Answers are grounded in this paper and cite their sources.
Due to the limited amount of ground-truth substrate sampling data and the influence of sample imbalance, semi-supervised seafloor substrate classification methods based on pseudo-labels have received increasing attention. To improve the reliability of pseudo-labeled data, this paper proposes a method for acquiring substrate sample pseudo-labels with consideration of geospatial correlations. First, the geospatial correlation of existing substrate samples is analyzed. Then, the geospatial correlation constraint is combined with supervised classification probabilities to acquire pseudo-labels for augmenting the sample numbers and achieve sample balance across different substrate categories. Finally, different weights are assigned to the pseudo-labels and the ground-truth substrate sampling data for the construction of the semi-supervised seafloor classification model, which could predict the substrate categories in unsampled areas and generates a seafloor substrate distribution map. Experimental results demonstrate that while a fixed pseudo-label strategy can yield an optimistic overall accuracy of 89.58%, a strict leakage-free and fold-wise validation reveals a more realistic accuracy of 79.17%. Notably, this leakage-free performance falls below the supervised baseline with an accuracy of 85.42%, and the bedrock F1-score decreases from 50.00% to 36.36%. These findings highlight the severe optimistic bias inherent in fixed pseudo-labeling and demonstrate that semi-supervised expansion may not surpass a robust supervised baseline in small-sample scenarios, emphasizing the critical need for rigorous leakage-free validation in future studies.
No takes yet. Share an insight, caveat, or question.
Shang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: