Abstract:
Large area monitoring plays an important role in the Maritime Domain Awareness initiative. To effectively monitor long coastlines and further out at sea, a multitude of monitoring techniques are necessary. One of these techniques, satellite Synthetic Aperture Radar (SAR), can monitor large areas independent of weather or time of day. SAR imagery is particularly useful in the tracking of ships at sea as ships are highly reflective objects and become visible against the dark ocean background. Two novel SAR ship detection techniques are proposed, both of which are tested against a newly created 46 image, medium resolution SAR imagery dataset. The first method extends the conventional Constant False Alarm
Rate prescreening method to allow per-pixel thresholding so thresholds can be adjusted for specific areas. The method makes further use of Simulated Annealing to help identify areas of probable ships using an auxiliary transponder dataset. This technique improves the flexibility of previous Constant False Alarm Rate-based methods and provides a mean DA of 87.51%, a mean FAR of 5.644 × 10^−8 and mean MCC of 0.80. The second method uses unique ship-like, rapidly calculable features known as Haar-like features to describe ships. These features are used to train a cascaded classifier to create a ship discrimination step. The combination of these aspects and an advanced training technique known as AdaBoost creates a method which can be efficiently applied to medium resolution SAR imagery to provide significant false alarm reduction whilst maintaining a high ship detection accuracy which provides the best results of the methods investigated with a mean DA of 88.71%, mean FAR of 1.940 × 10^-8 and an MCC of 0.89. These two methods are evaluated against a range of other ship detection methods using various standardized metrics and multiple test scenarios.
