<p>Maritime ship detection plays an important role in maritime surveillance. Conventional ship detection methods based on Synthetic Aperture Radar (SAR) images often rely on manual feature extraction, which can lead to limited detection performance and low computational efficiency. With the development of deep learning, new methods have emerged that can improve ship detection in SAR imagery. To address challenges such as complex background interference and scale variation, this paper proposes a SAR ship detection model named DA-YOLO. To improve adaptability to varying backgrounds, a Dynamic Adaptive Backbone Network (DABN) is introduced. This component adjusts convolution kernel parameters dynamically according to input features, aiming to improve the flexibility and effectiveness of feature extraction. To enhance the model’s ability to extract multi-scale features, a Multi-Scale Kernel Attention (MSKA) module is added. It uses large-size separable convolution kernels to capture both global and local information. Additionally, a Global Dynamic Feature Pyramid Network (GDFPN) is designed to fuse features from different layers and retain fine-grained target details through an adaptive upsampling strategy. Experiments are conducted on the LS-SSDD-v1.0 dataset, with additional evaluations on the HRSID and SSDD datasets to verify generalization performance. The results show that DA-YOLO achieves mAP improvements of 2.8% on LS-SSDD-v1.0, 1.1% on HRSID, and 0.5% on SSDD compared to the baseline model, indicating more effective detection and improved cross-dataset adaptability. Moreover, the F1-score on LS-SSDD-v1.0 increases by 3.4%, further demonstrating an improvement in overall model performance.</p>