Testing the reliability of geometric morphometric and computer vision methods to identify carnivore agency using Bi-Dimensional information

Abstract

Bidimensional information of tooth marks and other bone surface modifications (BSM) presents limitations, as highlighted in this study. Here, we establish a methodological comparison on a controlled experimentally-derived set of BSM generated by four different types of carnivores, using geometric morphometric (GMM) and computer vision (CV) methods. We highlight that previous generalizations of high accuracy on tooth marks using GMM are heuristically incomplete, because only a small range of allometrically-conditioned tooth pits have been used. Biased replication and exclusion of the most widely represented forms of non-oval tooth pits from such analyses have compromised the published results and their ensuing generalizations. Here, we document bidimensionally a much wider range of tooth pits, using their outlines (and not a limited set of non-reproducible idem locus semi-landmarks), through Fourier analyses. The resulting tooth mark sets show low accuracy (and resolution) in the classification of tooth marks per carnivore modifying agent. This low resolution is also reproduced when using a semi-landmark approach. In contrast, our study demonstrates that CV approaches, through Deep Learning (DL), using convolutional neural networks (DCNN), and Few-Shot Learning (FSL) models, classify experimental tooth pits with 81% and 79.52% accuracy, respectively, being equally efficient at classification. However, a limitation in CV methods occurs when applied to the fossil record, as BSM undergo dynamic transformations over time. The most impactful processes occur early in taphonomic history, altering the original BSM properties. Consequently, no objective referents exist for marks combining original and subsequent diagenetically or biostratinomically modifying processes. However, in well-preserved contexts, such as the 1.8 Ma tooth marks from some of the Olduvai sites, confidence in interpretations can be high with convergent CV models indicating high agent attribution probability. While GMM shows potential in 3D, its current bidimensional application yields limited discriminant power (<40%). Thus, future research should utilize complete 3D topographical information for more complex GMM and CV analyses, potentially resolving current interpretive challenges. Despite necessary cautions, these new methods offer an unprecedented objective means of classifying BSM to taxon-specific agency with confidence indicators. Continued research should refine these approaches, enhancing the reliability of prehistoric interpretations.