Comparing different calibration methods (WA/WA-PLS regression and Bayesian modelling) and different-sized calibration sets in pollen-based quantitative climate reconstruction

Abstract

We compare a Bayesian modelling-based technique with weighted averaging (WA) and weighted averaging-partial least squares (WA-PLS) regression in pollen-based summer temperature transfer function calibration. We test the methods using a new, 113-sample calibration set from Estonia, Lithuania and European Russia, and a Holocene fossil pollen sequence from Lake Kharinei, a previously studied lake in northeast European Russia. We find WA-PLS to outperform WA, probably because of smaller edge-effect biases in the ends of the calibration set gradient. The Bayesian-based calibration models show further improved performance compared with WA-PLS in leave-one-out cross-validation, while additional h-block cross-validation shows the Bayesian method to be little affected by spatial autocorrelation. Comparison with independent climate proxies reveals, however, some clear biases in the Bayesian palaeotemperature reconstructions, likely reflecting in part some specific limitations of our calibration set. As the selected prior parameters can significantly affect both Bayesian cross-validation performance and reconstructions, there is a clear need to further test the Bayesian method in different geographic contexts and over different timescales, with special attention given to the selection of the most realistic priors in each situation. In general, our finding that statistically well-performing transfer functions may produce clearly differing palaeotemperature reconstructions urges caution in transfer functionbased inferences. We additionally test a spatially restricted, 58-sample subset of the full 113-sample calibration set. We find some reduced biases with the smaller set, likely because of complex, partially bimodal responses of several taxa along the longer temperature gradient, ill-suited for calibration methods assuming unimodal responses to climate.

Major cooling intersecting peak Eemian Interglacial warmth in northern Europe

“…The calibration of the WA transfer functions was based on previously published calibration datasets on lake surface sediment pollen and chironomid assemblages: for pollen, the "central" calibration dataset used in Salonen et al (2013) (for further details about these samples see Sepp€ a et al, 2004;Bjune et al, 2010;Salonen et al, 2012), and for chironomids, the Norwegian calibration dataset described in Brooks and Birks (2001) and Self et al (2011). The performance of the transfer functions was tested using leave-one-out cross-validation, suggesting a rootmean-square error of prediction (RMSEP) of 0.86 C for the pollenbased inference model and 1.03 C for the chironomid-based inference model.…”

A Bayesian multinomial regression model for palaeoclimate reconstruction with time uncertainty

“…is the likelihood of the pollen data. Here, assuming conditional independence of training lakes and core sediment samples given temperatures and parameters, the likelihoods of Y f D h y Vasko et al (2000), Holmström et al (2015a), Salonen et al (2012), and Li et al (2015). In the present article, the priors for the Gaussian response model parameters are chosen as in Holmström et al (2015a) and Salonen et al (2012).…”

“…Salonen et al. () and Holmström et al. () applied modifications and extensions of Bummer to reconstruct temperature from pollen records, and Li et al.…”

“…Bummer was further discussed and modified in Erästö and Holmström (2006). Salonen et al (2012) andHolmström et al (2015a) applied modifications and extensions of Bummer to reconstruct temperature from pollen records, and Li et al (2015) used pollen to reconstruct past precipitation.In recent years, application of Bayesian modeling in paleoclimate reconstruction has become increasingly popular. One major advantage in the Bayesian approach is that many different sources of uncertainty can be relatively easily added to the model.…”

“…Bummer was further discussed and modified in Erästö and Holmström (2006). Salonen et al (2012) andHolmström et al (2015a) applied modifications and extensions of Bummer to reconstruct temperature from pollen records, and Li et al (2015) used pollen to reconstruct past precipitation.…”