Introduction
Aquaculture increasingly relies on precision monitoring to optimize feeding, size grading, and harvesting. Traditional physical sampling is invasive, labor-intensive, and induces acute stress, negatively impacting fish welfare and growth. While computer vision is transforming the sector (Yang et al., 2021), submerged systems are constrained in shallow or turbid land-based units by biofouling, light attenuation, backscatter, and maintenance demands. Above-water imaging avoids camera submersion but introduces a more complex optical problem because air-water refraction, ripples, foam, fluctuating water levels, and free-swimming behavior distort apparent fish geometry. This study presents an integrated above-water stereo vision methodology for estimating total length, biomass, and weight dispersion in European seabass (Dicentrarchus labrax) and rainbow trout (Oncorhynchus mykiss). The framework advances stock-level monitoring by combining optimized stereo calibration, refraction-aware three-dimensional reconstruction, deep-learning contour extraction, curved-body morphometry, biological consistency filtering, and allometric weight conversion.
Materials and Methods
The methodology is organized into a development pipeline and an operational pipeline. During development, stereo calibration was optimized using a custom cost function that minimized distance, boundary, and planarity errors from checkerboard images. The coordinate system was realigned so the XY-plane remained parallel to the water surface and the Z-axis represented water depth, reducing systematic reconstruction bias and enabling efficient two-dimensional refraction correction based on Snell's law instead of three-dimensional iterative ray tracing. YOLOv11x-seg models were trained for multi-fish and water-level instance segmentation using vertically stacked stereo pairs and iteratively refined annotations to preserve aspect ratios and improve contour recovery under occlusion and variable optical conditions. A Voronoi-based centerline algorithm measured naturally curved fish, overcoming straight-body assumptions common in earlier pipelines. Species-specific allometric models converted image-derived total length and body width into estimated weight.
During operation, synchronized stereo image pairs were acquired and processed with contrast-limited adaptive histogram equalization to reduce glare and improve visibility. The water-level model supplied dynamic depth parameters for refraction correction. Fish contours were encoded as translation-invariant polar descriptors anchored at the body center and matched between stereo views using K-Nearest Neighbors. After triangulation and optical correction, measurements were rejected when geometric or biological plausibility was violated, including unrealistic width-to-length ratios, excessive inclination, surface distortion, spatial inconsistency, partial occlusion, or merged silhouettes. Accepted detections were aggregated into population-level length and weight distributions rather than forcing estimates from unreliable frames.
Discussion and Conclusions
The system achieved high accuracy under realistic aquaculture conditions. Median total-length mean absolute percentage error reached 1.2% for rainbow trout and 2.3% for European seabass, while median-weight errors reached 4.9% and 6.9%, respectively. These results are comparable to leading underwater stereo systems(Tonachella et al., 2022) while avoiding submersion-related maintenance and visibility constraints. Optimized calibration was central to this performance, reducing measurement error from a maximum of 25% under standard OpenCV calibration to below 4.2% in all tanks. YOLOv11x-seg also surpassed the Mask R-CNN benchmark, reaching 98.2% mAP50 and improving contour recovery under partial overlap.
In European seabass trials, the method remained effective despite turbidity up to 200 NTU, surface foam, and ripples, conditions representative of estuarine earthen ponds and shallow marine units where underwater imaging may be impractical. Automated water-level detection and dynamic refraction correction also overcame limitations of above-water systems that require fixed camera-to-surface distances. The main contribution is a scalable, welfare-oriented framework that transforms above-water stereo images into reliable length and weight distributions for production management. By integrating 3D imaging, optical correction, robust segmentation, curved-body measurement, and morphological filtering, the method provides daily biometric data for feeding optimization, grading, growth modeling and harvest planning, which are key processes in Aquaculture 4.0 implementation.
Acknowledgment
This study was funded by the SAFE (SmartAqua4Future) project (101084549) under Horizon Europe, call HORIZON-CL6-2022-FARM2FORK-01. Luiz Navarro was funded by the doctoral grant (PRT/BD/154260/2022) financed by the Portuguese Foundation for Science and Technology (FCT) and the European Social Fund (ESF), under the MIT Portugal Program (MPP2030-FCT PhD Grants) for the project DOI: https://doi.org/10.54499/PRT/BD/154260/2022. The FCT partially supported this research within the scope of UID/04423/2025 (https://doi.org/10.54499/UID/04423/2025), UID/PRR/04423/2025 (https://doi.org/10.54499/UID/PRR/04423/2025), and LA/P/0101/2020 (https://doi.org/10.54499/LA/P/0101/2020).
References
Tonachella, N., Martini, A., Martinoli, M., Pulcini, D., Romano, A., & Capoccioni, F. (2022) An affordable and easy-to-use tool for automatic fish length and weight estimation in mariculture. Scientific Reports, 12, 15642. https://doi.org/10.1038/s41598-022-19932-9
Yang, L., Liu, Y., Yu, H., Fang, X., Song, L., Li, D., & Chen, Y. (2021) Computer vision models in intelligent aquaculture with emphasis on fish detection and behavior analysis: A review. Archives of Computational Methods in Engineering, 28, 2785-2816. https://doi.org/10.1007/s11831-020-09486-2