Scalable computing for earth observation - Application on Sea Ice analysis

Abstract

<p>In recent years, Deep learning (DL) networks have shown considerable improvements and have become a preferred methodology in many different applications. These networks have outperformed other classical techniques, particularly in large data settings. In earth observation from the satellite field, for example, DL algorithms have demonstrated the ability to learn complicated nonlinear relationships in input data accurately. Thus, it contributed to advancement in this field. However, the training process of these networks has heavy computational overheads. The reason is two-fold: The sizable complexity of these networks and the high number of training samples needed to learn all parameters comprising these architectures. Although the quantity of training data enhances the accuracy of the trained models in general, the computational cost may restrict the amount of analysis that can be done. This issue is particularly critical in satellite remote sensing, where a myriad of satellites generate an enormous amount of data daily, and acquiring in-situ ground truth for building a large training dataset is a fundamental prerequisite. <p>This dissertation considers various aspects of deep learning based sea ice monitoring from SAR data. In this application, labeling data is very costly and time-consuming. Also, in some cases, it is not even achievable due to challenges in establishing the required domain knowledge, specifically when it comes to monitoring Arctic Sea ice with Synthetic Aperture Radar (SAR), which is the application domain of this thesis. Because the Arctic is remote, has long dark seasons, and has a very dynamic weather system, the collection of reliable in-situ data is very demanding. In addition to the challenges of interpreting SAR data of sea ice, this issue makes SAR-based sea ice analysis with DL networks a complicated process. <p>We propose novel DL methods to cope with the problems of scarce training data and address the computational cost of the training process. We analyze DL network capabilities based on self-designed architectures and learn strategies, such as transfer learning for sea ice classification. We also address the scarcity of training data by proposing a novel deep semi-supervised learning method based on SAR data for incorporating unlabeled data information into the training process. Finally, a new distributed DL method that can be used in a semi-supervised manner is proposed to address the computational complexity of deep neural network training.