A framework for the annotation and expert validation of Nephrops norvegicus UWTV imagery for machine learning purposes
This paper descibes the methodology and framework for selecting, annotating and expertly validating Nephrops norvegicus UWTV imagery for machine learning purposes. The application of machine learning (ML) and computer vision to biological imagery has increased in recent years, providing opportunities to improve the efficiency and consistency of image-based monitoring. However, the development of supervised computer-vision models remains dependent on high-quality annotated training datasets, the production of which can be labour-intensive and requires careful consideration of how biological features are represented. This is particularly challenging where target features have poorly defined boundaries, vary in appearance and require expert interpretation. These challenges are evident in underwater television (UWTV) surveys for Nephrops norvegicus, where burrow systems are identified and counted to estimate population abundance and density. This study presents a structured framework for developing an expert-validated image dataset for machine-learning-based detection of Nephrops burrow entrances. The framework combines embedding-based selection of visually diverse imagery, model-assisted initial annotation, explicit bounding-box annotation guidelines based on burrow entrance characteristics, and a tiered expert-validation workflow supported by contextual reference imagery. The annotation guidelines formalise the visual information used by expert observers to identify Nephrops burrow entrances, including entrance shape, excavated material and angle of descent, while the validation workflow incorporates two expert reviews with additional review of unresolved cases. The resulting approach provides a transparent and reproducible method for translating specialist biological knowledge into consistently annotated training data. The framework provides a basis for continued development of automated Nephrops burrow detection and is transferable to other image-based machine learning applications requiring specialist interpretation.
Authors
- Mikel Arístegui
- Damian Smyth (ORCID: https://orcid.org/0000-0003-3104-7429)
- Eva Linehan
- Jennifer Doyle (ORCID: https://orcid.org/0000-0002-5012-0946)
- Haleigh Joyce (ORCID: https://orcid.org/0000-0002-3302-8128)
Institutions
- Marine Institute (IE)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24314695
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Marine Institute