Public defence in Nonlinear Dynamics of Sea Ice, MSc Matias Uusinoka

Title of the thesis: Nonlinear Dynamics of Sea Ice at Intermediate Scales via Deep Learning Optical Flow

Thesis defender: Matias Uusinoka
Opponent:  Dr. Nils Hutter, GEOMAR Helmholtz Centre for Ocean Research Kiel, Germany
Custos: Prof. Arttu Polojärvi,Aalto University School of Engineering

Deformation of ice covers occurs through a mix of slow drift and sudden, spatially localized events over a vast variety of different spatial and temporal scales. Understanding this behavior is especially challenging as it sits between the two fundamentally different viewpoints of detailed mechanical descriptions at small scales and large-scale continuum and statistical descriptions. This doctoral thesis addresses the regime in between by combining modern data-driven deep learning methodology with statistical-physics style analysis of deformation across scales.

This thesis, firstly, develops a deep learning optical flow framework that converts noisy sequential radar data into dense displacement and deformation fields. The method is then applied to ship radar data from the Arctic to study how deformation statistics change with spatial and temporal scale using structure functions, scale invariance, and multifractal analysis. The results show that classic mean scaling behavior has a clear lower limit in winter conditions, around the order of 100 meters. The thesis further shows that multifractal behavior is not automatically observable, but rather it requires sufficiently large windows in both space and time. The findings are used to hypothesize a theoretical concept of threshold domain sizes that couples these statistical-physics style characterization to measured properties of the ice cover.

The results provide quantitative guidance for connecting mechanical interpretations of localized failure with the large-scale continuum descriptions, and for judging when scale-based statistical tools are meaningful. Further major implications are yield from this work, including a potential possibility of reproducing large-scale statistical properties in laboratory conditions. The developed deep learning workflow is also broadly transferable to other sequential imaging problems where accurate motion and strain estimation is needed under strong noise and small displacements.

Keywords: Scale invariance, Multifractality, Deep learning, Optical flow, Sea ice dynamics  

Thesis available for public display 7 days prior to the defence at Aaltodoc. 

Contact information
E-mail: matias.uusinoka@aalto.fi 
LinkedIn: https://www.linkedin.com/in/matias-uusinoka/