Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.775806
Title: Enhancing the geological understanding of south west England using machine learning algorithms
Author: Yeomans, C. M.
ISNI:       0000 0004 7962 961X
Awarding Body: University of Exeter
Current Institution: University of Exeter
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
The geology of SW England is characterised by three geological phenomena which include the Permian granite batholith, complex structural geology and a world-class tin-tungsten-copper orefield. At present, interest in exploration for metals such as tungsten and tin is high because of their designation as "critical" to British and European economies. Furthermore, there is renewed interest in geothermal energy production and the potential for metal extraction, such as lithium, from geothermal brines. Many aspects of the resource-rich geology of SW England remain enigmatic, such as the distribution of different granite types, the fault network and the potential for mineralisation at depth (> 200 m). These three phenomena are intrinsically linked, and are key to finding new metalliferous resources and geothermal energy targets; however, the temperate climate and poor outcrop inland is a significant limitation. High-resolution airborne geophysical data collected over SW England as part of the Tellus South West project (2013-2014) provides a new opportunity to better understand these three geological phenomena. Furthermore, regional geochemical data has also been collected in 2012 as part of the Geochemical Baseline Survey of the Environment (G-BASE). These data have been investigated using machine learning algorithms for integrated data analysis. Machine learning incorporates computer science methods for pattern recognition into statistical analysis to create an objective model. The approach can use supervised or unsupervised algorithms; the former uses a priori data to train algorithms whereas the latter attempts to find natural clusters. These methods are used to generate new models of the granites, structural geology and tungsten prospectivity in the region. Granite classification models attempted to map the distribution of different granite types across the granite batholith. Unsupervised approaches, such as the k-means algorithm, found natural clusters of different granite types using airborne radiometric data and stream-sediment geochemistry, however, clusters were focused on the grouping of thorium-rich and thorium-poor areas. Whilst these showed some similarity to previously described granite types, it was not possible to define all expected granite types. Data from Simons et al. (2016) were used to train a number of supervised algorithms and proved more effective due to their ability to classify non-linearly separable data. The Support Vector Machine algorithm was found to be the most effective classifier, closely followed by Random Forest. A refined geological map of the granite types helps illustrate the potential for occurrences of different granite types that may not be exposed at surface. Unsupervised image segmentation algorithms within an Object-Based Image Analysis framework are applied for regional lineament detection using the high-resolution airborne geophysical data. Data preparation using the Tilt Derivative transform for magnetic, radiometric and LiDAR data was found to generate comparable, normalised input datasets that preserved subtle structural features. Two new stand-alone algorithms are presented that allow the integration of multi-sourced datasets for geological lineament detection. The new lineament datasets derived herein are a significant improvement on pre-existing structural datasets across the region that were either incomplete or subjectively analysed. Tungsten mineralisation is explored for using the Random Forest algorithm to generate a new prospectivity model for the region. The model incorporates high-resolution airborne geophysics and regional geochemistry with the geological lineament data derived in this thesis. Additional data are included on the proximity to granite at surface and at depth (based on gravity modelling by Willis-Richards and Jackson (1989)). Classical prospectivity modelling techniques such as Fuzzy Logic and Weights-of-Evidence, validated using the Receiver Operator Characteristics curve were both found to have accuracies of 0.891. In contrast, the Random Forest model had an accuracy of 0.989. New exploration targets are refined by the derivation of a Confidence Metric which clearly identifies three new areas for follow-up work. Metadata from the Random Forest model has been used to measure variable importance and identify superfluous data which can guide future exploration campaigns. The three geological phenomena are synthesised to demonstrate their relationship where the identification of geological lineaments in granite areas corresponds with highly variable granite distribution which is often found in close proximity to tungsten mineralisation targets. The research therefore enhances our understanding of granite distribution and structural geology with practical applications for mineral exploration. Machine learning is applied widely throughout to stream-line workflows and aid more objective decision-making. Methodologies developed herein are highly applicable to other temperate regions and could assist exploration for tungsten and other metals.
Supervisor: Shail, R. ; Lusty, P. ; Grebby, S. Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.775806  DOI: Not available
Keywords: Machine Learning ; SW England ; Granite ; Lineament Detection ; Object-Based Image Analysis ; Mineralisation ; Cornwall ; Devon ; Cornubian Batholith ; Geological Modelling ; Prospectivity ; Tungsten
Share: