May 2026 – Palevoprim

Références

Al Riaydh, M. H., Merceron, G., & Lehmann, T. (2026). Decoding As-Sahabi’s bovid postcranial fossils: A machine learning approach to palaeoecology and taxonomy. Palaeobiodiversity and Palaeoenvironments, 1-38. DOI: 10.1007/s12549-026-00704-6

Decoding As-Sahabi’s bovid postcranial fossils: A machine learning approach to palaeoecology and taxonomy

The Late Miocene-Pliocene fossil site of As-Sahabi in northeastern Libya contains a significant but understudied collection of bovid postcranial elements. These remains provide crucial evidence for understanding palaeoenvironmental conditions during an important phase of climatic transition in North Africa. Our study employed supervised machine learning approaches, to analyse 230 postcranial specimens.

The analysed material includes 159 astragali and 58 proximal phalanges from the stratigraphic Unit U1 (dated to ~6.8 Ma), which we classified to the subfamily level to reconstruct their palaeohabitat associations based on a comparative dataset of an extant African Bovidae. The models are trained using the comparative dataset and then choose the highest achieving Machine Learning model to be applied on our data. The models achieved near perfect classification separation (Mean Area Under the Receiver Operating Characteristic Curve) 90-95% at the taxonomic level and 98-99% for habitat prediction. Results reveal a bovid community dominated by Antilopinae (59%), followed by Bovinae (20%), Reduncinae (8%), Hippotraginae (7%), and Alcelaphinae (6%). Habitat reconstructions indicate a heterogenous ecosystem comprising predominantly open environments (48%), with lesser representation from lightly wooded (29%) and forested areas (20%), supporting a mosaic palaeo-landscape contradicting prior studies that suggested the absence of forest-dwelling bovids.

Body mass estimates show niche partitioning, with forest-dwelling Bovinae (mean= 49 kg, SD=27.3 kg) being markedly smaller than open-habitat taxa such as Hippotraginae (mean= 127 kg, SD=121 kg). These findings validate postcranial ecomorphology as a robust proxy for palaeoenvironmental inference and underscore the potential of machine learning to improve taxonomic resolution in fragmentary fossil datasets.

This integrative framework establishes a scalable, replicable approach for palaeoecological reconstruction at taphonomically biased sites, enhancing our understanding of mammalian evolution and environmental change in the Afro-Arabian fossil record during the Late Miocene.