Mining Goes Digital – Mueller et al. (Eds)
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
655
Lithological hyperspectral characterization for UAV sensor
selection
F.S. Beretta, A.L. Rodrigues, R.L. Peroni, S.B. Rolim & J.F. Costa
Federal University of Rio Grande do Sul, Porto Alegre, RS, Brazil
ABSTRACT: The definition between lithologies is vital for a successful mine planning.
new field data is constantly collected to feed the reconciliation process. Whilst performing
with the traditional approach, the professional is exposed to weather conditions, dust, risks
with the moving equipment and bench highs. This task can benefit from the automation
of boundaries detection between materials. The use of Unmanned Aerial Vehicles (UAVs)
and automated classification using Machine Learning (ML) techniques improved in the last
few years. However, non-visible wavelengths are still rarely used for materials classification.
Other electromagnetic frequencies can be captured by specific sensors so the automatization
process could use these data as training and test in a supervised classification. The characterization of materials that are similar in visible wavelengths can be useful to understand the
behavior of the electromagnetic bands to clearly differentiate materials, so the correct sensor
can be plugged onto the aerial vehicle.
1 INTRODUCTION
The identification of contacts between lithologies and ore/waste short term mapping is the
basis to short term mine planning, geological model updating and reconciliation between the
projected and the actual production. Automated lithological classification can grant safety
to the geologist, increases the model updating frequency and accuracy of the georeferenced
contacts. The input models are created from photogrammetric air-born point clouds, whose
points carry any local information detected by unmanned aerial vehicles (UAVs) and their
sensors.
Photogrammetric topographic reconstitution is already established in the mining industry, as presented by Salvani et al. (2016), Hugenholtz et al. (2013) Silva et al. (2016), Peroni
(2016), Westoby et al. (2012) and Beretta et al. (2017). Using machine learning (ML) techniques on 2-D images, Dalm et al. (2017), Ehrenfeld et al. (2017), and Bamford et al. (2017)
presented mineralogical differentiation, metallurgical classification and granulometric characterization, respectively. Automated lithological classification already being used based on
3-D point clouds on the visual specter range, such as red, green and blue (RGB) components
acquired by ordinary cameras, as presented by Beretta et al. (2019). Figure 1 shows the input
RGB point cloud along with the training/test selected dataset feeding the resulting classified
model for the selected ML algorithm and the sample size.
When the materials are visually similar, sensors that detect other electromagnetic wavelengths are necessary to detect differences that the human eye cannot distinguish. For example, long-wave infrared or any other wavelength that highlights the contrast between materials
for ML supervised classification of the 3-D point cloud. Dense point clouds are generally
huge, as such, excessive data can be laborious, slow and computationally time demanding.
In order to assess the importance of information for every kind of material of interest, a
prior characterization with exploratory data analysis using feature engineering is essential to
select the correct electromagnetic wavelength bands that behave differently for each pair of
lithologies.
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
655
Lithological hyperspectral characterization for UAV sensor
selection
F.S. Beretta, A.L. Rodrigues, R.L. Peroni, S.B. Rolim & J.F. Costa
Federal University of Rio Grande do Sul, Porto Alegre, RS, Brazil
ABSTRACT: The definition between lithologies is vital for a successful mine planning.
new field data is constantly collected to feed the reconciliation process. Whilst performing
with the traditional approach, the professional is exposed to weather conditions, dust, risks
with the moving equipment and bench highs. This task can benefit from the automation
of boundaries detection between materials. The use of Unmanned Aerial Vehicles (UAVs)
and automated classification using Machine Learning (ML) techniques improved in the last
few years. However, non-visible wavelengths are still rarely used for materials classification.
Other electromagnetic frequencies can be captured by specific sensors so the automatization
process could use these data as training and test in a supervised classification. The characterization of materials that are similar in visible wavelengths can be useful to understand the
behavior of the electromagnetic bands to clearly differentiate materials, so the correct sensor
can be plugged onto the aerial vehicle.
1 INTRODUCTION
The identification of contacts between lithologies and ore/waste short term mapping is the
basis to short term mine planning, geological model updating and reconciliation between the
projected and the actual production. Automated lithological classification can grant safety
to the geologist, increases the model updating frequency and accuracy of the georeferenced
contacts. The input models are created from photogrammetric air-born point clouds, whose
points carry any local information detected by unmanned aerial vehicles (UAVs) and their
sensors.
Photogrammetric topographic reconstitution is already established in the mining industry, as presented by Salvani et al. (2016), Hugenholtz et al. (2013) Silva et al. (2016), Peroni
(2016), Westoby et al. (2012) and Beretta et al. (2017). Using machine learning (ML) techniques on 2-D images, Dalm et al. (2017), Ehrenfeld et al. (2017), and Bamford et al. (2017)
presented mineralogical differentiation, metallurgical classification and granulometric characterization, respectively. Automated lithological classification already being used based on
3-D point clouds on the visual specter range, such as red, green and blue (RGB) components
acquired by ordinary cameras, as presented by Beretta et al. (2019). Figure 1 shows the input
RGB point cloud along with the training/test selected dataset feeding the resulting classified
model for the selected ML algorithm and the sample size.
When the materials are visually similar, sensors that detect other electromagnetic wavelengths are necessary to detect differences that the human eye cannot distinguish. For example, long-wave infrared or any other wavelength that highlights the contrast between materials
for ML supervised classification of the 3-D point cloud. Dense point clouds are generally
huge, as such, excessive data can be laborious, slow and computationally time demanding.
In order to assess the importance of information for every kind of material of interest, a
prior characterization with exploratory data analysis using feature engineering is essential to
select the correct electromagnetic wavelength bands that behave differently for each pair of
lithologies.