Sensitivity to Target Behaviour in Automatic Classification on Kinematic Track Features

During anti-submarine warfare operations, the use of high resolution broadband sonars often result in a false alarm cluttered sonar image. Correctly distinguishing between a sonar contact and a false alarm is a vital and complicated task for the sonar operator, which is only made more challenging in a cluttered environment. The origins of false alarms may be diverse and display a variety of different sources, from seamounts and underwater ridges to man-made objects, such as pipelines. Commonly today, the unmanagable amount of false alarms is handled by setting a detection threshold on the signal-to-noise ratio, removing a large number of false alarms, but unfortunately also lowering the detection probability of a valid target. Recent studies have shown that at track level, false alarms from topographical features, such as seamounts, display very different kinematic properties when compared to tracks originating from pipelines. This can be further exploited with the use of machine learning in order to automatically separate and classify tracks from different sources. Here we apply the mentioned method of kinematic track features and machine learning to a dataset which contains a variety of different environments and is populated by twelve synthetic submarines. Predictions on the detections of the synthetic submarines was done by a sound propogation model, Lybin. These were added to the original dataset, which was then processed up to track level before the kinematic properties of both false tracks and target tracks were extracted. Due to the confidential nature of actual submarine tracks, these are removed from the dataset before the addition of synthetic submarines. An important find was that training on a variety of navigational patterns resulted in a more robust set of classifiers, giving the best overall classification performance across all submarine types. In addition, we see that when training on a single navigational pattern, training on a more complex pattern will yield a better classification performance than training on a simple maneuvering pattern. Finally, several results emerging from this study are in agreement with previous results from studies on automatic classification and kinematic features of pipelines in various environments.