Enhancing ship recognition in satellite imagery: a YOLO-centric framework based on the updated ShipRSImageNet benchmark

To better monitor human maritime activities, this study presents an updated version of the ShipRSImageNet benchmark and evaluates recent fine-grained ship recognition models on high-resolution optical satellite imagery. The updated benchmark provides corrected and refined annotations with enhanced consistency. Approximately one-third of the samples (1206 out of 3435 images) have undergone correction, thereby improving label reliability for both model training and performance benchmarking. Based on this updated dataset, You Only Look Once (YOLO)-based models improved mean average precision (mAP) from the previously reported 0.66–0.70 to approximately 0.80, with improved precision and recall. Comparison experiments show that this gain mainly results from more reliable training supervision enabled by improved annotation quality. The normalized confusion matrices further indicate that the models can distinguish most ship categories under complex coastal backgrounds and diverse viewing geometries. Remaining errors are mainly caused by fine-grained class confusion, false positives in harbor environments, and missed detections of small vessels in crowded nearshore scenes. The updated ShipRSImageNet benchmark and ship recognition models developed in this study can support maritime activity monitoring, port vessel analysis, and non-cooperative vessel surveillance.