Ship Detection in Sentinel-1 Image through Non-Parametric Estimation

Ship Detection in Sentinel-1 Image through Non-Parametric Estimation
Jeon, Ho-Kun; Cho, Hong Yeon
KIOST Author(s)
Cho, Hong Yeon(조홍연)
Alternative Author(s)
전호군; 조홍연
Publication Year
A basic approach to detect ships in coastal waters using SAR images is to determine a threshold value that discriminates ship pixels with relatively strong signals from weak background seawater pixels. If the density distribution is presumed uni-modal, it can be estimated through K-distribution parameter estimation. Otherwise, multiple parameters can be estimated using the Gaussian Mixture Model in multi-modal. However, repetitive numerical computations are required in both cases since parameters are estimated through the maximum likelihood estimation. In addition, for the latter case, the number of modals must be specified as hyperparameters. Consequently, they require high computational costs. On the other hand, if the given data is sufficient, the non-parametric density estimation gives the advantage of estimating the density distribution regardless of uni-modal or multi-modal. In this study, the density distribution of seawater and ship pixels is estimated using kernel density estimation, a representative non-parametric estimation method. The kernel density function requires kernel type and bandwidth as hyperparameters instead of parameters. In order to obtain an appropriate continuous distribution, four combinations from the two kernels(Gaussian and Epanechnikov) and two bandwidth computation methods(Silverman's rule and Scott's rule ) are compared. Ship pixels are detected using a threshold value from the appropriate false alarm rate obtained after sensitivity analysis, along with the kernels and the bandwidths.
Bibliographic Citation
International Conference on Aquatic Science & Technology (i-CoAST) 2023, pp.260, 2023
international Conference on Aquatic Science & Technology (i-Coast)
Files in This Item:
There are no files associated with this item.


Items in ScienceWatch@KIOST are protected by copyright, with all rights reserved, unless otherwise indicated.