Biome classification across global vegetation models reveals consistent biome shifts under future climate change

Climate change is altering ecosystems and will reshape the global distribution of biomes. These shifts can significantly influence biodiversity, ecosystem functions and services that are essential for human livelihoods. Robust assessments of future biome dynamics are therefore urgently needed. Here, we aimed to robustly quantify and map future biome shifts under climate change using a reproducible, multi-model ensemble approach. We employed random forest models to classify outputs from five global vegetation models (GVMs) into 31 observation-based biome maps representing land cover under current climate conditions. Model-derived biome maps showed strong agreement with observation-based maps (average κ =0.77 for 31 maps), with higher agreement for biomes with well-known temperature constraints. Then we used the random forest models and GVM simulations for future climate conditions to infer future biome distributions, and we evaluated potential biome shifts for each GVM-biome map combination under three climate change scenarios (RCP2.6, RCP6.0, RCP8.5, 403 maps in total). Across all scenarios, GVMs projected biome shifts until the end of the century, where the likelihood of change increased with the level of climate change in RCP scenarios. Between 4 % and 56 % of the land surface were projected to undergo biome transitions in the full model ensemble of 403 different combinations of GVMs, RCPs and observation-based biome maps used to create biome maps. Broad spatial patterns of biome change were consistent across models. Biomes in cold regions were most susceptible to biome shifts, as boreal and temperate biomes shifted poleward, following temperature change. Equatorial rainforests remained largely stable, while other studies found forest dieback. These findings highlight regions and biomes most susceptible to future climate change, even under the low-emission scenario RCP2.6. Overall, we developed a multi-model GVM ensemble of future biome projections, based on a unified and reproducible biome classification approach. This approach allows quantification of uncertainties related to biome classification schemes across multiple GVMs and RCP scenarios, and it can be applied to any vegetation model. We provide critical insights for targeted climate mitigation and adaptation strategies and conservation of the remaining natural vegetation.