New AI framework cuts marine species labeling effort by 10x with 10-20 images

A new paper by Alzayat Saleh and Mostafa Rahimi Azghadi on arXiv tackles a critical bottleneck in ecological monitoring: how much labeling effort is needed to deploy automated image recognition at new sites. The authors propose a decision framework that quantifies the trade-off between annotation cost and recognition accuracy when transferring vision systems across marine habitats. Their benchmark spans five datasets, three oceans (Great Barrier Reef, French Polynesia, temperate Danish fjord), and three taxonomic groups (fish, corals, invertebrates). They evaluated four recognition models (DINOv2, CLIP, ResNet-50, EfficientNet-B4) under four adaptation strategies (linear probing, LoRA, Visual Prompt Tuning, full fine-tuning) across three protocols: within-habitat transfer (240 runs), cross-dataset geographic transfer (40 runs), and few-shot adaptation with 0-100 labeled samples per class (648 runs).

The key finding: frozen self-supervised foundation features from DINOv2 paired with a simple linear classifier (just 1,538 trainable parameters) consistently generalized to unseen reef sites as well as or better than fully fine-tuned convolutional baselines that were four orders of magnitude larger. Critically, DINOv2 learned species-diagnostic, habitat-invariant representations, while the baselines learned habitat-specific shortcuts that failed at new locations. The few-shot curves show that just 10-20 labeled images per species enabled reliable recognition at a new site, cutting annotation effort by roughly an order of magnitude. This means ecologists could annotate as few as 10-20 images per species (about 1-4 hours per site) and deploy reliable recognition using an open foundation model. The framework provides a practical budgeting tool for conservation programs expanding to new sites, ecosystems, or platforms.