Target Detection (Continuation)

Icebergs are pieces of freshwater ice that have broken off from marine glaciers and ice sheets. Calving of icebergs is part of the mass loss of glaciers and ice sheets in Polar regions and play an important role in the global freshwater cycle by delivering freshwater to the regions that are very far from the glaciers and ice sheet margins. Studying the regional distribution of icebergs, their volumes, motions, and their interactions with the ocean, atmosphere, and cryosphere is of interest. This interest is driven by the need for: 1) short-term iceberg drift predictions as a key to prevent damages to ships or oil rigs and 2) more knowledge about the role of icebergs in long-term climate change. Since spaceborne synthetic aperture radars (SARs) can image the vast Polar regions independently of light during the polar night and harsh weather conditions, they are the preferred sensors for iceberg detection and monitoring.  Many studies in the literature have demonstrated the potential of single-channel SAR images for icebergs detection and characterization. However, RADARSAT-2 (RS-2) and the recently launched Sentinel-1 provide quad- and dual-polarization SAR data, respectively. It has been shown in several studies that the additional information contained in multipolarization data improves characterization, classification, and detection of icebergs.

In this study, we first demonstrate the potential of different polarimetric features with respect to improved iceberg discrimination from the surrounding sea ice or open water. To find the feature set to best separate icebergs (large in size) from sea ice and open water, we run a Sequential Forward Feature Selection (SFFS) with a Bayesian classifier. Training regions for different classes (icebergs, sea ice, and open water) are manually selected from the radar polarimetric images. We estimate the probability density functions of the features using Parzen windows with a multivariate Gaussian kernel function and calculate classification accuracies with cross-validation, such that the result is independent of the training. Based on the classification accuracy, we add features one by one  to the chosen feature set, until the accuracy starts to decrease.  The SFFS automatically provides the best feature set for the separation of icebergs from other classes. We use 5 RS-2 quad-pol images in the Bellingshausen Sea and southern Weddell Sea regions for iceberg characterization. However, the detection of small icebergs floating in nonhomogenous sea clutter environments is a challenging task. Based on the best feature set from the SFFS for iceberg characterization, we use our segmentation-based iceberg detection algorithm to show its performance for detecting small icebergs in real complex situations. We test the algorithm with a series of quad-pol RS-2 images containing numerous icebergs broken off from glaciers in Kongsfjorden on Svalbard which cover different sea states, wind conditions, and incidence angles in open and ice-infested water background.

The opportunity to use satellite-based SAR sensors for the monitoring of ship traffic has been researched extensively during the last decades. Most commercial ships are well visible on SAR images due to their excellent backscattering capabilities. However, the direct detection of any kind of maritime vessel is not possible, as non-metallic maritime objects made of materials like rubber or wood are nearly invisible on SAR images. The reason for this is the inherent property of SAR of being most sensible to the availability of conductive materials. Nevertheless, the anomalies on the imaged conductive sea surface structure induced by moving ships can be recognized. For this reason, the consideration of ship wake signatures for the monitoring of ship traffic is of importance. While most automatic ship detection methods search for strong backscattering on the SAR images to identify possible positions of ships, the automatic detection of the ship wake signatures is a more complicated task. Simultaneously to automatic ship detection in the recent decades also the detection of the signatures of ship wakes has been studied. Published approaches often apply Hough transform or image convolution by filter banks.

The executed study elaborates on the detectability of ship wake signatures on TerraSAR-X, RADARSAT-2 and Sentinel-1 images. A binary logistic regression classifier is applied to build a new data-driven detectability model. The classifier is used to calculate probabilities of the visibility wake signatures in the surroundings of verified positions of moving ships, by parameters describing environmental conditions, image acquisition settings and ship properties. As the parameters describing environmental conditions and image acquisition settings are available for each SAR images, the wake detectability model can be applied to control the sensibility of operator-based and automatic wake detection methods. Also drawbacks to the ship velocity can be estimated by reversing the model. An example application will be presented.

Similar to the wake detectability the detectability of ship signatures can be modelled. Such a data-driven approach for ship detectability is also presented and compared to the state-of-the-art simulation-based ship detectability model, which was developed by Vachon. For high and low resolution images the models show similar dependencies of ship detectability.

Maritime surveillance is a very worthwhile application for space remote sensing systems. Indeed, it typically interests wide areas and needs a frequent observation of those areas. In particular, SAR systems fit perfectly these requirements being active sensors that work in any weather and lighting condition and because the resolution of current systems is really impressive even if the swath of the observed scene can be several km wide. Within this framework, the estimation of the moving target velocity in maritime environment represents a point that is still partially open. Despite the existence of several methods and algorithms that encompass both signal and image processing, all of them have different pros and cons. For sure, from the user point of view a single SAR product is generally used for several uses (objects detection, target coordinates extraction, classification…) and therefore the image domain methods can be more affordable. The author developed an innovative method to retrieve the vessel speed that is based on the observation and analysis of the Kelvin wake, that is the hydrodynamic perturbation generated by any target sailing in deep sea conditions (no second order perturbation caused by waves that are reflected by the seabed). The proposed algorithm works efficiently also when the wake structure is partially observable in real SAR images. The literature state of the art in this field was Zilman’s technique based on the application of the FFT on a general line chosen within the Kelvin cone, possibly in proximity of the edges. The uncertainty in the line positioning and sometimes the difficulty in clearly localizing the Kelvin cusps generates a low accuracy in the velocity estimation because the waves distances change significantly with slightly different line directions. Moreover, the well assessed relationship among the vessel speed and the Kelvin wake is based on the measure of the Kelvin wavenumber or, as Zilman does, the cusps distance. The problem of Zilman’s approach is that in real SAR image the aim of measuring the cusps distance, since real SAR images are not always clear, leads to a general peaks distance. On the contrary the proposed algorithm relates the general peak distance with the real wavenumber, that is generally masked by the ship turbulent wake, by the application of a developed model that takes into account the shape of the Kelvin waves. The ongoing activity is the comparison with other existing algorithms that works in image domain (especially the ship-wake displacement) and with the ground truth (AIS data), possibly in different observation conditions (incidence, polarimetry, ship speed, sea state…). In particular, since each image is linked with the corresponding AIS datum, the position and the velocity are interpolated (with good agreement with real values especially when the time gap is little) and then compared with the outcome of different image analysis methods. Preliminary results show that incidence angle appears very critical for Kelvin arms detection, whereas very low vessel velocities implies the absence of turbulent wake, being critical for the application of the ship-wake displacement algorithm.