The transfers of carbon between land, ocean and atmosphere, and their relation to temperature variability over glacial and interglacial cycles continue to intrigue the scientific community. Over the past four decades, many have focused on the role of the Southern Ocean to explain the atmospheric carbon dioxide (CO₂) patterns seen in ice core records, but recent advances also include mentions of a possible terrestrial component. We quantify important terrestrial organic soil carbon (C) stocks north of 23˚, using palaeo-data and modeled climate to reconstruct terrestrial C dynamics from the Last Glacial Maximum until present at millennial time steps. During the deglaciation, C storage declined to reach a minimum around 10 kyr BP, a trend which then turned and led to progressively higher soil C stocks during the Holocene. Net changes in mineral soil C stocks are small, even though significant geographic shifts occurred; instead, deglacial and interglacial terrestrial C stock dynamics are dominated by losses from permafrost loess, inundation of continental shelves and gains in peatlands, processes commonly overlooked in complex Earth System Models.

Vegetation is an important feature 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 reconstruct past and present vegetation with a data driven approach, to test if this allows us to create robust global and regional vegetation patterns. The motivation for this stems from the possibility of avoiding circular arguments when studying past time periods where vegetation is used to reconstruct climate, and climate is used to construct vegetation. By using the Random Forest machine learning tool, we train the vegetation reconstruction with available biomized pollen data of present and past conditions and are able to produce reasonable broad-scale vegetation patterns for the Pre-Industrial and the Mid-Holocene together with a few modeled climate variables. We test the methods robustness by introducing a systematic temperature bias based on existing climate model spread and compare the result with that of LPJ-GUESS, a process-based dynamic global vegetation model. Results prove that the Random Forest approach is able to produce robust patterns for periods and regions well constrained by evidence, but fails when evidence is scarce. The robustness is achieved by sacrificing a dynamic vegetation response to a changing climate.