A comparative study of vision foundation models for few-shot classification in benthic imagery
Frontiers in Marine Science · 20.11.2026 · 13
DOI: 10.3389/fmars.2026.1925773 ↗Kurz gefasst
Fachartikel in der Zeitschrift Frontiers in Marine Science, online erschienen am 20.11.2026. Verfasst von Hassan Elkholy und 2 weiteren Autor:innen. Die Arbeit nennt 44 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Benthic imagery constitutes a primary data source for studying seafloor habitats, providing rich visual information essential for ecological assessment and marine ecosystem monitoring. Recent advances in underwater robotic platforms have substantially increased the capacity for large-scale benthic image acquisition, shifting the primary bottleneck from data collection to image annotation and analysis. However, manual annotation of benthic imagery remains labor-intensive, time-consuming, and dependent on domain expertise, limiting the scalability of marine habitat analysis. Although artificial intelligence (AI)-assisted annotation tools have shown promise, their application to benthic imagery remains challenging due to the substantial domain gap between underwater imagery and conventional terrestrial computer vision datasets. In addition, annotated benthic datasets are often limited in size, geographically sparse, and highly imbalanced. Few-shot learning (FSL) offers a promising solution by enabling rapid adaptation to new datasets using only a small number of labeled examples. In this work, we propose a foundation-model-based few-shot classification (FSC) framework for benthic imagery annotation. The proposed framework uses pretrained visual representations and systematically investigates the influence of backbone family, model scale, support-set construction strategy, and adaptation algorithm under limited labeled-data conditions. Furthermore, we investigate a prioritized support-set construction strategy based on embedding-space clustering to improve annotation efficiency in low-shot regimes. Experimental results across three benthic imagery datasets demonstrate that DINOv3 ViT-S/16 provides the strongest overall classification performance among the evaluated backbone families, while increasing backbone size does not consistently improve accuracy. LinearSVC also demonstrates strong and consistent adaptation performance across the datasets. In addition, the proposed prioritized labeling strategy improves classification accuracy compared with class-balanced random sampling on Datasets B and C, with average gains of 7.62 and 5.02 percentage points, respectively, and corresponding 95% confidence intervals excluding zero, while the improvement on Dataset A is inconclusive. However, these accuracy gains do not consistently translate into higher macro-averaged F1 scores, highlighting the dataset-dependent effects of representative sample selection. Beyond classification performance, this work introduces an annotation-time modeling framework to analyze the practical deployment trade-offs of few-shot benthic annotation workflows. Under the evaluated annotation-time assumptions, FSC is estimated to reduce overall annotation time by 6.7% and 21.7% on Datasets A and B, respectively, compared with traditional fine-tuning approaches. On Dataset C, traditional fine-tuning shows only a small and uncertain estimated advantage of 1.3%. Overall, the results suggest that DINOv3-based few-shot classification provides an effective low-label annotation-support approach, with potential annotation-time advantages that depend on dataset size, classification accuracy, and annotation costs.
Zitieren
APA
Elkholy, H., Pizarro, O., & Ludvigsen, M. (2026). A comparative study of vision foundation models for few-shot classification in benthic imagery. Frontiers in Marine Science, 13. https://doi.org/10.3389/fmars.2026.1925773
Vancouver
Elkholy H, Pizarro O, Ludvigsen M. A comparative study of vision foundation models for few-shot classification in benthic imagery. Frontiers in Marine Science. 2026;13. doi:10.3389/fmars.2026.1925773
BibTeX
@article{Elkholy2026,
author = {Elkholy, Hassan and Pizarro, Oscar and Ludvigsen, Martin},
title = {A comparative study of vision foundation models for few-shot classification in benthic imagery},
journal = {Frontiers in Marine Science},
year = {2026},
volume = {13},
doi = {10.3389/fmars.2026.1925773}
}
EUVIMED