Multi-stage processing for effective segmentation of SAR sea ice images

  • Soumitra Ravindra Sakhalkar

Student thesis: Master's Thesis

Abstract

Remote Sensing and Earth Observation became a reality, ever since the launch of NASA's first satellite; Landsat-1 in 1972. Subsequently, numerous other satellites were launched such as the TerraSAR, Sentinel-1 etc, which made possible to acquire High-Resolution imagery of remote areas such as the Arctic region. Consequently, with the accessibility to such large amounts of data, it becomes necessary to develop fast, robust and automated image segmentation algorithms to extract key information as opposed to still relying on time consuming manual expert analysis.As a result, in this thesis, effective algorithms are proposed for efficiently segmenting and extracting information from the Synthetic Aperture Radar (SAR) Sea Ice imagery.Initially, the contributions for improving the quality of the SAR images itself are introduced. Inspired by the advantages of the Adaptive Median filter (AMF) and the Wiener filter, the Modified Adaptive Median filter (MAMF) is proposed. The MAMF uses local image statistics to identify speckle regions and the Minimum Mean Square Error (MMSE) estimator to suppress speckle. The MAMF is applied to various image types, to test its efficiency and robustness and subsequently compared with other existing techniques such as the Bilateral and Local Sigma filters.Furthermore, additional region-based filtering is suggested, which is based on user-defined threshold values for the quantitative parameters used to determine the performance of the filter.Another important part of extracting key information from the SAR Sea Ice imagery is '€Segmentation'. A Region and Condition based post processing is proposed for the established algorithm, Kernel Graph Cuts (KGC), for acquiring further improved segmentation results. The post processing incorporates algorithms such as Skeletonisation, Morphology and Active Contours. The proposed algorithm is compared against existing techniques such as the Closeness Degree Cut (CDCut) and Level Sets with Distance Regularisation (DRLSE).Furthermore, a novel Quantitative Analysis technique is proposed which accurately compares the Regional Accuracy of the various segmented regions.By the use of local statistics of the SAR images used, the MAMF filter effectively suppresses speckle noise without over-compensating for image features. Similarly, the post processing technique uses a combination of Conditional Morphology and Active Contours, to effectively improve the segmented result obtained with the established KGC algorithm. The results are validated against recently-used algorithms, on realworld and sample images acquired through various datasets by performing objective as well as subjective analysis.
Date of Award1 Jun 2017
LanguageEnglish
Awarding Institution
  • University Of Strathclyde
SupervisorJinchang Ren (Supervisor) & Stephen Marshall (Supervisor)

Cite this