DEEP LEARNING FOR VESSEL DETECTION AND IDENTIFICATION FROM SPACEBORNE OPTICAL IMAGERY

• Main Authors: G. Matasci, J. Plante, K. Kasa, P. Mousavi, A. Stewart, A. Macdonald, A. Webster, J. Busler
• Format: Article
• Language: English
• Published: Copernicus Publications 2021-06-01
• Series: ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
• Online Access: https://www.isprs-ann-photogramm-remote-sens-spatial-inf-sci.net/V-3-2021/303/2021/isprs-annals-V-3-2021-303-2021.pdf
• ISSN: 2194-9042; 2194-9050

We present a deep learning-based vessel detection and (re-)identification approach from spaceborne optical images. We introduce these two components as part of a maritime surveillance from space pipeline and present experimental results on challenging real-world maritime datasets derived from WorldView imagery. First, we developed a vessel detection model based on RetinaNet achieving a performance of 0.795 F1-score on a challenging multi-scale dataset. We then collected a large-scale dataset for vessel identification by applying the detection model on 200+ optical images, detecting the vessels therein and assigning them an identity via an Automatic Identification System association framework. A vessel re-identification model based on Twin neural networks has then been trained on this dataset featuring 2500+ unique vessels with multiple repeated occurrences across different acquisitions. The model allows to naturally establish similarities between vessel images. It returns a relevant ranking of candidate vessels from a database when provided an input image for a specific vessel the user might be interested in, with top-1 and top-10 accuracies of 38.7% and 76.5%, respectively. This study demonstrates the potential offered by the latest advances in deep learning and computer vision when applied to optical remote sensing imagery in a maritime context, opening new opportunities for automated vessel monitoring and tracking capabilities from space.