FROG: A global machine-learning temperature calibration for branched GDGTs in soils and peats

wBranched glycerol dialkyl glycerol tetraethers (brGDGTs) are a family of bacterial lipids which have emerged over time as robust temperature and pH paleoproxies in continental settings. Nevertheless, it was previously shown that other parameters than temperature and pH, such as soil moisture, thermal regime or vegetation can also influence the relative distribution of brGDGTs in soils and peats. This can explain a large part of the residual scatter in the global brGDGT calibrations with mean annual air temperature (MAAT) and pH in these settings. Despite improvements in brGDGT analytical methods and development of refined models, the root-mean-square error (RMSE) associated with global calibrations between brGDGT distribution and MAAT in soils and peats remains high (similar to 5 degrees C). The aim of the present study was to develop a new global terrestrial brGDGT temperature calibration from a worldwide extended dataset (i.e. 775 soil and peat samples, i.e. 112 samples added to the previously available global calibration) using a machine learning algorithm. Statistical analyses highlighted five clusters with different effects of potential confounding factors in addition to MAAT on the relative abundances of brGDGTs. The results also revealed the limitations of using a single index and a simple linear regression model to capture the response of brGDGTs to temperature changes. A new improved calibration based on a random forest algorithm was thus proposed, the so-called random Forest Regression for PaleOMAAT using brGDGTs (FROG). This multi-factorial and non-parametric model allows to overcome the use of a single index, and to be more representative of the environmental complexity by taking into account the non-linear relationships between MAAT and the relative abundances of the individual brGDGTs. The FROG model represents a refined brGDGT temperature calibration (R-2 = 0.8; RMSE = 4.01 degrees C) for soils and peats, more robust and accurate than previous global soil calibrations while being proposed on an extended dataset. This novel improved calibration was further applied and validated on two paleo archives covering the last 110 kyr and the Pliocene, respectively. (C) 2021 Elsevier Ltd. All rights reserved.

Automated summary

The study developed a new global terrestrial brGDGT temperature calibration using a machine learning algorithm. This improved model is more accurate and robust than previous global soil calibrations, and takes into account non-linear relationships to better represent environmental complexity.