How many labels do you need? A decision framework for cross‐habitat marine biodiversity monitoring
Ecological Solutions and Evidence · 2026-10 · 7 (4)
DOI: 10.1002/2688-8319.70348 ↗Kurz gefasst
Fachartikel in der Zeitschrift Ecological Solutions and Evidence, online erschienen am 06.11.2026. Verfasst von Alzayat Saleh und 1 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Ecological monitoring programmes increasingly rely on automated image recognition to scale beyond what human annotators can process, yet ecologists lack evidence‐based guidance on how much labelling effort is required to deploy these systems reliably at new sites. We present a decision framework that quantifies the trade‐off between labelling effort and recognition accuracy when transferring vision recognition systems across marine habitats; the benchmark spans five datasets across three oceans and three broad organism groups (fish, corals and other benthic invertebrates), from tropical reefs in the Great Barrier Reef and French Polynesia to a temperate Danish fjord. We evaluated four recognition models (DINOv2, CLIP, ResNet‐50, EfficientNet‐B4) under four adaptation strategies (linear probing, low‐rank adaptation, visual prompt tuning, full fine‐tuning) across three protocols: within‐habitat transfer across 20 reef sites (240 runs), cross‐dataset transfer along a difficulty gradient (40 runs) and few‐shot curves with 0–100 labelled target samples per class (648 runs). Frozen self‐supervised foundation model features, using DINOv2 with a linear classifier and only 1538 trainable parameters, generalised to unseen reef sites at least as well as fully fine‐tuned convolutional baselines four orders of magnitude larger. Feature analysis showed why: the foundation models learned category‐discriminative representations that are habitat‐invariant, whereas the convolutional baselines encoded habitat‐specific visual shortcuts that fail at new sites. The labelling budget scales with transfer difficulty. Within taxon and biome, 10–20 labelled images per class sufficed (balanced accuracy 0.90 at ), an order‐of‐magnitude saving; across taxa or biomes the same budget reached only 0.79 and 0.68, still rising at . Solution . Monitoring programmes expanding to new sites can deploy reliable automated image recognition by pairing a frozen, openly available foundation model (DINOv2) with a simple linear classifier and annotating 10–20 images per class, roughly 1–4 h per site, where the new site keeps the same categories within the same taxon and biome; wider moves need 50 or more. The framework lets programmes budget labelling effort against expected accuracy and applies to any monitoring programme that must generalise across sites with limited labelled data.
Zitieren
APA
Saleh, A., & Azghadi, M. R. (2026). How many labels do you need? A decision framework for cross‐habitat marine biodiversity monitoring. Ecological Solutions and Evidence, 7(4). https://doi.org/10.1002/2688-8319.70348
Vancouver
Saleh A, Azghadi MR. How many labels do you need? A decision framework for cross‐habitat marine biodiversity monitoring. Ecological Solutions and Evidence. 2026;7(4). doi:10.1002/2688-8319.70348
BibTeX
@article{Saleh2026,
author = {Saleh, Alzayat and Azghadi, Mostafa Rahimi},
title = {How many labels do you need? A decision framework for cross‐habitat marine biodiversity monitoring},
journal = {Ecological Solutions and Evidence},
year = {2026},
volume = {7},
number = {4},
doi = {10.1002/2688-8319.70348}
}
EUVIMED