Skill (or lack thereof) of data-model fusion techniques to provide an early warning signal for an approaching tipping point

Many coupled human-natural systems have the potential to exhibit a highly nonlinear threshold response to external forcings resulting in fast transitions to undesirable states (such as eutrophication in a lake). Often, there are considerable uncertainties that make identifying the threshold challenging. Thus, rapid learning is critical for guiding management actions to avoid abrupt transitions. Here, we adopt the shallow lake problem as a test case to compare the performance of four common data assimilation schemes to predict an approaching transition. In order to demonstrate the complex interactions between management strategies and the ability of the data assimilation schemes to predict eutrophication, we also analyze our results across two different management strategies governing phosphorus emissions into the shallow lake. The compared data assimilation schemes are: ensemble Kalman filtering (EnKF), particle filtering (PF), pre-calibration (PC), and Markov Chain Monte Carlo (MCMC) estimation. While differing in their core assumptions, each data assimilation scheme is based on Bayes' theorem and updates prior beliefs about a system based on new information. For large computational investments, EnKF, PF and MCMC show similar skill in capturing the observed phosphorus in the lake (measured as expected root mean squared prediction error). EnKF, followed by PF, displays the highest learning rates at low computational cost, thus providing a more reliable signal of an impending transition. MCMC approaches the true probability of eutrophication only after a strong signal of an impending transition emerges from the observations. Overall, we find that learning rates are greatest near regions of abrupt transitions, posing a challenge to early learning and preemptive management of systems with such abrupt transitions.