Revolutionizing Igneous Rock Classification: Proportion-Based Deep Learning Analysis of Petrographic Thin Section Photomicrographs

In our study, we present a segmented-based, rule-driven classification of igneous rocks through the analysis of thin section photomicrographs, representing a significant advancement over traditional petrographic methods. This deep learning-based approach is especially innovative in its recognition that the naming of rocks is intrinsically linked to the proportion of minerals they contain, a vital aspect frequently overlooked in conventional classification techniques. By focusing on accurately quantifying these mineral proportions, our method effectively addresses the subjectivity and observer variability inherent in traditional petrography. Utilizing semantic image segmentation on 963 petrographic thin section photomicrographs, we have successfully identified 29 distinct minerals and classified 15 types of igneous rocks. This showcases the precision and scope of our approach, which automates the quantification of mineral proportions, thus ensuring a more objective and precise rock classification. The development of our proprietary dataset mask, despite its labor-intensive nature and the challenges with incomplete labelling, was crucial for achieving accurate segmentation based on the proportional regions of each mineral within the photomicrographs. This segmentation, key to our rule-driven classification, streamlines the rock naming process. Our method not only sets new standards in igneous rock classification but also signifies a transformative leap in geological research. By integrating advanced image processing with deep learning, we are opening new frontiers in Earth sciences, highlighting the transformative impact of technology in refining traditional geological methodologies. Considering the dataset's incomplete and highly imbalanced mask scenario, our method achieves an accuracy of 73.32%, significantly surpassing the baseline method using VGG16 as the backbone, which attains only 63.64% classification accuracy.