Robust Wake Detection for Enhanced Maritime Monitoring: Transfer Learning Through a Dedicated Multispectral Ship Wake Data Set

Ship detection has been a topic of great interest since the development of the first remote sensing techniques. The wake generated by moving ships is rich in information regarding the vessel and their instantaneous state. This article tackles the problem of ship wake detection in multispectral (MS) data by deep learning techniques for object detection. This study introduces the multispectral ship wake data set, the first data set of its kind to combine MS imagery with annotated ship attributes, including identity, size, and velocity derived from automatic identification system data. Then, several different architectures are presented, fine-tuned, and benchmarked on the developed data set. The detection analysis is performed on four spectral bands, i.e., B2 (blue), B3 (green), B4 (red), and B8 (near-infrared). These bands are selected due to their spatial resolution, which is the highest available for Sentinel-2 data. Results demonstrate that tailored approaches achieve up to 8% missed detections and below 2% false positives, significantly advancing maritime surveillance by enabling precise wake detection. These improvements facilitate enhanced operational capabilities, such as the identification of dark vessels and bolstering maritime security in real-world scenarios.