Research on Acoustic Detection Methods for Submerged Buried Targets Using Big Data Transfer Learning-based Full Waveform Inversion

Acoustic detection of submerged buried targets using CFWI suffers from low computational efficiency and poor inversion performance. To address these limitations, this study develops a BFWI method based on big data transfer learning and deep learning algorithms. Theoretically, this paper introduces the fundamental principles and overall architecture of BFWI. Using 49,100 data samples from the SEG/EAGE salt model dataset and 9,110 data samples from a synthetic model dataset, the model undergoes big data transfer learning for training and validation, achieving optimal CNN parameter weights. Generalization capability of BFWI is tested with 4,000 data samples under three typical scenarios: layered structures, intermediate dense models, and bottom protruding models. In numerical simulations and field experiments, both CFWI and BFWI methods are implemented using three iterative approaches: L-B, G-N-1, and G-N-2. Results demonstrate that BFWI achieves faster computation, better robustness, and an overall improvement in computational efficiency by approximately one-third compared to CFWI. Furthermore, BFWI produces fewer sidelobe artifacts and yields clearer inversion images of submerged buried targets, providing a novel perspective and methodology for acoustic detection of submerged buried objects.