic and dolomitic
carbonate units, map the qualitative abundance of clay/mica minerals, and thus characterise seismic
scale facies architecture.
Keywords: infrared; hyperspectral; uncrewed aerial vehicle; calcite; dolomite
1. Introduction
Uncrewed aerial vehicles (UAVs) have revolutionised geological mapping by facilitat-
ing detailed and quantitative outcrop characterisation. Recent reviews highlight a variety
of applications, including geotechnical investigations, geohazard assessment, structural
analysis, stratigraphic mapping and teaching [
1
4
]. Most of these studies relied on an
interpretation of outcrop colour captured using conventional lightweight RGB cameras,
and geometric information from light detection and ranging (LiDAR) or structure from
motion multi-view stereo (SfM-MVS) workflows [5].
Recently, miniaturised hyperspectral cameras covering the shortwave infrared range
(SWIR; 1000–2500 nm) provided an opportunity for similarly detailed outcrop studies, but
with an added hyperspectral dimension that allows the distribution and abundance of
many rock-forming minerals (e.g., carbonates, clays, micas, chlorites, amphiboles and other
minerals with diagnostic features in the SWIR; [6]) to be mapped.
Remote Sens. 2022,14, 5. https://doi.org/10.3390/rs14010005 https://www.mdpi.com/journal/remotesensing
Remote Sens. 2022,14, 5 2 of 21
Previous studies using tripod-mounted sensors have demonstrated the significant
potential of SWIR data for geological mapping applications (e.g., [
7
10
]), but have been
limited in resolution by access constraints and oblique viewing angles. These constraints
can be especially challenging when scanning cliffs or mines, as (1) upward looking viewing
angles result in highly oblique images and associated distortions, and (2) rockfall hazards
and limited access to scree slopes prevents deployment of (generally heavy) equipment for
close-up mapping.
To avoid these limitations, hyperspectral sensors are increasingly being deployed on
UAV platforms. A plethora of UAV compatible visible-near infrared (VNIR) sensors are
available (cf. [
11
]); however, weight limitations present significant challenges for SWIR
cameras, which require heavy and energy intensive heat pumps to cool the detector to
<
100
C before spectrally accurate data can be acquired. For this reason, SWIR sensors
have only recently become small and lightweight enough for deployment on UAVs [
12
14
].
Most hyperspectral sensors do not acquire an entire data cube simultaneously, but
either acquire bands sequentially (e.g., the Senop Oy Rikola) or by splitting a narrow slit of
incoming light across a 2D sensor to acquire 1-pixel wide lines of continuous spectral data
(pushbroom sensors; [
11
]). Currently available UAV-compatible SWIR cameras all use push-
broom acquisition modes. As a result, sensor movement during data acquisition causes
complex geometric distortions that must be corrected before usable data can be obtained.
Gimbal systems can be employed to minimise the influence of vibrations, but distortions
due to the movement of the UAV along its flight path are unavoidable (
especially
for push-
broom sensors, which require this movement to cover the area of interest; Figure 1). These
distortions are larger for UAV platforms than for conventional aircraft, due to their lighter
weight and the higher resolution ground sampling associated with low flight altitudes
(see [15] for a detailed description of the challenges associated with pushbroom sensors).
For frame-sensors, image matching and computer vision techniques can be used to co-
register hyperspectral bands and so correct for geometric distortions [
10
,
16
,
17
]. Pushbroom
sensors are more challenging; an inertial processing unit (IMU) and global positioning
system (GPS) is needed to accurately quantify the movement of the sensor and remove
distortions by back-projection (e.g., [
15
]). Despite these challenges, several authors have
successfully used pushbroom VNIR sensors to collect extremely high-resolution
(2–10 cm)
data with reasonable levels of spatial accuracy (5–100 cm; [
15
,
18
20
]). All of these authors
corrected distortions resulting from UAV movement using closed-source geometric cor-
rections implemented in the PARGE software [
21
]. This approach applies a ray-tracing
algorithm to locate the centre point of each hyperspectral pixel on a high-resolution digital
elevation model (DEM) and then interpolates between these to fill gaps as necessary. Due to
uncertainties in measuring sensor position and orientation using an inertial processing unit
(IMU), most authors then warp the resulting hyperspectral orthomosaic to fit measured
ground control points [
21
] or a high-resolution photogrammetric RGB orthomosaic [
18
,
20
].
Due to their reliance on raster datasets, this correction method performs poorly in
topographically complex areas, which cannot be compressed into a 2.5D raster [
15
].
Areas
of geological interest and good exposure, such as cliffs, quarries and mines, often in-
clude sub-vertical faces with multiple orientations, so cannot be projected onto a 2D grid
(DEM/orthomosaic) without introducing projective distortions that significantly degrade
data quality [
10
,
22
,
23
]. To mitigate these issues and facilitate the collection of geological
data from potentially important but otherwise inaccessible outcrops, a rapidly increasing
number of methods are being developed to fuse remotely sensed hyperspectral data with
dense 3D point clouds to create hyperclouds [
8
,
10
,
17
,
23
,
24
] that objectively record outcrop
geometry and composition.
Remote Sens. 2022,14, 5 3 of 21
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Figure 1. Data acquisition procedure using a UAV-mounted pushbroom sensor. The sensor is
mounted on a UAV, which moves over the target of interest (a) and captures light from within a
thin scanline (green) to progressively construct a hyperspectral data cube (b). This data cube must
then be geometrically corrected to a spatially accurate hyperspectral swath (c) that can be analysed
or interpreted to map, e.g., mineralogy.
In this contribution, we present (1) a novel and open-source python workflow for
applying robust geometric and radiometric corrections to hyperspectral data, regardless
of acquisition and target geometry, and (2) results from a demonstration survey con-
ducted to map the distribution of dolomitic, calcitic and argillaceous carbonate lithologies
exposed along cliffs near Passo Giau, Italy (Figure 2). This transect is large enough (>500
× 100 m) to be directly analogous to geological structures resolvable in subsurface seismic
data.
2. Location
The study area is located on the south-eastern side of the Gusela del Nuvolau outcrop
[25], north of Passo Giau in the central-eastern Italian Dolomites (Figure 2). This 600 m
long and 300 m high cliff (Figure 3) exposes remnants of an exhumed Cassian carbonate
platform of the Upper Ladinian to Lower Carnian age, including the transition from plat-
form (Cassian Dolomite Formation) to basinal (San Cassiano Formation) facies [26]. The
Cassian Dolomite is a greyish-white fine grained dolostone with sucrose texture that, due
to the pervasive dolomitisation of the Cassian platform, preserves few original character-
istics of the rock (e.g., original textures, fossils or sedimentary structures; [27]). The San
Cassiano Formation comprises alternating marls and limestones with oolitic/bioclastic
grainstone [26,27]. In general, the geometries of the Cassian platform and the San Cassiano
Formation (Figure 3) suggest a north-east (NE) progradation of the slope deposits onto
the basinal unit [25,28].
Figure 1.
Data acquisition procedure using a UAV-mounted pushbroom sensor. The sensor is
mounted on a UAV, which moves over the target of interest (
a
) and captures light from within a thin
scanline (green) to progressively construct a hyperspectral data cube (
b
). This data cube must then
be geometrically corrected to a spatially accurate hyperspectral swath (
c
) that can be analysed or
interpreted to map, e.g., mineralogy.
In this contribution, we present (1) a novel and open-source python workflow for
applying robust geometric and radiometric corrections to hyperspectral data, regardless of
acquisition and target geometry, and (2) results from a demonstration survey conducted to
map the distribution of dolomitic, calcitic and argillaceous carbonate lithologies exposed
along cliffs near Passo Giau, Italy (Figure 2). This transect is large enough (>500
×
100 m)
to be directly analogous to geological structures resolvable in subsurface seismic data.
2. Location
The study area is located on the south-eastern side of the Gusela del Nuvolau out-
crop [
25
], north of Passo Giau in the central-eastern Italian Dolomites (Figure 2). This
600 m
long and 300 m high cliff (Figure 3) exposes remnants of an exhumed Cassian carbonate
platform of the Upper Ladinian to Lower Carnian age, including the transition from plat-
form (Cassian Dolomite Formation) to basinal (San Cassiano Formation) facies [
26
]. The
Cassian Dolomite is a greyish-white fine grained dolostone with sucrose texture that, due
to the pervasive dolomitisation of the Cassian platform, preserves few original character-
istics of the rock (e.g., original textures, fossils or sedimentary structures; [
27
]). The San
Cassiano Formation comprises alternating marls and limestones with oolitic/bioclastic
grainstone [
26
,
27
]. In general, the geometries of the Cassian platform and the San Cassiano
Formation (Figure 3) suggest a north-east (NE) progradation of the slope deposits onto the
basinal unit [25,28].
Remote Sens. 2022,14, 5 4 of 21
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Figure 2. Location map (a) of the study site and simplified geological map of the surrounding area
(b; cf. [27]). The Gusela del Nuvolau outcrop is indicated by the yellow dashed oval. This exposes
Middle/Upper Triassic stratigraphy (c) of the central-eastern Italian Dolomites [26]: DP, Dolomia
Principale; TVZ, Travenanzes Formation (Fm.); HKS, Heiligkreutz Fm.; DCS, Cassian Dolomite Fm.;
SCS, San Cassiano Fm.; WEN, Wengen Fm.; SCI, Sciliar Fm.; IMF, Mt. Fernazza Fm.
Figure 3. Photograph of the study area (Gusela del Nuvolau cliff). The carbonate body contains
dolomitic and calcitic units, but these cannot be distinguished using the naked eye. Cliff is 250300
m high and 600 m long. The grey rectangle shows the approximate extent of the hyperspectral sur-
vey.
Figure 2.
Location map (
a
) of the study site and simplified geological map of the surrounding area
(
b
; cf. [
27
]). The Gusela del Nuvolau outcrop is indicated by the yellow dashed oval. This exposes
Middle/Upper Triassic stratigraphy (
c
) of the central-eastern Italian Dolomites [
26
]: DP, Dolomia
Principale; TVZ, Travenanzes Formation (Fm.); HKS, Heiligkreutz Fm.; DCS, Cassian Dolomite Fm.;
SCS, San Cassiano Fm.; WEN, Wengen Fm.; SCI, Sciliar Fm.; IMF, Mt. Fernazza Fm.
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Figure 2. Location map (a) of the study site and simplified geological map of the surrounding area
(b; cf. [27]). The Gusela del Nuvolau outcrop is indicated by the yellow dashed oval. This exposes
Middle/Upper Triassic stratigraphy (c) of the central-eastern Italian Dolomites [26]: DP, Dolomia
Principale; TVZ, Travenanzes Formation (Fm.); HKS, Heiligkreutz Fm.; DCS, Cassian Dolomite Fm.;
SCS, San Cassiano Fm.; WEN, Wengen Fm.; SCI, Sciliar Fm.; IMF, Mt. Fernazza Fm.
Figure 3. Photograph of the study area (Gusela del Nuvolau cliff). The carbonate body contains
dolomitic and calcitic units, but these cannot be distinguished using the naked eye. Cliff is 250300
m high and 600 m long. The grey rectangle shows the approximate extent of the hyperspectral sur-
vey.
Figure 3.
Photograph of the study area (Gusela del Nuvolau cliff). The carbonate body contains
dolomitic and calcitic units, but these cannot be distinguished using the naked eye. Cliff is 250–300 m
high and 600 m long. The grey rectangle shows the approximate extent of the hyperspectral survey.
Remote Sens. 2022,14, 5 5 of 21
The outcrop is accessible today due to the Alpine orogenic event, during which the
Dolomites underwent only mild tectonic deformation [
29
31
]. This allowed preservation
of seismic scale depositional geometries of the Middle Triassic platform [25] at Gusela del
Nuvolau (Figure 3).
Geological determination of lithofacies partitioning (i.e., quantification of spatial juxta-
position and proportions of geological rock types) are typically obscured by a combination
of access limitations and subjective, non-reproducible interpretation by geologists. The
primary objectives to deploy the UAV and VNIR-SWIR camera are to provide an accurate
and objective method for mapping lithological and mineralogical variations (here calcite,
dolomite and clay), which benefits the reconstruction of realistic and accurate models in the
subsurface for the energy industry (geothermal, CO
2
sequestration and hydrocarbons). The
Passo Giau outcrop was selected as a technical test site to demonstrate the method’s poten-
tial, test correction procedures and validate the remotely sensed hyperspectral imagery.
3. Methods
3.1. Hyperspectral Data Acquisition
We used a HySpex Mjolnir VS-620 system (Table 1), which allows the concurrent
acquisition of hyperspectral data in the VNIR range (400–1000 nm, HySpex Mjolnir V-1240
sensor with 1024 spatial pixels) and the SWIR range (970–2500 nm, HySpex Mjolnir S-620
sensor with 620 spatial pixels). These sensors operate as pushbroom scanners to create
a spectral data cube by constant acquisition and movement along the target of interest
(Figure 1).
Unlike typical nadir-oriented sensor setups, we mounted the Mjolnir sideways in a
Gremsy gStabi H16 XL gimbal so that the camera faced the cliff face and the scan line
was sub-vertical. This setup was carried by a custom-built octocopter UAV (Figure 4)
which, with the equipped payload of ~8 kg, can achieve flight times of 15–20 min under
ideal conditions.
In the current study, two swaths of hyperspectral data (hereafter referred to as Swath
A and Swath B) were acquired by flying gently descending flight lines (parallel to bedding)
at 140 (Swath A) and 90 m (Swath B) distance from the cliff (Figure 5). Due to generally
sunny conditions, a frame rate of 50 fps and resulting integration times of 10 (VNIR) and
20 ms
(SWIR) was sufficient to ensure adequate image exposure, and a velocity of ~2 m/sec
chosen to give approximately square pixels. This configuration resulted in a theoretical
ground sampling distance of 8 cm (Flight A) and 5 cm (Flight B) for the SWIR sensor
(Table 1).
High winds (~14 m/s) and poor GPS signal presented significant challenges during
data acquisition, as pushbroom sensors require flight lines to be conducted in as smooth a
fashion as possible. These were (partially) mitigated by orienting the scanner such that it
faced the cliff while the UAV was flying forwards, as opposed to orienting the scanner in a
forward-facing direction and flying the UAV sideways. This configuration was found to be
more stable, because the UAV does not need to constantly manoeuvre to maintain a fixed
heading. The gently descending flight lines also helped to stabilise the UAV.
Table 1.
Sensor parameters of the HySpex Mjolnir VS-620 combined hyperspectral imaging and
photogrammetry system deployed in the study. Data from the V-1240 were spatially downsampled
(binned) to match the S-620 sensor for this study.
HySpex V-1240 HySpex S-620 Sony Alpha 6400
Spectral range 400–1000 nm 970–2500 nm RGB true color
Spatial pixels 1240 620 6000 ×4000
Spectral channels/sampling 200/3 nm 300/5.1 nm 3/-
FOV 20 deg (0.27 mrad/px) 20 deg (0.54 mrad/px) 40.6 ×60.9 deg (0.18 mrad/px)
Weight 7.253 kg (without cables)
Remote Sens. 2022,14, 5 6 of 21
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Table 1. Sensor parameters of the HySpex Mjolnir VS-620 combined hyperspectral imaging and
photogrammetry system deployed in the study. Data from the V-1240 were spatially downsampled
(binned) to match the S-620 sensor for this study.
HySpex V-1240
Sony Alpha 6400
Spectral range
4001000 nm
RGB true color
Spatial pixels
1240
6000 × 4000
Spectral channels/sampling
200/3 nm
3/-
FOV
20 deg (0.27 mrad/px)
40.6 x 60.9 deg (0.18 mrad/px)
Weight
7.253 kg (without cables)
Figure 4. Image of the UAV and Hyspex Mjolnir-VS mounted sideways in the Gremsy gStabi H16
XL gimbal. Lens caps for the three sensors (VS-1240, S-620 and Sony Alpha) can be seen on the right
side of the camera.
Figure 4.
Image of the UAV and Hyspex Mjolnir-VS mounted sideways in the Gremsy gStabi H16 XL
gimbal. Lens caps for the three sensors (VS-1240, S-620 and Sony Alpha) can be seen on the right side
of the camera.
3.2. Photogrammetry
The HySpex Mjolnir camera was also fitted with a co-aligned 24.2 megapixel mirrorless
DSLR camera (Sony Alpha 6400) that captured an RGB image every 200 hyperspectral
frames, resulting in 205 images with a horizontal overlap of ~80%. These images were
supplemented by an additional 144 images captured using a DJI Spark and 132 images from
a Parrot Anafi to ensure the entire outcrop was adequately covered, and then processed
using Agisoft Metashape (v. 1.6.3 build 10732) to generate a sparse cloud containing
249,329 points
using structure from motion (SfM). This sparse cloud was manually filtered
to remove outliers and then georeferenced using 14 ground control points surveyed using
a PPK GNSS rover (8 points; 3–5 cm accuracy) and Total Station (6 points; ~10 cm accuracy)
before re-running the bundle adjustment to mitigate potential photogrammetric distortions.
Finally, a dense cloud containing 99,141,154 points was created using the multi-view stereo
technique implemented in Agisoft Metashape and exported to CloudCompare [
32
], where
it was subsampled by removing closely spaced points until they had an average spacing of
5 cm, and clipped to the region of interest. Details of the photogrammetric reconstruction
can be found in Supplementary Materials 1.
Remote Sens. 2022,14, 5 7 of 21
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Figure 5. Post processed flight paths for flight A and B draped over the photogrammetric model.
Coloured points show the camera position at each frame (line of hyperspectral data), with blue at
the start of each survey and red at the end. The closely spaced red vectors show the camera viewing
direction at each of these positions. The transparent grey area shows the map-view extent of the
surveyed section of cliff. The location of the Zenith calibration panels (Section 3.5) and ground-truth
spectra (Section 3.7) are also overlain for reference.
3.2. Photogrammetry
The HySpex Mjolnir camera was also fitted with a co-aligned 24.2 megapixel mirror-
less DSLR camera (Sony Alpha 6400) that captured an RGB image every 200 hyperspectral
frames, resulting in 205 images with a horizontal overlap of ~80%. These images were
supplemented by an additional 144 images captured using a DJI Spark and 132 images
from a Parrot Anafi to ensure the entire outcrop was adequately covered, and then pro-
cessed using Agisoft Metashape (v. 1.6.3 build 10732) to generate a sparse cloud contain-
ing 249,329 points using structure from motion (SfM). This sparse cloud was manually
filtered to remove outliers and then georeferenced using 14 ground control points sur-
veyed using a PPK GNSS rover (8 points; 35 cm accuracy) and Total Station (6 points;
~10 cm accuracy) before re-running the bundle adjustment to mitigate potential photo-
grammetric distortions. Finally, a dense cloud containing 99,141,154 points was created
using the multi-view stereo technique implemented in Agisoft Metashape and exported
to CloudCompare [32], where it was subsampled by removing closely spaced points until
Figure 5.
Post processed flight paths for flight A and B draped over the photogrammetric model.
Coloured points show the camera position at each frame (line of hyperspectral data), with blue at
the start of each survey and red at the end. The closely spaced red vectors show the camera viewing
direction at each of these positions. The transparent grey area shows the map-view extent of the
surveyed section of cliff. The location of the Zenith calibration panels (Section 3.5) and ground-truth
spectra (Section 3.7) are also overlain for reference.
3.3. Geometric Corrections
As with all airborne pushbroom scanners, complex geometric distortions caused by
changes in position and orientation of the camera during flight require correction before a
usable image can be obtained. These corrections rely on precise position and orientation
data recorded by an Applanix APX-20 inertial processing unit (IMU) mounted within the
Hyspex Mjolnir camera to model and remove geometric distortions. However, established
correction workflows (e.g., [
11
,
15
]) rely on 2.5D representations of scene geometry and so
cannot correct highly oblique surveys covering e.g., cliffs or open-pit mines. For this reason
we have developed and implemented a novel geometric correction procedure within the
open-source hylite toolbox (cf. [10]).
First, as was previously described by [
15
], we processed the APX-20 data using
Trimble’s PosPAC software to apply post processed kinematic (PPK) corrections to the
GNSS position data and fuse this with accelerometer and orientation measurements from
the IMU using the Applanix Inertial Navigation System (INS) fusion toolbox. Data from
the onboard GNSS was corrected using a Trimble base station that we set up on site
Remote Sens. 2022,14, 5 8 of 21
using a Zephyr 2 GNSS Antenna and R5 receiver. The position of our base station was
determined using static measurement with an occupation time of 5 h and corrected using
Trimble Business Centre and permanent base station data from Bolzano, Bozen, Cercivento,
Rovereto, Padova and Venezia.
The cliffs obscured a significant part of the sky during data acquisition, meaning the
onboard GNSS receiver only recorded data from 5–12 satellites during the flight, causing
highly degraded GNSS solutions (PDOP < 4 for ~30% of the flight time). Because of this,
we processed each flight line separately and manually removed the degraded sections. This
lowered the overall quality of the processing, especially because it meant only one IMU
calibration figure was available for each survey, creating discrepancies between the two
flight lines. Each flight was processed twice using the INS fusion algorithm in POSPac; the
first solution was used to identify the lever-arm between the GNSS receiver and the IMU,
and the second to obtain the final smoothed best estimate (SBET) trajectory.
This trajectory was exported and processed using software provided by Hyspex
(Hyspex Nav) to derive camera orientation and position measurements for the pixel-centres
of each hyperspectral line (Figure 5). A boresight correction was also applied using HySpex
Nav and values determined from ground control points surveyed during a calibration
flight conducted by the sensor manufacturer.
We then used hylite [
10
] to project the photogrammetric point cloud onto the hy-
perspectral image using the measured position and orientation of the sensor during the
acquisition of each line of hyperspectral pixels, and so built a sparse mapping matrix (
M
)
that relates points in the scene to pixels in each swath.
M
has dimensions such that each
row corresponds to a point in the point cloud and each column the (flattened) index of
a hyperspectral pixel.
Mij
thus represents the visibility of point i(with position
ui
) in
pixel jgiven the camera position vector
c
and orientation matrix
R
at the start (n) and end
(n + 1)
of the hyperspectral frame. Note that unlike conventional frame sensors, pushbroom
acquisition modes often image the same point multiple times (due to rotation of the sensor
during acquisition), so
M
represents a many to many mapping between points and pixels.
For pushbroom sensors, a point can be considered to be visible from pixel jif it crossed
the sensor scan line during the exposure of this pixel, resulting in a change in sign (sgn) of
the dot product between the camera motion vector and relative point position vector (in
camera coordinates), satisfying:
sgn([(ucn)·Rn]1)6=sgn([(ucn+1)·Rn+1]1)(1)
This will be true only if the pixel crossed the sensor plane between frames nand
n + 1. Points outside the sensor field of view are then culled (based on a perspective
projection and focal length f), so that points are only considered visible if they satisfy both
Equation (1) and:
1<f×[(ucm)·Rm]0
[(ucm)·Rm]2
<1 (2)
where either m=nor m=n+ 1 (cf. [
33
] for an overview of the mathematics behind
perspective projection).
Constructing
M
for large clouds (10–100 million points) and hyperspectral swaths
(>1000 s of lines) would require prohibitively expensive computations, as
Equations (1) and (2)
need to be evaluated for every point and every line of pixels. Our algorithm reduces the
number of required computations by dividing each hyperspectral image into 200 pixel wide
chunks and, for each chunk, considering only points that cross the sensor plane between
its first and last line, and are within the cross-track field of view. This form of frustum
culling (a commonly applied technique for rapid 3-D rendering; [
33
]) significantly reduces
the computations required by projecting only a small subset of the total point cloud onto
each line of pixels, by assuming the camera position and orientation vary smoothly over
time (a requirement for collecting usable data).
Remote Sens. 2022,14, 5 9 of 21
To facilitate the identification and removal of occluded points (that map onto the
sensor after a perspective projection but were not visible due to a light-blocking object in
the foreground), the inverse of the distance between each point-pixel pair were stored as the
elements of the mapping matrix
M
. This allows efficient querying of point-pixel depths (by
computing the argmax of
M
along its columns) and the construction of a depth buffer [
33
]
that can be used to identify and remove elements in Mthat result from occluded points.
Different normalisations or filtering can then be applied to
M
to derive a normalised
mapping matrix, denoted
Mf
, which can be multiplied by a vector of pixel spectra (
Ps
) or
point attributes (
Va
) to derive back-projected per-point point spectra (
Vs
) or a projected
image of point attributes (Pa):
Mf·Ps=Vs(3)
MT
f·Va=Pa(4)
The normalisation method will determine how many to many relationships between
points and pixels will be handled. If
Mf
is calculated by normalising
M
such that its
columns sum to one, then the spectra in all of the pixels that a point is visible from will be
averaged (weighted by their proximity to the sensor). Alternatively, the number of points
contained in each pixel (which is proportional to the on-ground pixel size, or footprint) can
be calculated by summing the columns of
M
and used to derive a filtered
Mf
that maps the
spectra from the closest (smallest footprint) pixel directly to each point (with no averaging),
so that maximum spatial resolution is retained.
Code for constructing
M
and performing these filtering operations to rapidly transfer
data between a SfM-MVS point clouds and pushbroom hyperspectral imagery has been
implemented in python as part of the open-source hylite package and is available at https:
//github.com/samthiele/hylite (accessed on 16 December 2021).
3.4. Boresight Optimisation
Despite the accuracy of the Applanix IMU under optimal conditions, a combination
of GPS error, IMU orientation uncertainty, and the boresight between the IMU and each
hyperspectral sensor, resulted in a relatively large offset (1–2 m/10–20 pixels) between
projected hyperspectral pixels and their actual location on the point cloud (identified using
RGB colours from the SfM-MVS). These errors were not ameliorated by applying a stan-
dard boresight calibration using values estimated using SfM during the earlier calibration
flight (cf. [
15
]), probably because they result from rotations of the local coordinate system
induced by poor GPS signal during the survey and IMU calibration figure (cf. Section 3.3),
rather than the physical sensor boresight. This unacceptably large error was reduced by
developing a novel boresight optimisation algorithm that finds the boresight values which
result in the highest correlation between projected RGB colours and equivalent colours in
the SfM-MVS point cloud.
First, we assume that the combined IMU alignment errors can be approximated by
a single rigid body rotation between the IMU and the SfM-MVS point cloud coordinate
systems. Position errors most likely also result in a lever-arm offset, but this constant offset
will have less effect on projected pixel location than orientation error (which increases
proportionally to the sensor–target distance).
An optimal boresight adjustment was then found to correct this rotation (IMU mis-
alignment) for each flight line, by repeatedly projecting the red, green and blue bands of
each of the VNIR images onto a sub-sampled point cloud and maximising the correlation
with point colours derived from the SfM-MVS model. This optimisation was achieved using
the iterative least-squares algorithm implemented in SciPy (scipy.optimise.minimize; [
34
]).
Following this procedure, we were able to co-register the two hyperspectral swaths
to within ~1 pixel of each other, and ~2 pixels of the SfM-MVS point cloud (based on a
visual assessment of the projected point colours; Figure 6). A quantitative assessment of the
reprojection error was not possible, as surveyed ground control points could not be placed
on the cliffs. The projection-mapping matrix calculated using the boresight values was then
Remote Sens. 2022,14, 5 10 of 21
saved for each flight for subsequent use during radiometric correction (Section 3.5) and
fusion of the swaths (Section 3.6).
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Section 3.3), rather than the physical sensor boresight. This unacceptably large error was
reduced by developing a novel boresight optimisation algorithm that finds the boresight
values which result in the highest correlation between projected RGB colours and equiv-
alent colours in the SfM-MVS point cloud.
First, we assume that the combined IMU alignment errors can be approximated by a
single rigid body rotation between the IMU and the SfM-MVS point cloud coordinate sys-
tems. Position errors most likely also result in a lever-arm offset, but this constant offset
will have less effect on projected pixel location than orientation error (which increases
proportionally to the sensortarget distance).
An optimal boresight adjustment was then found to correct this rotation (IMU misa-
lignment) for each flight line, by repeatedly projecting the red, green and blue bands of
each of the VNIR images onto a sub-sampled point cloud and maximising the correlation
with point colours derived from the SfM-MVS model. This optimisation was achieved
using the iterative least-squares algorithm implemented in SciPy (scipy.optimise.mini-
mize; [34]).
Following this procedure, we were able to co-register the two hyperspectral swaths
to within ~1 pixel of each other, and ~2 pixels of the SfM-MVS point cloud (based on a
visual assessment of the projected point colours; Figure 6). A quantitative assessment of
the reprojection error was not possible, as surveyed ground control points could not be
placed on the cliffs. The projection-mapping matrix calculated using the boresight values
was then saved for each flight for subsequent use during radiometric correction (Section
3.5) and fusion of the swaths (Section 3.6).
Figure 6. Geometrically corrected true-colour hyperspectral datasets before (a) and after (b)
boresight optimisation for each flight. Outlines of the 1 m2 grey and black calibration panels as
mapped on the SfM-MVS point cloud (c) are overlain for reference. While there is significant room
for improvement, coregistration error dropped from ~2.5 m before optimisation to ~0.2 m after-
wards. Interestingly the optimal boresight values (roll, pitch, yaw in degrees) are substantially dif-
ferent between the two flights (d), even though these were conducted within an hour of each other.
This suggests that they represent virtual’ boresight values that correct for a misalignment between
the global coordinate system and the IMU coordinate system caused by poor GPS signal during
IMU calibration, rather than the real sensor boresight.
Figure 6.
Geometrically corrected true-colour hyperspectral datasets before (
a
) and after (
b
) boresight
optimisation for each flight. Outlines of the 1 m
2
grey and black calibration panels as mapped on the
SfM-MVS point cloud (
c
) are overlain for reference. While there is significant room for improvement,
coregistration error dropped from ~2.5 m before optimisation to ~0.2 m afterwards. Interestingly
the optimal boresight values (roll, pitch, yaw in degrees) are substantially different between the two
flights (
d
), even though these were conducted within an hour of each other. This suggests that they
represent ‘virtual’ boresight values that correct for a misalignment between the global coordinate
system and the IMU coordinate system caused by poor GPS signal during IMU calibration, rather
than the real sensor boresight.
3.5. Radiometric Corrections
The hyperspectral data for each flight were preprocessed using dedicated software
provided by HySpex (HyspexRad) to convert raw digital numbers into radiance and cor-
rect for smile and keystone distortions. Calibration spectra from two Zenith calibration
panels
(R = 5%
and 50%) placed at the base of the cliff in each scene (cf. Figure 5) were
then extracted by back-projection onto a subset point cloud containing only the panels.
Additionally
, indirect illumination was quantified by averaging pixels of a shaded Spec-
tralon R90 panel (R = 90%) that was placed in front of the camera following each survey
(after the UAV had landed).
Following the method of [
35
], these spectra were used to characterise the downwelling
sunlight, skylight and path radiance in each scene. The measured radiance (r) of each
panel with known reflectance spectra Rcan be described by the approximately Lambertian
reflection of a mixture of diffuse skylight and sunlight [35]:
rpanel =Rpanel ×apanel ·S+cosθpanel ·I+P(5)
where Iand Sare the incident sunlight (direct illumination) and skylight (indirect illumi-
nation) spectra, respectively; ais the fraction of visible sky from the panel,
θ
is the angle
Remote Sens. 2022,14, 5 11 of 21
between the panel normal and downwelling sunlight and Pis the path radiance. We find
the three unknowns in this equation (S,Iand P) by solving the following linear system:
aARAcos(θA)RA1
aBRBcos(θB)RB1
aCRC0 0
S
I
P
=rArBrC(6)
where A,Band Ccorrespond to each of the three panels used (5%, 55% and the (shaded)
90% panel, respectively). The panels’ sky view factor (a) and incidence angle (
θ
) were
calculated using the SfM-MVS model, as described in the following paragraph. Note that
the inclusion of a fully shaded panel (such that for this panel Iis known to be 0) ensures
proper characterisation of approximately omnidirectional skylight (cf. [35]).
These three illumination spectra (Figure 7c,d) allowed us to model and correct for the
influence of atmospheric and topographic effects. As described in detail by [
35
], incident
and viewing angles for each pixel were calculated by projecting the 3-D point cloud onto
each hyperspectral image (cf. Equation (4)) and computing an Oren-Nayar [
36
] bidirectional
reflectance distribution function (BRDF) that simulates the fraction αof sunlight reflected
towards the camera from each pixel. Sky view factors (a) were calculated for each point
using the PCV plugin [
37
] in CloudCompare, and projected from the point cloud onto each
pixel in the hyperspectral image. This allowed per-pixel reflectance (R) to be estimated
from measured radiance (r) using the following equation [35]:
R=rP
α·I+a·S(7)
where the numerator represents the at-target reflected radiance, and the denominator the
combined direct and indirect illumination across the scene.
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Figure 7. Measured calibration panel spectra (a,b) for each flight and derived illumination source
spectra (c,d). These spectra were used to model and correct for illumination effects (atmospheric
and topographic) separately in each flight. Note that the path radiance is essentially negligible, as
expected due to the short sensortarget distance, but that skylight accounts for 25% to 50% of the
total illumination during both flights.
3.6. Fusion and Minimum Wavelength Mapping
Due to the differing survey distances (cf. Figure 5), the swath from Flight A covered
a wider extent than Flight B, but at lower resolution. Thus, after applying the radiometric
and geometric corrections described in Sections 3.4 and 3.5, they were fused to derive a
combined dataset with optimal spatial coverage and resolution. The spectra from each
swath were projected onto two separate hyperclouds using a filtered mapping matrix Mf
that back-projects the smallest-footprint pixel onto each point in the SfM-MVS cloud (cf.
Section 3.3). Then, the per-point spectra from each swath were averaged using a weighting
factor proportional to each pixel’s ground-sampling distance, such that the two surveys
were averaged to combine their extent and (slightly) reduce spectral noise, but without
significantly compromising spatial resolution.
Finally, a hull-correction from 21002500 nm was applied to this merged hypercloud
to emphasise absorption features related to clay and mica minerals in the shales (~2200
nm), calcite (~2345 nm) and dolomite (~2325 nm) and create the false-colour composite
presented in Section 4. Minimum wavelength mapping of absorption feature position and
depth (cf. [3841]) was then performed using the multi-gaussian absorption feature-fitting
algorithm implemented in hylite. This was used to simultaneously fit the three main min-
eralogical absorption features (AlOH, FeOH and CO3) expected in the range 21502360
nm, and thus allow discrimination between the different minerals at Passo Giau.
3.7. Collection of Validation Spectra
Point spectra measurements were conducted using a portable spectrometer to pro-
vide ground truth data for validating the remotely sensed spectra. Ten sampling sites
were selected based on an on-site evaluation of the geology and stratigraphy, access lim-
itations and to ensure spatial coverage of the extent of the outcrop. Ten measurements
were collected at each of these sites, covering the range of dominant facies or micro-facies
(as interpreted by the geologist), using a FieldSpec spectrometer equipped with a contact
Figure 7.
Measured calibration panel spectra (
a
,
b
) for each flight and derived illumination source
spectra (
c
,
d
). These spectra were used to model and correct for illumination effects (atmospheric
and topographic) separately in each flight. Note that the path radiance is essentially negligible, as
expected due to the short sensor–target distance, but that skylight accounts for 25% to 50% of the
total illumination during both flights.
Remote Sens. 2022,14, 5 12 of 21
3.6. Fusion and Minimum Wavelength Mapping
Due to the differing survey distances (cf. Figure 5), the swath from Flight A covered a
wider extent than Flight B, but at lower resolution. Thus, after applying the radiometric
and geometric corrections described in Sections 3.4 and 3.5, they were fused to derive a
combined dataset with optimal spatial coverage and resolution. The spectra from each
swath were projected onto two separate hyperclouds using a filtered mapping matrix
Mf
that back-projects the smallest-footprint pixel onto each point in the SfM-MVS cloud
(cf. Section 3.3). Then, the per-point spectra from each swath were averaged using a
weighting factor proportional to each pixel’s ground-sampling distance, such that the two
surveys were averaged to combine their extent and (slightly) reduce spectral noise, but
without significantly compromising spatial resolution.
Finally, a hull-correction from 2100–2500 nm was applied to this merged hypercloud to
emphasise absorption features related to clay and mica minerals in the shales
(~2200 nm)
,
calcite (~2345 nm) and dolomite (~2325 nm) and create the false-colour composite pre-
sented in Section 4. Minimum wavelength mapping of absorption feature position and
depth (cf. [
38
41
]) was then performed using the multi-gaussian absorption feature-fitting
algorithm implemented in hylite. This was used to simultaneously fit the three main miner-
alogical absorption features (AlOH, FeOH and CO
3
) expected in the range 2150–2360 nm,
and thus allow discrimination between the different minerals at Passo Giau.
3.7. Collection of Validation Spectra
Point spectra measurements were conducted using a portable spectrometer to provide
ground truth data for validating the remotely sensed spectra. Ten sampling sites were
selected based on an on-site evaluation of the geology and stratigraphy, access limitations
and to ensure spatial coverage of the extent of the outcrop. Ten measurements were
collected at each of these sites, covering the range of dominant facies or micro-facies (as
interpreted by the geologist), using a FieldSpec spectrometer equipped with a contact probe
that allowed measurements of the reflectance spectra from areas of ~1 cm
2
. These sampling
sites were then marked using spray-on chalk such that they could be easily identified in the
photogrammetric model of the area. Of these sites, 5 overlapped with the extent of the UAV
survey and could be identified in the photogrammetric point cloud, allowing equivalent
spectra to be sampled from within
±
1 m of the spray-on chalk markers and within the
relevant microfacies (as identified based on field observations and interpretation of the
photogrammetric model).
4. Results
The geometric correction and boresight optimisation routine described in
Sections 3.3 and 3.4
allowed successful correction and georeferencing of the raw hyper-
spectral swaths (Figure 8), despite collection under adverse conditions (significant wind
and poor GPS signal) to produce a geometrically and radiometrically corrected hyper-
cloud. Jupyter notebooks showing the application of these methods are available in the
Supplementary Material.
The resulting point-spectra closely match those measured in situ using the FieldSpec
(Figure 9). The Mjolnir and FieldSpec spectra were compared by calculating the mean
average error (MAE; sensitive to spectra shape and magnitude) and spectral angle (SA;
sensitive to spectra shape only) between the Mjolnir and closest FieldSpec spectra at each
of the reference sites (Figure 10). These suggest that the majority of the remotely sensed
Mjolnir spectra are accurate to 2–4% (MAE) and 3.5–5.5
(SA). Considering the significant
challenges associated with remotely sensing accurate reflectance spectra, we consider these
errors to be acceptable.
Figures 9and 10 also show that the Mjolnir and FieldSpec spectra consistently identify
variations in the position and depth of the AlOH and CO
3
absorption features (at ~2200 and
2320–2345 nm, respectively). These subtle differences were be used to map the dominant
mineralogy across the outcrop, and so discriminate the fine-grained dolostone of the
Remote Sens. 2022,14, 5 13 of 21
Cassian Dolomite Formation (DCS; Figure 9) from the interbedded micritic limestone
and marls of the San Cassiano Formation (SCS; Figure 9), and their dolomite equivalents
(
SCS-D
; Figure 9). SCS and SCS-D both show an AlOH feature indicating the presence of
clay minerals, while the position of the CO
3
feature distinguishes calcitic SCS (2345 nm)
from dolomitic SCS-D and DCS (2325 nm).
The false colour visualisation (Figure 11a) created using the absorption depth at 2200,
2320 and 2340 nm highlights these differences, clearly distinguishing dolostone (green)
from limestone (blue), as well as shale-rich argillites (purple to red). Similarly, the minimum
wavelength map shows variations in the position of the CaCO
3
absorption feature from
2320 nm (dolomite) to 2340 nm (calcite), providing a visualisation of the calcite to dolomite
ratio (Figure 11b). The calcitic limestone (2340–2345 nm) can be clearly distinguished
from overlying dolomites (2315–2325 nm), and correlates with argillites (San Cassiano Fm)
identified based on their distinct texture (fine laminations) in the photogrammetric model
and abundant clay or mica minerals identified by mapping the depth of the 2200 nm AlOH
absorption feature (Figure 11c).
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Figure 8. True colour uncorrected (a,b) and geometrically corrected (c,d) visualisations of each hy-
perspectral swaths. Although spatial accuracy is limited to ~2 pixels (Figure 5), the extensive geo-
metric distortions in the original images have been almost entirely corrected. The 3-m spaced hori-
zontal contours are overlain in white for reference. Areas in the original images occluded by the leg
of the drone have been masked to prevent projection onto the point-cloud. Note that the duplication
of pixels in the raw hyperspectral swaths means that this did not leave gaps in the geometrically
corrected results.
Figure 8.
True colour uncorrected (
a
,
b
) and geometrically corrected (
c
,
d
) visualisations of each
hyperspectral swaths. Although spatial accuracy is limited to ~2 pixels (Figure 5), the extensive
geometric distortions in the original images have been almost entirely corrected. The 3-m spaced
horizontal contours are overlain in white for reference. Areas in the original images occluded by
the leg of the drone have been masked to prevent projection onto the point-cloud. Note that the
duplication of pixels in the raw hyperspectral swaths means that this did not leave gaps in the
geometrically corrected results.
Remote Sens. 2022,14, 5 14 of 21
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Figure 9. Comparison of corrected HySpex Mjolnir spectra and ground truth FieldSpec measure-
ments (plotted in grey) from the Cassian Dolomite Formation (DCS; green), dolomitised San Cassi-
ano Formation (SCS-D; red), and calcitic San Cassiano Formation (SCS; blue). The location of these
validation sites is plotted in Figures 4 and 11. Note that these measurements were used to check the
spectral accuracy of the UAV data, not to validate a classification or lithological map of the outcrop;
so lithological unit labels provided above are for reference only. Blue vertical lines show the ex-
pected positions of the AlOH, dolomitic CO3 and calcitic CO3 absorption features mapped in Figures
11 and 12.
Figure 10. Quantitative error metrics showing the mean absolute error (a) and spectral angle (b)
between the HySpex Mjolnir and corresponding FieldSpec measurements. Vertical black line shows
the median error and vertical dotted lines the 25th and 75th percentiles. A comparison of the car-
bonate absorption position measured using the different instruments is shown in (c), and shows
that the ~20 nm shift in absorption feature position between calcite and dolomite can be resolved in
both datasets, allowing calcitic lithotypes to be distinguished from dolomitic ones (point colour cor-
responds to the different stratigraphic units; cf. Figure 9). Vertical blue lines show the expected po-
sition for dolomite (2325 nm) and calcite (2345 nm).
Figure 9.
Comparison of corrected HySpex Mjolnir spectra and ground truth FieldSpec measurements
(plotted in grey) from the Cassian Dolomite Formation (DCS; green), dolomitised San Cassiano
Formation (
SCS-D
; red), and calcitic San Cassiano Formation (SCS; blue). The location of these
validation sites is plotted in Figures 4and 11. Note that these measurements were used to check
the spectral accuracy of the UAV data, not to validate a classification or lithological map of the
outcrop; so lithological unit labels provided above are for reference only. Blue vertical lines show
the expected positions of the AlOH, dolomitic CO
3
and calcitic CO
3
absorption features mapped in
Figures 11 and 12.
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Figure 9. Comparison of corrected HySpex Mjolnir spectra and ground truth FieldSpec measure-
ments (plotted in grey) from the Cassian Dolomite Formation (DCS; green), dolomitised San Cassi-
ano Formation (SCS-D; red), and calcitic San Cassiano Formation (SCS; blue). The location of these
validation sites is plotted in Figures 4 and 11. Note that these measurements were used to check the
spectral accuracy of the UAV data, not to validate a classification or lithological map of the outcrop;
so lithological unit labels provided above are for reference only. Blue vertical lines show the ex-
pected positions of the AlOH, dolomitic CO3 and calcitic CO3 absorption features mapped in Figures
11 and 12.
Figure 10. Quantitative error metrics showing the mean absolute error (a) and spectral angle (b)
between the HySpex Mjolnir and corresponding FieldSpec measurements. Vertical black line shows
the median error and vertical dotted lines the 25th and 75th percentiles. A comparison of the car-
bonate absorption position measured using the different instruments is shown in (c), and shows
that the ~20 nm shift in absorption feature position between calcite and dolomite can be resolved in
both datasets, allowing calcitic lithotypes to be distinguished from dolomitic ones (point colour cor-
responds to the different stratigraphic units; cf. Figure 9). Vertical blue lines show the expected po-
sition for dolomite (2325 nm) and calcite (2345 nm).
Figure 10.
Quantitative error metrics showing the mean absolute error (
a
) and spectral angle (
b
)
between the HySpex Mjolnir and corresponding FieldSpec measurements. Vertical black line shows
the median error and vertical dotted lines the 25th and 75th percentiles. A comparison of the
carbonate absorption position measured using the different instruments is shown in (
c
), and shows
that the
~20 nm
shift in absorption feature position between calcite and dolomite can be resolved
in both datasets, allowing calcitic lithotypes to be distinguished from dolomitic ones (point colour
corresponds to the different stratigraphic units; cf. Figure 9). Vertical blue lines show the expected
position for dolomite (2325 nm) and calcite (2345 nm).
Remote Sens. 2022,14, 5 15 of 21
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Figure 11. False-colour visualisation of the merged and radiometrically corrected hypercloud (a)
and minimum wavelength maps showing the distribution of calcitic (blue) and dolomitic (green)
carbonates (b) and clay/mica minerals (purple; c). The location of the validation spectra shown in
Figure 9 are overlain on (c) in yellow for reference.
Figure 11.
False-colour visualisation of the merged and radiometrically corrected hypercloud (
a
)
and minimum wavelength maps showing the distribution of calcitic (blue) and dolomitic (green)
carbonates (
b
) and clay/mica minerals (purple;
c
). The location of the validation spectra shown in
Figure 9are overlain on (c) in yellow for reference.
Interestingly, some of this limestone has been dolomitised to a distance of ~2–5 m
along its top contact (Figure 12), an observation that would likely have been missed if
the unit was mapped using only traditional methods. A geoscience interpretation of
these results is beyond the scope of this current work, but we present some preliminary
interpretations in the following section (Section 5). The false colour visualisations and
minimum wavelength map results can downloaded from https://tinyurl.com/nuvolau
(accessed on 16 December 2021).
Remote Sens. 2022,14, 5 16 of 21
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
Figure 12. Close-up of the dolostone-limestone contact (dashed white line) showing the stratigraph-
ically crosscutting transition zone containing dolomitised limestone in the true-colour SfM-MVS
model (a), and carbonate (b) and AlOH (c) absorption feature minimum wavelength maps. Note
that the image has been rotated to make bedding sub-horizontal; thin white lines show 3-m spaced
horizontal contour lines. Example reflectance spectra (d) sampled from the coloured circles in (ac)
are plotted for reference. These spectra (dotted lines) were smoothed (solid lines) using a Savitzky
Golay filter with a window size of 7 and polynomial order of 2, to reduce noise and highlight the
Figure 12.
Close-up of the dolostone-limestone contact (dashed white line) showing the stratigraph-
ically crosscutting transition zone containing dolomitised limestone in the true-colour SfM-MVS
model (
a
), and carbonate (
b
) and AlOH (
c
) absorption feature minimum wavelength maps. Note that
the image has been rotated to make bedding sub-horizontal; thin white lines show 3-m spaced hori-
zontal contour lines. Example reflectance spectra (
d
) sampled from the coloured circles in (
a
c
) are
plotted for reference. These spectra (dotted lines) were smoothed (solid lines) using a Savitzky–Golay
filter with a window size of 7 and polynomial order of 2, to reduce noise and highlight the clear shift
in absorption position from 2345 nm (calcite) to 2325 nm (dolomite). Refer to Figure 11 for the colour
scales used in (b,c).
Remote Sens. 2022,14, 5 17 of 21
5. Discussion
Our results highlight the significant potential of UAV-compatible SWIR hyperspectral
cameras for creating detailed, accurate and objective geological maps in inaccessible ter-
rain. As shown in Figures 911, the remotely mapped SWIR spectra are accurate and can
distinguish the lithologies and mineral alteration (dolomitisation) at a scale and resolution
that would be difficult with ground-based sensors or field mapping. While care must
be taken to fully understand potential weathering effects [
42
,
43
], these data significantly
supplement the high resolution geometric and textural information captured using con-
ventional digital outcrop mapping techniques, and specifically contribute to provide a
detailed map of mineralogy across the slope to basin transition at Gusela de Nuvolau.
Potential applications of this approach are far broader than mapping carbonate mineralogy
in reservoir analogues, and include, e.g., mapping clays, mica and chlorite associated
with hydrothermal alteration around mineral deposits or volcanic edifices for minerals
exploration or geohazard assessment.
A few geological studies have employed UAV hyperspectral SWIR cameras (e.g., [
14
]),
but we suggest that significant potential for the development and innovative application of
both methods remains. We hope that our novel processing approach helps to realise this
potential, by (1) facilitating accurate corrections in areas of steep terrain (e.g., cliffs or open
pit mines) and (2) retaining true 3-D information by avoiding the need to project data into
a 2-D orthomosaic. The former is significant as areas of exceptional geological exposure are
rarely flat, while the latter avoids gridding and reprojection artefacts associated with DEM-
based workflows (cf. [
23
]), as real pixel footprints are retained in
M
, and also results in point
cloud data suited for digital outcrop analysis (e.g., LIME, CloudCompare, VRGS; [
32
,
44
,
45
])
or 3-D modelling.
To demonstrate this, we used VRGS [
45
] to perform a preliminary interpretation of the
hyperspectral mineral mapping results and associated SfM-MVS point cloud (Figure 13).
This synthesises the mineral-mapping results with high resolution textural information
captured by the SfM-MVS model and regional knowledge to help better constrain the
stratigraphic nature of the slope to basin relationship (e.g., slope breccia downlapping on
basin wedge or interfingering), as well as the spatial reach of the dolomitisation fluids from
coarse slope sediments to basinal, thin bedded, fine-grained limestone intervals alternating
with shales and marls (Figure 13). Part of the dolomitising fluids are interpreted here to
have circulated throughout faults and connected fractures (Figure 13), laterally invading
porous and permeable rock volumes. Adjacent non-porous and impermeable basinal
facies acted as a barrier to dolomitising fluids, preserving the associated rock type from
dolomitisation. Dolomitisation processes are complex and multiphase in the Dolomites
region, and are not in the scope of this study. Instead, this contribution proves the reliability
of the developed technology and workflow, opening opportunities to acquire similar,
unbiased and accurate, geometrical and mineralogical information along vertical outcrops,
with major implications for the evaluation of subsurface analogues for geothermal, CO
2
sequestration and conventional hydrocarbon exploration and production.
However, as with any remote sensing technology employing UAVs, data acquisition is
particularly sensitive to field conditions (e.g., weather and GPS signals). The high winds and
poor GPS reception we experienced during data acquisition emphasise the need for robust
processing solutions that can accurately correct data from an erratically moving platform.
Our novel geometric correction and boresight optimisation routines are a step in this
direction, and allowed us to extract accurate data in spite of the adverse conditions, but there
remains significant room for improvement. Tools for planning automated UAV surveys that
use non-nadir camera angles (due to topographic complexity) are also lacking, especially if
high-resolution topographic information is unavailable. Small-scale commercial solutions
are beginning to emerge from the engineering inspection industry (Skydio; www.skydio.
com, accessed on 16 December 2021), but these currently cannot be employed for high-
payload drones.
Remote Sens. 2022,14, 5 18 of 21
Finally, we suggest that good quality analysis and control (QAQC) procedures that
can be applied while in the field to ensure sufficient data quality are needed. Some progress
has been made on applying geometric corrections in real-time [
46
], but more needs to be
done to make these solutions available, robust and field-ready. Appropriate field QAQC is
essential for many geological studies, as field locations are often inaccessible and visiting
multiple times is impractical or prohibitively expensive. Real-time processing capabilities
also allow adaptation of survey plans and acquisition parameters when areas of interest are
identified, and potentially validated or ground-truthed while onsite. By implementing our
method in the open-source hylite toolbox we hope to facilitate advances in these directions
to realise the potential of rapidly developing UAV hyperspectral imaging technologies.
Remote Sens. 2022, 13, x FOR PEER REVIEW 7 of 22
(Skydio; www.skydio.com, accessed on 18 December, 2021), but these currently cannot be
employed for high-payload drones.
Finally, we suggest that good quality analysis and control (QAQC) procedures that
can be applied while in the field to ensure sufficient data quality are needed. Some pro-
gress has been made on applying geometric corrections in real-time [46], but more needs
to be done to make these solutions available, robust and field-ready. Appropriate field
QAQC is essential for many geological studies, as field locations are often inaccessible and
visiting multiple times is impractical or prohibitively expensive. Real-time processing ca-
pabilities also allow adaptation of survey plans and acquisition parameters when areas of
interest are identified, and potentially validated or ground-truthed while onsite. By im-
plementing our method in the open-source hylite toolbox we hope to facilitate advances
in these directions to realise the potential of rapidly developing UAV hyperspectral im-
aging technologies.
Figure 13. Digital outcrop interpretation of the Gusela del Nuvolau cliff based on the SfM-MVS
model and hyperspectral survey data. The minimum wavelength map (cf. Figures 11 and 12) has
been overlain in (a) as a reference. This clearly shows the gradational and stratigraphically
Figure 13.
Digital outcrop interpretation of the Gusela del Nuvolau cliff based on the SfM-MVS model
and hyperspectral survey data. The minimum wavelength map (cf. Figures 11 and 12) has been
overlain in (
a
) as a reference. This clearly shows the gradational and stratigraphically crosscutting
dolomitisation of the otherwise calcitic San Cassiano Formation (
b
), and that the Cassian Dolomite
Formation has been completely dolomitised.
Remote Sens. 2022,14, 5 19 of 21
6. Conclusions
We provide an open-source toolbox for the accurate correction of non-nadir pushbroom
hyperspectral data. Our robust algorithm allows the processing of VNIR and SWIR data
acquired obliquely and targeting outcrops with complex geometries, such as cliffs or
mines. In a first step, this algorithm was developed for the Hyspex NEO Mjolnir, but
we expect users to be able to use our tools to process data from any pushbroom sensor,
provided a detailed photogrammetric outcrop model is also available. The procedure is
included in the open-source hylite package (https://github.com/samthiele/hylite, accessed
on
16 December 2021
), and the authors welcome further contributions. We demonstrated
the potential of this toolbox using data collected in conditions representative of operational
acquisitions. Despite non-optimal weather and GPS reception, distortions and related
artefacts could be successfully corrected to derive a coherent and spectrally accurate dataset.
This contribution opens the door to new UAV hyperspectral acquisition types and supports
the development of UAV-based imaging in Earth sciences.
Supplementary Materials:
The following are available online at https://www.mdpi.com/article/
10.3390/rs14010005/s1, S1: Agisoft Metashape processing report, N1: IMU Visualisation notebook,
N2a: Flight A geometric corrections notebook, N2b: Flight B geometric corrections notebook, N3a:
Flight A illumination corrections notebook, N3b: Fight B illumination corrections notebook, N4: Data
fusion notebook, N5: Minimum wavelength mapping notebook
Author Contributions:
Conceptualisation by R.G., J.K., S.T.T., E.D. and S.L.; methodology, S.T.T., Z.B.
and S.L.; software, S.T.T., Z.B. and Y.M.; validation, S.T.T., N.M., A.B., E.D. and J.K.; writing—original
draft preparation, S.T.T., S.L., Z.B., A.B. and N.M.; writing—review and editing, S.T.T., S.L., A.B., R.G.,
E.D. and J.K.; project administration, R.G. and J.K.; funding acquisition, R.G. and J.K. All authors
have read and agreed to the published version of the manuscript.
Funding:
The authors would like to acknowledge funding from TotalEnergies, The European Social
Fund (ESF) and the State of Saxony.
Data Availability Statement:
Data presented in this study are available at https://doi.org/10.6084/
m9.figshare.17126321.v1, accessed on 16 December 2021.
Conflicts of Interest: The authors declare no conflict of interest.
References
1.
Bemis, S.; Micklethwaite, S.; Turner, D.; James, M.R.; Akciz, S.; Thiele, S.T.; Bangash, H.A. Ground-based and UAV-Based
photogrammetry: A multi-scale, high-resolution mapping tool for structural geology and paleoseismology. J. Struct. Geol.
2014
,
69, 163–178. [CrossRef]
2.
Dering, G.M.; Micklethwaite, S.; Thiele, S.T.; Vollgger, S.A.; Cruden, A.R. Review of drones, photogrammetry and emerging
sensor technology for the study of dykes: Best practises and future potential. J. Volcanol. Geotherm. Res.
2019
,373, 148–166.
[CrossRef]
3.
Janiszewski, M.; Uotinen, L.; Merkel, J.; Leveinen, J.; Rinne, M. Virtual Reality Learning Environments for Rock Engineering,
Geology and Mining Education. In Proceedings of the 54th U.S. Rock Mechanics/Geomechanics Symposium (ARMA-2020-1101),
Golden, CO, USA, 28 June–1 July 2020.
4.
Nesbit, P.; Durkin, P.R.; Hugenholtz, C.H.; Hubbard, S.; Kucharczyk, M. 3-D stratigraphic mapping using a digital outcrop model
derived from UAV images and structure-from-motion photogrammetry. Geosphere 2018,14, 2469–2486. [CrossRef]
5.
Marques, A.; Horota, R.K.; de Souza, E.M.; Kupssinskü, L.; Rossa, P.; Aires, A.S.; Bachi, L.; Veronez, M.R.; Gonzaga, L.;
Cazarin, C.L.
Virtual and digital outcrops in the petroleum industry: A systematic review. Earth-Sci. Rev.
2020
,208, 103260.
[CrossRef]
6.
Laukamp, C.; Rodger, A.; LeGras, M.; Lampinen, H.; Lau, I.; Pejcic, B.; Stromberg, J.; Francis, N.; Ramanaidou, E. Mineral
Physicochemistry Underlying Feature-Based Extraction of Mineral Abundance and Composition from Shortwave, Mid and
Thermal Infrared Reflectance Spectra. Minerals 2021,11, 347. [CrossRef]
7.
Kurz, T.H.; Buckley, S.J.; Howell, J.A. Close-range hyperspectral imaging for geological field studies: Workflow and methods.
Int. J. Remote Sens. 2013,34, 1798–1822. [CrossRef]
8.
Kurz, T.H.; Buckley, S.J.; Howell, J.A.; Schneider, D. Integration of panoramic hyperspectral imaging with terrestrial lidar data.
Photogramm. Rec. 2011,26, 212–228. [CrossRef]
Remote Sens. 2022,14, 5 20 of 21
9.
Lorenz, S.; Salehi, S.; Kirsch, M.; Zimmermann, R.; Unger, G.; Vest Sørensen, E.; Gloaguen, R. Radiometric Correction and 3D
Integration of Long-Range Ground-Based Hyperspectral Imagery for Mineral Exploration of Vertical Outcrops. Remote Sens.
2018
,
10, 176. [CrossRef]
10.
Thiele, S.T.; Lorenz, S.; Kirsch, M.; Acosta, I.C.C.; Tusa, L.; Herrmann, E.; Möckel, R.; Gloaguen, R. Multi-scale, multi-sensor data
integration for automated 3-D geological mapping. Ore Geol. Rev. 2021,136, 104252. [CrossRef]
11.
Aasen, H.; Honkavaara, E.; Lucieer, A.; Zarco-Tejada, P.J. Quantitative Remote Sensing at Ultra-High Resolution with UAV
Spectroscopy: A Review of Sensor Technology, Measurement Procedures, and Data Correction Workflows. Remote Sens.
2018
,
10, 1091. [CrossRef]
12.
Kim, J.; Chi, J.; Masjedi, A.; Flatt, J.E.; Crawford, M.M.; Habib, A.F.; Lee, J.; Kim, H. High-resolution hyperspectral imagery from
pushbroom scanners on unmanned aerial systems. Geosci. Data J. 2021. [CrossRef]
13.
Goldstein, N.; Wiggins, R.; Woodman, P.; Saleh, M.; Nakanishi, K.; Fox, M.E.; Tannian, B.E.; Ziph-Schatzberg, L.; Soletski, P.
Compact visible to extended-SWIR hyperspectral sensor for unmanned aircraft systems (UAS). In Proceedings of the Algorithms
and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XXIV, Bellingham, WA, USA, 17–19 April 2018;
Volume 10644, p. 106441G.
14.
Barton, I.F.; Gabriel, M.J.; Lyons-Baral, J.; Barton, M.D.; Duplessis, L.; Roberts, C. Extending geometallurgy to the mine scale with
hyperspectral imaging: A pilot study using drone- and ground-based scanning. Min. Met. Explor. 2021,38, 799–818. [CrossRef]
15.
Arroyo-Mora, J.; Kalacska, M.; Inamdar, D.; Soffer, R.; Lucanus, O.; Gorman, J.; Naprstek, T.; Schaaf, E.; Ifimov, G.; Elmer, K.; et al.
Implementation of a UAV–Hyperspectral Pushbroom Imager for Ecological Monitoring. Drones 2019,3, 12. [CrossRef]
16.
Booysen, R.; Jackisch, R.; Lorenz, S.; Zimmermann, R.; Kirsch, M.; Nex, P.A.M.; Gloaguen, R. Detection of REEs with lightweight
UAV-based hyperspectral imaging. Sci. Rep. 2020,10, 17450. [CrossRef] [PubMed]
17.
Kirsch, M.; Lorenz, S.; Zimmermann, R.; Tusa, L.; Möckel, R.; Hödl, P.; Booysen, R.; Khodadadzadeh, M.; Gloaguen, R. Integration
of Terrestrial and Drone-Borne Hyperspectral and Photogrammetric Sensing Methods for Exploration Mapping and Mining
Monitoring. Remote Sens. 2018,10, 1366. [CrossRef]
18.
JuanManuel, J.R.; Padua, L.; Hruska, J.; Feito, F.R.; Sousa, J.J. An Efficient Method for Generating UAV-Based Hyperspectral
Mosaics Using Push-Broom Sensors. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021,14, 6515–6531. [CrossRef]
19.
Turner, D.; Lucieer, A.; McCabe, M.; Parkes, S.; Clarke, I. Pushbroom Hyperspectral Imaging from an Unmanned Aircraft System
(UAS)—Geometric Processingworkflow and Accu-racy Assessment. ISPRS Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci.
2017,XLII-2/W6, 379–384. [CrossRef]
20.
Angel, Y.; Turner, D.; Parkes, S.; Malbeteau, Y.; Lucieer, A.; McCabe, M.F. Automated Georectification and Mosaicking of
UAV-Based Hyperspectral Imagery from Push-Broom Sensors. Remote Sens. 2020,12, 34. [CrossRef]
21.
Schlaepfer, D.; Schaepman, M.E.; Itten, K.I. PARGE: Parametric geocoding based on GCP-calibrated auxiliary data. SPIE’s Int.
Symp. Opt. Sci. Eng. Instrum. 1998,3438, 334–344. [CrossRef]
22.
Ghamisi, P.; Shahi, K.R.; Duan, P.; Rasti, B.; Lorenz, S.; Booysen, R.; Thiele, S.; Contreras, I.C.; Kirsch, M.; Gloaguen, R. The
Potential of Machine Learning for a More Responsible Sourcing of Critical Raw Materials. IEEE J. Sel. Top. Appl. Earth Obs. Remote
Sens. 2021,14, 8971–8988. [CrossRef]
23.
Inamdar, D.; Kalacska, M.; Arroyo-Mora, J.P.; Leblanc, G. The Directly-Georeferenced Hyperspectral Point Cloud: Preserving the
Integrity of Hyperspectral Imaging Data. Front. Remote Sens. 2021,2, 9. [CrossRef]
24.
Buckley, S.; Kurz, T.H.; Howell, J.A.; Schneider, D. Terrestrial lidar and hyperspectral data fusion products for geological outcrop
analysis. Comput. Geosci. 2013,54, 249–258. [CrossRef]
25.
Bosellini, A. Progradation geometries of carbonate platforms: Examples from the Triassic of the Dolomites, northern Italy.
Sedimentology 1984,31, 1–24. [CrossRef]
26.
Neri, C.; Gianolla, P.; Furlanis, S.; Caputo, R.; Bosellini, A. Note Illustrative Della Carta Geologica d’Italia. In Foglio Cortina
D’ampezzo 029; ISPRA: Rome, Italy, 2007.
27.
Inama, R.; Menegoni, N.; Perotti, C. Syndepositional fractures and architecture of the lastoni di formin carbonate platform:
Insights from virtual outcrop models and field studies. Mar. Pet. Geol. 2020,121, 104606. [CrossRef]
28.
Blendinger, W.; Blendinger, E. Windward-leeward effects on Triassic carbonate bank margin facies of the Dolomites, northern
Italy. Sediment. Geol. 1989,64, 143–166. [CrossRef]
29.
Bosellini, A.; Neri, C. The Sella Platform. Dolomieu Conference on Carbonate Platforms and Dolomitization; Guidebook Excursion B;
KARO-Druck: Ortisei, Italy, 1991.
30.
Cadrobbi, L.; Nobile, G.; Lutterotti, G. Studio Di Supporto Alla Stesura Del Piano Regolatore Generale Del Comune Di Canazei, Studio
Associato Di Geologia Applicata; Canazei, Italy, 1995; Volume 1955-1.
31.
Mollema, P.N.; Antonellini, M. Development of strike-slip faults in the dolomites of the Sella Group, Northern Italy. J. Struct. Geol.
1999,21, 273–292. [CrossRef]
32. Girardeau-Montaut, D. CloudCompare. 2020. Available online: https://cloudcompare.org/ (accessed on 16 December 2021).
33.
Akenine-Moller, T.; Haines, E.; Hoffman, N. Real-Time Rendering; AK Peters/crc Press: Natick, MA, USA, 2019; ISBN 1-315-36545-6.
34.
Virtanen, P.; Gommers, R.; Oliphant, T.E.; Haberland, M.; Reddy, T.; Cournapeau, D.; Burovski, E.; Peterson, P.; Weckesser, W.;
Bright, J.; et al. SciPy 1.0: Fundamental algorithms for scientific computing in Python. Nat. Methods
2020
,17, 261–272. [CrossRef]
[PubMed]
Remote Sens. 2022,14, 5 21 of 21
35.
Thiele, S.T.; Lorenz, S.; Kirsch, M.; Gloaguen, R. A Novel and Open-Source Illumination Correction for Hyperspectral Digital
Outcrop Models. IEEE Trans. Geosci. Remote Sens. 2021,14, 1–12. [CrossRef]
36.
Oren, M.; Nayar, S.K. Seeing beyond Lambert’s law. In Proceedings of the Computer Vision—ECCV ’94; Springer: Berlin/Heidelberg,
Germany, 1994; pp. 269–280.
37.
Berthaume, M.A.; Winchester, J.; Kupczik, K. Ambient occlusion and PCV (portion de ciel visible): A new dental topographic
metric and proxy of morphological wear resistance. PLoS ONE 2019,14, e0215436. [CrossRef]
38.
Huguenin, R.L.; Jones, J.L. Intelligent information extraction from reflectance spectra: Absorption band positions. J. Geophys. Res.
Space Phys. 1986,91, 9585–9598. [CrossRef]
39.
Van der Meer, F. Analysis of spectral absorption features in hyperspectral imagery. Int. J. Appl. Earth Obs. Geoinf.
2004
,5, 55–68.
[CrossRef]
40.
Van Ruitenbeek, F.; Bakker, W.H.; van der Werff, H.; Zegers, T.E.; Oosthoek, J.H.; Omer, Z.A.; Marsh, S.; van der Meer, F.D.
Mapping the wavelength position of deepest absorption features to explore mineral diversity in hyperspectral images. Planet.
Space Sci. 2014,101, 108–117. [CrossRef]
41.
Sunshine, J.; Pieters, C.M.; Pratt, S.F. Deconvolution of mineral absorption bands: An improved approach. J. Geophys. Res. Space
Phys. 1990,95, 6955–6966. [CrossRef]
42.
Beckert, J.; Vandeginste, V.; McKean, T.J.; Alroichdi, A.; John, C.M. Ground-based hyperspectral imaging as a tool to identify
different carbonate phases in natural cliffs. Int. J. Remote Sens. 2018,39, 4088–4114. [CrossRef]
43.
Kurz, T.H.; San Miguel, G.; Dubucq, D.; Kenter, J.; Miegebielle, V.; Buckley, S.J. Quantitative Mapping of Dolomitization Using
Close-Range Hyperspectral Imaging: Kimmeridgian Carbonate Ramp (Alacón, NE Spain). Geosphere 2022, in press.
44.
Buckley, S.J.; Ringdal, K.; Naumann, N.; Dolva, B.; Kurz, T.H.; Howell, J.A.; Dewez, T.J. LIME: Software for 3-D visualization,
interpretation, and communication of virtual geoscience models. Geosphere 2019,15, 222–235. [CrossRef]
45.
Hodgetts, D.; Gawthorpe, R.L.; Wilson, P.; Rarity, F. Integrating Digital and Traditional Field Techniques Using Virtual Reality
Geological Studio (VRGS). In Proceedings of the 69th EAGE Conference and Exhibition Incorporating SPE EUROPEC 2007; EAGE
Publications BV: Utrecht, The Netherlands, 2007; p. cp-27-00300.
46.
Koirala, P.; Løke, T.; Baarstad, I.; Fridman, A.; Hernandez, J. Real-time hyperspectral image processing for UAV applications,
using HySpex Mjolnir-1024. In Proceedings of the Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral
Imagery XXIII; SPIE: Bellingham, WA, USA; Washington, DC, USA, 2017; Volume 10198, p. 1019807.
... Outcrops in a lithium-pegmatite mine in Northern Portugal were scanned using ground-based and UAS-based SWIR [113], and underground HSI outcrop mapping has been performed in several instances, e.g., to map clay materials to assess sealing properties in the Swiss Mont Terri rock laboratory [114] and in a Li-underground mine in Zinnwald/Cìnovec [115]. The first UAS-based, fully corrected, oblique, hyperspectral SWIR survey of an outcrop was published for limestone lithology [116]. A review of close-range, ground-based hyperspectral studies from 2019 [33] covers many of these studies in detail. ...
... The SWIR wavelength range introduces dependency on accurate measurements of atmospheric conditions, as well as the geometric and radiometric correction and processing of the data. Only a limited number of publications using UAS-mounted SWIR cameras are currently available (e.g., [116,118,[177][178][179][180]), meaning that this research field is only just emerging, and our understanding of good methodologies remains nascent. Similarly, the best-practice guidelines for the acquisition and correction of coaligned VNIR-SWIR datasets, and the accurate georeferencing and geometrical correction of such combined datasets, remain a topic of current research. ...
... This and the lack of robust equipment to place into an active mine site might explain the reluctance to adopt HSI in active mines. Standard methods of reflectance retrieval and correction approaches for ground-based HSI include empirical line correction based on ground-truth spectra, terrain and geometric correction, and incident light correction using photogrammetry or LiDAR-based 3D models for closeand long-range applications [35,43,[115][116][117][118][119][120][121]. Sophisticated UAS-based atmospheric and geometric correction is currently established for nadir-based data acquisition. ...
Article
Full-text available
Hyperspectral imaging technology holds great potential for various stages of the mining life cycle, both in active and abandoned mines, from exploration to reclamation. The technology, however, has yet to achieve large-scale industrial implementation and acceptance. While hyperspectral satellite imagery yields high spectral resolution, a high signal-to-noise ratio (SNR), and global availability with breakthrough systems like EnMAP, EMIT, GaoFen-5, PRISMA, and Tanager-1, limited spatial and temporal resolution poses challenges for the mining sectors, which require decimetre-to-centimetre-scale spatial resolution for applications such as reconciliation and environmental monitoring and daily temporal revisit times, such as for ore/waste estimates and geotechnical assessments. Hyperspectral imaging from drones (Uncrewed Aerial Systems; UASs) offers high-spatial-resolution data relevant to the pit/mine scale, with the capability for frequent, user-defined re-visit times for areas of limited extent. Areas of interest can be defined by the user and targeted explicitly. Collecting data in the visible to near and shortwave infrared (VNIR-SWIR) wavelength regions offers the detection of different minerals and surface alteration patterns, potentially revealing crucial information for exploration, extraction, re-mining, waste remediation, and rehabilitation. This is related to but not exclusive to detecting deleterious minerals for different processes (e.g., clays, iron oxides, talc), secondary iron oxides indicating the leakage of acid mine drainage for rehabilitation efforts, swelling clays potentially affecting rock integrity and stability, and alteration minerals used to vector toward economic mineralisation (e.g., dickite, jarosite, alunite). In this paper, we review applicable instrumentation, software components, and relevant studies deploying hyperspectral imaging datasets in or appropriate to the mining sector, with a particular focus on hyperspectral VNIR-SWIR UASs. Complementarily, we draw on previous insights from airborne, satellite, and ground-based imaging systems. We also discuss common practises for UAS survey planning and ground sampling considerations to aid in data interpretation.
... On the contrary imaging spectroscopy, hyperspectral remote sensing, samples the electromagnetic spectrum in a large number of contiguous, narrow intervals, while the spectral bandwidth is often under 10 nm (Meyer et al., 2022). This imaging setup can generate large volumes of information (Kruse, 1998), making hyperspectral data a rich source of information (Huber et al., 2008;Pullanagari, 2012;Kopačková, 2014;Notesco et al., 2014a;Cheng and Sun, 2016;Lawley, 2016;Lorenz et al., 2021;Thiele et al., 2021a). Electromagnetic radiation interacts uniquely with every object depending on the object's chemical structure/compositions and physical properties (e.g., grain size). ...
... The spectral analysis includes a diverse range of techniques, such as classification, and quantification of materials (Figure 1.4) (Chang, 2007;Khalili et al., 2015;Asadzadeh and de Souza Filho, 2016;Dalm et al., 2017;Pullanagari et al., 2017a;Kereszturi et al., 2018;Chang, 2021;Shrestha et al., 2022). The advances in machine learning and pattern recognition have made identifying the correlation between the hyperspectral data and the desired target variable, such as element concentration, alteration extent, biomass magnitude and water contamination, and providing robust results compared to many classical techniques (Ghamisi et al., 2017b;Ghamisi et al., 2021;Parsa, 2021;Thiele et al., 2021a;Booysen et al., 2022). Machine learning-based approaches are known for their robustness in handling variability in the ground truth and spectral data. ...
Thesis
Full-text available
Preliminary steps of mineral exploration have traditionally included drilling and other destructive, expensive, and time-consuming techniques. To meet the ever-increasing demand for mineral resources pertaining to the increase in population and technological demand, it is very important to develop environmentally friendly, faster, and cheaper prospecting methods. In this study, we have targeted three known regions of mesothermal gold mineralisation in the South Island, New Zealand to develop hyperspectral remote sensing-based prospect models combined with biogeochemical data. The three study sites have geological similarities around the gold mineralisation including the major pathfinder elements. On the contrary the environmental settings, and other surficial and near-surface processes including the soil and groundwater interactions with the host rock, are vastly different. This led to a wide variation in the physico-chemical properties of the soil cover and the subsequent uptake by the overlying vegetation. The Pinus radiata plantation at the Hyde-Macraes Shear Zone was the first study site to test the feasibility of using biogeochemical responses overlying the gold mineralisation through hyperspectral remote sensing for gold prospecting. Pinus radiata is known to be an accumulator of metals and metalloids with roots reaching as deep as the shear zone beneath it. The data showed a good spatial elemental trend along the shear zone for both the bark and the needle samples although the regression models performed much better with R 2 CV >0.7 for the bark samples. After confirming the feasibility of utilising the vegetation cover as a medium, the second site in the Rise and Shine Shear Zone was examined to assess the limits of the airborne hyperspectral data over variably exposed soil. The potentially high As anomalies indicating the gold mineralisation were classified coupled with a thorough understanding of the soil cover and its relation to the lithology. The orthogonal total variation component analysis transformed data produced the best-performing models using random forest classification with an accuracy ~50% for the high concentration As zonation. Finally, the third study site in Reefton exhibited a multi-species natural forest overlying the gold mineralisation. Apart from varying elemental responses among the different species the Reefton study area also manifested regions contaminated by previous mining activity which likely impacted the elemental uptake in the overlying vegetation. The regression models performed poorly but the spatial predictions rendered some valid correlations based on ground knowledge from previous studies.
... These measurements have been used to predict various plant biophysical and biochemical traits, as well as biodiversity metrics such as plant functional diversity and species composition from UASs and other airborne platforms (e.g., [3][4][5][6]). Beyond the study of ecological systems, UAS imaging spectroscopy is being increasingly utilised across a range of industries, including agricultural [7,8], and geological applications [9,10]. Given the modular nature of UAS platforms, payload weight and size limitations often necessitate 2 of 21 the use of smaller, specialised sensors or multiple flights with different imaging spectrometers. ...
Article
Full-text available
Remote sensing from unoccupied aerial systems (UASs) has witnessed exponential growth. The increasing use of imaging spectroscopy sensors and RGB cameras on UAS platforms demands accurate, cross-comparable multi-sensor data. Inherent errors during image capture or processing can introduce spatial offsets, diminishing spatial accuracy and hindering cross-comparison and change detection analysis. To address this, we demonstrate the use of an optical flow algorithm, eFOLKI, for co-registering imagery from two pushbroom imaging spectroscopy sensors (VNIR and NIR/SWIR) to an RGB orthomosaic. Our study focuses on two ecologically diverse vegetative sites in Tasmania, Australia. Both sites are structurally complex, posing challenging datasets for co-registration algorithms with initial georectification spatial errors of up to 9 m planimetrically. The optical flow co-registration significantly improved the spatial accuracy of the imaging spectroscopy relative to the RGB orthomosaic. After co-registration, spatial alignment errors were greatly improved, with RMSE and MAE values of less than 13 cm for the higher-spatial-resolution dataset and less than 33 cm for the lower resolution dataset, corresponding to only 2–4 pixels in both cases. These results demonstrate the efficacy of optical flow co-registration in reducing spatial discrepancies between multi-sensor UAS datasets, enhancing accuracy and alignment to enable robust environmental monitoring.
... The work also showed the difference in the use of the obtained data in VNIR, SWIR and LWIR ranges. In the Dolomites (Italy), calcite, dolomite carbonate formations, clay and mica minerals were distinguished using hyperspectral data from UAVs during mineralogical mapping [21]. ...
Article
Full-text available
Effective and sustainable resource management is critical to the mining industry, which plays a significant role in the Kazakhstan economy. To achieve this goal, the use of remote sensing technologies has become necessary. These advanced tools are now integral to activities such as mineral mapping, research and environmental monitoring. Therefore, the purpose of this comprehensive review article is to scrutinize how remote sensing data is used specifically in the context of mining in Kazakhstan. The comprehensive review covers a wide range of aspects, including the use of remote sensing for mapping and mineral resource studies. Key findings highlight the growing importance of remote sensing technologies in Kazakhstan's mining industry as they facilitate the precise delineation of ore deposits, enable effective monitoring of environmental impacts and promote optimal use of natural resources. The article discusses the difficulties and limitations encountered in the application of remote sensing data. Particular attention is paid to the need to improve methods of data processing, correction of atmospheric influences and spectral analysis. The overview presented in this paper is of great importance to those involved in research, policy development and industrial sectors who are interested in using remote sensing data to enhance mining activities and promote environmental sustainability in Kazakhstan. By providing a comprehensive analysis of the subject, this review offers valuable information that can help researchers deepen their understanding of the topic and guide them toward more effective strategies.
... The eastern wall of the Nuvolau outcrop ( Figure 10d) represents a spectacular example of the platform-to-basin transition: the rock wall is mainly represented by the massive northeastward-dipping clinoforms of the slope facies, that prograde over the coeval basinal deposits, composed by marly limestones (San Cassiano Formation) F I G U R E 7 (a)-(c) Zoom onto the σ 1 results of modelling for the second stage of carbonate platform growth with PA ratios equal to 0, 1 and 25. EAGE MENEGONI et al. (Blendinger & Blendinger, 1989;Bosellini, 1984;Thiele et al., 2021). The inner platform facies is visible only in a limited area at the upper and southern part of the Nuvolau outcrop. ...
Article
Full-text available
A two-dimensional numerical analysis based on the finite element method and linear elasticity is used to demonstrate how the differential compaction of the basinal unit can cause the early deformation of a prograding and/or aggrading carbonate platform. Our model investigates the modification of the carbonate platform stratal architecture and stress field driven by the process of differential compaction. We compared the results of our model with observations from two Triassic carbonate platforms in the Italian Dolomites: Lastoni di Formin and Nuvolau Mts. (Passo Giau, Italy). We show that the model can explain the modification of stratal architecture, as well as fault and fracture patterns observed on these platforms. In particular, we show that (1) the slope and slope-to-basin transition regions are expected to experience most of the brittle deformation and, differently from what was suggested by previous numerical studies, the formation of platform-ward dipping faults and major fractures with dip angles that tend to decrease moving dip-ward. In addition, (2) the inner platform region can exhibit a slightly tensile regime, which may lead to the formation of syndepositional and/or syndiagenetic fractures. Moreover, (3) in the case of predominantly prograding platforms, the results of the model show a general tilting and thickening of the inner platform strata towards the shelf-slope break.
Article
The Dead Sea is one of the most saline terminal lakes on Earth, and few organisms survive in this harsh environment. In some onshore spring pools, active and diverse microbial communities flourish. In the geological past, microbial-rich environments left their marks in the form of stromatolites. Stromatolites are studied to better understand the appearance of life on Earth and potentially on other planets. Hyperspectral methodologies have been shown to be useful for detecting structures in stromatolites. In an effort to characterize the biosignatures and chemical composition inherent to stromatolites, we created a spectral classification scheme for distinguishing between stromatolites and their bedrock environment—typically carbonatic rocks, mostly dolomites. The overarching aim comprises the development of an automated hyperspectral reflectance method for detecting the presence of stromatolites. We collected and measured 82 field samples with an ASD spectrometer and used our spectral dataset to train three machine learning algorithms (linear regression, K-Nearest Neighbor, XGBoost). The results show the successful detection of stromatolites, with all three prediction methods giving high accuracy rates (stromatolite > 0.9, bedrock dolomite > 0.8). The continuum removal and spectral ratio technique results identified two significant spectral regions, ~1900 nm (water) and ~2310–2320 nm (carbonates), that allow one to differentiate between stromatolites and dolomites. This study establishes the grounds for the automated detection of a fossilized livable environment in a carbonatic terrain based on its hyperspectral reflectance data. The results have significant implications for future mapping efforts and emphasize the feasibility of automated mapping, extending the data acquisition to airborne or satellite-based hyperspectral remote sensing technologies to detect life forms in extreme environments.
Article
Full-text available
Accurate and efficient crop maps are essential for decision-makers to improve agricultural monitoring and management, thereby ensuring food security. The integration of advanced artificial intelligence (AI) models with hyperspectral remote sensing data, which provide richer spectral information than multispectral imaging, has proven highly effective in the precise discrimination of crop types. This systematic review examines the evolution of hyperspectral platforms, from Unmanned Aerial Vehicle (UAV)-mounted sensors to space-borne satellites (e.g., EnMAP, PRISMA), and explores recent scientific advances in AI methodologies for crop mapping. A review protocol was applied to identify 47 studies from databases of peer-reviewed scientific publications, focusing on hyperspectral sensors, input features, and classification architectures. The analysis highlights the significant contributions of Deep Learning (DL) models, particularly Vision Transformers (ViTs) and hybrid architectures, in improving classification accuracy. However, the review also identifies critical gaps, including the under-utilization of hyperspectral space-borne imaging, the limited integration of multi-sensor data, and the need for advanced modeling approaches such as Graph Neural Networks (GNNs)-based methods and geospatial foundation models (GFMs) for large-scale crop type mapping. Furthermore, the findings highlight the importance of developing scalable, interpretable, and transparent models to maximize the potential of hyperspectral imaging (HSI), particularly in underrepresented regions such as Africa, where research remains limited. This review provides valuable insights to guide future researchers in adopting HSI and advanced AI models for reliable large-scale crop mapping, contributing to sustainable agriculture and global food security.
Thesis
Full-text available
L'osservazione planetaria eseguita con satelliti remote sensing richiede spesso la convergenza dell'informazione proveniente da un sistema di imaging con quella ottenuta da uno spettrometro, con conseguenti problemi di disomogeneità originati dalla diversità degli strumenti. Lo strumento HYPSOS (HYPerspectral Stereo Observing System) è il primo strumento in grado di produrre dati 4D, cioè che contengono sia l'informazione spaziale in tre dimensioni sia quella spettrale del target osservato, ed è quindi in grado di minimizzare questa criticità. Lo strumento consiste di 2 canali ottici affiancati che consentono una visione stereoscopica del target e che alimentano un singolo spettrometro. Uno dei problemi da risolvere nella configurazione ottica di HYPSOS è evitare il "cross-talk" tra i due canali, cioè che la luce che entra da un canale illumini il piano focale dell'altro. In questa tesi si è studiato come evitare questo problema inserendo degli opportuni setti lungo il cammino ottico e si è analizzato con un opportuno software di simulazione ottica il contributo di luce diffusa che ne è conseguito, per valutare il rapporto segnale-rumore così ottenuto.
Article
Full-text available
Hyperspectral imaging (HSI) is an advanced imaging technique that captures detailed spectral information across multiple fields. This review explores its applications in counterfeit detection, remote sensing, agriculture, medical imaging, cancer detection, environmental monitoring, mining, mineralogy, and food processing, specifically highlighting significant achievements from the past five years, providing a timely update across several fields. It also presents a cross-disciplinary classification framework to systematically categorize applications in medical, agriculture, environment, and industry. In counterfeit detection, HSI identified fake currency with high accuracy in the 400–500 nm range and achieved a 99.03% F1-score for counterfeit alcohol detection. Remote sensing applications include hyperspectral satellites, which improve forest classification accuracy by 50%, and soil organic matter, with the prediction reaching R2 = 0.6. In agriculture, the HSI-TransUNet model achieved 86.05% accuracy for crop classification, and disease detection reached 98.09% accuracy. Medical imaging benefits from HSI’s non-invasive diagnostics, distinguishing skin cancer with 87% sensitivity and 88% specificity. In cancer detection, colorectal cancer identification reached 86% sensitivity and 95% specificity. Environmental applications include PM2.5 pollution detection with 85.93% accuracy and marine plastic waste detection with 70–80% accuracy. In food processing, egg freshness prediction achieved R2 = 91%, and pine nut classification reached 100% accuracy. Despite its advantages, HSI faces challenges like high costs and complex data processing. Advances in artificial intelligence and miniaturization are expected to improve accessibility and real-time applications. Future advancements are anticipated to concentrate on the integration of deep learning models for automated feature extraction and decision-making in hyperspectral imaging analysis. The development of lightweight, portable HSI devices will enable more on-site applications in agriculture, healthcare, and environmental monitoring. Moreover, real-time processing methods will enhance efficiency for field deployment. These improvements seek to enhance the accessibility, practicality, and efficacy of HSI in both industrial and clinical environments.
Conference Paper
We argue that traditional 2D hyperspectral imaging is not adapted to many modern challenges. With the rise of high spatial resolution, hyperspectral sensors mounted on different platforms (e.g. drones, terrestrial, satellites) and innovative applications (e.g. urban mapping, mining monitoring), projections, occlusions, perspective effects and data processing limit the use of 2D hyperspectral imaging. We propose that 3D hyperclouds, in which Lidar or photogrammetric point clouds are augmented with hyperspectral attributes, can address numerous of these challenges. We demonstrate the benefits of hyperclouds and dedicated machine learning architectures with several realistic examples.
Article
Full-text available
The digitization and automation of the raw material sector is required to attain the targets set by the Paris Agreements and support the sustainable development goals defined by the United Nations. While many aspects of the industry will be affected, most of the technological innovations will require smart imaging sensors. In this review, we assess the relevant recent developments of Machine Learning for the processing of imaging sensor data. We first describe the main imagers and the acquired data types as well as the platforms on which they can be installed. We briefly describe radiometric and geometric corrections as these procedures have been already described extensively in previous works. We focus on the description of innovative processing workflows and illustrate the most prominent approaches with examples. We also provide a list of available resources, codes, and libraries for researchers at different levels, from students to senior researchers, willing to explore novel methodologies on the challenging topics of raw material extraction, classification, and process automatization.
Article
Full-text available
Hyperspectral data are gaining popularity in remote sensing and signal processing communities because of the increased spectral information relative to multispectral data. Several airborne and spaceborne hyperspectral datasets are publicly available, facilitating the development of various applications and algorithms. However, hyperspectral data are usually limited by their narrow, highly correlated and contiguous spectral bands in both processing and analysis. Moreover, the resolution of available hyperspectral datasets is not sufficiently high for the identification of small objects. Nevertheless, with the rapidly advancing technology, hyperspectral imaging systems can now be mounted on small aerial vehicles for detecting small objects at low altitude. To properly handle these high spectral and spatial resolution data, new or redesigned data processing or analysis pipelines must be developed, but such datasets are limited. In this study, we describe two hyperspectral datasets acquired by a drone and evaluate their radiometric and geometric quality. Based on appropriate data acquisition and processing approaches, our datasets are expected to be useful as testbeds for new algorithms and applications. Hyperspectral remote sensing has been developed to detect individual absorption features related to specific chemical bonds in soils, liquids or gases; however, because UAV‐based pushbroom hyperspectral sensor technologies are relatively new, no public datasets are currently available. To address this gap, we acquired hyperspectral imagery over a permafrost area, and provided two hyperspectral datasets from GNSS/INS‐assisted co‐aligned pushbroom hyperspectral scanners on a drone. Our datasets are expected to provide a new perspective for improving UAV‐based hyperspectral data processing and analysis algorithms.
Article
Full-text available
The widespread application of drones and associated miniaturization of imaging sensors has led to an explosion of remote sensing applications with very high spatial and spectral resolutions. The 3-D ultrahigh-resolution digital outcrop models created using drones and oblique imagery from ground-based sensors are now commonly used in the academic and industrial sectors, while the generation of spatially accurate models has been greatly facilitated by the development of computer vision tools, such as structure from motion, and the correction of spectral attributes to achieve material reflectance measurements remains challenging. Following the development of a topographical correction toolbox (mephysto), we now propose a series of new tools that can leverage the detailed geometry captured by digital outcrop models to correct for illumination effects caused by oblique viewing angles and the interaction of light with complex 3-D surfaces. This open-source code is integrated into hylite, a python toolbox for the full 3-D processing and fusion of digital outcrop models with hyperspectral imaging data. We validate the performance of our novel method using a case study at an open-pit mine in Tharsis, Spain, and demonstrate the importance of accurate illumination corrections for quantitative spectral analyses. Significantly, we show that commonly applied spectral analysis techniques can yield erroneous results for data corrected using current state-of-the-art approaches. Our proposed method ameliorates many of the issues with these established approaches.
Article
Full-text available
Hyperspectral sensors mounted in unmanned aerial vehicles (UAVs) offer new opportunities to explore high-resolution multi-temporal spectral analysis in remote sensing applications. Nevertheless, the use of hyperspectral data still poses challenges mainly in postprocessing to correct from high geometric deformation of images. In general, the acquisition of high-quality hyperspectral imagery is achieved through a time-consuming and complex processing workflow. However, this effort is mandatory when using hyperspectral imagery in a multi-sensor data fusion perspective, such as with thermal infrared imagery or photogrammetric point clouds. Push-broom hyperspectral sensors provide high spectral resolution data, but its scanning acquisition architecture imposes more challenges to create geometrically accurate mosaics from multiple hyperspectral swaths. In this article, an efficient method is presented to correct geometrical distortions on hyperspectral swaths from push-broom sensors by aligning them with an RGB photogrammetric orthophoto mosaic. The proposed method is based on an iterative approach to align hyperspectral swaths with a RGB photogrammetric orthophoto mosaic. Using as input pre-processed hyperspectral swaths, the workflow is fully automatic and consists of: (1) adaptive swath subdivision into multiple fragments; (2) detection of significant image features; (3) estimation of valid matches between individual swaths and the RGB orthophoto mosaic; and (4) calculation of the best geometric transformation model to the retrieved matches. This methodology provides an expedite solution able to produce a hyperspectral mosaic, enabling the hyperspectral data integration with data from other sensors for multiple applications.
Article
Full-text available
Enhanced digital outcrop models attributed with hyperspectral reflectance data, or hyperclouds, provide a flexible, three-dimensional medium for data-driven mapping of geological exposures, mine faces or cliffs. This approach allows the collection of spatially contiguous information on exposed mineralogy and so provides key information for understanding mineralising processes, interpreting 1-D drillhole data, and optimising mineral extraction. In this contribution we present an open-source python workflow, hylite, for creating hyperclouds by seamlessly fusing geometric information with data from a variety of hyperspectral imaging sensors and applying necessary atmospheric and illumination corrections. These rich datasets can be analysed using a variety of techniques, including minimum wavelength mapping and spectral indices to accurately map geological objects from a distance. Reference spectra from spectral libraries, ground or laboratory measurements can also be included to derive supervised classifications using machine learning techniques. We demonstrate the potential of the hypercloud approach by integrating hyperspectral data from laboratory, tripod and unmanned aerial vehicle acquisitions to automatically map relevant lithologies and alterations associated with volcanic hosted massive sulphide (VHMS) mineralisation in the Corta Atalaya open-pit, Spain. These analyses allow quantitative and objective mineral mapping at the outcrop and open-pit scale, facilitating quantitative research and smart-mining approaches. Our results highlight the seamless sensor integration made possible with hylite and the power of data-driven mapping approaches applied to hyperclouds. Significantly, we also show that random forests (RF) trained only on laboratory data from labelled hand-samples can be used to map outcrop scale data.
Article
Full-text available
The raster data model has been the standard format for hyperspectral imaging (HSI) over the last four decades. Unfortunately, it misrepresents HSI data because pixels are not natively square nor uniformly distributed across imaged scenes. To generate end products as rasters with square pixels while preserving spectral data integrity, the nearest neighbor resampling methodology is typically applied. This process compromises spatial data integrity as the pixels from the original HSI data are shifted, duplicated and eliminated so that HSI data can conform to the raster data model structure. Our study presents a novel hyperspectral point cloud data representation that preserves the spatial-spectral integrity of HSI data more effectively than conventional square pixel rasters. This Directly-Georeferenced Hyperspectral Point Cloud (DHPC) is generated through a data fusion workflow that can be readily implemented into existing processing workflows used by HSI data providers. The effectiveness of the DHPC over conventional square pixel rasters is shown with four HSI datasets. These datasets were collected at three different sites with two different sensors that captured the spectral information from each site at various spatial resolutions (ranging from ∼1.5 cm to 2.6 m). The DHPC was assessed based on three data quality metrics (i.e., pixel loss, pixel duplication and pixel shifting), data storage requirements and various HSI applications. All of the studied raster data products were characterized by either substantial pixel loss (∼50–75%) or pixel duplication (∼35–75%), depending on the resolution of the resampling grid used in the nearest neighbor methodology. Pixel shifting in the raster end products ranged from 0.33 to 1.95 pixels. The DHPC was characterized by zero pixel loss, pixel duplication and pixel shifting. Despite containing additional surface elevation data, the DHPC was up to 13 times smaller in file size than the corresponding rasters. Furthermore, the DHPC consistently outperformed the rasters in all of the tested applications which included classification, spectra geo-location and target detection. Based on the findings from this work, the developed DHPC data representation has the potential to push the limits of HSI data distribution, analysis and application.
Article
Full-text available
Reflectance spectroscopy allows cost-effective and rapid mineral characterisation, addressing mineral exploration and mining challenges. Shortwave (SWIR), mid (MIR) and thermal (TIR) infrared reflectance spectra are collected in a wide range of environments and scales, with instrumentation ranging from spaceborne, airborne, field and drill core sensors to IR microscopy. However, interpretation of reflectance spectra is, due to the abundance of potential vibrational modes in mineral assemblages, non-trivial and requires a thorough understanding of the potential factors contributing to the reflectance spectra. In order to close the gap between understanding mineral-diagnostic absorption features and efficient interpretation of reflectance spectra, an up-to-date overview of major vibrational modes of rock-forming minerals in the SWIR, MIR and TIR is provided. A series of scripts are proposed that allow the extraction of the relative intensity or wavelength position of single absorption and other mineral-diagnostic features. Binary discrimination diagrams can assist in rapidly evaluating mineral assemblages, and relative abundance and chemical composition of key vector minerals, in hydrothermal ore deposits. The aim of this contribution is to make geologically relevant information more easily extractable from reflectance spectra, enabling the mineral resources and geoscience communities to realise the full potential of hyperspectral sensing technologies.
Article
Full-text available
Abstract Rare earth elements (REEs) supply is important to ensure the energy transition, e-mobility and ultimately to achieve the sustainable development goals of the United Nations. Conventional exploration techniques usually rely on substantial geological field work including dense in-situ sampling with long delays until provision of analytical results. However, this approach is limited by land accessibility, financial status, climate and public opposition. Efficient and innovative methods are required to mitigate these limitations. The use of lightweight unmanned aerial vehicles (UAVs) provides a unique opportunity to conduct rapid and non-invasive exploration even in socially sensitive areas and in relatively inaccessible locations. We employ drones with hyperspectral sensors to detect REEs at the earth’s surface and thus contribute to a rapidly evolving field at the cutting edge of exploration technologies. We showcase for the first time the direct mapping of REEs with lightweight hyperspectral UAV platforms. Our solution has the advantage of quick turn-around times (< 1 d), low detection limits (< 200 ppm for Nd) and is ideally suited to support exploration campaigns. This procedure was successfully tested and validated in two areas: Marinkas Quellen, Namibia, and Siilinjärvi, Finland. This strategy should invigorate the use of drones in exploration and for the monitoring of mining activities.
Article
Geological models from outcrop analogues are often utilized as a guide, or soft constraint, for distributing reservoir properties in subsurface models. In carbonate outcrops, combined sequence stratigraphic, sedimentological, and petrographic studies constrain the heterogeneity of geobodies and diagenetic processes, including dolomitization, at multiple scales. High-resolution digital outcrop modeling further aids geometric mapping, geobody definition, and statistical analysis, though its usefulness for detailed mineralogical and lithological mapping is limited. Hyperspectral imaging offers enhanced spectral resolution for mapping subtle mineralogical differences. In both outcrops and subsurface, differences in carbonate composition can provide key information for distributing porosity and permeability, yet this mapping is highly challenging in field studies due to access difficulties, visible material differences and sampling resolution. Spectral analysis of limestone– dolomite ratios conducted in laboratory studies indicates theoretical measures for quantitative identification and mapping of dolomite degrees within carbonate rocks. In this study, close-range hyperspectral imaging is applied to outcrops of the Alacón Member, Barranco del Mortero, northeastern Spain, to identify exposed limestone–dolomite geobodies and to quantify the degree of dolomitization across outcrop faces. Hyperspectral imaging is supplemented with photogrammetric outcrop modeling, field spectroscopy, and laboratory sample analysis for empirical validation and uncertainty analysis. Hyperspectral mapping shows that earlier fieldwork utilizing visual inspection of difficult to access outcrop surfaces had overestimated the amount of dolomite in the outcrop. Results indicate that hyperspectral imaging identified dolomite bodies more accurately and reliably than conventional field methods and facilitates the mapping of dolomite contribution in areas modified by dedolomitization, where dolomite content changes by more than ~20%.
Article
Geometallurgical assessment of orebodies in the mining industry typically relies on bench-scale or lab-based characterization techniques. In this study, we investigate drone- and tripod-based field hyperspectral imaging as a potential addition to the geometallurgy toolkit in multiple applications. This pilot study tests hyperspectral imaging for large-scale mineral mapping in and around the active Lisbon Valley copper mine, including natural exposures, previously producing U-V mines, highwalls, dumps, and leaching sites. Tests include different (supervised and unsupervised) mineral data classification methods, varying mineral spectral reference libraries, comparison with ground-truth geological and spectroscopic mapping and sampling, and integration with LiDAR data. The results show that hyperspectral scans can produce spatially registered maps of the distribution of different spectrally active mineral types over dumps, highwalls, leach pads, and natural outcrops. Clays, other phyllosilicates, carbonates, and sulfates showed up particularly well. The sensor was also able to distinguish dry from lixiviant-saturated areas and map different clay types on the leach pads, and shows promise for differentiating types and health of vegetation. These results suggest that hyperspectral imaging, if coupled with robust ground-truthing, can be a useful complement to existing geometallurgical techniques in the mining industry, such as geological mapping, blast hole sampling and automated mineralogy identifications, and handheld spectrometry. In particular, hyperspectral imaging has promise for mapping the distribution of acid-consuming minerals; mapping the distribution of swelling, sliming, and heap-blinding clays; and pinpointing problem areas on heap leach pad surfaces.