A MACHINE LEARNING APPROACH FOR INFERRING PAST VEGETATION CHANGES THROUGH TIME AND SPACE

Geological and paleontological evidence have long been used to infer paleovegetation changes, which are crucial to understanding the evolution of environments and ecosystems. Reconstructions of paleovegetation are typically developed for specific sites, using various data sources, including phytoliths, pollen, and fossil assemblages. These reconstructions typically represent specific points in space and time and their extrapolation to larger spatiotemporal scales is non-trivial due to the patchy and discontinuous nature of individual data sources. Here we develop a new model based on deep learning (an Artifical Intelligence method) to produce temporally and spatially continuous maps of past vegetation based on multiple data sources. We train the model on existing paleovegetation reconstructions from individual sites as well as modern vegetation, using a neural network to learn the associations between the vegetation at a given site and multiple biotic and abiotic factors. These factors include mammal fossil occurrences, plant macrofossils, temperature and precipitation from models and measurements, latitude, and the effects of spatial and temporal autocorrelation. The trained model can be used to infer the most likely vegetation for any given point in space and time. We demonstrate this full-evidence machine learning approach by inferring paleovegetation maps of North America throughout the past 30 million years. Our reconstructions track the origination and expansion of open habitats in North America and showcase the potential of machine learning in developing robust hypotheses about the dynamics of paleoenvironmental changes.