Computer Vision Framework for Recreational Boat Speed Estimation and Corresponding Noise Mapping from Satellite Imagery

Abstract

Recreational boats are an unnegligible source of underwater radiated noise in coastal areas, especially during the summer season. Traditional noise mapping methods mainly rely on AIS-based ship position and speed information, which works better for merchant ships than for small recreational vessels. However, most small recreational boats are not equipped with reliable Automatic Identification System. This creates a data gap for vessel monitoring and underwater noise assessment. This thesis investigates whether AI-based computer vision methods and machine learning can be used to detect recreational boat wakes from optical satellite imagery and support speed-based noise mapping.

The proposed framework is based on Sentinel-2 L2A optical imagery with 10 m spatial resolution. A wake detection dataset was built from satellite images and annotated for YOLO-based object detection. The detection model was designed for small and weak wake features, using multi-scale inference and sliced image prediction. To support speed estimation, detected wakes were linked for ships with AIS records using geolocation and imaging timestamps. The matched samples were then processed into a speed-wake dataset, including denoised, feature-enhanced and augmented wake images. A CNN-based regression model was used to estimate vessel speed from visual wake features.

The results show that YOLO-based models, as machine learning based computer vision methods, can be applied to detect wake-like vessel traces that are missing from AIS-based observations. In one case study, the YOLO-based detector identified substantially more vessel wake objects than the available AIS records in the same satellite scene. A high percentage of the vessel wake objects identifiable with the human eyes were detected. The detected objects were initially matched with AIS records to extract vessel speed information, where only a small proportion of the corresponding AIS records were found. The speed of the vessel wake objects without available AIS records were then estimated using a CNN-based speed regression model. The complete vessel speeds were further used to produce a simplified underwater noise map, using an empirical noise model and an acoustic propagation model. These results suggest that AI-based wake detection and vessel speed estimation can provide a useful complementary data source for monitoring recreational boat activity and assessing its potential noise impact in coastal waters. The approach is still limited by image resolution, cloud cover, sea clutter, wake overlap, and the availability of reliable speed labels. However, it provides a practical basis for future research on small-vessel monitoring from remote sensing data.