Seminar in applied mathematics and statistics
SPEAKER: Bo Christiansen (Research and Development, Danish Meteorological Institute).
TITLE: The challenges of temperature reconstructions of the past two millennia.
ABSTRACT: Knowledge of the temperature variability during the last 1-2 millennia is important for, e.g., providing a perspective to present-day climate variations. Since systematic instrumental temperature records only extend back to the 19th century such knowledge mainly relies on climate-sensitive proxy data such as tree-ring width. Reconstruction methods assume a relationship between proxies and temperature which is calibrated using the recent period of overlap between proxies and instrumental temperature data.
Previous large-scale reconstructions are dominated by variants of (multiple and multivariate) linear regression models combined with some kind of regularization scheme. Here we critically review such reconstructions and assess some of the many related challenges. Proxy records often only have a weak and unstable climate signal. The temperature field has a complicated spatial covariance structure which depends strongly on the considered time-scales and which may also be non-stationary. The consequences of this are enhanced by the fact that the network of proxy records is sparse and spatially heterogeneous. Also, proxy records can have coarse temporal resolution, temporal correlations, and dating uncertainties. Another problem is the brief period of overlap between proxies and instrumental observations which furthermore is dominated by a strong warming trend.
We also discuss how the choice of an unphysical regression model may result in an underestimation of low-frequency variability in the reconstructions. The proxy-climate relationship is probably not linear in reality as generally assumed. Recently we have been able to employ non-linear methods for probing the proxy-climate relationship -- so called Generalized Additive Methods (GAM). We will describe a recent application of GAM in the field of climate reconstructions from proxies in Europe.
Although, the potential extension of the above linear and non-linear methods using machine-learning methods seems evident we are not aware of any such attempts.
This is joint work with Peter Thejll.
----
Upcoming events (after October 2):
Wednesday, November 13 at 15.15: Helene Rytgaard
Wednesday, November 20 at 15.15: Sascha Desmettre
Wednesday, November 27 at 15.15: Jesper Møller
Wednesday, February 19 at 15.15: Budhi Surya