The Application Of Subspace Clustering Algorithms In Drill-Core Hyperspectral Domaining

Diamond drilling is used in the mining industry to extract drill-cores for characterising mineral deposits. Traditionally, drill-cores are visually analysed by an on-site geologist, subjected to geochemical analyses, and then, few representative samples subjected to additional high-resolution mineralogical studies. However, the choice in samples is frequently subjective and the mineralogical analyses are highly time-consuming. In order to optimize the choice of samples and accelerate the analyses, drill-cores can be partitioned into domains, and then, laboratory analyses can be carried out on selected domains. Nevertheless, in the mining industry, automatic drill-core domaining still remains a challenge. Recently, hyperspectral imaging has become an important technique for the analysis of drill-cores in a non-invasive and non-destructive manner. Several clustering algorithms of hyper-spectral data are proposed for automatic drill-core domaining. In this paper, we suggest using advanced subspace clustering algorithms (i.e., sparse subspace clustering algorithm, spectral-spatial sparse subspace clustering algorithm). These algorithms work based on the self-representation property of the hyperspectral data. The clustering methods are tested on two drill-core samples which present different mineralogical and structural features. The subspace clustering algorithms are compared with the result of the K-means clustering algorithm. Our experimental results show that subspace clustering algorithms provide accurate drill-core domains and it is shown that including spatial information significantly improves the clustering results.