Deep learning method for lithology classification of sandstone uranium deposits based on imaging spectroscopy

This article first used the HySpex imaging spectrometer to scan the rock of ZKH3 of sandstone uranium deposits in the southwestern Ordos Basin. After that, the data preprocessing was conducted and the reflectance data which can reflect rock structure and composition information was obtained. Then, based on the lithological classification rules of sandstone uranium deposits, the ability of HySpex imaging spectral data to depict rock particle size and reflect mineral composition was analyzed. It is believed that the spatial resolution of this image data can only depict the characteristics of rock particles above the coarse sandstone level, while the image spectrum can not only invert mineral compositions but also contain structural information of rocks, making it more suitable for lithology recognition. Subsequently, the characteristics of different deep learning algorithms, models, and parameters were analyzed, and a CNN model consisting of 7 one-dimensional convolutional layers, 2 pooling layers, 1 one-dimensional CBAM, and 3 fully connected layers was constructed. In addition, a total of 26877 spectral samples from 7 types of rocks were collected, and model optimization and training were completed. Finally, the performance of the model was evaluated through comparison with Support Vector Machine (SVM) and its application in the whole borehole.The results showed that its overall accuracy (OA) reached 94.6%, among which the producer’s accuracy (PA) of mudstone, fine sandstone, siltstone, medium sandstone, coarse sandstone, glutenite and background were 95.07%, 72.02%, 97.50%, 97.37%, 96.65%, 97.33%, and 99.01%, respectively. The Kappa coefficient was 0.94, which was better than SVM overall and achieved comparable results to geological logging. This indicates that the deep learning model based on imaging spectral data has good lithology classification performance, can provide reference for research on digitalization and automation of core logging.