Feature-Enriched Deep Learning: CAPSE-Based Extraction with ResNet50 for Underwater Acoustic Classification

Abstract

Passive acoustic classification acts as a major function in automated ship identification, where deep learning models are used to recognize ship types from radiated noise. A major challenge lies in embedding domain-specific knowledge into these models to improve feature discrimination. This study proposes a classifier that integrates Coherently Averaged Power Spectral Estimation (CAPSE) method with a ResNet50-based classifier. Initially, ships' frequency spectrums are processed using CAPSE analysis, enabling the extraction of key machinery characteristics, and are transformed into LOFAR grams. These time-frequency representations are subsequently processed by a ResNet50 network, which leverages deep convolutional architectures to capture hierarchical feature representations. Taking advantage of transfer learning with ResNet50 deep feature extraction capabilities, the proposed model effectively learns complex patterns within the data, leading to improved classification performance. Evaluation on the standard DeepShip dataset showed that the proposed methodology attained an average classification accuracy and F1-score of 89.93%, while maintaining good generalization and robustness. tSNE feature visualization demonstrated improved class separation by the trained model, with most samples correctly clustered. This study improves underwater acoustic classification by showing that combining deep learning with better feature extraction can lead to highly accurate results.