AI-assisted benthic monitoring at an artificial reef: BRUV platform design and RF-DETR model for Baltic fauna

Summary

This publication presents a comprehensive technical report and trained computer vision model on six different taxa, demonstrating a scalable, cost-effective approach to automated biodiversity monitoring in offshore wind farms using Baited Remote Underwater Video systems (BRUVs) and AI.

Deployed at the "BioBuzReef" artificial reef, built as part of a collaboration between OX2, Under Ytan, and Nemo Seafarms near Björkskär in the Åland archipelago. Recordings were taken in August and September 2025. This study successfully demonstrates that consumer-grade hardware combined with machine learning can achieve broad ecological goals. Our RF-DETR object detection model, trained on Roboflow, achieved 88.8% mean average precision (mAP@50) while reducing video analysis time by over 90%, from days to minutes, through integration with the Swedish Platform for Subsea Image Analysis (SUBSIM).

Eight taxonomic groups were detected over 19 hours of footage, revealing unexpected benthic diversity and enabling quantification of key ecological variables including species richness, temporal abundance patterns, community composition, and species co-occurrence. These metrics provide a baseline for characterising this reef system and form a foundation for assessing biodiversity-positive effects at offshore wind farms by comparison with other areas.

This work demonstrates that AI-boosted BRUV systems offer a practical pathway toward autonomous, non-invasive, and scalable biodiversity monitoring, paving the way for operationalising "cabled observatories" that provide continuous data streams for long-term ecological surveillance, restoration and enhancement in marine renewable energy installations.

Study Site

Location: BioBuzReef artificial reef, Björkskär, Åland archipelago, Baltic Sea
Duration: August - September 2025
Deployments: Nine baited camera deployments in total, of which five were retained for quantitative analysis
Total footage analyzed: 19 hours 

Contents

1. Technical Report (AI_Based_Biodiversity_Monitoring_at_BioBuzReef_Åland_2025_v2.pdf, 19 pages) The main document details the development process including:

• Low-cost BRUV hardware design, costs, and specifications 
• Field deployment protocols and operational procedures
• Semi-automated analysis pipeline to reduce video analysis time drastically
• Baseline ecological assessment and community analysis
• Behavioral observations and community dynamics analysis

2. Trained RF-DETR Computer Vision Model (Baltic_RFDETR_Model.zip)

• Complete model architecture and trained weights
• Annotated training dataset (1,344 frames)
• Performance metrics and validation results
• ROC curves and confusion matrices
• Example detection videos in real-time with annotations

Model Performance

Species Detection Performance

Note: Only six of eight detected taxonomic groups were successfully tracked by the computer vision model.

Impact & Next Steps

This publication provides a ready-to-deploy blueprint for marine biodiversity monitoring, bridging the gap between research prototypes and operational systems. A major barrier to scaling underwater monitoring across the rapidly expanding offshore wind sector or other sectors is achieving cost efficiency and ease of use.

By combining a trained computer vision model with standardized field protocols and an integrated analysis pipeline, this work delivers a fast, accurate, and efficient system for automated biodiversity monitoring. Applied to artificial reef structures, such systems can document how these installations attract and sustain marine life, and when integrated with offshore wind farms, may help offset or even reverse negative ecological impacts.

The openly shared dataset, GitHub resources, and reproducible methods enable replication at other sites and accelerate the development of next-gen monitoring systems. As offshore wind farms expand across European waters, standardized, AI-powered monitoring becomes not only feasible but essential for assessing ecosystem functions and biodiversity-positive outcomes. 

Ultimately, this work moves underwater monitoring toward being fast, replicable, precise, and efficient - qualities essential for large-scale, long-term observation. By automating the process, it reduces the time and effort needed for analysis while opening pathways toward real-time ecological intelligence for adaptive management of marine renewable energy installations.