Toward Real-Time Shipwreck Detection for Autonomous Underwater Vehicles Using Deep Learning: A Model Evaluation Using High-Resolution Bathymetry Data

Autonomous Underwater Vehicles (AUVs) equipped with multibeam echosounders (MBESs) are deployed in oceans in expeditions worldwide to find shipwrecks, as they can survey the seafloor at the resolution required to identify such objects. Due to the severely constrained acoustic bandwidth of underwater communication, it is fundamental to transmit compressed information between the AUV and the support vessel when objects of interest are detected during a mission. This paper presents a systematic evaluation of six YOLO-based configurations combining three architectures with two optimizers, along with selected hyperparameters and bathymetric visualization methods, to identify the optimal model for shipwreck detection suitable for deployment on a deep-water AUV. Such an application has the potential to optimize mapping operations, allowing the mission to be terminated early or to employ adaptive route replanning to maximize survey efficiency. We developed and evaluated an object detection model trained on high-resolution shallow-water open-source data, identifying over 630 shipwrecks along the coast of England to fine-tune the model. Six experiments were conducted and compared across the three most recent YOLO architectures by Ultralytics (v8, v11, v26) and specific hyperparameter configurations. The best-performing model achieved scores above 0.91 in precision, recall, F1, and mAP50, while successfully detecting all prominent shipwrecks in the dataset. Three additional models trained on different data visualizations (hillshade, color scale, shaded relief) demonstrated similar performance. The best-performing model was further tested on a small AUV dataset from the Gulf of America, where it successfully detected the shipwreck in deep-water.