Deep State-Space Models for Understanding Changes in Ecosystems

• Applicant: Dr.-Ing. Maha Shadaydeh, Ph.D.
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 455121326

Data is ubiquitous in climate and ecosystems sciences. Despite the abundance of data to process, data science has not had a considerable impact on research for a better understanding of the complex dynamics of the climate-ecosystem exchange system (CEES). Accurate modelling of such dynamics is necessary for the attribution of climate change, prediction of future data, as well as for planning future actions. Ecological processes exhibit variation over different spatial and temporal scales. Over at least some scales, these processes are nonlinear with possible unobserved causes/confounders. To deal with all these challenges, we need to develop data-driven methodologies that are guided by expert knowledge to constrain search and produce more accurate models. Probabilistic graphical state-space models (SSMs) are a rich framework for learning since they can deal with our imperfect knowledge of the world and explicitly take into account the uncertainty in the observations. However, the use of SSMs for modelling nonlinear high dimensional processes has numerous challenges and efficient parametrisation of these models remains a critical challenge for their implementation. Deep learning approaches are more flexible with better capabilities in learning nonlinear inference and capturing complex long term dependency in the data, but lack interpretability and uncertainty representation. The goal of this proposal is to combine the areas of probabilistic graphical SSMs and deep learning to model the nonlinear dynamics of the CEES. We also aim to build on the obtained dynamical system to examine the causal-effect interaction patterns between its subsystems (different pairs of its variables) on multiple time scales. We would like then to investigate whether the use of multiscale causal inference knowledge can lead to a better understanding of the impact of extreme events on ecosystem functioning when compared to statistical association methods. The computer vision group at FSU Jena aims to work on these goals in close collaboration with the MPI-BGC Jena. As an example application, we intend to start with the model formulation of the net ecosystem exchange of CO2 flux partitioning, which is a major problem for understanding the climate impact on ecosystem functioning. As a data source, we plan to use the Eddy Covariance measurement.

• DFG Programme: Research Grants