PolSAR Ship Detection Using the Superpixel-Based Neighborhood Polarimetric Covariance Matrices

In order to detect ships from the imagery of polarimetric synthetic aperture radar (PolSAR), a neighborhood polarimetric covariance matrix (for simplicity, we call it [
$N$
] hereinafter) was recently constructed. However, its calculation process is time-consuming and the backscattering heterogeneity near ship edges is also not well considered. For curing these shortcomings, we here propose two novel superpixel-based neighborhood polarimetric covariance matrices. In brief, the first matrix denoted by [SN] uses the simple linear iterative clustering (SLIC) to yield superpixels, whereas in the second matrix denoted by [GN], the gradient operator Sobel is adopted to obtain superpixels. Based on these two different kinds of superpixels, then, two different feature vectors
$v_{\text {SN}}$
and
$v_{\text {GN}}$
are separately built to compute [SN] and [GN]. Experiments performed on the real PolSAR datasets show that, compared to [
$N$
], [SN] and [GN] can improve the performance of the polarimetric whitening filter (PWF) more significantly and the time consumptions of calculating [SN] and [GN] are both much less.