Taxonomy-aware deep learning for hierarchical marine species classification in underwater imagery
Dan Zimmerman, Dimitris A. Pados, George Sklivanitis
Read on arXiv →
Key claim
Taxonomy-aware framework improves marine species classification accuracy.
In plain English
Imagine you're trying to identify different species of fish from underwater images. This is crucial for monitoring ocean health, but it’s tough because many species look alike, and the way we collect images can vary a lot. Current methods often fail when the conditions change or when the species are very similar, leading to misidentifications. This is known as domain shift and fine-grained visual similarity issues.
To tackle these problems, the authors developed a new framework that takes into account the hierarchical structure of biological classification. Instead of treating all species equally, it uses a special loss function that weighs the importance of different taxonomic ranks and employs a smart inference method that minimizes risk based on this structure. This means that the model learns to classify species not just as individual entities but as part of a larger family tree, which helps it generalize better across different conditions.
In practice, this new approach was tested on a dataset with 79 marine classes and showed a mean taxonomic distance very close to the top-performing solution. This indicates that it can effectively handle the challenges of species classification in real-world scenarios, making it a valuable tool for anyone involved in marine conservation and biodiversity monitoring.
The approach introduces a taxonomy-aware framework that aligns with biological classification, which is a significant extension in the domain of species classification.
The evaluation on a comprehensive dataset with competitive results supports the claims made about the framework's effectiveness.
Deep reliability assessment
The methodology supports the narrower claim that metric-aligned inference and simple taxonomy-aware components improve mean taxonomic distance on FathomNet-style hierarchical marine classification under domain shift. The broader claim that these design lessons generalize to other hierarchical imaging domains like ATR/ISR is plausible but overextended because both main evaluation sets still come from FathomNet and the backbone comparison is partly confounded by different fine-tuning recipes.
Reproducibility
Partial: the FathomNet 2025 dataset is public via Kaggle, and the authors state that source code and configuration files will be released at https://github.com/C2A2-at-Florida-Atlantic-University/fathomnet-taxonomy upon publication.
Key figure
The key architecture is a DINOv2-Base image backbone with multi-scale feature encoding feeding independent classification heads for different taxonomic ranks, trained with taxonomy-weighted loss and decoded using minimum-risk inference over the hierarchy.