Supervised classification of the Magdalena VMS deposit using Support Vector Machines. Drillcore scanning using SWIR hyperspectral imagery
de Waard, Robbert
Promotor(s) : Pirard, Eric
Date of defense : 20-Aug-2019 • Permalink : http://hdl.handle.net/2268.2/8517
Details
Title : | Supervised classification of the Magdalena VMS deposit using Support Vector Machines. Drillcore scanning using SWIR hyperspectral imagery |
Author : | de Waard, Robbert |
Date of defense : | 20-Aug-2019 |
Advisor(s) : | Pirard, Eric |
Committee's member(s) : | Nguyen, Frédéric
Gloaguen, Richard |
Language : | English |
Number of pages : | 128 |
Keywords : | [en] ANCORELOG [en] Automated drillcore scanning' [en] Hyperspectral [en] SWIR [en] Support vector machines [en] SVM |
Discipline(s) : | Engineering, computing & technology > Geological, petroleum & mining engineering |
Name of the research project : | ANCORELOG |
Target public : | Researchers Professionals of domain Student |
Institution(s) : | Université de Liège, Liège, Belgique |
Degree: | Master en ingénieur civil des mines et géologue, à finalité spécialisée en "geometallurgy (EMERALD)" |
Faculty: | Master thesis of the Faculté des Sciences appliquées |
Abstract
[en] The development of automated core-logging holds great potential for the improvement of core-logging procedures within the mining industry. This thesis was part of ANCORELOG, a research project that is developing automated cores-scanning technology using hyperspectral imaging, XRF, XRT and LIBS.
This work was focussed on conducting lithological classifications using support vector machine algorithms based on SWIR images of a large set of cores provided from the Magdalena mine of MATSA in the Iberian pyrite belt of Spain. Average classification accuracies of up to 73% have been achieved at a theoretical logging speed of approximately 30 meters of core per hour. It was concluded that SVM classification of SWIR images is not able to accurately classify all lithologies of Magdalena and that a simpler class-system should be developed based on properties discriminatory with SWIR. To improve results, future developments of the core-logging system should include additional sensor techniques. Furthermore, different machine learning algorithms such as artificial neural networks are more likely to achieve better results than SVM.
Secondly, efforts to create a direct correlation between alteration mineralogy and hyperspectral response were inconclusive. SEM EDX scans that were made did not produce accurate representations of chlorite or sericite concentrations and consequently made regressions meaningless. Additionally, template matching procedures to co-registrate SEM images with hyperspectral images proved to be more problematic than initially estimated. It required large amount of user-input and only 1 out of 4 samples could be matched with reasonable accuracy.
Finally, the wavelength position of the Al-OH absorption of sericite within samples was compared to its positioning relative to the orebody. A minor indication of the wavelength position shifting to shorter wavelengths when proximal to the orebody was found. A more detailed study to truly verify this correlation is recommended.
Cite this master thesis
The University of Liège does not guarantee the scientific quality of these students' works or the accuracy of all the information they contain.