Climatic drivers and biogeophysical feedbacks: a causal inference approach over multiple temporal scales

Earth system models (ESMs) need to correctly simulate the impact of climate on vegetation, as well as the feedback of vegetation on climate. Improving the skill of ESMs in representing climate—biosphere interactions is crucial to enhance predictions of climate and ecosystem functioning. Correlation and regression techniques are commonly used to study these interactions statistically, but these methods lack the ability to unravel the bidirectional nature of the climate–biosphere system. Here, we explore these interactions across multiple temporal scales by adopting a spectral Granger causality framework that allows identifying potentially inter-dependent variables. Multi-decadal remotely-sensed records are used to analyse the impact of key climatic drivers (precipitation, radiation and temperature) on vegetation (Leaf Area Index, LAI), as well as the biophysical feedback on local climate. These observational results are in turn used to benchmark a set of Coupled Model Intercomparison Project Phase 5 (CMIP5) members at the global scale.

Results show that the climate control on LAI variability increases with longer temporal scales, being the highest at inter-annual scales. Globally, precipitation is the most dominant driver of vegetation at monthly scales, particularly in (semi-)arid regions, as expected. The seasonal LAI variability in energy-driven latitudes is mainly controlled by radiation, while air temperature controls vegetation growth and decay in northern latitudes at inter-annual scales. ESMs have a tendency to over-represent the climate control on LAI dynamics, and especially the role of precipitation at inter-annual scales. Likewise, the widespread effect of LAI variability on radiation, as observed over the northern latitudes due to albedo changes, is also overestimated by the CMIP5 models. Overall, our experiments emphasise the potential of benchmarking the representation of climate—biosphere interactions in online ESMs using causal statistics in combination with observational data.