Design and Verification of an Explainable Target Recognition Architecture for Water Surface Situation Awareness

Objectives To address the problems of the lack of a unified integration framework, the limited effectiveness of stepwise optimization, and the lack of traceability in the recognition process in existing water surface target recognition methods, a layered and decoupled explainable target recognition architecture is proposed. Methods A four-layer architecture consisting of perception input, feature enhancement, intelligent recognition, and decision interpretation is constructed, with multiple functional modules coordinated through unified interface specifications. An enhancement–recognition collaborative optimization method is designed by embedding the lightweight AOD-Net dehazing network into the architecture and jointly training it with YOLOv3, where the parameters of both networks are jointly constrained by the recognition loss. IA-YOLO is introduced as an architectural instance with a different enhancement–recognition implementation to evaluate the adaptability of the proposed architecture to different module combinations. Furthermore, a spatial-response visualization and quantitative analysis method for target recognition results is developed based on Grad-CAM to provide traceable analysis of the water surface target recognition process. Results Experiments are conducted on the SeaShips7000 dataset under different fog-density conditions. The collaborative model achieves an overall mAP@0.5 of 0.5893, representing an improvement of 1.88 percentage points over the YOLOv3 baseline. Across different fog-density conditions, the EBPG (Energy-based Pointing Game) of the collaborative model remains above 0.915, reaching 0.9268 under heavy-fog conditions, indicating that its spatial response is well concentrated within the target region. Conclusions The proposed architecture integrates environmental enhancement, target recognition, and result interpretation into a unified framework. While improving recognition performance, it also provides explainable support for decision making, providing a feasible technical approach for the modular development, deployment, and continuous optimization of water surface situation awareness systems.