Reef-fish RF-DETR: reproducible computer-vision workflow for spatiotemporal reef-fish monitoring

Abstract

This record provides the versioned reproducibility package supporting the manuscript “Scaling spatiotemporal reef-fish monitoring with computer vision in the southern Gulf of Mexico.” The study develops and evaluates a computer-vision workflow for processing sustained underwater imagery of reef-fish communities. An RF-DETR Medium object detector was trained to recognize 51 reef-fish taxon/morphotype categories plus an auxiliary unidentifiable category, representing 52 operational detection categories. The validation-selected model was evaluated on a held-out image-level test set of 520 images containing 1,846 ground-truth objects, achieving mAP@50 = 0.741, mAP@50:95 = 0.545, precision = 0.794, and recall = 0.650. The selected model was subsequently applied to the sustained monitoring archive, processing 198,965 field-image records and generating 81,601 saved detections. The resulting workflow addresses the image-processing bottleneck associated with sustained, high-volume underwater monitoring and provides a reproducible basis for scaling spatiotemporal reef-fish observations. This Zenodo record provides the versioned reproducibility package for the study. It contains the v1.0.0 repository snapshot with the software, documentation, operational class ontology, evaluation outputs, diagnostic results, provenance metadata, archive-scale inference products, and derived data used to reproduce manuscript analyses, together with the validation-selected RF-DETR model checkpoint. The corresponding development repository is maintained on GitHub under edlinguerra/reef-fish-rfdetr. The project also uses cam2model, a companion workflow developed by Edlin José Guerra Castro and Arturo Sanchez-Porras for organizing and supporting reproducible image-to-model processing; cam2model is maintained separately in the arturoSP/cam2model GitHub repository. Third-party reference photographs are not redistributed in this archive. Their provenance is documented through source identifiers and URLs where available. Raw SAMP image collections are not included in this release. This research was supported by the Dirección General de Asuntos del Personal Académico (DGAPA), Universidad Nacional Autónoma de México (UNAM), through the Programa de Apoyo a Proyectos de Investigación e Innovación Tecnológica (PAPIIT), grant IN208124, awarded to Edlin Guerra-Castro.Licensing. Original project code and scripts are distributed under the PolyForm Noncommercial License 1.0.0. Eligible author-generated derived data and metadata are distributed under CC BY-NC-SA 4.0. RF-DETR upstream software and Apache-designated model components are subject to Apache License 2.0. File-specific licensing and attribution information is provided within the archived package.