Examining methods for maximising ship classifications in maritime surveillance

Both Royal Australian Air Force (RAAF) and Border Protection Command (BPC) aircraft fly maritime surveillance missions for the Australian Government on a regular basis. These missions involve searching particular areas of interest (AIs) for illegal fishing or people smuggling activities. In this work a square AI is considered, with an aircraft tasked to search for ships using its sensors (e.g., radar, electrooptical). Waypoints (points that must be visited) are included in the AI to ensure that the entire AI is covered. The aim for a particular search is to detect and classify as many ships as possible, while doing so in the shortest time. Depending on the ship density, the aircraft may not have time to search the entire AI. In this paper an augmentation of the traditional Travelling Salesman Problem (TSP) is considered, where the ships (cities) move with random velocities (dynamic TSP), have different start and end points (open TSP) and there is incomplete a priori knowledge of the problem space (on-line TSP), making this problem much more challenging. Earlier work (Marlow et al, 2007) considered a “baseline” case of an S-shaped search pattern and a default heading (the route flown when there are no ships currently detected) direct to the next waypoint. This paper considers three extensions with an aim to increase the level of ship classifications: these are 1) alternative initial flight paths, 2) alternative default headings and 3) including “ghost ships” in the search. The principle behind the S-shaped pattern is that the aircraft will cover the entire AI with its sensors, giving aircrew the best chance of detecting all ships in the AI. This paper considers alternative spiral waypoint patterns, both an “inspiral” (from one corner of the AI, spiraling towards the centre) and an “outspiral” (the reverse). These approaches are theoretically more likely to detect ships that enter the AI during the mission. The direct-to-waypoint default heading will minimise the travel time, but it also may potentially result in the aircraft not covering the entire AI with its sensors, particularly if it has already been diverted significantly from the “wayline” (the direct line between waypoints). In this work, a perpendicular return to the wayline is considered, which increases the distance travelled but is also likely to increase the probability of detecting more ships. A third option is also considered in which the aircraft continually aims for the midpoint on the wayline of the perpendicular intercept point and the waypoint. The object of ghost ships is to direct the aircraft to fly to areas of the AI that it may not otherwise visit. This may particularly be the case in low-density environments where the aircraft has to substantially divert from the wayline to classify ships and, in returning, inadequately cover other areas where ships may be present. Ghost ships remain in the current tour until they are “detected”, whereupon they are removed. Results suggest that, at lower densities, the perpendicular default heading and including ghost ships provide an overall improvement in classifications of the order of a few percentage points, which in real terms translates to an extra 1-2 ships on average. Significantly it also translates to an increase in the percentage of cases where 100% classifications are achieved. These improvements generally come at a cost of increased distance travelled by the aircraft and thus greater fuel consumption. In the case of ghost ships, there is an additional cost in computational time due to the requirement to include them in the tour. Beyond the critical ship density (where classifications cannot physically reach 100%), these variations offer no real advantage and in some cases are counter-productive, so the baseline pattern is more appropriate in these circumstances.