Self-supervised vessel detection from low resolution satellite imagery

Maritime surveillance is a growing field of application for optical imaging satellites. In this paper, we aim to increase the practical effectiveness of vessel detection algorithms by developing a computationally inexpensive method that can discern vessels in low resolution optical satellite imagery. This creates an opportunity to host the algorithm on-board small, low cost nanosatellites, generating feasibility for large surveillance constellations. The presented algorithm, which is based upon a deep convolutional autoencoder utilized as an anomaly detector, was found to achieve a detection precision of 98%, and a recall of 79%. Data used to test and train the algorithm was produced by augmenting ESA Sentinel-2 imagery, reducing the resolution to mimic that of a nanosatellite’s. The results presented here are compared with comparable papers in the field, where it is demonstrated that the proposed algorithm is capable of outperforming classical techniques in terms of detection precision. For an equivalent resolution, the proposed algorithm provides a 34% increase in precision over the Constant False Alarm Rate (CFAR) technique.