Reconstructing Past Global Vegetation With Random Forest Machine Learning, Sacrificing the Dynamic Response for Robust Results

Vegetation is an important component in the Earth system, providing a direct link between the biosphere and atmosphere. As such, a representative vegetation pattern is needed to accurately simulate climate. We attempt to model global vegetation (biomes) with a data-driven approach, to test if this allows us to create robust global and regional vegetation patterns. This not only provides quantitative reconstructions of past vegetation cover as a climate forcing, but also improves our understanding of past land cover-climate interactions which have important implications for the future. By using a Random Forest (RF) machine learning tool, we train the vegetation reconstruction with available biomized pollen data of present and past conditions to produce broad-scale vegetation patterns for the preindustrial (PI), the mid-Holocene (MH, ∼6,000 years ago), and the Last Glacial Maximum (LGM, ∼21,000 years ago). We test the method's robustness by introducing a systematic temperature bias based on existing climate model spread and compare the result with that of LPJ-GUESS, an individual-based dynamic global vegetation model. The results show that the RF approach is able to produce robust patterns for periods and regions well constrained by evidence (the PI and the MH), but fails when evidence is scarce (the LGM). The apparent robustness of this method is achieved at the cost of sacrificing the ability to model dynamic vegetation response to a changing climate