An autonomous imaging system for reef fish monitoring in the southern Gulf of Mexico

Introduction Reef fish assemblages are widely used as indicators of coral reef condition. Yet, diver-based surveys provide limited temporal resolution and are constrained by logistical demands. We evaluated an autonomous underwater imaging system based on cameras programmed for unattended image acquisition to monitor reef fish assemblages in the southern Gulf of Mexico. Methods Five terrestrial trail cameras adapted for underwater deployment were operated at Bajo de Diez reef during two seasonal deployments in 2024 and 2025, yielding 191,758 valid underwater images at a nominal 3-min interval. Image quality, camera view, species diversity, and daily assemblage dynamics were evaluated using expert annotations of 27,471 images, iNEXT-4steps, multivariate analyses, and an RF-DETR Medium object-detection model applied to the complete valid underwater record. Results Across the expert-sampled record, the mean daily proportion of images containing at least one fish was 17%, and 65 species were recorded. Laplacian variance distinguished lower- from higher-quality images. Assemblage composition differed clearly between the two deployment contexts and showed smaller but statistically detectable heterogeneity among fixed camera views within each deployment, while daily observations revealed structured temporal variation. On the held-out image-level test set, RF-DETR Medium achieved mAP@50 = 0.741, mAP@50:95 = 0.545, precision = 0.794, and recall = 0.650. Across 68 dates shared by the expert and CV analyses, expert- and CV-derived Bray–Curtis temporal structure was positively correlated (ρ = 0.63; Mantel permutation test, p ≤ 0.0001), with similar concordance for daily richness and the image-based relative abundance index. Applied to nearly seven times the number of manually annotated images, RF-DETR expanded temporal support from 68 to 110 dates. Discussion Autonomous imaging can extend reef fish monitoring beyond discrete surveys through repeated acquisition, while computer vision can reduce manual processing within a human-supervised workflow. Within the site and deployment periods evaluated, combining sustained fixed imaging, expert annotation, and computer vision made fuller use of the image archive and provided daily information on reef fish assemblage structure and its variation through time.