Arctic Sea Ice property retrieval from synthetic aperture radar with deep learning methods

• Main Author: Kortum, Karl
• Other Authors: Spreen, Gunnar, Singha, Suman, Haas, Christian
• Format: Doctoral or Postdoctoral Thesis
• Language: English
• Published: Universität Bremen 2024
• Subjects: Sea Ice; Machine Learning; Deep Learning; Synthetic Aperture Radar; Physics-informed Neural Networks; Altimetry; 530; 530 Physics; ddc:530; Arctic; Climate change; Sea ice
• Online Access: https://media.suub.uni-bremen.de/handle/elib/7803; https://doi.org/10.26092/elib/2885; https://nbn-resolving.org/urn:nbn:de:gbv:46-elib78032

Current climate models are not capturing the feedback mechanisms driving the accelerated warming of the Arctic. A central challenge is the sparsity of observations. Satellite-borne synthetic aperture radar (SAR) instruments have the capability of monitoring Earth's sea ice masses at high resolution, unhampered by cloud coverage or the Arctic night. The measurements are made at scales of 10's of metres whilst still covering the Arctic in a matter of days. However, interpreting the radar signal to retrieve relevant sea ice information is difficult because of the complex interactions of the ice with the electromagnetic radar signal. Conventional neural network algorithms leverage contextual image data to make accurate predictions of surface ice properties comparable to those made by human experts. They are, however, dependent on large amounts of high-quality ground truth that is rare in these regions. Thus, the full potential of the SAR data is yet to be unlocked. With the advent of the MOSAiC mission, large timeseries of SAR data and near-coincident ground measurements were acquired for the first time. This thesis uses the unique opportunity provided by these data to analyse the behaviour of deep learning models. Seven months of data from the campaign is classified and analysed, using newly developed techniques to enable robust predictions across the timeseries. Core features are identified to facilitate robust and high-resolution classification. The final challenge of ground truth sparsity is then overcome using innovative network configurations that enable the training of 99.99%$ of the model parameters without any ground truth data. The techniques open up sea ice property retrieval to big data technologies, relying only on the abundantly available SAR data. These techniques enable the extrapolation of sparse reference data to a large space of sea ice conditions and enable high resolution mapping of the Earth's region most affected by human-made climate change.