Feature clustering based discrimination of ship targets for SAR images

Synthetic aperture radar (SAR) is one of the most important and widely‐used tools of large scale marine surveillance. Constant false alarm rate is tailored for maritime metallic target detection in SAR images. However, it is very difficult to distinguish ship targets from other metallic targets. Some manual features have been designed to discriminate ship and non‐ship target. These manual features often overfit some scenarios but fail to other situations. Recently, deep convolutional networks have achieved impressive success in optical image classification, which encourages scholars of remote sensing society to leverage this powerful tool to extract robust features of SAR ship targets. This study adopts VGG‐16 to extract features of ship and non‐ship targets. Then, unsupervised clustering is conducted to group ship and non‐ship targets, respectively, using the extracted features, which can effectively tackle intraspecies variants of ship and non‐ship targets. Finally, these representative feature vectors are utilised to differentiate ship and non‐ship targets of other scenarios. The proposed algorithms are extensively tested undertaken on 11 ALOS‐2 SAR images, which demonstrate good performance of the methods.