Effective segmentation of Arctic sea ice floes from high-resolution optical images

  • Siyuan Chen

Student thesis: Master's Thesis


Arctic sea ice, as an essential environmental component that is closely connected to the Arctic ecosystem, plays an important role in the global weather and climate system. In addition, the Marginal Ice Zone (MIZ) which is the region close to the open water, plays a significant role in studying the physical and dynamic processes in Arctic area as it reflects the intense interactions between the atmosphere, open water, and sea ice. In addition, the MIZ can provide a physical buffer to protect the inner larger pack ice from being breakup by the ocean waves, which accordingly prevents the sea ice retreat that may occur due to the increasing Arctic open water area. Floe Size Distribution (FSD), a keystone indicator of the MIZ which has great impact on multiple sea ice processes such as the lateral melt rate and the propagation of waves underneath the sea ice, is particularly beneficial for the weather prediction, and the management of the Arctic region. Until now, remote sensing data is one of the most popular and very often the only sources of information regarding sea ice conditions in the Arctic. During the last few decades, many efforts have been made for sea ice segmentation from the Synthetic Aperture Radar (SAR) images and optical images. However, separating the touching floes is still a main obstacle for accurate FSD retrieval. In recent years, the High-Resolution Optical (HRO) images with less speckle noise compared to the SAR images have provided an alternative solution to accurately delineate the floe boundaries and extract FSD. Nevertheless, traditional floe separation approaches require manual interactions. Meanwhile, the data annotation of sea ice images requires domain knowledge and can be labor extensive, resulting in the deep learning based methods hard to be applied. In this thesis, a multi-stage segmentation and floe separation model is proposed to effectively investigate ice pixels and separate the touching ice floes automatically from the HRO images. For ice pixels investigation, a novel segmentation framework is proposed, where a combination of superpixel K-means clustering is employed for identifying the ice, water-ice mixed, and open water regions. Afterwards, the contrast enhancement technique is applied on the water-ice mixed regions to improve the performance of the subsequent thresholding process. For floe separation, the marker controlled watershed transformation based method is proposed, where different strategies are employed to generate the markers for preventing the floes from being over-segmented. The robustness of the segmentation framework is validated on the image dataset and compared to two state-of-the-art methods. The result shows that the proposed framework has yielded the highest performance in terms of the accuracy, Matthews correlation coefficient (MCC) and F-1 score. For FSD retrieval, the proposed method outperforms the traditional distance transformation based watershed segmentation with the closer power law exponents and less mean square error compared to the ground truth. In addition, the experiments for optimising the model by evaluating the scenarios using different algorithms and parameter settings are conducted. As a result, for the proposed model, a combination of bilateral filter for pre-processing, Simple Linear Iterative Clustering (SLIC) for superpixel generation, and Top-bottom-hat transformation for contrast enhancement is recommended.
Date of Award22 Nov 2022
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SponsorsUniversity of Strathclyde
SupervisorStephen Marshall (Supervisor) & Jaime Zabalza (Supervisor)

Cite this