Synthetic aperture radar (SAR) is an important instrument for oceanographic observations, providing detailed information of oceans’ surface and artificial floating structures. Due to advances in SAR technology and deployment of new SAR satellites, an increasing amount of data is available, and the development of efficient classification systems based on deep learning is possible. A deep neural network has improved the state of the art in classification tasks of optical images, but its use in SAR classification problems has been less exploited. In this paper, a full workflow for SAR maritime targets detection and classification on TerraSAR-X high-resolution image is presented, and convolutional neural networks (CNNs) recently proposed in the literature are cross evaluated on a common data set composed of five maritime classes, namely, cargo, tanker, windmill, platform, and harbor structure. Based on experiments and tests, a multiple input resolution CNN model is proposed and its performance is evaluated. Our results indicate that CNNs are efficient models to perform maritime target classification in SAR images, and the combination of different input resolutions in the CNN model improves its ability to derive features, increasing the overall classification score.

The Sentinel-1A is the first of two satellites that composes the Sentinel-1 radar mission. Both satellites operate a C-band synthetic aperture radar (SAR) system to give continuity to the European SAR program. SAR is a flexible sensor able to fulfil users/applications requirements in terms of resolution and coverage thanks to different operational modes and polarizations. With the in-orbit availability of very-high-resolution X-band SAR sensors, the Sentinel-1 satellites have been designed to achieve wide coverage at medium to high resolution. The interferometric wide swath (IWS) mode implemented with the terrain observation with progressive scan (TOPS) technique is the standard acquisition mode over European waters and land masses. IWS in dual-polarization (VV/VH) combination offers 250-km swath at 5 m × 20 m (range × azimuth) spatial resolution. These specifications are in line with the needs of the European Maritime and Security Agency (EMSA) for oil spill and ship detection applications included in the CleanSeaNet program. The main goals of this paper are: assessment of medium-to-high-resolution C-band Sentinel-1 data with very-high-resolution X-band TerraSAR-X data for maritime targets detection; synergetic use of multiplatforms satellite SAR data for target features extraction; evaluation of polarimetric target detectors for the available co-polarization and cross-polarization Sentinel-1A IWS VV/VH products. The objectives are achieved by means of real, almost coincident C-band and X-band SAR data acquired by Sentinel-1A and TerraSAR-X satellites over Gulf of Naples and Catania (South Italy). Furthermore, the obtained results are supported by recorded ground truth vessel reports via terrestrial automatic identification system (AIS) stations located in the area.

Synthetic aperture radar (SAR) ship detection is an important application in the field of maritime security. Azimuth ambiguities caused by the aliasing of the Doppler phase history of each point are often visible in SAR images particularly in ocean areas of low wind speed condition, e.g., in coastal areas, in harbors, etc. The main sources of azimuth ambiguities are man-made metallic structures over the ocean, e.g., ships, oil platforms etc., and over land near the coast, e.g., big tanks, bridges' pylons etc., that have a high SAR backscatter responses. Although the ambiguities' backscatter is generally low, in many situations, it is above the surrounding ocean clutter and are mistaken by classic detection techniques, like constant false alarm rate, as real targets causing false positives. This paper addresses both the discrimination of real targets from non-trivial false positives, namely those due to azimuth ambiguities and the detection itself using a Generalized- K distribution approach. The methodology is firstly proposed and demonstrated over a significant data set of full polarimetric X-band SAR data, which have been acquired by the German satellite TerraSAR-X during the experimental dual receive antenna campaign in April and May 2010. It is based on the intrinsic configuration of monostatic two-channel PolSAR systems and relies on the different signature of azimuth ambiguities in cross-polarized channels. Automatic Identification System messages collected and collocated with the data set analyzed are used as ground truth to evaluate and validate the proposed methodology.

A physical dual-polarimetric model to observe man-made metallic targets at sea in dual-polarimetric coherent X-band synthetic aperture radar (SAR) data is proposed. The model exploits the intrinsic different symmetry properties of man-made targets and sea surface and is tested over actual StripMap TerraSAR-X HH-HV and VV-VH dual-polarimetric SAR data and colocated ground truth measurements. Then, an operational physically based filter to observe targets at sea is proposed. The filter is very attractive in terms of both detection performances and processing time. A typical SAR scene is processed in seconds by a conventional PC processor.

Synthetic aperture radar (SAR) images are operationally used for the detection of oil spills in the marine environment, as they are independent of sun light and weather-induced phenomena. Exploitation of radar polarimetric features for operational oil spill detection is relatively new and until recently those properties have not been extensively exploited. This paper describes the development of a oil spill detection processing chain using coherent dual-polarimetric (copolarized channels, i.e., HH-VV) TerraSAR-X images. The proposed methodology focuses on offshore platform monitoring and introduces for the first time a combination of traditional and polarimetric features for object-based oil spill detection and look-alike discrimination. A total number of 35 feature parameters were extracted from 225 oil spills and 26 look-alikes and divided into training and validation dataset. Mutual information content among extracted features have been assessed and feature parameters are ranked according to their ability to discriminate between oil spill and look-alike. Extracted features are used for training and validation of a support vector machine-based classifier. Performance estimation was carried out for the proposed methodology on a large dataset with overall classification accuracy of 90% oil spills and 80% for look-alikes. Polarimetric features such as geometric intensity, copolarization power ratio, span proved to be more discriminative than other polarimetric and traditional features.