Adaptive Sampling for Seabed Identification from Ambient Acoustic Noise

We study the problem of adaptively obtaining am-bient acoustic measurements via an autonomous underwater vehicle, with the goal of characterizing the geoacoustic properties of the seabed. In contrast to the traditional adaptive sampling scenario, we are provided with sets of snapshots associated with each spatial location, making the problem one of unsupervised learning. We demonstrate how sets of snapshots can be used to obtain noisy pairwise similarities between locations, which can then be used to perform level set estimation to separate the seabed into two highly-distinct types. We propose an adaptive sampling policy that aims to directly reduce the number of locations whose level set membership is uncertain, as well as an approach to minimizing the distance traveled while sampling. Results on synthetic and real-world sediment data demonstrate the benefits of our approach in terms of both accuracy and distance traveled.