Sea ice segmentation in SAR images using Deep Learning

Abstract

Sea ice covers over seven percent of the world's oceans. Due to the effect of global warming, Arctic's ice extent has decreased significantly in the past decades. This reduction in sea ice cover is opening new pathways for the international shipping community through the Arctic. Due to the lengthening of the open water season, the Canadian Arctic has also observed a three-fold increase in the shipping traffic in the past few years. Although the ice extent has reduced, the risks and hazards involved in shipping through these regions are still significant. To promote safe and efficient maritime activities in the Canadian Arctic, Canadian Ice Service (CIS) provides information about ice in Canada's navigable waters. CIS uses Synthetic Aperture Radar (SAR) images as one of the prominent sources to gain insights about the ice conditions in Canadian waters. Automated SAR image interpretation is a complex task and requires algorithms to learn complex and rich features. Convolutional neural networks (CNNs) have demonstrated their ability to learn such features and have been used in various image classification, segmentation and object detection tasks. In this thesis, we first propose a method to detect marginal ice zones (MIZs) in SAR images. This method uses transfer learning combined with a multi-scale patch technique to detect the MIZs. The multi-scale patch technique involves generating the segmentation masks over different patch sizes for the same image. These masks are later stacked together and thresholded to generate the final MIZ prediction mask for an image. Later we dive deep into the MIZs and focus on segmenting sea ice floes. We propose a segmentation model optimized for the task of ice floe segmentation in SAR images. The model is based on a fully convolutional architecture with residual connections. In addition to this, a conditional random field is also used as a post-processing step. The whole network is trained end-to-end using a dual loss function. Qualitative and quantitative analysis suggests that our model beats the conventional segmentation architectures for the task of ice floe detection.