Deep learning-based detection of buried objects in marine sediments using time-frequency spectrograms from sub-bottom profiling data

• This paper presents a deep learning-based approach for detecting buried objects and obtaining their information in the marine sediments. • A deep learning framework (SBONet) is proposed for detecting buried objects in marine sediments. • A trace-correlated data augmentation strategy enhances weak spectrogram features and mitigates small-sample limitations. • A FEM-simulated dataset provides reliable training data for effective model learning and validation. Accurate detection of buried objects in marine sediments remains a challenging task due to the complex seafloor conditions and the low signal-to-noise ratios of target echo signals. To address this issue, this paper presents a deep learning-based approach for detecting buried objects and obtaining their information in the marine sediments. In this method, time-frequency spectrograms generated from sub-bottom profiling acoustic pressure signals are utilized as input data, and the SBONet architecture is proposed to accurately identify targets. The SBONet employs data augmentation strategies on time-frequency spectrograms to enhance deep-level feature extraction. It incorporates a Transformer encoder for global feature extraction and integrates a Multi-Layer Perceptron for multi-task classification of target material properties, geometric shapes, and burial depths. To address the limited availability of field data, this paper constructs a comprehensive simulated dataset using Finite Element Method modeling of multi-layered marine sediments containing various buried targets. The effectiveness of the proposed method is validated using field data collected from the southern waters near the mouth of Hangzhou Bay. Experimental results demonstrate that SBONet achieves superior prediction accuracy on both noise-free and Gaussian noise-augmented datasets, outperforming existing mainstream models. Additionally, the method exhibits high prediction accuracy when applied to actual sub-bottom profiling data of buried pipelines in the southern waters near the mouth of Hangzhou Bay. The research findings validate the feasibility and effectiveness of integrating physical modeling with deep learning approaches for buried objects detection in complex marine environments.