Marine Ship Detection from GF-2 High-Resolution Remote Sensing Images with improved YOLOv13 model

Liwei Zhang¹Shang Zhao^1*Da An¹Pengfei Wang¹Bokai Guo¹Haonan Sun^2*

• ¹China Satellite Network Digital Technology Co., Ltd, Hebei, China
• ²China University of Geosciences, Wuhan, China

With the rapid development of global maritime trade and the rising demand for real-time, accurate marine ship monitoring, satellite image-based ship detection has become crucial for marine management and national defense. However, it faces two core challenges: complex backgrounds in high-resolution marine remote sensing images, and great variations in ship sizes—especially difficult small ship extraction. To address these, this study proposes an enhanced method based on improved YOLOv13, using China's Gaofen-2 (GF-2) satellite images. First, GF-2 image data is preprocessed, including radiometric correction to eliminate atmospheric effects, orthorectification to correct image distortion, and fusion of multispectral and panchromatic images to improve spatial resolution and enrich spectral information. Then, three key optimizations are made to the YOLOv13 model: 1) In the backbone network, the A2C2f module is modified by introducing a single-head attention mechanism. By parallelly fusing global and local feature information, it avoids multi-head redundancy and improves the recognition accuracy of small ship targets; 2) In both the backbone and neck networks, the DS C3K2 module is modified by integrating a lightweight attention mechanism, which enhances the model's feature extraction capability in complex backgrounds while reducing channel and spatial redundancy; 3) In the head network, a path-fused Global Feature Pyramid Network (GFPN) is introduced, which leverages skip-layer and cross-scale connections to strengthen cross-scale feature interaction, refine the representation of small ship features, and effectively address the issues of insufficient deep supervision and feature information loss in multi-scale ship detection. Additionally, the improved YOLOv13 model is pre-trained using the open-source DOTA dataset (rich in non-ship negative samples) to enhance its ability to distinguish between ship foreground and background clutter, and then applied to ship detection in segmented sub-images of GF-2 remote sensing images; finally, the detected sub-images are stitched to restore complete regional images. Experiments show that the accuracy rate reaches 96.9%, the recall rate reaches 91.4%, the mAP50 reaches 95.5%, and the mAP50-95 reaches 75.9%, all of which are higher than the mainstream target detection models. It provides a high-performance solution for complex marine ship detection and has important practical significance for both civilian and military fields.