Seabed Sediment Classification Using Multibeam Backscatter Data Based on LSC Superpixel Segmentation and GWO-XGBoost—A Case Study of the Bohai Strait Cross-Sea Passage

Accurate high-resolution information on seabed sediment distribution is important for marine engineering investigations, submarine pipeline planning, and cross-sea passage assessment. However, multibeam seabed sediment classification is often constrained by the limited number and spatial representativeness of field samples, as well as ambiguous acoustic responses in heterogeneous seabed environments. This study presents a classification framework integrating multibeam backscatter features, topographic features, Linear Spectral Clustering (LSC) superpixel segmentation sample augmentation, and grey wolf optimizer-enhanced XGBoost (GWO-XGBoost), and applies it to the proposed Bohai Strait cross-sea passage area. Nine backscatter texture features and seven topographic features were extracted from 5 m resolution bathymetric and backscatter data acquired using a SeaBat T50-R multibeam system. LSC-based augmentation generated 12,061 pixel-level samples from 52 original grab locations. To prevent information leakage between training and validation data, 51 of the 52 original grab locations were assigned to five folds before augmentation, with all pixels derived from the same grab kept within the corresponding fold; the single sM grab was retained as training support. Under this grab-grouped validation framework, LSC combined with GWO-XGBoost achieved a mean overall accuracy of 72.38 ± 6.33% and a mean Kappa coefficient of 0.6353 ± 0.0877 at the augmented-pixel level. A complementary grab-level evaluation, in which each validation-eligible grab contributed one prediction at its original coordinate, yielded a mean overall accuracy of 58.91 ± 10.29% and a mean Kappa coefficient of 0.4850 ± 0.1341. The classification map indicates that silty sand and slightly gravelly muddy sand are the most widely mapped sediment types in the study area. Gravel-bearing sediments were more frequently mapped in areas associated with relatively steep slopes and channel- or trough-like topography, whereas sandy silt and silty sand were more widely distributed in relatively gentle areas. The results indicate the potential of the integrated LSC–GWO-XGBoost workflow for regional seabed sediment characterization under limited field sampling. Nevertheless, the augmented pixels were derived from only 52 original grab locations, of which 51 were independently evaluated, and therefore do not constitute an equivalent number of independent sediment observations. The classification map should consequently be interpreted as preliminary regional-scale sediment information rather than design-stage engineering ground truth.