Newfoundland Marine Refuge Fish Classification Dataset (N-MARINE) (Papers Track) Spotlight
Kameswari Devi Ayyagari (Dalhousie University); Maurice Drautz (Dalhousie University); Daniel Porter (Fisheries and Oceans Canada); Joshua Barnes (National Research Council Canada); Corey Morris (Fisheries and Oceans Canada); Christopher Whidden (Dalhousie University)
Abstract
Scaling marine ecosystem monitoring is increasingly urgent as ocean warming accelerates ecological change. Current techniques like trawling are invasive, harmful to habitats, and infrequent, leading to missed shifts in species distributions such as the snow crab collapse in the Bering Sea after the 2018–2019 marine heatwave and the spread of invasive European green crab in Atlantic Canada. Machine learning offers a pathway to scalable, automated monitoring, but progress in regions such as the North Atlantic has been constrained by the scarcity of annotated data. We present N-MARINE, the first open-source North Atlantic underwater image dataset for groundfish detection and classification. It contains 23,936 ecologist-annotated ocean-floor camera images across nine species, with bounding boxes, standardized splits, and YOLOv7 baselines (best $\mathrm{mAP}@0.5=0.808$). N-MARINE provides a foundation to build upon and advance research on generalizable visual representations and transfer learning for regional underwater ecosystems.