Abstract

The mining operations are part of the industry 4.0 revolution, and there is a need in developing new ways to produce a flow of information among all the processes of a plant. In this context, the lithological classification of the rocks, just after being extracted, provides information related to their chemical composition and physical properties. Hyperspectral imaging is an exceptional tool for acquiring information to perform this characterization. We present a method based on deep learning and hyperspectral images, within the short-wavelength infrared range of 900-2500 nm, to perform lithological classification. The method performs an instance segmentation of the rocks, thus segmenting and classifying the rocks at the same time. A transfer learning methodology was applied by using a deep neural network pretrained with millions of color images to classify the rocks. To use this network, the dimensionality of the hyperspectral images is reduced from 268 to only 3 channels by another neural network. In addition, these 3-channels images can be used for human interpretation. We compare various deep network architectures and classical methods for performing dimensionality reduction. The method was tested on our hyperspectral image database with 13 different lithological classes, obtaining an F1-score that was above 96% and 98% in the instance and pixel-wise performance, respectively.