Using 3D photogrammetry to monitor aquaculture impacts on hard substrate seabeds

BACKGROUND

Sustainable aquaculture development relies on robust environmental regulation that effectively protects sensitive marine ecosystems. Traditional seabed monitoring techniques, such as grab sampling, are often ineffective - or indeed impossible to use - in areas with hard substrates or where Priority Marine Features (PMFs) are present. These limitations can hinder our understanding of how such habitats respond to aquaculture-related pressures, creating knowledge gaps that restrict evidence-based regulation and protection.

3D photogrammetry, a technique that creates detailed three-dimensional models from overlapping photographic images, offers a solution. By enabling repeated, high-resolution surveys of the seafloor, it can improve accuracy, provide permanent records, and increase visibility in turbid or low-light conditions. Its ability to reveal fine-scale benthic structures makes it especially valuable for monitoring sensitive or complex habitats where traditional tools fall short.

AIMS

This project sought to validate 3D photogrammetry as a regulatory tool for monitoring aquaculture impacts on hard substrate seabeds. Specifically, it aimed to:

• Determine whether models generated from ROV (remotely operated vehicle) imagery are accurate and consistent enough to detect environmental changes.
• Evaluate the feasibility, quality, and cost-effectiveness of using ROVs in underwater habitat assessments.
• Identify measurable indicators of seabed stability, guided by input from regulators and industry stakeholders.
• Compare diver- and ROV-based photogrammetry methods and assess their respective strengths and limitations for regulatory application.

Datasheets and modelling workflows were also evaluated and refined, with academic partners conducting independent validation to support the method’s use in environmental decision-making.

OVERVIEW

A steering group including the Scottish Environment Protection Agency (SEPA), NatureScot, Salmon Scotland and industry partners, coordinated and supported by SAIC, oversaw the project. Three sites were chosen for diver- and ROV-based photogrammetric surveys, with key ecological indicators defined early in the process.

Locations

The following three sites were chosen as essential for evaluating the quality and accuracy of our diver- and ROV-based photogrammetric models:

Table 1. Description of sites where research for this study took place

Survey methodology

Two main survey methods were employed: diver-based and ROV-based photogrammetry. The ROV, fitted with a GoPro HERO10 camera, captured images along predefined transects at half-second intervals while logging GPS and depth data. Meanwhile, divers used a high-resolution Nikon D850 to take still images along parallel transects, ensuring image overlap and recording metadata such as depth and orientation.

Repeat surveys were conducted at selected sites to evaluate the consistency and temporal resolution of photogrammetric models. Both image sets were processed using Agisoft Metashape software to generate 3D models, point clouds, and orthomosaic maps.

Tritonia Scientific Ltd led model generation using a structure-from-motion (SfM) workflow, which included aligning images, applying scale, building dense point clouds, meshing the data, and applying textures to create detailed, navigable 3D environments.

Model evaluation and statistical analysis

Model quality was assessed using several metrics, including reprojection error, Ground Sample Distance (GSD), Ground Control Point Root Mean Square Error (GCP RMSE), and scaling accuracy. Statistical analyses, including linear regression, ANOVA, and Pearson correlation, were conducted to identify relationships between survey method, photo density, model resolution, and point density. Validation steps included comparisons to ground-truth data and replicate surveys to assess repeatability and internal consistency.

At the Hellisay fish farm, photogrammetric models were used to investigate habitat structure and quantify physical impacts from nearby aquaculture activities. Key metrics included surface complexity, terrain ruggedness and curvature, derived using open-source tools like GDAL and QGIS. These were analysed alongside maerl and biota cover, using grid-based sampling and statistical modelling in RStudio.

Table 2. Definitions of metrics used throughout this study

RESULTS

Photogrammetric model performance

ROV surveys generated more tie points - key features that align images during 3D reconstruction - thanks to higher image overlap and varied viewing angles. This led to more consistently aligned images compared to diver data. However, diver models had higher point densities, driven by superior image quality and lighting from the Nikon D850.

Point density decreased as survey size increased, particularly for diver-based methods. For example, diver point density dropped from 6.54 to 3.51 points/mm² between 5×5m and 20×10m surveys, while ROV density declined more gradually (1.58 to 1.32 points/mm²). Diver models were better suited for fine-scale detail, while ROVs offered broader, more scalable coverage.

Textured mesh models were analysed, using face count as a proxy for surface detail. Diver models had higher face counts in smaller surveys, but ROVs outperformed divers in larger, more complex environments. At site 1.3, for instance, ROVs generated significantly higher face counts due to favourable conditions and a greater number of captured images.

Although high face counts indicate granularity, they do not always translate to accuracy if the data is noisy. The optimal survey method depends on specific site conditions and the intended regulatory application.

Table 3. Comparison of Reprojection Error (pixels), GCP RMSE (metres), and GSD (millimetres/pixel) for each respective site

Spatial accuracy and scaling

Diver-based models showed consistently lower reprojection errors, often under the one-pixel benchmark considered ideal. ROV models exceeded this threshold across all sites, largely due to limitations in the ROV’s Short Baseline (SBL) geolocation system, image quality, and lighting. At Site 1.1, for example, diver models achieved reprojection errors of 0.35–0.41 pixels, compared to 1.42–1.64 pixels for ROV models.

GSD values confirmed that diver models delivered higher spatial resolution, especially at smaller scales. At Site 1.1, the average GSD difference between methods was 0.2194 mm/pixel. Nonetheless, both methods achieved millimetre-scale resolution, suitable for detecting subtle seabed changes such as sediment accumulation or benthic shifts.

GCP RMSE values were higher for ROV models but remained within acceptable bounds, averaging around 1.5m, better than the theoretical 2m accuracy of the SBL system. High image redundancy and effective post-processing contributed to this performance.

CloudCompare analysis showed low mean distances (around 1cm) between diver and ROV point clouds in 5x5m surveys, though differences grew at larger scales due to inherent differences in model construction.

Scaling accuracy improved at larger grid sizes. Relative error at 5x5m reached 18% but dropped substantially at 10x10m and 20x10m. The 10x10m grid offered the best balance of resolution, accuracy, and repeatability. Stereo-camera scaling at the Hellisay site further reduced error and variability, suggesting its promise for high-precision monitoring.

Repeat ROV surveys showed high internal consistency, with mean differences of 0.63–2.66 cm and standard deviations under 2.63 cm.

Statistical trends

Linear regression models revealed that increasing photo density consistently reduced GSD and GCP RMSE, while slightly increasing reprojection error. Diver methods significantly outperformed ROVs on all three metrics. The models explained over 94% of the variance in reprojection error and GCP RMSE, indicating strong predictive power.

Survey method and photo density also strongly influenced point density and model resolution. ROV surveys produced lower point densities, while higher photo densities increased detail. However, denser point clouds were associated with reduced model resolution, highlighting a trade-off between completeness and clarity.

Habitat analysis at Hellisay

Five photogrammetric surveys around the Hellisay fish farm enabled analysis of aquaculture impacts on benthic habitats. Structural metrics such as slope, aspect, curvature, and terrain ruggedness were derived from 3D models.

Surface complexity and Terrain Ruggedness Index (TRI) served as indicators of ecological condition. Lower TRI values, particularly in impacted areas, suggested sediment smothering. In contrast, higher TRI at reference sites signalled intact habitat structure. Correlation analysis revealed consistent relationships between slope and surface complexity, and between curvature metrics.

Grid-based analysis also quantified benthic cover, including maerl beds and associated species. Systematic mapping and visualisation provided regulators with quantitative evidence of aquaculture impacts.

IMPACT

The integration of 3D photogrammetry with ROV surveys offers a powerful tool for monitoring aquaculture impacts while supporting environmental protection. Compared to diver-based photogrammetry, ROVs provide broader coverage and higher face counts, making them ideal for large-scale surveys. However, diver methods continue to yield higher data quality, with denser point clouds and better resolution in smaller, more detailed areas.

A key limitation of ROV photogrammetry is its lower georeferencing accuracy, due to constraints in GNSS and SBL sensors. The project team aims to overcome this by achieving centimetre-level precision through further innovation projects funded by the Seafood Innovation Fund (SIF), further supported by SAIC, exploring technologies like RTK and DVL. Scaling accuracy improves with survey size, and the 10×10 metre scale was found to offer an effective balance between coverage and detail. The introduction of stereo-camera systems further enhanced scaling precision, particularly in complex or smaller environments.

This approach has proven effective in assessing aquaculture impacts on sensitive hard-substrate habitats such as maerl beds, capturing changes in surface complexity and habitat condition. These findings underscore the importance of sustainable aquaculture practices.

Looking forward, key recommendations include adopting a hybrid ROV/diver approach, improving georeferencing accuracy, enhancing scaling with larger survey areas, and integrating advanced camera systems. Expanding these methods to other sensitive habitats will support more informed environmental management and reinforce the role of ROV photogrammetry in sustainable marine monitoring. Such insights amplify our capability to oversee, nurture, and shield these vital marine habitats while also accommodating for important food production systems to continue operating.

Further reading

Quantitative Comparison of ROV and Diver-Based Photogrammetry to Reconstruct Maerl Bed Ecosystems

(Iona L. R. Paterson, Kathryn E. Dawson, Andrew O. M. Mogg, Martin D. J. Sayer, Heidi L. Burdett. First published: 25 November 2024)