An explainable context-adaptive fusion and expert-in-the-loop evaluation framework for underwater sonar image classification

Sonar image interpretation is critical for identifying submerged objects in underwater exploration. However, conventional deep learning models often fail to capture sonar-specific features such as acoustic shadows and highlights, and they typically operate as black boxes, limiting both performance and trust. To address these challenges, we propose an explainable sonar image classification system that fuses specialized classifiers for shadows, highlights, and general features through a context-adaptive fusion mechanism. The system further incorporates expert-in-the-loop evaluation via the Augmented QUality Assessment for eXplainability (AQUA-X) framework to ensure interpretability and trust. The contributions of this study are threefold: first, the development of a novel Context-Adaptive Fusion Framework (CAFF) that integrates Naive, ShadowNet, and HighlightNet classifiers using an attention-based fusion mechanism; second, the incorporation of Gradient-weighted Class Activation Mapping, SHapley Additive exPlanations, and Local Interpretable Model-agnostic Explanations to provide detailed interpretability of sonar-specific features; and third, expert validation through AQUA-X, enabling comparative assessment and refinement of explanations and identification of the most suitable explainable artificial intelligence techniques for sonar interpretation. Extensive analysis on both real and S3Simulator+ datasets further supports the robustness of the proposed framework. Together, CAFF and AQUA-X promote trustworthy and transparent artificial intelligence solutions for real-world sonar applications.