Identification of Zooarchaeological Assemblages using Image Classification of Small Skeletal Remains from Ardales Cave, Malaga, Spain

Carmina Baylon1*, Chara Punzal1, Patricia Cabrera2, José Ramos-Muñoz3, Gerd-Christian Weniger4, Pedro Cantalejo Duarte5, Juan Rofes2,6, and Giovanni Tapang1,7

Zooarchaeological assemblages from cave sites can yield tens of thousands of small skeletal remains, making manual taxonomic identification by experts both time-consuming and difficult to scale. This paper presents an image-based machine learning pipeline for classifying small mammal bones into taxonomic orders — Rodentia, Lagomorpha, Chiroptera, and Eulipotyphla — using specimens from Ardales, Malaga, Spain, a cave with Paleolithic rock art, used by both Neanderthals and, later, modern humans as a sacred space for 58,000 years. Digital photographs of each bone were taken and preprocessed with grayscale conversion, inverted binary thresholding, masking, connected component analysis, and morphological operations to isolate individual bone elements. The labeled training set includes 1,716 bones (Rodentia: 1393, Eulipotyphla: 37, Chiroptera: 32, Lagomorpha: 254). The class imbalance was addressed using SMOTE and data augmentation prior to model training. The unlabeled test set contains 2,355 images. The classification methods evaluated include Convolutional Neural Networks (CNNs), Support Vector Machines (SVMs), and other benchmark classifiers. Results demonstrate the feasibility of automated order-level classification even in fragmentary assemblages, offering a reproducible, scalable supplement to expert morphological analysis for large zooarchaeological collections.