Most mining companies have registered important amounts of drill core composite spectra using different acquisition equipment and by following diverse protocols. These companies have used classic spectrography based on the detection of absorption features to perform semi-quantitative mineralogy. This methodology requires ideal laboratory conditions in order to obtain normalized spectra to compare. However, the inherent variability of spectral features—due to environmental conditions and geological context, among others—is unavoidable and needs to be managed. This work presents a novel methodology for geometallurgical sample characterization consisting of a heterogeneous, multi-pixel processing pipeline which addresses the effects of ambient conditions and geological context variability to estimate critical geological and geometallurgical variables. It relies on the assumptions that the acquisition of hyperspectral images is an inherently stochastic process and that ore sample information is deployed in the whole spectrum. The proposed framework is basically composed of: (a) a new hyperspectral image segmentation algorithm, (b) a preserving-information dimensionality reduction scheme and (c) a stochastic hierarchical regression model. A set of experiments considering white reference spectral characterization and geometallurgical variable estimation is presented to show promising results for the proposed approach.

Supervised analysis using spectral data requires a well-informed characterisation of the response variables and abundant spectral data points. The presented hyperspectral dataset comes from five sets of geometallurgical samples, each characterised by different methods. To provide the spectral data, all mineral samples were scanned with SPECIM VNIR and SWIR hyperspectral cameras. For each subset the following data are provided 1) hyperspectral reflectance images in the VNIR spectral range (400–1000 nm wavelength); 2) hyperspectral reflectance images in the SWIR spectral range (900–2500 nm wavelength); 3) hyperspectral reflectance images in the VNIR-SWIR range (merged to SWIR spatial resolution); 4) RGB images constructed from hyperspectral data using a Bilateral Filter based sensor fusion method; 5) response variables representing mineral sample characterisation results, provided as training and validation data. This dataset is intended for use in general regression and classification research and experiments. All subsets were validated using machine learning models with satisfactory results.

This paper introduces a new latent variable probabilistic framework for representing spectral data of high spatial and spectral dimensionality, such as hyperspectral images. We use a generative Bayesian model to represent the image formation process and provide interpretable and efficient inference and learning methods. Surprisingly, our approach can be implemented with simple tools and does not require extensive training data, detailed pixel-by-pixel labeling, or significant computational resources. Numerous experiments with simulated data and real benchmark scenarios show encouraging image classification performance. These results validate the unique ability of our framework to discriminate complex hyperspectral images, irrespective of the presence of highly discriminative spectral signatures.