Semester

Summer

Date of Graduation

2021

Document Type

Thesis

Degree Type

MS

College

Eberly College of Arts and Sciences

Department

Geology and Geography

Committee Chair

Graham Andrews

Committee Member

Chris Russoniello

Committee Member

Charlie Shobe

Abstract

Traditional methods of collecting quantitative petrographic data from thin sections (modal mineralogy, size distribution, shapes, etc.) are time- and labor-intensive, and rarely have sample sizes adequate to statistically describe complex rocks (i.e. volcanic rocks). Although manual counting and measurements are now routinely supplemented by digital image analysis, the majority of quantitative petrographic studies still go through a manual digitization stage where object classes are traced before further analyses. This is a major rate-limiting step that reproduces the same problems of small n-values resulting from significant effort. We have valuated the potential and limitations of using the Trainable Weka Segmentation (TWS) plugin within the commonly used ImageJ / Fiji digital image analysis and processing environment. Specifically, we have assessed their capacity to classify, segment, and threshold user-defined petrographic features from a suite of images of progressively more complex volcanic rocks to accelerate the collection of quantitative petrographic data.

TWS uses a fast-random-forest algorithm to classify an image based on a set of training pixels selected by the user - in this case different mineral phases, vesicles, etc. Training of the classifier is intuitive and fast. For example, three classes each with eleven training spots are classified in less than 1 minute for a medium to high-resolution image. Eight plane polarized light photomicrographs with increasing crystallinity and complexity were classified (i.e. trained) and automatically segmented using TWS. Samples where the assigned classes have distinct, homogeneous RGB values and sharp boundaries are successfully classified with TWS. However, samples where the classes are heterogeneous but similar, as a result of alteration for example, are not adequately classified. Once classified, two major efficiency gains are possible: (1) the classifier can be saved and applied again to any similar sample, and (2) the segmented image is immediately available for thresholding in ImageJ / Fiji (i.e. separating into class-specific images) without manual tracing or cut-and-paste. The thresholded images can then be measured using the image analysis tools in ImageJ / Fiji (e.g., dimensions, area, circularity, long-axis orientation, etc.).

Share

COinS