Title:Capsule and convolutional neural network-based SAR ship classification in Sentinel-1 data

Abstract:Synthetic Aperture Radar (SAR) constitutes a fundamental asset for wide-areas monitoring with high-resolution requirements. The first SAR sensors have given rise to coarse coastal and maritime monitoring applications, including oil spill, ship and ice floes detection. With the upgrade to very high-resolution sensors in the recent years, with relatively new SAR missions such as Sentinel-1, a great deal of data providing a stronger information content has been released, enabling more refined studies on general targets features and thus permitting complex classifications, as for ship classification, which has become increasingly relevant given the growing need for coastal surveillance in commercial and military segments. In the last decade, several works focused on this topic have been presented, generally based on radiometric features processing; furthermore, in the very recent years a significant amount of research works have focused on emerging deep learning techniques, in particular on Convolutional Neural Networks (CNN). Recently Capsule Neural Networks (CapsNets) have been presented, demonstrating a notable improvement in capturing the properties of given entities, improving the use of spatial informations, in particular of spatial dependence between features, a severely lacking feature in CNNs. In fact, CNNs pooling operations have been criticized for losing spatial relations, thus special capsules, along with a new iterative routing-by-agreement mechanism, have been proposed. In this work a comparison between Capsule and CNNs potential in the ship classification application domain is shown, by leveraging the OpenSARShip, a SAR Sentinel-1 ship chips dataset; in particular, a performance comparison between capsule and various convolutional architectures is built, demonstrating better performances of CapsNet in classifying ships within a small dataset.