Analysis of petrographic thin sections is crucial to determine rock properties such as texture, crystal size distribution, porosity and structure of rocks at a microscopic level. However, an accurate determination of minerals and textures in petrographic thin sections is a time-consuming process. Despite efficiency improvements through intensity-based segmentation, thin section petrography remains slow and manual, leaving a significant amount of sections unexamined due to time constraints. This project aims to determine whether object-based segmentation can be used to segment similar mineral phases in petrographic thin sections using the AI-driven ZEISS arivis Cloud (formerly APEER).

Samples used in this project were collected by the authors of Schmiedel et al. (2021) across a traverse of the high viscosity Sosa Dyke (37°2’S, 68°52’W) in Argentina. Ten thin sections of trachyte/trachydacite composition were scanned under bright light (BL), circularly polarized light (CPL), cross-polarized light (XPL) and plane-polarized light (PPL) using the ZEISS Axioscan 7 and provided to be used in this project.

The effects of annotation quantity, image resolution and the combination of CPL and BL images in a single model were investigated through the creation of multiple AI segmentation models using arivis Cloud. Additionally, the time spent on segmentation was compared to intensity-based segmentation, point counting, and manual segmentation. The classes used to segment images were amphiboles, plagioclase, and opaque minerals.

Results from this project show that ZEISS arivis Cloud can be used to segment thin sections using object-based AI models. Object-based segmentation with ZEISS arivis Cloud is shown to be more time-efficient than other manual and intensity-based segmentation methods. Larger datasets of similar thin sections result in greater time savings, and the platform also generates valuable data which can be used for geologic interpretation. BL images produced more accurate models than CPL images, and increased annotation of images improves model accuracy. Models trained on higher resolution images more accurately differentiated between plagioclase and amphiboles. Conversely, higher resolution models were less consistent in identifying large plagioclase crystals. Annotation bias and arivis Cloud performing optimally when objects are no larger than 320*320 could be reasons for this. Although contradictory to the recommendations of the arivis Cloud documentation, combining BL and CPL images into a single model may improve segmentation accuracy when applied to BL images.

If fully optimized, ZEISS arivis Cloud could become an important tool for analyzing thin sections that were previously deemed too time consuming, potentially unlocking new geologic insights