An interpretable hydrodynamic intelligence framework for underwater species detection using biomimetic sensing and hybrid deep learning models

Underwater marine organism detection is essential for biodiversity monitoring, ecological assessment, and autonomous ocean exploration; however, conventional vision-based systems often experience performance degradation in low-visibility and dynamic underwater environments. Existing studies primarily rely on image-based detection models and provide limited integration of hydrodynamic-inspired sensing mechanisms for intelligent species monitoring. To address these limitations, this study proposes a novel biomimetic hydrodynamic sensing framework inspired by the biological lateral line system of fish. The framework introduces a Hydrodynamic Signature Encoding Layer (HSEL) for structured feature representation and integrates ConvNeXt for feature extraction, TabNet for interpretable feature selection and classification, and K-Means clustering for exploratory pattern discovery. The model was implemented using Python with TensorFlow, PyTorch, Scikit-learn, and OpenCV. Experiments were conducted using the Oceanic Life Dataset containing 7,990 marine organism images and the UCI Yacht Hydrodynamics Dataset as auxiliary hydrodynamic information. Results demonstrated an accuracy of 97.23%, precision of 96.80%, recall of 97.10%, and F1-score of 97.00%, achieving a 0.49% accuracy improvement over the best baseline model (96.74%) and a 9.24% improvement over conventional CNN-based approaches. The findings indicate that the proposed framework provides an effective and scalable solution for intelligent marine organism monitoring, biodiversity assessment, and digital twin-enabled ecosystem observation.