Deep neural networks have achieved promising results in sonar image classification, but the dynamic nature of underwater environments necessitates class incremental learning (CIL) capabilities to recognize newly emerging targets. Sonar images often suffer from low contrast and high inter-class similarity, which obscure class boundaries and amplify catastrophic forgetting—a challenge that existing CIL methods designed for natural images fail to address effectively. Our analysis reveals that exemplar-based CIL methods better capture subtle inter-class distinctions in sonar data, significantly outperforming exemplar-free alternatives. To tackle these challenges, we propose a Decoupled Feature Adaptive CIL framework that employs multi-scale foundational layers for shared knowledge representation and dynamically generates task-specific feature branches for new classes. The framework further integrates a sonar-specific noise-resistant attention module and achieves a balance between knowledge retention and adaptability through end-to-end joint training. Extensive experiments on the SN-CIL dataset show that our method reduces confusion among similar classes while preserving accuracy on old classes. Additional validation on fine-grained underwater fish classification confirms its superior generalization to high-similarity underwater tasks.