Abstract
We test several quantitative algorithms as palaeoclimate reconstruction tools for North American and European fossil pollen data, using both classical methods and newer machine-learning approaches based on regression tree ensembles and artificial neural networks. We focus on the reconstruction of secondary climate variables (here, January temperature and annual water balance), as their comparatively small ecological influence compared to the primary variable (July temperature) presents special challenges to palaeo-reconstructions. We test the pollen-climate models using a novel and comprehensive cross-validation approach, running a series of h-block cross-validations using h values of 100-1500 km. Our study illustrates major benefits of this variable h-block cross-validation scheme, as the effect of spatial autocorrelation is minimized, while the cross-validations with increasing h values can reveal instabilities in the calibration model and approximate challenges faced in palaeo-reconstructions with poor modern analogues. We achieve well-performing calibration models for both primary and secondary climate variables, with boosted regression trees providing the overall most robust performance, while the palaeoclimate reconstructions from fossil datasets show major independent features for the primary and secondary variables. Our results suggest that with careful variable selection and consideration of ecological processes, robust reconstruction of both primary and secondary climate variables is possible.