Sea ice classification from RADARSAT constellation mission images using normalizer-free ResNet

Lyu, Hangyu (2022) Sea ice classification from RADARSAT constellation mission images using normalizer-free ResNet. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (76MB)

Abstract

Sea ice monitoring plays a vital role in climate study, maritime navigation and offshore industries. Sea ice monitoring consists of different applications, such as ice classification, concentration and thickness retrieval. As one of the branches of sea ice monitoring, sea ice classification is an essential task in sea ice mapping and the premise to obtain other sea ice parameters. Satellite images are the primary source for sea ice classification due to the broad coverage, the extremely harsh environment in the polar regions and the near real-time requirements of some applications. Spaceborne Synthetic Aperture Radar (SAR) has been widely used as an effective tool for sea ice sensing for decades because it can collect data day and night and in all weather conditions. As a typical representative of the next generation SAR mission, the RADARSAT Constellation Mission (RCM) provides three C-band SAR satellites with shorter revisit time and broader spatial coverage, which will be widely used in various earth observation applications including sea ice sensing. The Sentinel-1 mission comprises two C-band SAR satellites with dual-polarized imaging capability, providing open and free data from the European Space Agency (ESA). Both RCM and Sentinel-1 C-band SARs operate at a center frequency of 5.405 GHz. In addition, RCM provides more spatial coverage and a shorter revisit time than Sentinel-1. However, actual RCM data have not been used for sea ice classification, and no study for comparing the sea ice classification performances of RCM and Sentinel-1 has been conducted. Deep convolutional neural networks (CNN) have been extensively employed in sea ice monitoring applications in the last decade. An example of deep CNN, Normalizer-Free ResNet (NFNet) was proposed by DeepMind in early 2021 and achieved a new state-of-the-art accuracy on the ImageNet dataset. In this thesis, a NFNet based approach has been proposed for sea ice classification using dual-polarized SAR data. In the first part of this study, the RCM data are utilized for sea ice detection and classification using NFNet for the first time. HH (horizontal transmit and horizontal receive), HV (horizontal transmit and vertical receive) and the cross-polarization ratio are extracted from the dual-polarized RCM data with a medium resolution (50 m) for NFNet-F0 model. Experimental results from the eastern Arctic show that destriping in the HV channel is necessary to improve the quality of sea ice classification. A two-level random forest (RF) classification model is also applied as a conventional technique for comparisons with NFNet. The sea ice concentration, estimated based on the classification result from each region, was validated with the corresponding polygon of the Canadian weekly regional ice chart. The overall classification accuracy confirms the superior performance of the NFNet model over the RF model for sea ice monitoring and the sea ice sensing capacity of RCM. The second part of this study focuses on comparing sea ice classification results from the two C-band SAR missions (RCM and Sentinel-1) with the state-of-the-art convolutional neural network, NFNet. HH, HV and the cross-polarization ratio are extracted from the overlapping area of dual-polarized RCM and Sentinel-1 images acquired on similar dates. The sea ice classification results show that the RCM Medium Resolution 50m mode performs better than the Sentinel-1 EW GRD Medium Resolution 90m mode.

Actions (login required)

View Item