Enhancing seabed sediment classification with multibeam echo-sounding and self-training: a case study from the East Sea of South Korea

Introduction Accurate classification of seabed sediments is essential for marine spatial planning, resource management, and scientific research. While direct sampling yields precise sediment information, it is costly and spatially limited. Multibeam echo-sounding systems (MBES) offer broad coverage but lack detailed sediment characterization, creating a need for an integrated, data-driven approach. Methods We developed a machine-learning framework that fuses MBES backscatter data with limited seabed samples. Missing MBES values were first interpolated using a U-Net model to create a complete raster dataset. Advanced texture and spectral descriptors—Gray-Level Co-occurrence Matrix, Law’s texture filters, and discrete wavelet transforms—were extracted from the backscatter imagery. Five classifiers (Random Forest, Support Vector Machine, Deep Neural Network, Extreme Gradient Boosting, Light Gradient-Boosting Machine) were trained to predict four sediment classes (gravel, sand, clay, silt). To mitigate sample scarcity and class imbalance, a semi-supervised self-training loop iteratively added high-confidence pseudo-labels to the training set. Results Field validation in the East Sea (Republic of Korea) showed that the Extreme Gradient Boosting model achieved the highest accuracy. Overall prediction accuracy increased from 60.81 % with the baseline workflow to 72.73 % after applying data interpolation, enhanced feature extraction, and self-training. Discussion The proposed combination of U-Net interpolation, multi-scale texture features, and semi-supervised learning significantly improves sediment classification where MBES data are incomplete and sediment samples are sparse. This integrated workflow demonstrates the potential of machine-learning techniques to advance seabed mapping and support informed marine resource management.