3D photogrammetry and deep‐learning deliver accurate estimates of epibenthic biomass

Accurate biomass estimates are key to understanding a wide variety of ecological functions. In marine systems, epibenthic biomass estimates have traditionally relied on either destructive/extractive methods that are limited to horizontal soft‐sediment environments, or simplistic geometry‐based biomass conversions that are unsuitable for more complex morphologies. Consequently, there is a requirement for non‐destructive, higher‐accuracy methods that can be used in an array of environments, targeting more morphologically diverse taxa, and at ecological relevant scales. We used a combination of 3D photogrammetry, convolutional neural network (CNN) automated taxonomic identification, and taxa‐specific biovolume:biomass calibrations to test the viability of estimating biomass of three species of morphologically complex epibenthic taxa from in situ stereo 2D source imagery. Our trained CNN produced accurate and reliable annotations of our target taxa across a wide range of conditions. When incorporated into photogrammetric 3D models of underwater surveys, we were able to automatically isolate our three target taxa from their environment, producing biovolume measurements that had respective mean similarities of 99%, 102% and 120% of those obtained from human annotators. When combined with taxa‐specific biovolume:biomass calibration values, we produced biomass estimates of 88%, 125% and 133% mean similarity to that of the ‘true’ biomass of the respective taxa. Our methodology provides a highly reliable and efficient method for estimating epibenthic biomass of morphologically complex taxa using non‐destructive 2D imagery. This approach can be applied to a variety of environments and photo/video survey approaches (e.g. SCUBA, ROV, AUV) and is especially valuable in spatially extensive surveys where manual approaches are prohibitively time‐consuming.


| INTRODUC TI ON
Biomass is a critical metric in our understanding of ecosystem structure and functioning.It is a crucial component in the assessment of standing stock and flows of energy (Möller et al., 1985;Oevelen et al., 2009), carbon storage (Fang et al., 2001) and a key predictor of metabolic activity (Raven & Kübler, 2002) and secondary production (e.g.Degen et al., 2016).In marine systems, the measurement of epibenthic biomass is hindered by sub-tidal access difficulties and is largely reliant on destructive and extractive methods such as trawl surveys (e.g.Fredriksen et al., 2020;Gage & Bett, 2006).Trawl surveys are usually limited to sedimentary, near-horizontal substrates and, as a destructive practice, are particularly unsuitable for surveying fragile taxa (e.g.corals), in marine protected areas, or in long-term monitoring programmes where assemblage preservation is required.
While divers or ROVs can also be used to remove benthos on hard or vertical substrata (Gambi et al., 2000;Sant et al., 2017), they share the same destruction-based limitations of trawl-sampling, as well as their own respective constraints.
In response to the limitations of extractive sampling, ecologists have often turned to the use of non-destructive underwater imagery to qualify and quantify benthic assemblages.While this has the potential to increase underwater data-gathering efficiency, subsequent analysis, particularly at ecosystem scales, can be lengthy, and methods may overlook more cryptic taxa (Kornder et al., 2021).Furthermore, abundance metrics are often restricted to frequency and/or planer-viewderived percentage cover (e.g.Beazley et al., 2013;Cánovas-Molina et al., 2016;Marlow et al., 2020), which can lead to substantial biases towards low-relief taxa over more erect, morphologically complex taxa (Kornder et al., 2021).Underwater imagery can also be used to estimate epibenthic biomass but requires a conversion factor.This usually is based upon (1) the product of species abundance and average biomass of individuals for a given taxon (e.g.Rowden et al., 2010), (2) conversion factors from 2D information such as surface area or length (e.g. De Clippele et al., 2021;Durden et al., 2016;Fillinger et al., 2013) or (3) conversion from volume based on assumed simple geometric shapes (Benoist et al., 2019;Kramer et al., 2014).However, these approaches have limitations; extrapolation from average individual biomass may under-or overestimate the biomass of non-heterogenous populations and both 2D conversation ratios and assumed geometry are less applicable for more morphologically complex taxa (e.g.Durden et al., 2016).Accurate measurements of volume offer the potential for more reliable conversions to biomass, but acquiring volumetric data for complex benthic morphologies has usually only been possible with extensive in situ measurements (e.g.Wulff, 2001).The development of structure-from-motion (SfM) photogrammetry technology offers the opportunity to change this.SfM photogrammetry uses overlapping images taken at varying angles to accurately produce 3D models of a given habitat or taxon; the specific geometry is digitally reconstructed by detecting common features in multiple images and uses their relative positions to produce a 3D point cloud of the photographic scene (Carrivick et al., 2016;D'Urban Jackson et al., 2020).Underwater SfM surveys can be spatially extensive (Palma et al., 2018;Pizarro et al., 2017), conducted using affordable non-specialist camera equipment (Bayley & Mogg, 2020;Raoult et al., 2016) and (when appropriately scaled) produce mm resolution photogrammetric outputs (Rossi et al., 2020).To date, SfM has been used from tropical reefs (e.g.Fukunaga et al., 2019), to temperate (Spyksma et al., 2022), polar (Piazza et al., 2019) and deep-water environments (De Oliveira et al., 2021) and has been used to assess habitat complexity (Aston et al., 2022;Bayley et al., 2019;Price et al., 2019), growth rates (Olinger et al., 2019;Prado et al., 2021), bioerosion (Morais et al., 2022), habitat provision (Urbina-Barreto et al., 2021) and benthic cover (Raoult et al., 2016).Furthermore, Palma et al. (2018) used SfM-derived surface area measurements to estimate biomass of Mediterranean sea fans and Ríos et al. (2020) estimated biomass of the deep-sea sponges from SfM-derived sponge perimeter.However, to the best of our knowledge, no studies have used SfM to estimate biomass of complex benthic species using 'biovolume' as a direct proxy.
One important advantage of image-based surveys, including SfM, is the capacity to cover large spatial areas, especially if images are acquired using towed, remotely operated or autonomous equipment (e.g.Ríos et al., 2020;Robert et al., 2017).While large-scale SfM surveys have the potential to inform at scales relevant to assessing ecosystem services, function and delivery, they also have high data analysis requirements.Machine learning, particularly the use of automated species identification, is seen as an essential tool in increasing the efficiency of SfM survey analysis (Chirayath & Instrella, 2019;Mohamed et al., 2020).However, very few studies have applied automated species identification directly to 3D models, with the majority focused on application to 2D orthomosaics, (Mizuno et al., 2017;Pavoni et al., 2020Pavoni et al., , 2021Pavoni et al., , 2022)).Those studies that have applied automated annotation to 3D models have used differing techniques; (Hopkinson et al., 2020) annotated coral reef meshes directly while (Pierce et al., 2021)

| Target species
We selected three benthic species that are conspicuous taxa on rocky reefs in the Northeast Atlantic.Selection was based on both high abundance in these environments and a requirement for the selected species to fulfil a variety of morphological types and profile elevations.The species selected were the soft coral Alcyonium digitatum, the common sea urchin Echinus esculentus and the plumose anemone Metridium dianthus (Figure 1).

| Species biomass-biovolume calibration
Representative individuals of the target taxa, encompassing a range of body sizes, were collected by SCUBA from a rocky reef (56° 27′ 30.4″N, 5° 26′ 4.71″ W; WGS84 datum) and were stored in flowthrough aquarium facilities at the Scottish Association for Marine Science (SAMS), UK.All diving work was conducted in line with UK Health and Safety Executive (HSE) Diving at Work regulations, no specific licence was required for the collection of our target species, and the SAMS aquarium facilities are registered premises (site number: SS074) under Marine Scotland-Science.
For each species, biovolume was determined using photogrammetric volume.This method was selected rather than more traditional techniques (such as displacement) as this was our proposed field method and photogrammetrically derived volume can differ from true volume.Differences are due to the presence of unseen spaces; internal voids within individual taxa (e.g.branching corals) and/or concealed gaps between individuals within assemblages, all of which are unaccounted in photogrammetric models.Photogrammetric models were repeated for each species across a range of numerical abundances, groupings and individual sizes (min n = 15 per species) to assess how photogrammetric biovolume (and consequently biomass conversions) changed with body size and aggregations of multiple individuals.Imagery was collected using video (1080 p resolution) on GoPro™ Hero 9 and Olympus™ Tough TG-6 cameras in 0.4m 3 aquaria, with in situ static coded targets acting as scale bars (Figure 2a).3D models and biovolume measurements were subsequently generated in Agisoft Metashape™, the workflow for which is listed in Table 1.
Metridium dianthus and A. digitatum required additionally density measurements due to their temporally variable morphology; both species are able to substantially change their morphology (and thus biovolume) over short time scales without concurrent changes to their biomass.Metridium dianthus (Figure 2) responds to increased food availability and water flow by significantly elongating its body column and expanding its oral disk (Batham & Pantin, 1950).both.However, previous ex situ experiments (Fabricius et al., 1995;Migné & Davoult, 2002) have failed to induce these morphological changes through ex situ manipulations of flow or food availability; individuals in our tanks predominantly had their polyps retracted.
To capture this variation, additional 3D models (n = 21) of A. digitatum were conducted in situ by SCUBA on a local reef and modelled taxa were collected for subsequent biomass measurements.
Biomass of each taxon (or groups of) from each model was measured using drip-free wet-weight (g) for E. esculentus and A.
digitatum, and constant dry-weight (g) for M. dianthus.Dry-weight was selected as the preferred method for M. dianthus as initial wetweight measurements demonstrated a significant inconsistency in internal fluid retention for this species.Biomass conversion equations were calculated in R using ordinary least square (OLS) regressions; although neither our biomass nor biovolume were controlled variables, and hence using Model II ranged major axis (RMA) was considered (Legendre & Legendre, 2012;Sokal & Rohlf, 2010), we opted for using OLS regressions as our intention was to predict biomass from biovolume rather than to explicitly describe the relationship (Smith, 2009).For each species, model residuals were assessed for the assumptions of normality and homoscedasticity (Zuur et al., 2010).Due to the high variation in biovolume exhibited by M. dianthus, a single biomass conversion equation was deemed inappropriate, and regressions for this species were conducted three times to represent the lowest, highest and median densities from each of our repeated photogrammetric models.

| Machine learning
A semantic segmentation data set was generated using the Computer Vision Annotation Toolkit (CVAT), although many other annotation tools that provide support for semantic segmentation would also have been suitable, for example, Labelme, or Label Studio.We considered an approach similar to Pierce et al. (2020) using sparse annotations and Fast Multilevel Superpixel Segmentation (FAST-MSS).However, due to the fractal nature of the benthos and often multiple overlapping species, superpixelbased techniques performed poorly on our data set, with little time saving to be had over the more accurate dense-annotation approach.
The training data set was comprised of 738 images, split 80/10/10 for training, validation and testing respectively.In addition to this  The best-performing model was chosen to annotate the field imagery, and after the creation of a 3D dense cloud by SfM, the annotated images were selected as source images, replacing the original images (as per Pierce et al., 2021) and a classified point cloud produced (see Table 1 for steps).

| Fieldwork
In this study, five 15 m transect tapes (depth ~ 12 m) were laid out parallel to the reef slope and adjacent to each transect were positioned three 0.5 m quadrats at approximately 3 m intervals; quadrats were positioned to encompass as many of our target taxa as possible.Quadrats were manufactured out of aluminium composite and contained 8 Agisoft coded targets (12-bit; each target positioned 30 cm from the nearest adjacent target), thereby also acting as in situ scale bars.Photogrammetry surveys were conducted by divers using a customised camera setup consisting of a stereo pair of GoPro™ Hero 9 action cameras (with 0.3 m separation between lens centres) and a pair of SUPE™ V3K video lights attached to an aluminium plate (Figure 3).The stereo-cameras were used to scale the resultant photogrammetric model (Workflow in Table 1 and Supplementary Material 1) with the in situ scale bars acting as an alternate back-up scaling option.All cameras were using GoPro Labs firmware to enable GPS time sync, recorded video at a resolution of 1920 × 1080 pixels and at a frame rate of 60 fps.
Divers maintained a regular pace of approximately 0.15 m −1 s −1 , an elevation of approximately 1 m above the substrate and the cameras were angled 20° off the vertical.Transects were conducted in both directions, along the centre line of the transect tape and 1 m on either side of the transect tape (6 passes in total for each transect).After completion of the video surveys, all target taxa within each of the quadrats were collected and retained for biomass measurements.

| Biomass estimation and species abundance
Three-dimensional photogrammetric models were created for each transect using the workflow outlined in Table 1.The biovolume and surface area were calculated for each of the target taxa after isolation from the 3D models using the machine-learning annotation workflow (Table 1).All individuals of M. dianthus were near-to-fully retracted and consequently, the 'fully retracted' regression was used as the most appropriate biomass predictor.
In addition, the biovolume of the taxa within the subsampled quadrats was calculated separately.Conversion of biovolume to biomass was conducted for each of the target taxa across each transect and additionally within each subsampled quadrat using our species-specific conversion ratios.
Validation of field biomass estimates was achieved by comparing the 'true' biomass of the target taxa (measured from weighing subsampled quadrats) with that predicted from our biovolume conversions.For each of the quadrats, the biomass of each of the target taxa estimated from photogrammetric biovolume was normalised to the true biomass and plotted on an interval plot to visualise the mean difference (as a percentage of the true biomass) with associated 95% confidence intervals.

| Machine-learning annotation validation
To test the accuracy of the automated model annotation, we compared the biovolume from manually annotated meshes with the biovolume from meshes created using machine-learning annotated dense clouds.On each transect, three separate groups of each of the target taxa were selected to be annotated using each method (n = 15 pairs per species).For each of the paired models, the biovolume derived from machine learning was normalised to that of the manual annotation biovolume and plotted on an interval plot to visualise the mean difference (as a percentage of the manual annotation volume) with associated 95% confidence intervals.

| Species biovolume to biomass calibration
Photogrammetric biovolume was found to be a significant predictor (p < 0.001 for all) for biomass in all target taxa (Figure 4, Table 2).
The relationship was strongest for the urchin Echinus esculentus, followed by the soft coral Alcyonium digitatum, with the plumose anemone Metridium dianthus showing differing goodness of fit for the regression depending upon whether the anemone was fully retracted, fully extended or median biovolumes.

| Machine-learning annotation validation
Biovolume measurements derived from machine-learning annotated models were largely consistent with those from manually annotated models, with some inter-taxa differences (Figure 7).

| Biomass
When the actual biomass of the target taxa from the subsampled quadrats was compared to biomass estimates from our Predictor plots of ordinary least squares (OLS) regressions of biovolume as a predictor for biomass in Echinus esculentus (a), Alcyonium digitatum (b) and Metridium dianthus (c).Circles (blue) represent biomass: biovolume data, solid lines represent the regression and shaded areas 95% confidence intervals.Data for differing M. dianthus morphologies are presented as retracted (blue), median (green) and extended (red).

| DISCUSS ION
Reliable measurements of ecosystem biomass are key to our understanding of a suite of ecosystem functions and processes.Our study has demonstrated that 3D photogrammetry has the potential to enable accurate biovolume measurements of more complex morphologies, providing reliable biomass calibration for taxa that would have been unachievable with previous methods.Furthermore, we have shown that automated annotation of these photogrammetric models is accurate, with the potential to greatly increase efficiency, especially in the annotation of spatially large and taxonomically diverse 3D models.
Validation of photogrammetrically derived biovolume as a predictor of biomass was supported by the strength of our target taxa biomass regressions, although these also highlight the potential for inter-taxa differences.Any differences are likely due to variation in the complexity of the respective morphologies (and their photogrammetrically 'hidden' intra-and inter-individual spaces), and also the consistency of the subsequent biomass measurement.In this study, our greatest challenge was associated with Metridium dianthus, and its highly variable morphology.We elected to capture this variation by categorising the body forms into three different morphological types, with the obvious limitations that (1) this requires biomass calibration data that confidently captured the morphological variation and (2) field data were unlikely to fully conform to the proposed categories.Nevertheless, this degree of intra-specific morphological variation (up to 1000% difference in M. dianthus biovolumes) is unusual, and we believe most epibenthic taxa would be suitable for photogrammetric biovolume:biomass calibrations.
The use of machine learning to isolate taxonomically distinct portions of the dense cloud was largely very accurate, producing biovolume measurements that were consistent with those derived from manual annotations.Some difference would be expected given that a human annotator is able to incorporate both 2D and alternative software (e.g.Autodesk Meshmixer) may produce better mesh-closure results (Olinger et al., 2019), but the in case of encrusting or low-relief epibenthic taxa, the surface area is likely to be a better proxy for biomass than biovolume.Furthermore, given that mesh closure is likely to produce the most reliable volume estimates on level substrates, perhaps one of the most appropriate uses of this technique is on artificial substrates.On offshore structures such as oil and gas platforms and wind farms, the underlying substrate is not only likely to be uniform but often also of a known surface geometry; design drawings or digital computer-aided design (CAD) files.
As outlined above, the use of photogrammetry and machinelearning to estimate taxa-specific biomass from in situ photographic/ For instance, in the case of a coral growth study, a more intensive in situ survey design would likely be beneficial, but this would incur increased data analysis and/or reduced spatial coverage.

| CON CLUS IONS
We have demonstrated that photogrammetrically derived biovolume can be used as a reliable proxy for biomass in morphologically complex benthic taxa given appropriate application of taxa-specific conversion values.When combined with the outputs from a well-trained CNN architecture (Pierce et al., 2021), large-scale photogrammetric models can be used to efficiently and accurately estimate benthic biomass over ecologically relevant scales.The potential applications of this methodology are great, particularly when combined with stereo-camera scaling or other technologies that don't require in situ scale devices.Although some extractive sampling is required for initial biomass regression calculations, subsequent biomass estimates can be made in an entirely non-destructive manner.This allows for uses in fragile environments (e.g.coral reefs) and/or repetitive time-series for assessments of metrics such as recruitment, growth or mortality (Ferrari et al., 2017(Ferrari et al., , 2021;;Fukunaga et al., 2022;Lange & Perry, 2020).Furthermore, our method can be based on any underwater camera platform (SCUBA, ROV, AUV) and consequently, unlike trawls or grabs, can be used in inaccessible vertical environments such as seamounts and fjord walls, which would otherwise require less precise methods of biomass estimation (Durden et al., 2016;Rowden et al., 2010;Thresher et al., 2011).In these environments, initial sample collections for biomass regressions are likely to be logistically complex, especially when trying to preserve soft-bodied biovolume, but in situ photogrammetric biovolume estimates could mitigate for this.
We expect that as photogrammetry and machine-learning technologies mature and become increasingly accessible, automated collection of in situ 3D benthic data will become progressively more mainstream and provide information that was hitherto unattainable on the form and function of the earth's benthic environments.

AUTH
also produced annotated models of corals but reused the re-projection from the photo-alignment step of SfM to project labels onto the point cloud.The accuracy of the Pierce et al. (2021) methodology is reliant on both the segmented imagery input (in this case derived from sparse labels) and the quality of the 3D model in question.The combination of SfM photogrammetry with deep-learningbased image segmentation offers the potential to automatically categorise benthic geometry by target taxa.Here, we demonstrate how by combining these techniques with taxa-specific density conversion values, photogrammetric biovolume measurements can be converted into accurate estimates of benthic biomass across a variety of taxa of differing morphological complexity.
To account for this changeable biovolume, each 3D model of M. dianthus was repeated at least five times (mean = 10.1 models) under differing conditions of flow and food availability (mix of commercially available frozen daphnia, rotifers, artemia and copepods; BCUK Aquatics Ltd).Alcyonium digitatum also displays morphological variation, where polyps are either retracted, expanded or a mix of F I G U R E 1 Field images of the target species Alcyonium digitatum (a), Echinus esculentus (b) and Metridium dianthus (c).Photos credit: J. Marlow.F I G U R E 2 Example image of ex situ aquaria modelling of Metridium dianthus with coded-target scale bars (a), and photogrammetric models of M. dianthus when retracted (b) and fully extended (c).Images (b and c) are of the same anemones and represent photogrammetrically estimated biovolumes of 325 and 3370 cm 3 respectively.TA B L E 1 Photogrammetry workflow using Agisoft Metashape software, following initial image collection.Shaded areas represent differing steps/groups of related steps.
study's target taxa, it also contained annotations for the cold-water coral Desmophyllum pertusum and the blue mussel Mytilus edulis; our training data came for a wide selection of sources, including industry (offshore energy) ROV footage of offshore structures (on which D. pertusum and M. edulis are a common biofouling component).The wide selection of sources was chosen to train the model to generalise across images from different cameras, lighting conditions and substrates, etc.

For a baseline
semantic segmentation model, DeepLabv3+(Chen et al., 2018) was selected, from PyTorch Segmentation Models(Iakubovskii, 2019).This network combines the advantages of spatial pyramid pooling with a encode-decoder structure(Chen et al., 2018).DeepLabV3+ has been shown to offer excellent performance in the segmentation of similar data sets, for example,Pavoni et al. (2020), where its performance matched that of a human annotator for a binary segmentation task.We then tried a newer model architecture, from PyTorch Segmentation Models, based on a Mix Vision Transformer encoder (from SegFormer, Xie et al., 2021) combined with a simple U-net decoder.SegFormer was chosen because it offers state-of the-art efficiency, accuracy and, significantly for underwater image segmentation, robustness to image degradation similar to that found underwater (low visibility and backscatter).All training was performed using a Nvidia RTX A6000, with Tensorboard used to track relevant metrics during training and models trained to 400 epochs with a batch size of 4. The models were initialised with weights from pre-training on ImageNet and all parameters were left unfrozen.An Adam optimizer was used, with an initial learning rate of 0.0001.Due to the nature of benthic assemblages, our training data have a significant class imbalance, with some classes are being harder to classify than others.Some species, for example M. dianthus, grow in large encrusting areas, whereas others, for example E. esculentus, are only present as a few individuals and so are a difficult, minority class.During training, network performance is evaluated by a loss function and through careful selection of this loss function, the network can be trained to perform well on different issues in the training set.Here we used the focal loss, a generalisation of the commonly used Cross-Entropy loss.Focal loss reduces the loss for well-classified examples, causing the model to focus on hard, misclassified examples, which was ideal for this data set.The Albumentations library(Buslaev et al., 2020) was used for augmentation during model training.In addition to standard augmentation techniques (flip, rotate, random crop, etc.), an attempt was made to replicate the physical effects of the underwater environment on imaging, including random variation of the red channel (to simulate red light absorption) and blur.

Both
DeepLabV3+ and the SegFormer style model performed extremely well on the data set, with high Intersection over Union (IoU) and accuracy metrics (Figure 5).The worst-performing class, E. esculentus, is likely due to it being a minority class in the training data and also F I G U R E 3 Field camera equipment; stereo pair of action cameras positioned 30 cm apart with two underwater video lights.its behaviour of covering itself with shell fragments and algae.The performance of all classes will likely decrease when predictions are made on new geographic areas/imaging conditions, although since our data set contains many of these, it should have some ability to generalise.Examples from the test set are shown in Figure 6, comparing model predictions to ground truths and demonstrating the improved performance of the SegFormer style model over DeepLabv3+.

Figure 6
Figure 6 also illustrates the challenges of the data set; as images were often acquired for photogrammetric purposes, they were frequently taken at an oblique angle, resulting in the top of the image being further away than the bottom, with significant light attenuation, leading to challenges in species ID.In Figure 6 (areas a, b and c), three E. esculentus are labelled, all partially obscured, and in the case of (area b), missing entirely from the ground truth.DeepLabv3+ fails to detect any, but the SegFormer style model correctly segments all three.In Figure 6 (area d), M. dianthus is incorrectly segmented as A. digitatum by DeepLabv3+ but correctly segmented by the SegFormer style model.

F
Left, performance metrics from models and right, confusion matrix from SegFormer Backbone/U-Net.Far left) test images, (centre left) human annotated ground truths, (centre right) predictions from DeepLabv3+ model and (far right) predictions from SegFormer backbone/U-Net model.
3D information into their decision-making, and while incorporating 3D information (e.g.Runyan et al., 2022) into the CNN might improve 3D model annotation, it is also likely to greatly increase overall processing time.The largest divergence in biovolume from human-annotated meshes was associated with M. dianthus, but as this was not reflected in the CNN performance (as evidenced by the confusion matrix), this cannot be attributed to the annotated source imagery.Alternatively, we suggest that the greater error was a product of the mesh closure, that is, the process of filling the gap in the 3D model where the taxa are attached to the substrate, which is unavoidably photogrammetrically 'hidden'.Mesh closure is an automated process within Agisoft Metashape and has previously been associated with greater error in volume measurements(Raoult et al., 2017).The reliability of the automated mesh closing function decreases with increasing hole size (relative to the overall size of the mesh), irregularity in hole perimeter and complexity of the underlying surface.In the context of isolating and closing meshes of epibenthic taxa, more reliable mesh closure (and overall model geometry) is likely to increase on even, flat substrates, and with large erect taxa that have a low percentage of their morphology in contact with the underlying substrate.Consequently, low relief or encrusting taxa (such as retracted M. dianthus), especially on irregular substrates, are likely to produce less reliable mesh closures, leading to greater relative biovolume estimation error.Importing the mesh into

F
Interval plot of variation in biovolume measurements from meshes created using machine-learning annotated dense clouds vs manually annotated meshes.Variation is expressed as percentage similarity to biovolumes derived from manually annotated meshes.The blue dot represents mean similarity and error bars represent 95% confidence intervals.FI G U R E 8 Interval plot of biomass estimate accuracy of target taxa from subsampled quadrats.Accuracy is expressed as percentage similarity between biomass estimated from photogrammetric/machine learning methods and actual biomass values as measured ex situ using traditional weighing methods.Blue dots represent mean similarity and error bars represent 95% confidence intervals.videographydata is reliant upon the accuracy of (1) the biovolume:biomass conversion factors, (2) the machine-learning image annotation and (3) the in situ model in question.Field testing of our methods produced biomass predictions that ranged from a 32.8% overestimate in M. dianthus to an 11.8% underestimate in E. esculentus.Whether this level of variance from the 'true' biomass is considered acceptable will largely depend on the given research question and/or the accuracy of the alternative methods available.For instance, a 32.8% error might be considered acceptable if this method were used to estimate biofouling on offshore structures (where current methods produce overestimates of approximately 42%,Mallat et al., 2014) but unacceptable if used to study the growth of a very slow growing species (e.g.corals;Lange & Perry, 2020).All three components of the methodology can be manipulated to improve accuracy, but the extent to which this is necessary will depend on the respective research question and the acceptance of the trade-offs.
O R CO NTR I B UTI O N S Joseph Marlow, John Edward Halpin and Thomas Andrew Wilding conceived the ideas and methodology; Joseph Marlow collected the data; Joseph Marlow and John Edward Halpin analysed the data; Joseph Marlow led the writing of the manuscript.All authors contributed critically to the drafts and gave final approval for publication.

F
Images of underwater photogrammetric 3D models, showing textured mesh model of entire transect (a), machine-learning annotated mesh model of entire transect (b), focused textured mesh model (c), focused machine-learning annotated mesh model (d) and focused model of Alcyonium digitatum isolated from main model by colour (e).A copy of the photogrammetry model (obj.format) can be found in Supplementary Material 2.