A comparison of t-SNE, SOM and SPADE for identifying material type domains in geological data. (April 2019)

• Record Type: Journal Article
• Authors: Balamurali, Mehala; Silversides, Katherine L.; Melkumyan, Arman
• Is Part Of: Computers & geosciences. Volume 125(2019)
• Page Start: 78
• Page End: 89
• Publication Date: 2019-04
• Subjects: Algorithms -- Geology -- Computational methods -- Data processing; Environmental policy -- Periodicals; 550.5
• DOI: 10.1016/j.cageo.2019.01.011
• Languages: English
• ISSNs: 0098-3004

Abstract

Abstract: The standard mine modelling practice often involves investing significant effort into the interpretation of the deposit and identification of the geological domains using the chemical assays and geophysics. These domains and their accuracy play a key role in grade estimation using spatial modelling approaches, such as Gaussian processes and kriging. However, the domains developed for grade estimation do not always produce regions with well-defined correlations between the material types that are present in the deposit. The material type is based on the ore characteristics, such as mineralogy, texture and other visible petrological properties. Therefore a new and potentially more flexible methodology for domaining of material types is needed. This study applies t-SNE, SOM and SPADE to the material type data to embed high-dimensional data into low dimensions and thus facilitate clustering. These methods were tested on a banded iron formation hosted iron ore deposit located in the Hammersley region of Western Australia. All three methods produced clusters that were purer mixtures of the material types than the original domains. Due to the geologist input SPADE produced clusters closest to the original domains. However, this may not be the best clustering for material type modelling. Additionally, SPADE produced the best results for the ore, and highlighted how the user input can focus this method on the region of greatest interest. t-SNE and SOM are more automatic, but had results that were further from the original clusters. t-SNE identified clusters that were better spatially grouped than SOM, which was generally the most affected by the high variation in material within the detritals. Highlights: Accurate domaining is key for estimation using spatial modelling approaches. T-SNE, SOM and SPADE were applied to cluster geological iron ore data. High-dimensional data was embedded into low dimensions to facilitate grouping. SPADE had the best results for ore, highlighting how users can focus the results. t-SNE and SOM are more automatic, but the results were further from the domains.