Resumen
Hydrothermal alteration minerals, which are important as indicators in the exploration of ore deposits, exhibit diagnostic absorption peaks in the short-wavelength infrared region. We propose an approach for the identification of alteration minerals that uses a deep learning method and compare it with conventional identification methods which use numerical calculation. Inexpensive spectrometers often tend to show errors in the wavelength direction, even after wavelength calibration, which causes erroneous mineral identification. In this study, deep learning is applied to extract features from reflectance spectra to remove such errors. Two typical deep learning methods?a convolutional neural network and a multi-layer perceptron?were applied to spectral reflectance data, with and without hull quotient processing, and their accuracy rates and f-values were evaluated. There was an improvement in mineral identification accuracy when hull quotient processing was applied to the learning data.