Reconstructing climatic niches of South African Early Pleistocene hominins using rodent and bovid fossil assemblages with machine learning

Résumé En

Quaternary climatic changes have exerted a significant in uence on African landscapes and faunal communities. As this period coincides with important events in human evolution, it is assumed that climatic and environmental changes played a major role in the morphological and behavioural evolution of the human lineage. The "Cradle of Humankind" in South Africa preserves a rich sequence of fossils that document the evolution of fauna, including hominins, and environments over the last million years. Among fossil faunas, rodents and bovids are particularly abundant and can serve as useful palaeoenvironmental indicators that complement other biological and geological proxies. However, they are often treated separately in analyses due to their distinct modes of accumulation and preservation. We developed a new taxonomic method for reconstructing the climate of African Quaternary terrestrial ecosystems, enabling accurate predictions based on the composition of the rodent and bovid palaeocommunity. In the initial step, we used neo-ecological data to establish the framework and methodology for building machine learning regression models. We aggregated modern species distribution with climatic raster layers and used this dataset to predict climate parameters based on presence/absence data through random forest regression algorithms. Supra-specific taxonomic ranks, including genus, subfamily, and family, were also recorded as explanatory variables, allowing faunal lists with heterogeneous taxonomic determination to be used for predictions. To address spatial dependence in our data,we introduced a geographical block validation strategy for model validation and selection. In the subsequent step, we investigated the impact of sampling/preservation bias and taxonomic indeterminacy on palaeoenvironmental reconstructions by introducing false absences in the training data and measuring the loss of accuracy with increasing numbers of undetermined or unseen taxa. The new models, more robust to the loss of taxa, were ultimately used to infer climate conditions at various early Homo and Paranthropus robustus-bearing fossil sites from the Cradle of Humankind. The predicted climate variables allow us to compare the climatic niches of the two hominin genera and formulate hypotheses about the ecological niches they occupied.