A critical assessment of lacustrine branched glycerol dialkyl glycerol tetraether (brGDGT) temperature calibration models

Abstract

Branched glycerol dialkyl glycerol tetraethers (brGDGTs) are membrane-spanning lipids produced by bacteria that are ubiquitous in natural sedimentary archives and preserved over geologic timescales. The main influence on their distributions in the environment appears to be temperature, thus making them a potentially powerful proxy for paleotemperature reconstruction. Application of recent lacustrine brGDGT-based temperature calibration models to specific regions results in inaccurate reconstructed temperatures suggesting that regional or site-specific temperature calibration models may be necessary.Using an extended data set of 692 lake sediment samples from across the globe we determined whether brGDGT distributions in samples from the same regions or sites are significantly different from one another via hierarchical agglomerative clustering analysis (HAC). Results of HAC analysis showed four significant clusters with varying geographic distributions. Cluster 1 samples are mainly located at high latitudes (Mean Annual Air Temperature; MAAT = 3.10 ± 5.91 °C). Cluster 2 samples are concentrated in the Tibetan Plateau (MAAT = 1.54 ± 5.91 °C). Cluster 3 samples span temperate-tropical latitudes (MAAT = 17.26 ± 8.16 °C). Cluster 4 samples are mainly located in Central and South America (MAAT = 24.56 ± 4.01 °C). The clustering led us to develop random forest regression models to predict temperature (MAAT and Months Above Freezing, MAF, air temperature) based on samples within each cluster (cluster-specific temperature models). Model performance was the highest for Cluster 3 (MAF: R2 = 0.78, RMSE = 2.85 °C, n = 261; MAAT: R2 = 0.76, RMSE = 4.07 °C, n = 270), followed by Cluster 1 (MAF: R2 = 0.54, RMSE = 1.67 °C; n = 219; MAAT: R2 = 0.62, RMSE = 3.62 °C, n = 226) and Cluster 4 (MAF: R2 = 0.58, RMSE = 2.46, n = 67; MAAT: R2 = 0.59, RMSE = 2.52 °C, n = 67). The Cluster 2 model had the lowest model performance (MAF: R2 = 0.42, RMSE = 4.51, n = 38; MAAT: R2 = 0.51, RMSE = 1.74 °C, n = 129).We also developed a random forest classification model to predict the cluster assignment for new samples (cluster prediction model with an overall accuracy of 95%), which informs the user as to which cluster-specific temperature model(s) to apply to their samples. Finally, we applied our approach (i.e., cluster assignment followed by cluster-specific temperature reconstruction) to seven published lacustrine (paleo)records and illustrate pitfalls among the temperature reconstructions from brGDGT-based temperature calibration models. Overall, our study defines broad geographic relationships among lacustrine brGDGT distributions and air temperature while underscoring model limitations for paleotemperature reconstruction and subsequent interpretation.