Semantic segmentation of petrographic thin sections using deep neural networks for automated mineral mapping

Accurate mineral identification in petrographic thin sections is a fundamental task in geosciences, traditionally performed through manual optical microscopy. This study presents a deep learning-based approach for semantic segmentation of minerals in thin section images using a ResNeXt101-UNet architecture. The model targets eleven semantic classes: ten distinct mineral phases (biotite, chlorite, K-feldspar, muscovite, olivine, plagioclase, quartz, serpentine, talc, titanite) and one background class, representing a diverse set of silicate and phyllosilicate minerals commonly found in basement rocks. By leveraging the feature representation capabilities of ResNeXt101 as an encoder and the spatial precision of the U-Net decoder, our approach achieves superior performance in delineating fine mineral boundaries and differentiating between visually similar phases. Quantitative evaluations demonstrate that the ResNeXt101-UNet outperforms the baseline U-Net and ResNet-UNet models across multiple segmentation metrics, particularly in handling mineralogical heterogeneity and subtle textural variations. Over the 4,440,900 annotated test pixels, the model reaches macro-averaged precision, recall and F1-score of 0.842, 0.811 and 0.813 and weighted-averaged values of 0.874, 0.860 and 0.863, respectively. Performance is highest for optically distinctive phases (olivine F1 = 0.95, serpentine = 0.91, biotite = 0.90) and lowest for the background class (F1 = 0.44) and for sparsely represented accessory minerals such as titanite (0.76) and chlorite (0.78), which remain the principal sources of error. These results highlight the potential of advanced deep learning frameworks to support and accelerate petrographic analysis with high accuracy and reliability.

Earth Science Informatics

Earth Science Informatics

19

1865-0481

N/A

N/A

18