Classifying meteorites with MetNet: A deep learning approach using reflectance spectroscopy

Meteorites, remnants of asteroids that successfully survive their passage through the Earth's atmosphere, hold critical information about the evolution and history of the solar system. Traditional methods of analyzing these rare and precious specimens often involve destructive geochemical techniques, which deplete the sample and limit subsequent analyses. The accurate classification of meteorites, typically determined through petrological examination, is crucial before any further analytical steps. Reflectance spectroscopy, which interprets a sample's characteristics by analyzing reflected light, has emerged as a nondestructive alternative with significant potential for meteorite classification. In this technique, apparently, sometimes we do not need to process the sample. This technique allows for the examination of spectral features such as absorption bands, symmetry, band centers, inflection points, and overall slope. In this study, we employed spectral reflectance data from 1781 meteorite samples to develop and fine-tune a deep learning model capable of accurate classification. The model was trained on 75% of the dataset and validated on the remaining 25%, achieving a validation accuracy of 93%. These results demonstrate the efficiency of using deep learning and reflectance spectroscopy for meteorite classification, offering a nondestructive and accurate alternative to traditional methods.