Subspace clustering for situation assessment in aquatic drones

We propose a novel methodology based on subspace clustering for detecting, modeling and interpreting aquatic drone states in the context of autonomous water monitoring. It enables both more informative and focused analysis of the large amounts of data collected by the drone, and enhanced situation awareness, which can be exploited by operators and drones to improve decision making and autonomy. The approach is completely data-driven and unsupervised. It takes unlabeled sensor traces from several water monitoring missions and returns both a set of sparse drone state models and a clustering of data samples according to these models. We tested the methodology on a real dataset containing data of six different missions, two rivers and four lakes in different countries, for about 5.5 hours of navigation. Results show that the methodology is able to recognize known states "in/out of the water", "upstream/ downstream navigation" and "manual/autonomous drive", and to discover meaningful unknown states from their data-based properties, enabling novelty detection.