Ship Wake Detection Method Based on Dynamic Graph Convolutional Network

Optical remote sensing images are susceptible to interference from environmental conditions and sea state variations, often resulting in incomplete structural information of ship wakes. This limitation hinders comprehensive modeling of contextual semantic relationships within features, thereby severely constraining the accurate detection of wake targets. To address these challenges, this paper proposes a dynamic graph convolutional network-based method for ship wake detection. By introducing a graph structure modeling mechanism, our method designs graph convolution and graph reconstruction modules to effectively enhance the model's capability to characterize the spatial structural features of wakes. Simultaneously, we construct a cross-layer and cross-space attention module that incorporates an inter-scale feature interaction mechanism within the feature pyramid, enabling dynamic fusion of multi-scale and spatial semantic information. This approach significantly improves wake perception and recognition capabilities in complex backgrounds. Experimental results demonstrate that compared to existing mainstream detection methods, the proposed method achieves higher detection accuracy and stronger robustness across multiple evaluation metrics.