Semi-automatic segmentation of petrographic thin section images using a seeded-region growing algorithm with an application to characterize wheathered subarkose sandstone

Accurate imaging of minerals in petrographic thin sections using (semi)-automatic image segmentation techniques remains a challenging task chiefly due to the optical similarity of adjacent grains or grain aggregates rendering definition of grain boundaries difficult. We present a new semi-automatic image segmentation workflow for the quantitative analysis of microscopic grain fabrics. The workflow uses an automated seeded region growing algorithm, which is based on variance analysis of five or more RGB images. The workflow is implemented in the open-source Geographic Information System (GIS) software SAGA (System for Automated Geoscientific Analyses). SAGA provides all required tools for image analysis and geographic referencing. It also features a graphical user interface that allows the user to simultaneously display and link multiple images and, thus, facilitates manual post-processing of the images. SAGA's capabilities for vector data analysis offer instant calculation and visualization of the compiled geometric database within a GIS environment. Specifically, grain contacts are automatically identified by lines of intersection and manually classified by contact type to characterize the mineral fabric of petrographic thin sections. To demonstrate the effectiveness of the workflow, 39 transmitted light images of 13 weathered sandstone samples of the Buntsandstein Formation in northwestern Germany were analyzed. Based on the segmentation results obtained from the samples, a number of parameters, including modal composition, geometry of grain contacts, porosity, and grain size distribution were determined and statistically evaluated. The results of the image analysis are utilized to assess the weathering susceptibility of the sandstone samples and point to the importance of cementation determining the geotechnical properties of a given sandstone sample. (C) 2015 Elsevier Ltd. All rights reserved.