Reconstructing the palaeoclimate of the Cradle of Humankind in South Africa over the last 3.5 million years using machine learning and mammalian fossil assemblages

Hypotheses linking climate change to hominin evolution are based on the observation that many events of speciation, extinction, and morphological and behavioural evolution coincided with periods of increased climatic variability and habitat instability. Fossil fauna recovered from palaeoanthropological sites can be used to reconstruct past climatic and environmental conditions associated with hominin assemblages and to explore correlative changes through time. However, temporal correlations between global climate change patterns and local environmental proxies from hominin-bearing deposits are not always verifiable. This study explores the use of machine learning algorithms for faunal-based palaeoenvironmental reconstructions in the Cradle of Humankind (CoH) over the past 3.5 million years. We develop a new method based on random forest models using the presence/absence data of rodents and bovids, two clades commonly employed in palaeoenvironmental reconstructions, to assess the evolution of six temperature and precipitation variables in the region. In contrast to continental-scale trends, we found no evidence for a gradual transition toward more arid conditions. Instead, all reconstructions indicate lower precipitation levels accompanied by more moderate interannual precipitation variability in the past. Although interpretations differ slightly when faunal indicators are analysed separately, the combined evidence from rodents and bovids shows that habitats, despite considerable climatic variability through time, consistently occupied the ecotone between grassland and savanna, closely resembling present-day conditions in the CoH. These findings challenge hypotheses that propose a progressive shift from closed, mesic environments to open, arid landscapes during the Pliocene and Pleistocene in the region. • Palaeoclimatic reconstructions from faunal remains using machine learning (South Africa, Cradle of Humankind). • Reconstructed temperatures and precipitation over the last 3.5 Ma show no progressive shift toward aridification. • Annual precipitation would have been lower than today. • Environments may have remained relatively stable over time, with most sites at the ecotone between grassland and open savannah.