Limitations of early warning signals: evaluating the performance of resilience loss detection methods to predict forest die-back events from remote sensing data

Terrestrial ecosystems worldwide are under increasing stress due to changing climate and weather regimes, as well as direct anthropogenic influences such as land use changes. The combination of stressors can erode an ecosystem’s ability to resist and recover from external shocks and pressures.

Vegetation resilience loss is often assessed by applying temporal early warning signals (EWS) based on dynamical systems theory to remotely sensed time series of different vegetation indices. The global coverage and regular measurement intervals of the satellite data in combination with easily computable EWS such as temporal autocorrelation and variance make this an appealing approach. Recent studies have confirmed that common EWS are good indicators of recovery rates after small disturbances in global ecosystems. However, to be useful in real-world applications, EWS need also be able to provide warning signals before a major upcoming ecosystem collapse, as driven for example by drought or heat stress. This has been evaluated for local case studies of specific ecosystems, but a global assessment of EWS accuracy and sensitivity for predicting terrestrial ecosystem collapses is lacking.

Here, we evaluate the performance of different EWS in predicting forest dieback events recorded in situ and on manually assessed satellite data around the world. We compare different frequently used remote sensing datasets, vegetation indices, and a range of EWS. This work highlights limitations of commonly applied resilience loss assessment methods for real-world applications and aims to contribute to the discussion on how to reliably evaluate changes in large-scale ecosystem resilience.