Artificial intelligence in zooarchaeology: Convolutional neural networks to classify duck bones

In the last decade, the use of neural networks has accelerated the integration of Artificial Intelligence (AI) systems into archaeology. These systems assist archaeologists in fundamental yet time-consuming tasks, such as cataloguing and identifying artefacts. This study focuses on the application of Convolutional Neural Networks as a proof of concept for the systematic cataloguing of avifauna osteological remains, an especially challenging task in archaeozoology, due to the lack of systematic reference collections, morphological similarity and high taxonomic diversity of species. Differentiating species of duck is important because they exhibit diverse and highly specialised habitat requirements, which allow detailed palaeoenvironmental reconstructions. Furthermore, tracking changes in migration patterns over time provides insights into climatic and anthropogenic impacts on bird behaviour and ecology. Two neural networks were developed: the first designed to classify five types of skeletal elements to the species level and the second aimed at cataloguing 20 duck species (Anatinae, Aythyinae, Tadorninae, and Oxyurini) according to their skeletal elements. The CNNs developed in this study were trained on 1833 photographs, achieving 98% top-1 and 72% top-3 accuracy for the classification of element type and species, respectively. Overall, this work demonstrated that, even for especially challenging tasks, the integration of AI in archaeozoology expedites cataloguing processes.