Hybrid Deep Learning Framework for Underwater Bioacoustic Recognition with Multi-Feature Spectrogram Analysis and Augmentation

The effective identification of marine animal vocalizations is critical for advancing ecological conservation and underwater biodiversity monitoring. This study presents an innovative deep learning framework architecture that integrates a fine-tuned ResNet18-based CNN for spatial feature extraction and a temporal BiLSTM network for temporal acoustic information capture. with high accuracy and robustness. The framework is designed to leverage the complementary capability of convolutional layers (CNN s) to effectively extract spatial and spectral patterns from spectrogram inputs, while Bidirectional Long Short-Term Memory (Bi-LSTM) component is tasked with the temporal dynamics present within audio data. To cope with the challenges arising from the highly variable and noise-rich underwater environment, we incorporate a spectrogram-based data augmentation approach using natural ambient sounds, which effectively enhances the model's generalization to diverse acoustic conditions. Our system is trained on a comprehensive dataset comprising multiple marine species, with both original and augmented samples, to ensure scalability and adaptability across varied underwater soundscapes. Extensive experimental evaluations on real-world marine bioacoustic datasets demonstrate that the proposed hybrid model consistently outperforms baseline CNN-only and Bi-LSTM-only architectures with respect to accuracy, recall, and F1-score metrics. Moreover, the model exhibits strong resilience to environmental noise, validating its applicability for practical, real-time underwater acoustic monitoring tasks. This research contributes a robust, scalable, and introduces a unified deep learning framework that advances the current state-of-the-art in automated marine animal sound classification, enabling more efficient monitoring and conservation efforts in acoustically diverse and increasingly noisy marine habitats.