Detecting paleoclimate regime transitions: A novel approach
Most timeseries analysis techniques are predicated on assumptions of linearity. That works great on linear systems of course, but it’s rather well known that the climate system is nonlinear, so linear methods are inherently fraught. For some years now, nonlinear timeseries analysis techniques have begun to claim impressive accomplishments in detection and prediction tasks. They’ve been used to detect causality in complex dynamical systems (Sugihara et al. 2012), understand phase transitions in fluid dynamics (Gorski et al. 2015), identify early warning signs of risk during pregnancy (Ramirez-Avila et al. 2013), and much more. Notably, recurrence quantification analysis (RQA) has emerged as a robust way to handle all kinds of nonlinear signals, including climate ones.
These methods sound great on paper, but many methods that work well in other domains fail on the noisy, sparse and age-uncertain timeseries that are so common in paleoclimatology. We were curious as to whether any of these techniques would stand up to thorough stress-testing on data such as this.
To properly test an RQA based approach to paleoclimate analysis, we needed to find a method that was well suited to noisy, sparse timeseries data. A paper by Nishant Malik (RIT) describing a new method using Recurrence Plots to detect dynamical changes in time series data caught our attention. We call this technique Laplacian Eigenmaps for Recurrence Matrices (LERM). We decided to implement his method in a robust Python framework, and evaluate it on real and synthetic paleoclimate data. We targeted both abrupt and gradual transitions, using as our paragon the 8.2 ka event and the Mid-Pleistocene Transition, respectively.
LERM proved capable of identifying both types of transitions, even in fairly noisy data. However, it requires careful handling, as it does have a propensity for false positives, and is sensitive to uncertainties in the time axis (poor time control, etc.). That being said, it shows promise as a useful method within the analysis toolkit of paleoclimatologists, and suggests the untapped potential of other RQA based techniques for the study of paleoclimate (see Boers 2018, Westerhold et al. 2020, Donges et al. 2011, Marwan et al. 2021). The systematic work we did to document LERM’s strength and weaknesses should help automate the detection of dynamical transitions in all manner of paleoscientific series, including paleoclimatology, paleoceanography, and paleoecology.
For a more detailed account of this technique, its strengths and weaknesses, and the results it produced, please check out our recently published paper. To facilitate the community adoption of this technique, we developed an accompanying Python package, Ammonyte . This package is built on top of Pyleoclim, a Python package for the analysis of paleoclimate data. We hope that this infrastructure will be used to help answer longstanding paleo questions, and that Ammonyte will make it easy for anyone with access to the internet (and a half-way decent CPU) to do so.
The work was supported by the NSF P2C2 project as part of grant 2002556, which aimed to use data science tools to explore fundamental paleoclimate questions, and to democratize these tools via training events. In this context, LERM is now being applied to identifying hydroclimate regime transitions over the Holocene. Watch this space for how cyber-paleo tools may help answer longstanding questions about the causes of civilizational collapse around 4,200 years ago (hint: it probably wasn’t the dreamstone).
