HOW MUCH OF THE PALEOECOLOGICAL RECORD HAS NO ANALOGUE AND WHAT DOES THIS MEAN FOR ENVIRONMENTAL RECONSTRUCTION?

Environmental reconstruction using paleoecological proxies is commonplace, but key uncertainties remain in understanding how these methods behave beyond modern reference conditions. Environmental reconstructions using paleo-proxies are key to capturing ecological and environmental changes at longer time scales, but most methods require dependence on modern calibration data. As paleoecological records move deep in time the possibility of no-analogue conditions increases, but reported model error may remain constant if calculated through cross-validation with the modern datasets.

In this work we show the potential impacts of no-analogue conditions on environmental reconstruction using pollen-based climate models, including machine learning techniques such as Boosted Regression Trees and Random Forests, as well as traditional methods including weighted averaging and the modern analogue technique. We show that prediction from individual pollen assemblages are strongly dependent on the analogue distance to their closest neighbors. Prediction behavior in no-analogue space is strongly dependent on the model type – the machine learning techniques appear to provide the greatest performance while WAPLS shows the worst performance of any model. We then discuss the implications of recent anthropogenic land use change in eastern North America as a potential source of uncertainty in calibrating and understanding models of environmental prediction using pollen, by illustrating the rapid rise in near-neighbor distance over the last 250 years as a result of extensive land use change, and discussing the potential impact of these changes on reconstructions in the region.