Theses and Dissertations
ORCID
https://orcid.org/0000-0003-0758-6406
Advisor
Owens, Frank
Committee Member
Costa, Adriana
Committee Member
Wiedenhoeft, Alex
Committee Member
Shmulsky, Rubin
Date of Degree
5-15-2026
Original embargo terms
Embargo 2 years
Document Type
Dissertation - Open Access
Major
Forest Resources (Sustainable Bioproducts)
Degree Name
Doctor of Philosophy (Ph.D.)
College
College of Forest Resources
Department
Department of Sustainable Bioproducts
Abstract
Macroscopic digital images captured from the transverse surfaces of wood specimens with the XyloTron system provide sufficient anatomical detail for human-based macroscopic identification and are useful for training computer-vision wood identification (CVWID) models. To taxonomically classify a wood specimen, a wood anatomist identifies anatomical features in the image — primarily growth rings, vessels, rays, and parenchyma — and draws conclusions based on the patterns they exhibit. In contrast, image-level CVWID models classify woods by extracting pixel patterns from the image that do not necessarily correspond directly to anatomical features. As a first step toward assigning anatomically meaningful labels to these pixel patterns, this study developed a series of computer-vision wood feature detection (CVWFD) models to delineate growth ring boundaries (GRB) in XyloTron images. The objectives of the study are threefold: 1) to train three multiscale deep supervision edge detection models to delineate GRB in North American softwoods, diffuse-porous hardwoods, and ring-porous hardwoods, 2) to test the models’ ability to detect GRB across those different wood types, and 3) to evaluate the accuracy of the predicted GRB delineations. The results showed 100% true-positive ring detection for the softwood and ring-porous models and 98.5% for the diffuse-porous model. While the softwood model predicted 8.1% false positives, both hardwood models were perfect. This dissertation comprises four chapters on the following topics: an executive summary (Chapter 1); a softwood ring detection model trained and tested on 823 XyloTron images from 40 species (Chapter 2); a diffuse-porous model trained and tested on 322 XyloTron images from 23 species (Chapter 3); and a ring-porous model trained and tested on 496 XyloTron images from 33 species (Chapter 4). By demonstrating consistent GRB detection across three major wood types, this study lays the groundwork for universal GRB detection in any type of wood.
Sponsorship (Optional)
National Institute of Food and Agriculture (NIFA), U.S. Department of Agriculture, McIntire-Stennis project under accession number 7004014.
Recommended Citation
Ebeheakey, Alberta Asi, "Deep learning models for growth ring boundary detection in macroscopic images of North American woods" (2026). Theses and Dissertations. 6892.
https://scholarsjunction.msstate.edu/td/6892
