VHRV: Very High-Resolution Benchmark Dataset for Vessel Detection

10 saat ago Uluslararası Hakemli Dergilerdeki Yayınlar 13 Views

Remote Sensing Applications: Society and Environment 39 (2025) 101641

VHRV: Very High-Resolution Benchmark Dataset for Vessel Detection

Furkan Büyükkanber, Mustafa Yanalak, Nebiye Musaoğlu

Abstract: The veneration for vessel detection from remote sensing imagery is rapidly increasing, carrying immense significance and multifaceted implications across a spectrum of domains in marine applications, encompassing maritime traffic control, anti-illegal fishing measures, oil discharge monitoring, marine pollution prevention, and safety measures. The utilization of deep learning techniques with vessel detection practices inevitably enhances ship detection outcomes. This research proposes a novel and substantial dataset, named as Very-High Resolution Vessels (VHRV), which include wide variety of vessel types, differing vessel sizes acquired under several environmental conditions from inshore to offshore locations for ship detection based on deep learning approaches. State-of-the-art detection – both single stage and two stage – algorithms are evaluated on the proposed VHRV dataset to present a benchmark for deep-learning-based vessel detection methods. The models were trained with diverse hyperparameters and input sizes in both a local and cloud environment to assess the impact of computational resources on detection accuracy. Results indicate that YOLOv9 (0.823), YOLO11 (0.822), and YOLOv12 (0.816) achieved the highest mAP@0.50:0.95 score among single-stage models, while Cascade R-CNN with ResNet-50 (0.668) led the two-stage group. While the choice of hardware did not drastically alter overall detection accuracy, certain architectures, such as YOLOv5, YOLOv7, and YOLOv10, exhibited noticeable improvements in the cloud environment, suggesting that specific model structures benefit more from increased computational capacity. These outcomes emphasize the importance of model architecture and dataset design. The VHRV dataset, with its rich diversity in ship categories and environmental backgrounds, provides a valuable benchmark for training and evaluating deep learning models for maritime applications.