This project studies how hyperspectral imaging can be used in mineral exploration to investigate spectral signatures to classify minerals. Interpreting these spectra can be difficult and often depends on a reference library or absorption features that have been defined manually. The aim of this project was to investigate whether machine learning can be used as a supportive tool for mineral classification on drill core data.

The model, My_Model was trained using mineral measurements from a hyperspectral mineral library in the short wave infrared range. The data had to be cleaned, filtered and pretreated using Standard Normal Variate correction and the Savitzky Golay filter before training. The model was then evaluated on several different datasets provided by the company, Prediktera and then compared with an already existing USGS based classification model.

The results showed that My_Model often classified similarly to the USGS MICA model on drill core datasets, especially for data where muscovite and chlorite were the dominant minerals. On the second dataset, the model classified 5 out of 6 minerals correctly on minerals that were included in the training data. However the additional spectral library based validation test showed that the model did not generalize as well, with an overall accuracy of 35.4%.

The results indicate that neural networks can be useful for hyperspectral mineral classification, but that the model is most likely limited by the amount of variation in the training data. Before this method can reliably be used in practice, more independent samples and better validation data should be added.