Identification of geology and mineralogy in situ on the mine face is essential if mining processes are to be automated for tasks such as exploration, grading and reconciliation. Hyperspectral data acquired from field based platforms are ideally suited to these tasks, and are set to revolutionise the automated mining industry. Changing conditions of solar illumination introduces variation into the hyperspectral signal, which is unrelated to geology, complicating analysis of data. Thus, algorithms must be able to deal with changing illumination conditions and operate in a consistent and probabilistic framework, which is a prerequisite for data fusion with other sensors. To produce accurate classification maps of geology, algorithms must be developed that are effective under different conditions of illumination.The efficacy of three algorithms to classify hyperspectral data for geology mapping on vertical mine faces was assessed. These algorithms are the spectral angle mapper (SAM), and two machine learning methods operating within a fully probabilistic Gaussian process (GP) framework - the squared exponential (SE) and the observation angle dependent (OAD) covariance functions (kernel). Two experiments were done to assess the classification performances._x000D_
The first determined the performance of the algorithms by classifying hyperspectral data of known rock samples acquired using a field spectrometer._x000D_
The GP-based methods had a similar performance, with average accuracies and F-scores of 97 and 88 per cent (GP-OAD) and 95 and 81 per cent (GP-SE), respectively. Average accuracies and F-scores for SAM were significantly smaller (80 and 40 per cent respectively). The algorithms were then applied to hyperspectral imagery acquired from a vertical mine face at different times of the day, under very different conditions of illumination. In contrast with experiment1, SAM and GP-OAD gave results which were superior to GP-SE and were entirely consistent with geology mapped in the field. Compared with SAM, GP-OAD was less sensitive to the effects of shadow and produced classified images that were less noisy. The GP-SE method failed to delineate even gross changes in geology on the mine face. These results show conclusively that in order to use independent training data (spectral library) to classify imagery acquired under a different set of conditions, algorithms must take into account variability in incident illumination. Measures based on spectral distance are inferior to those based on spectral angle.The OAD kernel within the GP framework (GP-OAD) is superior to the other two methods because:it provides a probabilistic framework;it is insensitive to variability in illumination within and between imagery acquired under different conditions; andit can identify and map geological zones on a vertical mine face using an independent spectral library, without a priori knowledge.