High-accuracy fish species identification using transfer learning on vision foundation models

Citizen science initiatives play an important role in large-scale monitoring of marine biodiversity, engaging the public in ecological data collection and supporting long-term assessments of species distribution. The Mediterranean Sea, one of the most biodiverse yet heavily impacted marine ecosystems, faces growing pressures from climate change, invasive species, and habitat degradation. To enhance the reliability of observations contributed by non-expert participants, automated tools for species identification are becoming essential. In this study, we compile MEDFISH101, a carefully curated dataset of approximately 70,000 validated images covering 101 Mediterranean fish species. Using this resource, we developed and evaluated a series of deep learning pipelines based on transfer learning of pretrained vision foundation models to achieve accurate species recognition. Our best-performing model, DINOv2-G (self DIstillation with NO labels - Giant) trained via low rank adaptation, reached a Top-1 accuracy of 94.12%, demonstrating that current, state-of-the-art Artificial intelligence (AI) techniques can identify with high probability the correct fish species and thus robustly assist marine biodiversity monitoring. To facilitate transparency, reproducibility, and further community-driven research, a publicly accessible live demo is hosted on HuggingFace.