Detection of ice types in the Eastern Weddell Sea by fusing L- and C-band SIR-C polarimetric quantities

Abstract

This article discusses the use of spaceborne polarimetric L-band and C-band synthetic aperture radar (SAR) data for sea-ice detection and classification. The benefits of combining L-band with C-band polarimetric quantities for supervised sea-ice classification in the Eastern Weddell Sea, Antarctica, are investigated. In the experiments, we compared the performance of a maximum likelihood (ML) classifier when used with the combined preferred polarimetric parameters and the individual ones, respectively. The relation between the classification accuracy and the preferred number of polarimetric parameters for classification was examined as well as whether principal component analysis (PCA) and locally linear embedding (LLE) can be used to reduce the dimensionality of the parameter sets. Combining dual-frequency polarimetric quantities improves classification accuracy compared to using individual single-frequency polarimetric quantities. By increasing the dimensionality of the preferred polarimetric parameter sets, the classification using high dimensionality can either result in improvements over the smaller subsets or result in no significant differences. Therefore, using all available polarimetric quantities over the study region is recommended. Further, data fusion with a PCA-based approach is found to be beneficial for sea-ice detection and classification, and poor results have been produced with an LLEbased approach.

A Meta-Analysis of Sea Ice Monitoring Using Spaceborne Polarimetric SAR: Advances in the Last Decade

“…For example, L-band provides better discrimination between young ice and smooth FYI than C-band since the L-band's correlation coefficient is a vital feature for young ice and smooth FYI discrimination. In addition, sea ice observation using multifrequency SAR is also analyzed in [30], [139], and [140]. In [140], L-band SAR is found to be able to identify ice ridges with greater ease because longer wavelength data are less affected by microscale ice structures.…”

Deep-learning-based information mining from ocean remote-sensing imagery

“…The critical challenge is to develop a robust model that captures domain-specific expert knowledge for discriminating between ice and water using SAR backscatter characteristics. To achieve this goal, different types of sea-ice-detection models based on backscatter thresholding [ 91 ], regression techniques [ 92 ], expert systems [ 93 ], Bayesian techniques [ 94 ], gray-level co-occurrence matrix and the support vector machine (SVM) hybrid method [ 95 ], among others [ 96 ], are proposed.…”