Learning to Estimate Sea Ice Concentration from SAR Imagery

• Main Author: Wang, Lei
• Format: Doctoral or Postdoctoral Thesis
• Language: English
• Published: University of Waterloo 2016
• Subjects: Sea ice concentration; convolutional neural network; synthetic aperture radar; Arctic; Beaufort Sea; Sea ice
• Online Access: http://hdl.handle.net/10012/10954

Through the growing interest in the Arctic for shipping, mining and climate research, large-scale high quality ice concentration is of great interest. Due to the unavailability of suitable ice concentration estimation algorithms, ice concentration maps are interpreted from synthetic aperture radar (SAR) images manually by ice experts for operational uses. An automatic ice concentration estimation algorithm is required for accurate large-scale ice mapping. In this thesis, a set of algorithms are developed aiming at operational ice concentration estimation from SAR images. The major difficulty in designing a robust algorithm for ice concentration estimation from SAR images is the constantly changing SAR image features of ice and water in time and location. This difficulty is addressed by learning features instead of designing features from SAR images. A set of convolutional neural network based ice concentration estima- tion algorithms are developed to learn multi-scale SAR image features and simultaneously regress ice concentration from the learned image features. We first demonstrated the capa- bility of CNNs in ice concentration estimation from SAR images when trained using image analysis charts as ground truth. Then the model is further improved by taking into account the errors in the image analysis charts. Ice concentration estimates with improved robust- ness to training samples errors, accuracy and scale of details are obtained. The robustness of the developed methods are further demonstrated in the melt season of the Beaufort Sea, where reasonable ice concentration estimates are acquired. In order to reduce the model training time, it is desired to reuse existing models. The model transferability is evaluated and suggestions on using existing models to accelerate the training process are given, which is shown to reduce the training time by over 10 times in our case.