Biology subjects

Pung, A. J.

Publications and source records attributed to Pung, A. J..

2 recordsLinked to original sources

bioRxiv2025

Dental Age Assessment of White-tailed Deer via Computer Vision and Deep Learning

Accurate age estimation of white-tailed deer remains challenging for wildlife management, with existing computer vision methods limited to trail camera imagery and manual dental analysis requiring specialized expertise. This research presents the first computer vision approach to dental age assessment of white-tailed deer, eliminating the need for manual tooth replacement and wear evaluation. A dataset of 243 jawbone images is used to develop a multi-fold Con-volutional Neural Network ensemble, achieving 90.7% {+/-} 2.6% accuracy through 5-fold cross-validation. Attention map analysis confirms the model locates and focuses exclusively on dental features rather than imaging artifacts, successfully identifying the same wear patterns and eruption sequences that wildlife biologists rely upon in manual assessments. This implementation substantially outperforms computer vision models based on trail camera imagery and meets the accuracy standards required for scientific research and population management. The automated approach developed in this study enables rapid, objective age determination from dental specimens, dramatically reducing analysis time while offering immediate practical value for wildlife agencies, research institutions, and hunting programs requiring precise age data from harvested deer.

ecology↗

bioRxiv2025

Age Classification of White-tailed Deer Via Computer Vision and Deep Learning

Accurate age estimation of wild whitetail deer remains a significant challenge for wildlife management. This study presents the first application of computer vision to whitetail buck age estimation using trail camera imagery, evaluating over sixty classification algorithms from traditional machine learning to advanced deep learning techniques. Our approach utilizes transfer learning and CNN ensembles to achieve a breakthrough cross-validation accuracy of 76.7 {+/-} 5.9%, substantially outperforming traditional classifiers (57%), human expert assessment (60.6%), and morphometric methods (63%), and surpassing the 70% accuracy threshold required by professionals for wildlife management decisions. Furthermore, attention map analysis of the ResNet-18 ensemble reveals that the model learns to focus on the same anatomical features (neck, chest, and stomach regions) that human experts rely upon for age assessment. This biological validation demonstrates that the CNN identifies genuine age-related morphological changes rather than spurious correlations, lending credibility to its predictions. This breakthrough offers wildlife professionals a practical tool to dramatically reduce manual age assessment workload while exceeding current accuracy standards, potentially transforming how deer populations are monitored and managed across North America.

ecology↗