Deep-learning-based hybrid uncertainty-aware modelling of the coupled water and carbon cycle with Earth observation data
The aim of the DUKE project is to further develop and combine the research paths of hybrid modelling and uncertainty assessment. To this end, the Chair of Remote Sensing Technology is collaborating with the Max Planck Institute for Biogeochemistry in Jena.
A particular focus will be on ensuring optimal scalability in terms of the regions observed – from the field level to the global scale – as well as the data volumes to be processed. DUKE focuses on the real-world problem of modelling global matter transport cycles, specifically the currently still open question of coupled water and carbon cycles. Here, a data-driven perspective offers enormous potential, as the mostly rigid boundary conditions and parameterisations in process models lead to systematic deviations in the simulations. Furthermore, the quantification of uncertainties is essential for model development as well as for testing alternative hypotheses.
Within the framework of this project, the TUM Chair of Remote Sensing Technology (TUM-LMF) and the Max Planck Institute for Biogeochemistry (MPI-BGC) will work on complementary issues in close cooperation.
TUM: Uncertainty Quantification
The TUM-LMF would like to answer the following questions in particular:
- How can currently known methods for estimating the uncertainty of model outputs be mapped to the described problem domain? To what extent do the properties inherent in the context of Earth observation and Earth system research limit this applicability and thus necessitate the use of specialised approaches?
- How can formalised expert and domain knowledge be profitably integrated into these methods?
- How can the causes of these uncertainties be attributed, localised and thus circumvented?
MPI: Hybrid Modelling
The MPI-BGC is concerned with the applicability and implications of the aforementioned topics to the field of earth sciences. In particular, the following central questions arise:
- How can these methods be used to develop hybrid deep learning models that exhibit higher levels of performance, descriptivity, robustness and interpretability?
- At what conceptual level should such hybrid designs ideally come to fruition?
- How can the knowledge gained be used to model a complex example scenario, a global combined C/H2O cycle?