Assessing performance and seasonal bias of pollen-based climate reconstructions in a perfect model world

With transfer functions (numerical methods for estimating environmental conditions from species assemblages), it is trivially easy to make reconstructions of many different climate variables. If you want to reconstruct December sunshine from pollen assemblages, simply train your model with December sunshine and it will give you a reconstruction. And why stop at one variable when you can reconstruct as many as you want.

Whether the reconstructions are any good is an entirely different question that is too often overlooked. Of course, people look at the cross-validation performance statistics, but several biases can make these over-optimistic. In a few cases, we can compare reconstructions with instrumental data, but generally we have to rely on various reconstruction diagnostics and reconstruction significance tests, but we don’t know how helpful these are.

In an ideal world, we would be able to asses how well these diagnostics perform if reconstructions that pass the tests are more accurate than those that fail. The real world offers few opportunities for this, but model worlds do. Some global climate models have an interactive vegetation module – the vegetation both responds to and influences climate. Such models have been run over the Holocene to simulate Holocene climate and vegetation dynamics. In a paper now in open review at Climate of the Past Discussions, Kira Rehfeld and coauthors use one such model world to explore the behaviour of transfer function reconstructions. We train a transfer function on the modern model vegetation and then make climate reconstructions from the model’s Holocene vegetation which we compare with the model’s Holocene climate.

We find that cross-validation performance statistics are not a very useful guide as to which climate variables can be reconstructed (this should not surprise anybody who has thought much about transfer functions), but tests of how much variance in the fossil data is explained by the reconstruction have utility. We also find (again not surprisingly really), that reconstructions of climate variables that are not very ecologically important can be severely biased by changes in ecologically important variables.

This work gives some guidance about which climate variables can be reconstructed (and how to check this) and so should aid palaeoclimate data-model comparisons by helping us to focus on the reconstructions that we most trust.

Here is the abstract:

Reconstructions of summer, winter or annual mean temperatures based on the species composition of bio-indicators such as pollen, foraminifera or chironomids are routinely used in climate model-proxy data comparison studies. Most reconstruction algorithms exploit the joint distribution of modern spatial climate and species distribution for the development of the reconstructions. They rely on the assumption of “uniformitarianism”, which implies that environmental variables other than those reconstructed should not be important, or that their relationship with the reconstructed variable(s) should be the same in the past as in the modern spatial calibration dataset.

Here we test the implications of uniformitarianism on such reconstructions in an ideal model world, in which climate and vegetation are known at all times. The alternate reality is a climate simulation of last 6000 years with dynamic vegetation. Transient changes of plant functional types are considered as surrogate pollen counts, and allow to establish, apply and evaluate transfer functions in the modeled world.

We find that the transfer function cross-validation r2 is of limited use to identify reconstructible climate variables, as it only relies on the modern spatial climate/vegetation relationship. However, ordination approaches that assess the amount of fossil vegetation variance explained by the reconstructions are promising. We furthermore show that correlations between climate variables in the modern climate/vegetation relationship are systematically extended into the reconstructions. Summer temperatures, the most prominent driving variable for modelled vegetation change in the Northern Hemisphere, are accurately reconstructed. However, the amplitude of the winter and mean annual temperature cooling between the mid-Holocene and present day is overestimated, and similar to the summer trend in magnitude.

This effect occurs, because temporal changes of a dominant climate variable, such as summer temperature, are imprinted on a less important variable, leading to reconstructions biased towards the dominant variable’s trends. Our results indicate that reconstructions of multiple climate variables from the same bio-indicator dataset should be treated with caution. Expert knowledge on the eco-physiological drivers of the proxies, and statistical methods that go beyond the cross-validation on modern calibration datasets are crucial to avoid misinterpretation.

Rehfeld, K., Trachsel, M., Telford, R., and Laepple, T.: Assessing performance and seasonal bias of pollen-based climate reconstructions in a perfect model world, Clim. Past Discuss., doi:10.5194/cp-2016-13, in review, 2016.

About richard telford

Ecologist with interests in quantitative methods and palaeoenvironments
This entry was posted in transfer function and tagged , . Bookmark the permalink.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s