Biosphere-atmosphere interaction : inferring dependencies via causal graph reconstruction

Biosphere--atmosphere interactions determine a large fraction of the observed variability in carbon and water fluxes. Understanding observed variability thus goes hand in hand with understanding the dependencies of water and carbon flux regulating processes to atmospheric influences as well as their feedbacks. As large scale experimental studies are typically hindered by practical, financial and moral burdens and model assessments requires prior assumptions of the underlying mechanisms, empirical analyses can valuably complement the study of biosphere-atmosphere interactions utilising the growing availability of observational data. The complexity of biosphere-atmosphere interactions given by the number of involved components and a potential bidirectionality, not necessarily happening at similar time scales, renders conventional correlative approaches as insufficient. Causal inference has gained much momentum within the last decades achieving states of maturity enabling us to address real world questions. Causal graph learning algorithms thereby have become a major tool. The main objective of this thesis is to evaluate the ability of causal inference to improve our understanding of biosphere--atmosphere interactions with the focus on the dependencies of CO2 and water fluxes. Three minor objectives are derived to accomplish the main objective. As any statistical method, causal inference methods are linked to a set of assumptions. Under real world conditions, not all can always be fully guaranteed. The consequences of such shortcomings can be manifold. Thus, first I evaluate the potential of the causal graph discovery method PCMCI to reconstruct the causal dependency structure underlying biosphere-atmosphere interactions from corresponding data sets (chapter 2). Using these insights, I perform an ecosystem comparison. The ecosystems depict the available climate regions and vegetation types of the Earth. The objective is to test whether causal inference can challenge or extent our current perception on the functioning of ecosystems and their interactions with the atmosphere which currently are typically classified by their structural properties or climate conditions (chapter 3). To test whether the algorithm is suited to examine specific processes I investigate the hypothesis of a decoupling between photosynthesis and transpiration during heat stress. Thereby the objective is to test if limitations which arise by unmet assumptions can be overcome to gain insight in the existence and behaviour of a certain process (chapter 4). With this thesis I could show that causal inference improves the study of biosphere-atmosphere interactions. The main advantage lies in the focus on relevant dependencies which are detected from a potentially large set of variables and time lags. Detected dependencies are not necessarily causal and inference of directions is difficult given that the temporal resolution of the data exceeds the interaction time of processes. Here further research capable of inferring directions of contemporaneous dependencies may enable stronger causal interpretations in the future. Further, automated handling of autocorrelation and the choice of conditioning sets based on causal theory can homogenise analyses and improve comparisons. Combining spectral decomposition and causal inference would add further benefits along those lines and open up extra layers of interpretation