Modelling of yearly latent heat flux
Modelling of yearly GPP
Modelling of yearly NEE
Modelling of yearly sensible heat N
Our research addresses fundamental questions on land-surface/climate interactions from local to global scales and from sub-daily to centennial time-scales. Processes we study will link biophysiological/ecological processes with biophysical and biogeochemical mechanisms across time-scales and encompass uptake and release of climate relevant trace gasses and particles by ecosystems (e.g., CO2, CH4, nitrous oxides and other reactive nitrogen, aerosols), eco- hydrological dynamics, soil and vegetation carbon turnover and disturbance dynamics introduced by episodic disturbances such as fire, biotic agents, or climate extremes.
- Causal inference and modeling
Methods for causal inference from observation are key for a better understanding of the Earth system, because interventions and experiments across the whole Earth system are not feasible. Yet, for subsystems, e.g. ecosystems, experiments are possible, allowing validation of causal inference, by comparing modeled with experimentally observed intervention effects. A challenge for causal inference with Earth system science are the manifold feedback loops which preclude the “classical” modeling with directed acyclic graphs (DAGs), so that strategies have to be extended.
- Spatio-temporal modeling
Earth system dynamics is spatio-temporal dynamics par excellence. As such it closely relates to current machine learning/computer vision challenges like video prediction. Additional challenges are introduced by the multi-modality in Earth observations, by long-range interactions (“teleconnections”) and long memory effects, where temporal system evolution can depend on events long time ago, e.g. past fires or past (different) land-use.
- Dynamic system identification
This encompasses an important wide field essentially bringing together theory and system modeling with data science and machine learning, including inverse modeling, data assimilation, gradient free optimization, emulation, hybrid modeling, Bayesian uncertainty quantification. Major challenges include but are not restricted to unbalanced observation (sparse/rich), measurement uncertainties, curse of dimensionality, and equifinality of models. Integrating multi-modal observations and linking mechanistic modeling and machine learning into hybrid approaches appear to be promising avenues, yet still to be better founded from a learning theoretical perspective.