🧬 biology

Automated estimation of gadoid and flatfish abundance in stereo-BRUV footage using YOLO object detection

This study presents a YOLO-based automated pipeline that successfully detects gadoid and flatfish in stereo-BRUV footage from North Sea wind farms, introducing a custom metric to validate its ability to efficiently estimate fish abundance and demonstrating that combining dual camera viewpoints enhances the sensitivity of population estimates compared to single viewpoints.

Original authors: Mario Lambrette, Anthony W. J. Bicknell, Samuel Gierhart, Matthew J. Witt

Published 2026-10-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mario Lambrette, Anthony W. J. Bicknell, Samuel Gierhart, Matthew J. Witt

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The ocean floor is a vast, shifting landscape where life thrives in ways that are often invisible to the human eye. To understand how fish populations change over time, scientists need to count them, but doing so underwater is notoriously difficult. Traditional methods often involve dragging nets or traps across the seabed, which can harm the very ecosystems researchers are trying to study. A gentler alternative has emerged: baited remote underwater video. Scientists lower a camera system, weighted with a small amount of fish bait, to the ocean floor. The scent of the bait draws in local fish, and the camera records them for an hour or more. This method provides a non-invasive window into marine life, but it comes with a massive bottleneck. A single survey can generate hundreds of gigabytes of video footage. To get a count of the fish, a human expert must sit and watch every minute of that footage, frame by frame, to tally the maximum number of individuals seen at once. This manual process is slow, expensive, and limits how much data can ever be analyzed.

A team of researchers at the University of Exeter has tackled this bottleneck by teaching computers to do the counting. They developed a specialized artificial intelligence system designed to watch underwater video and automatically identify and count two specific groups of fish: cod-like fish, known as gadoids, and flatfish. The study took place in the northern North Sea, an area dotted with offshore wind farms where the team was monitoring how these massive structures affect local fish life. Instead of relying on a single camera angle, the team used a stereo system, which records the scene from two slightly different perspectives simultaneously, much like human eyes do. By training their AI on thousands of annotated video frames, they created a pipeline that could scan hours of footage in a fraction of the time it would take a human, producing reliable counts of fish abundance.

The researchers did not just build a detector; they built a system that understands the specific challenges of underwater counting. In a crowded video frame, fish often hide behind one another, making it hard to tell exactly how many are present. The team trained their model to recognize that as the number of fish increases, the difficulty of counting them perfectly also increases. They introduced a new way of measuring success that accounts for this reality, accepting that a small margin of error is natural when fish are packed tightly together. When they tested their system against the counts made by human experts, the results were striking. The computer's estimates matched the human counts with very high accuracy, correctly identifying the presence and approximate numbers of flatfish and cod-like fish in nearly every case. The system was so effective that it could replace a significant portion of the manual work, cutting the time needed to process survey data from months down to weeks.

One of the most surprising discoveries came from the way the system used the two camera angles. In traditional surveys, scientists usually pick just one camera view to count the fish, treating the second camera as a backup or for measuring fish length. The researchers found that by letting the computer look at both angles and simply taking the higher count from the two, they got a much clearer picture of the fish population. This approach, which they called the "maximum count," reduced the noise and uncertainty in the data. It made the survey results more sensitive to real changes in the environment. For instance, when analyzing how fish numbers changed with distance from the wind turbine bases, the two-camera method revealed a clear pattern that the single-camera method missed. The flatfish, which tend to stay low and blend in with the seabed, were particularly hard to spot from just one angle, but the dual-view approach captured them much more effectively.

The study also highlighted the practical benefits of this automation for long-term monitoring. While setting up the system required an initial investment of time to label the training data and tune the software, the payoff was immediate in the field. The team was able to process the 2024 survey data in just 68 days, compared to 134 days for the previous year's survey, which relied entirely on manual counting. This speed does not just save time; it allows scientists to ask more complex questions about how marine ecosystems respond to human activity. The researchers noted that while their system is currently tailored to the specific fish and conditions of the North Sea, the framework is open and accessible. It proves that with the right tools, the massive volumes of video data collected by marine scientists can be turned into actionable knowledge without being buried under the weight of manual labor.

The success of this project suggests a new path forward for marine conservation. By automating the tedious task of counting, scientists can focus their energy on interpreting the data rather than just collecting it. The ability to detect subtle changes in fish abundance, such as the decline in numbers further away from wind turbines, provides crucial evidence for managing these industrial sites responsibly. The researchers emphasize that while the computer is a powerful tool, it works best when it complements human expertise rather than replacing it entirely. The system is designed to handle the bulk of the counting, flagging the most complex scenes for human review if necessary. This partnership between human insight and machine speed offers a scalable solution for monitoring the health of our oceans, ensuring that as we build more infrastructure on the seabed, we can also keep a precise and watchful eye on the life that calls it home.

Technical Summary: Automated Estimation of Gadoid and Flatfish Abundance in Stereo-BRUV Footage Using YOLO Object Detection

Problem Statement
Marine ecosystem management requires robust monitoring to address anthropogenic pressures such as overfishing, pollution, and the proliferation of offshore renewable energy infrastructure. While Baited Remote Underwater Video (BRUV) surveys offer a non-invasive, cost-effective method for assessing fish populations, their utility is constrained by the labor-intensive nature of manual video analysis. This creates an analytical bottleneck, particularly as data volumes grow into terabytes. Furthermore, conventional BRUV analysis typically utilizes only a single camera viewpoint from stereo systems, discarding potentially valuable data from the second camera, which is traditionally reserved for length measurements. There is a need for automated, end-to-end pipelines that can efficiently enumerate specific taxa and leverage the full capacity of stereo-BRUV systems to improve the sensitivity of abundance estimates.

Methodology
The authors developed an end-to-end computer vision pipeline using the You Only Look Once (YOLOv8) object detection framework to automate the enumeration of Gadidae (cods) and Pleuronectiformes (flatfish) in BRUV footage collected at offshore wind farms in the northern North Sea.

Data Collection and Annotation: Video data was collected during surveys in 2022 (n=108 deployments) and 2024 (n=96 deployments). Training data was extracted from the 2022 survey, specifically the "MaxN frame" (the frame with the maximum number of individuals) and every 1000th frame thereafter. A total of 2,709 images were annotated, containing 16,679 organisms. To address class imbalance, species were grouped into three classes: 'Pleuronectiformes', 'Gadidae', and 'Benthic invertebrates'.
Model Development: The authors employed a two-stage optimization process. First, a coarse search evaluated 41 combinations of input image sizes (640×640 vs. 1280×1280), model architectures (YOLOv8n to YOLOv8x), and training data structures. The optimal configuration utilized 1280×1280 input images and the YOLOv8s (small) framework. Second, a genetic algorithm-based hyperparameter tuning was performed over 100 iterations.
Ensemble Approach: To mitigate variability introduced by random weight initialization, a model ensemble of five distinct models was trained using the optimized hyperparameters. During inference, the detection counts from all five models were averaged to determine the final abundance estimate for each frame.
MaxN Estimation: The pipeline analyzed one frame every 2 seconds of the video. The maximum number of detections for a specific class in any single frame was recorded as the model-derived MaxN ( $MaxN_{YOLO}$ ).
Evaluation Metrics: Standard metrics (mAP, precision, recall, RMSE, MAE) were calculated. Crucially, the authors introduced a custom metric, Mean Peak Success (MPS), to evaluate deployment-level accuracy. MPS calculates the proportion of predictions falling within an abundance-dependent error tolerance (ranging from 10% at low abundance to 20% at high abundance). The optimal confidence threshold for deployment was determined by maximizing the peak success rate for each taxon.
Stereo Vision Analysis: The pipeline was applied to both left and right camera views of the stereo-BRUV units. The authors compared the conventional single-viewpoint estimate (right camera) against a max-MaxN metric, defined as the higher abundance estimate between the two viewpoints for each deployment. Statistical tests (paired t-tests, permutation tests on Coefficient of Variation, and Friedman tests) were used to assess bias, variability, and sensitivity to distance from turbine bases.

Key Results

Detection Performance: The final model ensemble achieved high detection accuracy, with mean Average Precision (mAP) scores of $0.896 \pm 0.009$ for Gadidae and $0.814 \pm 0.009$ for Pleuronectiformes.
Abundance Estimation: $MaxN_{YOLO}$ showed very strong agreement with expert-annotated human counts ( $MaxN_{Human}$ ), with Spearman rank correlations ( $\rho$ ) ranging from 0.955 to 0.987 and coefficients of determination ( $R^2$ ) from 0.874 to 0.987 across both survey years. The model exhibited a slight, consistent negative bias (undercounting), which was most pronounced for Pleuronectiformes at high abundances, likely due to occlusion.
Stereo Vision Benefits: Comparing the single viewpoint (right camera) to the max-MaxN approach revealed that using both viewpoints significantly reduced the variability of abundance estimates (lower Coefficient of Variation).
Statistical Sensitivity: The max-MaxN metric demonstrated greater sensitivity to experimental variables (distance from turbine jackets) than the single viewpoint. While the single viewpoint failed to detect significant differences in abundance across distance bands for either taxon, the max-MaxN metric revealed significant differences for Pleuronectiformes ( $p = 0.043$ ) but not for Gadidae ( $p = 0.146$ ). Permutation testing indicated that the increase in effect size was statistically significant only for Pleuronectiformes ( $\Delta W = 0.092, p = 0.003$ ), whereas the difference for Gadidae was not significant ( $\Delta W = 0.061, p = 0.114$ ).
Efficiency: The implementation of the automated pipeline reduced the elapsed time from fieldwork completion to final MaxN estimates from 134 days (2022, manual) to 68 days (2024, automated with manual validation).

Significance and Claims
The paper claims that tailored object detection pipelines can effectively extend the efficiency and analytical scope of video-based marine monitoring within defined survey systems. The study demonstrates that:

Automation is Viable: YOLO-based pipelines can automate the enumeration of commercially important fish taxa with high accuracy, partially replacing manual annotation in multi-year monitoring programs.
Stereo Data Utility: Automated analysis makes it practical to utilize both viewpoints of stereo-BRUV systems. The "max-MaxN" approach, which selects the higher count from the two cameras, increases the sensitivity of abundance estimates to experimental variables and reduces variability, thereby improving statistical power without additional field effort.
Metric Innovation: The introduction of Mean Peak Success (MPS) provides a more appropriate evaluation framework for BRUV data than standard error metrics, as it accounts for the abundance-dependent nature of counting errors (e.g., occlusion).
Accessibility: The workflow relies on open-source frameworks and can be executed on mid-range hardware (NVIDIA RTX A4000), offering a lower-cost route for researchers to adopt automated analysis, provided they can invest in the initial training data annotation.

The authors maintain a modest stance, noting that while automated methods are effective for abundant taxa in multi-year monitoring, human annotators remain the benchmark for absolute accuracy. They also acknowledge that the models are tailored to specific conditions and that species-level identification was limited by training data availability, suggesting that broader generalizability may require larger, foundation-model-based approaches or more extensive annotation efforts.

More like this