Rapid Over-Horizon Awareness of AUV Image Datasets over Low Communication Bandwidths

We present a method for flexible over-horizon awareness of autonomous underwater vehicle (AUV) gathered seafloor imagery. The method uses self-supervised learning to create compact dataset summaries. These consist of georeferenced latent representations of images and a subset of human-viewable dataset representative images. The summaries are small enough to send over low-bandwidth satellite networks such as Iridium. This allows seafloor spatial patterns to be visualized and flexibly interpreted within tens of minutes of the AUV surfacing, without the need for physical recovery or proximity to the robot. The method is demonstrated using the University of Southampton’s Smarty200 AUV in a UK Marine Protected Area. Over 4000 images (140 GBytes) were summarised into 51 Iridium messages (102 kBytes), achieving a reduction to 1.4 millionth of the original size. Transmission of the summaries takes ~20mins once the AUV has surfaced, at which point the data can be remotely interpreted by geographically dispersed experts and operators.