Evaluating climate field reconstruction techniques using improved emulations of real-world conditions

Pseudoproxy experiments (PPEs) have become an important framework for evaluating paleoclimate reconstruction methods. Most existing PPE studies assume constant proxy availability through time and uniform proxy quality across the pseudoproxy network. Real multiproxy networks are, however, marked by pronounced disparities in proxy quality, and a steep decline in proxy availability back in time, either of which may have large effects on reconstruction skill. A suite of PPEs constructed from a millennium-length general circulation model (GCM) simulation is thus designed to mimic these various real-world characteristics. The new pseudoproxy network is used to evaluate four climate field reconstruction (CFR) techniques: truncated total least squares embedded within the regularized EM (expectationmaximization) algorithm (RegEM-TTLS), the Mann et al. (2009) implementation of RegEM-TTLS (M09), canonical correlation analysis (CCA), and Gaussian graphical models embedded within RegEM (GraphEM). Each method’s risk properties are also assessed via a 100-member noise ensemble. Contrary to expectation, it is found that reconstruction skill does not vary monotonically with proxy availability, but also is a function of the type and amplitude of climate variability (forced events vs. internal variability). The use of realistic spatiotemporal pseudoproxy characteristics also exposes large inter-method differences. Despite the comparable fidelity in reconstructing the global mean temperature, spatial skill varies considerably between CFR techniques. Both GraphEM and CCA efficiently exploit teleconnections, and produce consistent reconstructions across the ensemble. RegEM-TTLS and M09 appear advantageous for reconstructions on highly noisy data, but are subject to larger stochastic variations across different realizations of pseudoproxy noise. Results collectively highlight the importance of designing realistic pseudoproxy networks and implementing multiple noise realizations of PPEs. The results also underscore the difficulty in finding the proper bias-variance tradeoff for jointly optimizing the spatial skill of CFRs and the fidelity of the global mean reconstructions.