ADDRESSING SEABED CHARACTERIZATION AS FUZZY DEEP LEARNING SEGMENTATION TO MITIGATE AMBIGUOUS SYNTHETIC APERTURE SONAR DATA

Seabed characterization is critical for mine countermeasures planning and evaluation, and this study extends prior efforts addressing it as deep learning segmentation with synthetic aperture sonar data. Although traditional crisp annotations have yielded relatively reliable results, they fall short in capturing the complexity and diversity of the seabed, particularly in environments with mixed compositions, leading to cryptic errors. To address these challenges, this work introduces homogeneous patch resampling, an incremental improvement that balances the input data distribution by selecting more heterogeneous samples. Furthermore, a novel fuzzy label pre-processing approach is proposed, which approximates the density membership of each seabed class within a Region of Interest. This approach is compared against a benchmark of standard deep learning soft label training regularization strategies. Both methods outperform the baseline, and the benchmark highlights the efficacy of fuzzy label training strategies in managing ambiguous sonar data.