Learning Spatio-Temporal Correlations from Dynamic and Sparse Data : Deciphering AMOC Drivers from Sparse Ocean Observations

The amount of spatio-temporal measurements around Earth, its atmosphere, in human-made urban settings, and in its oceans increases with affordable and smaller sensor technologies. Improved communication technologies allow sensors to be not geostationary, changing their positions and measurement semantics over time, resulting in datasets with dynamic spatial distributions. Existing spatial learning approaches often fall short in capturing spatio-temporal patterns from these distributions. One of the affected environmental fields is physical oceanography, which is the second discipline, besides computer science, in this interdisciplinary work. The large extent of the world's oceans, and the limited communication of instruments, compared to more manageable urban settings, form a challenging environment in which nowadays drifting and floating autonomous instruments are deployed. This interdisciplinary thesis addresses this challenge by developing and evaluating spatio-temporal learning methods tailored to this kind of challenge. Besides these novel methods being evaluated in computer science, this thesis aims to transfer them to oceanographic data and assess them from an oceanographic point of view.
Methodologically, this thesis develops and adapts spatio-temporal learning approaches to explicitly handle dynamic spatial layouts, where the sensors' location, availability, and the semantics of their measurements change over time. For supervised learning, spatio-temporal Graph Neural Networks are redesigned by breaking the one-to-one mapping between sensors and graph nodes. Fixed sensor identities are replaced by contextual measurement representations that generalize across dynamic spatial distributions. In the unsupervised setting, spatial patterns are formulated independently of spatial distributions by embedding global and local structures directly in the spatial reference frame. 
This interdisciplinary thesis provides a novel perspective on how to connect spatio-temporal learning methods and dynamic environments, using oceanographic use cases as an example. This approach underscores the potential of interdisciplinary work connecting the ocean full of data with oceanographic knowledge and an advanced machine learning toolbox simultaneously.