Improving Zooarchaeological Methods for Classifying Fragmented Faunal Remains Using Differential Geometric Methods and Machine Learning
This is an abstract from the "SAA 2019: General Sessions" session, at the 84th annual meeting of the Society for American Archaeology.
Accurately identifying bone fragments and the agents that broke them is essential to site reconstruction and improving our understanding of human evolution and behavior. Here we implement geometric invariants and machine learning on digital 3D models of experimentally derived bone fragments to classify them by breakage agent. We characterize the surface with far more geometric properties (such as total volume, surface area, higher moments, torsion and surface curvatures) than used by traditional zooarchaeological methods, while also expanding our sample to include several taxa and modes of breakage. Using surface curvature, preliminary results of hominin and carnivore broken elk bones exhibit exemplary pairwise classification rates (>92% in all tests) according to skeletal element, actor of breakage and, in the case of hominin broken bones, method of breakage. Several methods of machine learning are employed including KNN, decision trees, random forests, and neural networks. These methods quickly and efficiently capture and exploit a rich amount of shape information used for evaluating competing hypotheses regarding hominin behavior, surpassing the current state-of-the-art in zooarchaeology and taphonomy. Given the success of the preliminary research, we expect that individual specimens, not simply assemblage level trends, will be classifiable to breakage agent.
Cite this Record
Improving Zooarchaeological Methods for Classifying Fragmented Faunal Remains Using Differential Geometric Methods and Machine Learning. Katrina Yezzi-Woodley, Jeff Calder, Peter Olver, Martha Tappen, Reed Coil. Presented at The 84th Annual Meeting of the Society for American Archaeology, Albuquerque, NM. 2019 ( tDAR id: 449472)
This Resource is Part of the Following Collections
Abstract Id(s): 26063