Lead Detection in the Arctic Ocean: Assessment of thresholding and machine learning classification methods on Sentinel-3 SRAL altimeter for lead detection in the Arctic Ocean

• Main Author: Martin, Ericka (author)
• Other Authors: Bij de Vaate, I. (mentor), Naeije, M.C. (graduation committee), Slobbe, D.C. (graduation committee), Delft University of Technology (degree granting institution)
• Format: Master Thesis
• Language: English
• Published: 2021
• Subjects: Arctic; Lead detection; SAR altimetry; Sentinel-3; Machine Learning; Arctic Ocean; Rus’; The Sentinel; ice pack; Sea ice
• Online Access: http://resolver.tudelft.nl/uuid:f89e569e-fd1b-4fcf-81e5-e355570fb586

Detection of the openings in the Arctic sea ice pack, or leads, allow to sample instantaneous sea surface height (SSH) and this information is crucial for quantifying the impact of sea ice melting. It is therefore important to correctly detect as many leads as possible to obtain more SSH references. This paper studies 12 different classification methods including supervised-, unsupervised machine learning methods and thresholding method, being applied to the Sentinel-3 Synthetic Aperture Radar (SAR) altimetry data collected in March/April of 2017-2020 and June/July of 2020 from areas all across the Arctic Ocean. These are compared and assessed with respect to images taken by Ocean and Land Color Instrument (OLCI), also on board Sentinel-3, ensuring a perfect temporal alignment between the two measurements. The supervised Adaptive Boosting, Artificial Neural Network and Linear Discriminant classifiers showed excellent and robust results in March/April with overall accuracies up to 91.82%. The unsupervised K-medoid classifier produced excellent results achieving up to 91.51% accuracy and it is an attractive classifier as it does not require ground truth data. The classifiers perform poorly in the summer months, as sea ice returns show more ambiguous reflections due to melting. Therefore on summer data, classifications that are solely based on waveform data from SAR altimetry is unsuitable and auxiliary information is required. Furthermore, this paper attempted to identify off-nadir leads (ONL) by adding an extra class in supervised learning methods, intending to reduce the falsely detected leads. Most classifiers failed to detect leads and did not improve their false lead rate. However, as RUS Boost classifier was able to identify 61.6% of total ONLs, this can be used to initially reject these points for more conservative lead detection. Aerospace Engineering