Object-based sample augmentation improves rare-class seabed sediment mapping from multibeam data in the East Sea (Sea of Japan)

This is the authors' abstract. We don't add key points for this paper.

Seabed-sediment classification from multibeam echo-sounder data is constrained by sparse, class-imbalanced ground truth, particularly for coarse gravel, which pixel-based classifiers struggle to resolve. This study maps the seabed sediments of the southwestern East Sea (Sea of Japan) continental shelf using object-based sample augmentation, which propagates field-sample labels across Simple Linear Iterative Clustering superpixels of calibrated backscatter and bathymetry. Under 20-seed repeated random cross-validation, object-based augmentation raised the rare-class (Gravel) F1-score from 0.036 for a pixel-only baseline to 0.436, and the gain persisted under spatial-block cross-validation, a within-scene spatial-robustness test on spatially held-out sites (0.024 to 0.311), where SLIC exceeded all tested fixed-neighborhood scales. These rare-class estimates rest on only eight independent Gravel field sites and should be interpreted in that context. Oversampling the original rare-class anchors to the same training size recovered only a minority of this gain, indicating that most of the improvement arose from newly propagated pixels rather than anchor replication. Feature-space and multi-sample diagnostics provided empirical support for the within-superpixel label-propagation assumption. The resulting gravel–sand–mud map reproduces a depth-controlled, fining-offshore facies pattern consistent with hydraulic sorting on the narrow, river-fed shelf. An expert-interpreted Rock assessment further showed that exposed bedrock could not be reliably separated from coarse gravel at an acceptable trade-off in the available single-frequency-per-location backscatter–bathymetry feature space. Overall, the approach recovers rare coarse-substrate facies that pixel-based mapping largely misses while identifying the limits of acoustic-only substrate discrimination, thereby strengthening the baseline seabed-sediment maps that underpin benthic habitat mapping, marine spatial planning, and seabed engineering.