On Automated Classification of Sea Ice Types in SAR Imagery

• Main Author: Lohse, Johannes
• Format: Doctoral or Postdoctoral Thesis
• Language: English
• Published: UiT Norges arktiske universitet 2021
• Subjects: VDP::Mathematics and natural science: 400::Physics: 430::Electromagnetism; acoustics; optics: 434; VDP::Matematikk og Naturvitenskap: 400::Fysikk: 430::Elektromagnetisme; akustikk; optikk: 434; VDP::Mathematics and natural science: 400::Information and communication science: 420::Simulation; visualization; signal processing; image processing: 429; VDP::Matematikk og Naturvitenskap: 400::Informasjons- og kommunikasjonsvitenskap: 420::Simulering; visualisering; signalbehandling; bildeanalyse: 429; VDP::Mathematics and natural science: 400::Mathematics: 410::Statistics: 412; VDP::Matematikk og Naturvitenskap: 400::Matematikk: 410::Statistikk: 412; DOKTOR-004; Arctic; Annals of Glaciology; Sea ice
• Online Access: https://hdl.handle.net/10037/20606

With the Arctic sea ice continuously decreasing in both extent and thickness, fast and robust production of reliable ice charts becomes more important to ensure the safety of Arctic operations. This thesis focuses on the development of automated algorithms for the mapping of sea ice from synthetic aperture radar (SAR) images. It presents a thorough background on the topics of sea ice observations and ice charting, sea ice image classification, and the appearance of sea ice in SAR imagery. Three papers present the scientific developments in the thesis. Paper 1 focuses on the topic of feature selection. The study investigates the benefits of splitting a multi-class problem into several binary problems and selecting different feature sets specifically tailored towards these binary problems. Using a combination of classification accuracy and sequential search algorithms, the best order of classification steps and the optimal feature set for each class are found and combined into a numerically optimized decision tree. The method is tested on various examples, including an airborne, multi-frequency SAR data set over sea ice, and compared to traditional classification approaches. Paper 2 and 3 focus on the classification of Sentinel-1 (S1) wide-swath SAR images. Both papers use a newly generated training and validation data set for different sea ice types, which is is based on the visual analysis of overlapping S1 SAR and optical data. A particular challenge for the automated analysis of wide-swath SAR images is the surface-type dependent variation of backscatter intensity with incident angle (IA). In Paper 2, a novel method to directly incorporate this per-class IA effect into a classification algorithm is developed. Paper 3 investigates the IA dependence of texture features and extends the algorithm from Paper 2 to include textural information, in order to solve the ambiguities inherent in a classifier based on intensity only.