Artificial intelligence for the identification of taphonomic bio-accumulator agents: an actualistic test of potential faunal accumulation agency applied to Tritons Cave (Lleida, Spain)

Abstract

Studying bone surface modifications (BSMs) in neotaphonomic research is an important aid to reconstruct agency in the archaeological and palaeontological record. The significance of correctly identifying BSMs has led to extensive debates about adequate methodological and interpretive frameworks to identify taphonomic agents and their bone-modifying patterns in dynamic taphonomic processes; especially those involving inter-agency interactions. Recent analytical innovations in the field have given rise to more updated and less biasing methods, including those rooted in multivariate statistics and artificial intelligence algorithms. Among the latter, convolutional neural networks have demonstrated a substantial potential in precision in the identification and characterization of BSMs. In this study, we present a successful application of these methods to reconstructing taphonomic agency in an archaeological context, specifically in Unit 2 of Tritons Cave (Lleida, Spain). Previous interpretations of this palaeontological layer suggested that it was a natural palimpsest created by carnivores, namely leopards. Through the present research, we objectively tested the original felid-accumulation hypothesis associated with the site's original taphonomic analysis. Our findings provide additional evidence of the primary leopard agency in the formation of the assemblage and underscores the significant potential of transfer learning for identifying taxon-specific carnivore agency.

Keywords: Panthera pardus; bone surface modifications; carnivores; convolutional neural networks; taphonomy.

Figure 1.

Examples of scores (Tr1, Tr30 and Tr38) documented from Unit 2 of Tritons Cave at 30× magnification.

Figure 2.

Model accuracy (left) and loss (right) for DenseNet 201, ResNet 50 and VGG19 (top to bottom).