Theses and Dissertations

ORCID

0009-0008-8304-1911

Advisor

Oewns, Frank

Committee Member

Costa, Adriana

Committee Member

Wiedenhoeft, Alex

Committee Member

Shmulsky, Rubin

Date of Degree

8-12-2026

Original embargo terms

Visible MSU Only 6 months

Document Type

Graduate Thesis - Campus Access Only

Major

Sustainable Bioproducts

Degree Name

Master of Science (M.S.)

College

College of Forest Resources

Department

Department of Sustainable Bioproducts

Abstract

Building on previous work in anatomical feature detection, this study developed deep-learning models for detecting vessels and rays in macroscopic images of Peruvian hardwoods. U-Net-based segmentation models were trained using 115 annotated XyloTron images representing 23 kinds of wood and evaluated on 46 test images. Probability maps were converted into binary masks using thresholds of 0.1–0.9. The vessel model achieved its highest mean IoU of 0.785 and smallest mean absolute error in area of approximately 10% at the threshold of 0.3. The mean F1 score across all test images was 0.880. In contrast, the ray model achieved its highest mean IoU of 0.477 at the threshold of 0.4 and its smallest mean absolute error in area at approximately 28% at the threshold of 0.6. These results led to several proposals for improving the performance of both models including making changes to model training, anatomical feature annotation, and model architecture.

Sponsorship (Optional)

U.S. Department of Agriculture, Agricultural Research Service (USDA ARS); U.S. Department of Agriculture, National Institute of Food and Agriculture (USDA NIFA)

Download

Share

COinS