Comparison of Oil Spill Classifications Using Fully and Compact Polarimetric SAR Images

TLDR: In this article, a comparison between several algorithms for oil spill classifications using fully and compact polarimetric satellite synthetic aperture radar (SAR) images is presented, where dimension reduction algorithms, including principle component analysis (PCA) and the local linear embedding (LLE) algorithm, are employed to learn low dimensional and distinctive information from quad-polarimetric SAR features.

Abstract: In this paper, we present a comparison between several algorithms for oil spill classifications using fully and compact polarimetric SAR images. Oil spill is considered as one of the most significant sources of marine pollution. As a major difficulty of SAR-based oil spill detection algorithms is the classification between mineral and biogenic oil, we focus on quantitatively analyzing and comparing fully and compact polarimetric satellite synthetic aperture radar (SAR) modes to detect hydrocarbon slicks over the sea surface, discriminating them from weak-damping surfactants, such as biogenic slicks. The experiment was conducted on quad-pol SAR data acquired during the Norwegian oil-on-water experiment in 2011. A universal procedure was used to extract the features from quad-, dual- and compact polarimetric SAR modes to rank different polarimetric SAR modes and common supervised classifiers. Among all the dual- and compact polarimetric SAR modes, the π/2 mode has the best performance. The best supervised classifiers vary and depended on whether sufficient polarimetric information can be obtained in each polarimetric mode. We also analyzed the influence of the number of polarimetric parameters considered as inputs for the supervised classifiers, onto the detection/discrimination performance. We discovered that a feature set with four features is sufficient for most polarimetric feature-based oil spill classifications. Moreover, dimension reduction algorithms, including principle component analysis (PCA) and the local linear embedding (LLE) algorithm, were employed to learn low dimensional and distinctive information from quad-polarimetric SAR features. The performance of the new feature sets has comparable performance in oil spill classification.