Mammalian faunas, ecological indices, and machine-learning regression for the purpose of paleoenvironment reconstruction in the Miocene of South America

Abstract

© 2019 Elsevier B.V. Reconstructing paleoenvironments has long been considered a vital component for understanding community structure of extinct organisms, as well as patterns that guide evolutionary pathways of species and higher-level taxa. Given the relative geographic and phylogenetic isolation of the South American continent throughout much of the Cenozoic, the South American fossil record presents a unique perspective of mammalian community evolution in the context of changing climates and environments. Here we focus on one line of evidence for paleoenvironment reconstruction: ecological diversity, i.e. the number and types of ecological niches filled within a given fauna. We propose a novel approach by utilizing ecological indices as predictors in two regressive modeling techniques—Random Forest (RF) and Gaussian Process Regression (GPR)—which are applied to 85 extant Central and South American localities to produce paleoecological prediction models. Faunal richness is quantified via ratios of ecologies within the mammalian communities, i.e. ecological indices, which serve as predictor variables in our models. Six climate/habitat variables were then predicted using these ecological indices: mean annual temperature (MAT), mean annual precipitation (MAP), temperature seasonality, precipitation seasonality, canopy height, and net primary productivity (NPP). Predictive accuracy of RF and GPR is markedly higher when compared to previously published methods. MAT, MAP, and temperature seasonality have the lowest predictive error. We use these models to reconstruct paleoclimatic variables in two well-sampled Miocene faunas from South America: fossiliferous layers (FL) 1–7, Santa Cruz Formation (Early Miocene), Santa Cruz Province, Argentina; and the Monkey Beds unit, Villavieja Formation (Middle Miocene) Huila, Colombia. Results suggest general concordance with published estimations of precipitation and temperature, and add information with regards to the other climate/habitat variables included here. Ultimately, we believe that RF and GPR in conjunction with ecological indices have the potential to contribute to paleoenvironment reconstruction.