Feasibility of automatic classification of sandstones using spectral data and an artificial neural network

Lithology and mineral classification are essential in many aspects of geology. The integration of reflectance spectroscopy with an artificial neural network (ANN) offers a promising approach for facilitating automatic lithology and mineral classification, thereby reducing reliance on expert supervision. In this feasibility study, we collected reflectance spectra from three sandstone drill cores and hyperspectral images from three sandstone thin sections. We also extracted sample spectra from these datasets to train ANNs. Specifically, when the ANN architecture consisted of 2 hidden layers with 20 neurons each, the learning rate was set to 0.01, and the training loops were set to 50,000, the recognition accuracy of the ANNs reached 85.4% for drill core lithology and 80.0% for thin section mineralogy – outperforming Support Vector Machine and Random Forest. These trained ANNs were then applied to classify lithologies in drill cores and minerals in thin sections. The classification results were compared with core photos and thin section micrographs, showing strong agreement with actual results. Additionally, we found that spectral data can effectively distinguish rocks with different colors and grain sizes, and can be utilized to construct borehole stratigraphic columns, thereby facilitating rapid drill core logging. Moreover, the spatial resolution of a hyperspectral image has a certain impact on mineral recognition results.