Matthew J. Cracknell

Geological mapping using remote sensing data: A comparison of five machine learning algorithms, their response to variations in the spatial distribution of training data and the use of explicit spatial information

Machine learning algorithms (MLAs) are a powerful group of data-driven inference tools that offer an automated means of recognizing patterns in high-dimensional data. Hence, there is much scope for the application of MLAs to the rapidly increasing volumes of remotely sensed geophysical data for geological mapping problems. We carry out a rigorous comparison of five MLAs: Naive Bayes, k-Nearest Neighbors, Random Forests, Support Vector Machines, and Artificial Neural Networks, in the context of a supervised lithology classification task using widely available and spatially constrained remotely sensed geophysical data. We make a further comparison of MLAs based on their sensitivity to variations in the degree of spatial clustering of training data, and their response to the inclusion of explicit spatial information (spatial coordinates). Our work identifies Random Forests as a good first choice algorithm for the supervised classification of lithology using remotely sensed geophysical data. Random Forests is straightforward to train, computationally efficient, highly stable with respect to variations in classification model parameter values, and as accurate as, or substantially more accurate than the other MLAs trialed. The results of our study indicate that as training data becomes increasingly dispersed across the region under investigation, MLA predictive accuracy improves dramatically. The use of explicit spatial information generates accurate lithology predictions but should be used in conjunction with geophysical data in order to generate geologically plausible predictions. MLAs, such as Random Forests, are valuable tools for generating reliable first-pass predictions for practical geological mapping applications that combine widely available geophysical data.

Mapping geology and volcanic-hosted massive sulfide alteration in the Hellyer–Mt Charter region, Tasmania, using Random Forests™ and Self-Organising Maps

The Hellyer–Mt Charter region of western Tasmania includes three known and economically significant volcanic-hosted massive sulfide (VHMS) deposits. Thick vegetation and poor outcrop present a considerable challenge to ongoing detailed geological field mapping in this area. Numerous geophysical and soil geochemical datasets covering the Hellyer–Mt Charter region have been collected in recent years. These data provide a rich source of geological information that can assist in defining the spatial distribution of lithologies. The integration and analysis of many layers of data in order to derive meaningful geological interpretations is a non-trivial task; however, machine learning algorithms such as Random Forests and Self-Organising Maps offer geologists methods for indentifying patterns in high-dimensional (many layered) data. In this study, we validate an interpreted geological map of the Hellyer–Mt Charter region by employing Random Forests™ to classify geophysical and geochemical data into 21 discrete lithological units. Our comparison of Random Forests supervised classification predictions to the interpreted geological map highlights the efficacy of this algorithm to map complex geological terranes. Furthermore, Random Forests identifies new geological details regarding the spatial distributions of key lithologies within the economically important Que-Hellyer Volcanics (QHV). We then infer distinct but spatially contiguous sub-classes within footwall and hangingwall, basalts and andesites of the QHV using Self-Organising Maps, an unsupervised clustering algorithm. Insight into compositional variability within volcanic units is gained by visualising the spatial distributions of sub-classes and associated statistical distributions of key geochemical data. Compositional differences in volcanic units are interpreted to reflect contrasting primary composition and VHMS alteration styles. We conclude that combining supervised and unsupervised machine-learning algorithms provides a widely applicable, robust means, of analysing complex and disparate data for machine-assisted geological mapping in challenging terranes.

Catchment-based gold prospectivity analysis combining geochemical, geophysical and geological data across northern Australia

The results of a pilot study into the application of an unsupervised clustering approach to the analysis of catchment-based National Geochemical Survey of Australia (NGSA) geochemical data combined with geophysical and geological data across northern Australia are documented. NGSA Mobile Metal Ion® (MMI) element concentrations and first and second order statistical summaries across catchments of geophysical data and geological data are integrated and analysed using Self-Organizing Maps (SOM). Input features that contribute significantly to the separation of catchment clusters are objectively identified and assessed. A case study of the application of SOM for assessing the spatial relationships between Au mines and mineral occurrences in catchment clusters is presented. Catchments with high mean Au code-vector concentrations are found downstream of areas known to host Au mineralization. This knowledge is used to identify upstream catchments exhibiting geophysical and geological features that indicate likely Au mineralization. The approach documented here suggests that catchment-based geochemical data and summaries of geophysical and geological data can be combined to highlight areas that potentially host previously unrecognised Au mineralization.

Construction and analysis of Hydrogeological Landscape units using Self-Organising Maps

The Hydrogeological Landscape (HGL) framework divides geographic space into regions with similar landscape characteristics. HGL regions or units are used to facilitate appropriate management actions tailored to individual HGL units for specific applications such as dryland salinity and climate-change hazard assessment. HGL units are typically constructed by integrating data including geology, regolith, soils, rainfall, vegetation and landscape morphology, and manually defining boundaries in a GIS environment. In this study, we automatically construct spatially contiguous regions from standard HGL data using Self-Organising Maps (SOM), an unsupervised statistical learning algorithm. We compare the resulting SOM-HGL units with manually interpreted HGL units in terms of their spatial distributions and attribute characteristics. Our results show that multiple SOM-HGL units successfully emulate the spatial distributions of individual HGL units. SOM-HGL units are shown to define subregions of larger HGL units, indicating subtle variations in attribute characteristics and representing landscape complexities not mapped during manual interpretation. We also show that SOM-HGL units with similar attributes can be selected using Boolean logic. Selected SOM-HGL units form regions that closely conform to multiple HGL units not necessarily connected in geographic space. These SOM-HGL units can be used to establish generalised land management strategies for areas with common physical characteristics. The use of SOM for the construction of HGL units reduces the subjectivity with which these units are defined and will be especially useful over large and/or inaccessible regions, where conducting field-based validation is either logistically or economically impractical. The methodology presented here has the potential to contribute significantly to land-management decision-support systems based on the HGL framework.

Turning geophysical data into geological informationorWhy a broader range of mathematical strategies is needed to better enable discovery

Abstract The goal of exploration geophysics is to infer the nature of buried structure and, in particular, generate drill targets that lead to a mineral deposit discovery or reserve delineation. As a profession, we aim to turn geophysical data into geological information. Most geophysical techniques enable inferences to be made from airborne, ground-based or bore-hole data through a deterministic process whereby a single model ‘answer’ is generated. Well-founded algorithms include uncertainty estimates for different parameters in the model and/or some form of model validation. This approach has been successful to date, and we advocate the continued use of deterministic algorithms. We also advocate that alternate strategies for extracting information from data are used alongside deterministic strategies. If we consider two general properties of data inference approaches, that of (1) assurance and (2) opportunity, deterministic approaches score poorly regarding opportunity: that is, useful answers may be missed. Alternate strategies can be computationally intensive, but several important classes of approach, summarised in this article are now tractable on workstation or high-specification notebook PCs. By using a range of strategies we can maximise both assurance and opportunity for a particular data inference goal and obtain extra, useful geological information from our data.

Expansion of landscape characterisation methods within the Hydrogeological Landscape Framework: application in the Australian Capital Territory

ABSTRACT The Hydrogeological Landscape (HGL) Framework is a landscape-characterisation tool that is used to discern areas of similar physical, hydrogeological, hydrological, chemical and biological properties, referred to as HGL Units. The HGL Framework facilitates prioritisation of natural-resource management investment by identifying current and potential hazards in the landscape. Within prioritised regions, on-ground management actions are tailored for specific Management Areas within individual HGL Units. The HGL Unit boundaries are determined through expert interpretation of spatial and field based datasets, such as climate, landform, geology, regolith, soil, stream network, groundwater flow systems, water quality and vegetation assemblages. The resulting HGL Units are validated by an interdisciplinary team using field assessment and biophysical testing. The use of the HGL Framework for new applications creates opportunities for refinement of the existing methodology and products for end users. This paper uses an application in the Australian Capital Territory as a case study to illustrate two enhanced techniques for the landscape characterisation component of the HGL Framework: use of an unsupervised statistical learning algorithm, Self-Organising Maps (SOM), to further validate HGL Units; and landform modelling to assist in delineation of Management Areas. The combined use of SOM and landform modelling techniques provides statistical support to the existing expert and field-based techniques, ensuring greater rigour and confidence in determination of landscape patterns. This creates a more refined HGL Framework landscape-characterisation tool, facilitating more precise hazard assessment and strategic natural-resource management by end users.

National Virtual Core Library HyLogging data and Ni-Co laterites: a mineralogical model for resource exploration, extraction and remediation

ABSTRACT The National Virtual Core Library (NVCL) HyLogging core-scanning system generates mineralogical information from visible, short-wave infrared and thermal infrared spectroscopic data. Currently, HyLogging data are freely available for more than 1500 drill holes via the AuScope Discovery Portal and various Geological Survey websites. With any new technology, there is commonly a lag between provision and take-up by users that can be aided by the publication of case studies in the scientific literature. This paper uses the Mt Davies nickel–cobalt (Ni–Co) laterite deposits, located in northwest South Australia, as a case study to assess the accessibility and representation of HyLogger data and provides an example of its application to all aspects of resource mining: exploration, extraction and processing, and remediation. In this study, we combine HyLogger-derived scalars indicating Fe-oxide and clay mineralogy with historical geological logs and assay data. In general, background Ni grades (<0.1 wt%) are linked to the presence of montmorillonite + hematite ± goethite, moderate grades (0.1–1.0 wt%) are associated with goethite ± nontronite ± saponite ± kaolinite ± montmorillonite, and higher grades (1.0–2.0 wt%) are coincident with goethite and minimal clay alteration, suggesting that goethite hosts Ni mineralisation. Gibbsite, where it occurs, is found immediately above zones of moderate to high Ni grades and may be an important proximal indicator mineral of nickeliferous laterite. Such a case study serves to suggest opportunities for further data modelling and search and query functionality that could facilitate increased use of this important digital geoscience data resource by the Australian minerals industry for all aspects of resource exploitation: exploration, extraction, processing, and environmental remediation.

Spatial-Contextual Supervised Classifiers Explored: A Challenging Example of Lithostratigraphy Classification

Spatial-contextual classifiers exploit characteristics of spatially referenced data and account for random noise that contributes to spatially inconsistent classifications. In contrast, standard global classifiers treat inputs as statistically independent and identically distributed. Spatial-contextual classifiers have the potential to improve visualization, analysis, and interpretation: fundamental requirements for the subsequent use of classifications representing spatially varying phenomena. We evaluate random forests (RF) and support vector machine (SVM) spatial-contextual classifiers with respect to a challenging lithostratigraphy classification problem. Spatial-contextual classifiers are divided into three categories aligned with the supervised classification work flow: 1) data preprocessing-transformation of input variables using focal operators; 2) classifier training-using proximal training samples to train multiple localized classifiers; and 3) postregularization (PR)-reclassification of outputs. We introduce new variants of spatial-contextual classifier that employ self-organizing maps to segment the spatial domain. Segments are used to train multiple localized classifiers from k neighboring training instances and to represent spatial structures that assist PR. Our experimental results, reported as mean (n = 10) overall accuracy ±95% confidence intervals, indicate that focal operators (RF 0.754 ±0.010, SVM 0.683 ±0.010) and PR majority filters (RF 0.705 ±0.010, SVM 0.607 ±0.010 for 11 × 11 neighborhoods) generate significantly more accurate classifications than standard global classifiers (RF 0.625 ±0.011, SVM 0.581 ±0.011). Thin and discontinuous lithostratigraphic units were best resolved using non-preprocessed variables, and segmentation coupled with postregularized RF classifications (0.652 ±0.011). These methods may be used to improve the accuracy of classifications across a wide variety of spatial modeling applications.

Geological knowledge discovery and minerals targeting from regolith using a machine learning approach

We identify and understand the diverse nature of Ni mineralisation across the Australian continent using Self-Organising Maps, an unsupervised clustering algorithm. We integrate remotely sensed, continental-scale multivariate geophysical/ mineralogical data and combine the outputs of our machine learning analysis with Ni mineral occurrence data. The resulting Ni prospectivity map identifies the location of Ni mines with an accuracy 92.58%. We divide areas of prospective Ni mineralisation into five clusters. These clusters indicate subtle but significant differences in regolith and bedrock geophysical/ mineralogical footprints of Ni sulphide and Ni laterite deposits. This information is used to identify and understand the nature of potential Ni targets in regions where prospective bedrock mineralisation is concealed by regolith materials. Our machine learning approach can be applied to the analysis of other mineral commodities and at local-/prospect-scales.

Insights into the continental structure of southeast Australia and Tasmania from passive seismic and magnetic datasets

The continental crust of southeast Australia is a complex and highly prospective area. Southeast Australia comprises the Delamerian and Lachlan Orogenies which, together with the Eastern Tasmania Terrane, are understood to have Phanerozoic basement. In contrast, the Western Tasmanian Terrane comprises areas of exposed Neoproterozoic basement which were assembled along the proto-Pacific margin of East Gondwana. In this study, the crustal structure across southeast Australia and Tasmania is considered using seismic and aeromagnetic methods. We use previous passive seismic results and present a new analysis of magnetic data. The Curie temperature, the temperature at which magnetic rocks lose their magnetisation, is investigated using spectral analysis of aeromagnetic data and the Curie point depth (CPD) is consequently determined. CPD is compared to the depth of the seismic Moho discontinuity throughout the study area. The Moho depth and newly calculated CPD throughout the study area vary from ~20 to >38 km and ~25 to >45 km, respectively. The CPD is slightly shallower than the Moho across the study area. The Delamerian and Lachlan Orogenies are underlain by a 30-35 km and -40-50 km deep Moho respectively, while average CPD depths are ~30 and ~28 km for these regions. A relatively shallow CPD is observed in the northeast of the study area and corresponds to Cainozoic volcanism in eastern Australia. The shallow Moho beneath Tasmania supports the idea of crustal thinning during Gondwana breakup. In Tasmania, CPD increases in depth from ~21 km in the northwest to >31 km in north. This is consistent with variations in the depth of the Moho from 25 km in the northwest to 37 km in the north.