Detecting abyssal megabenthos in the Great Atlantic Sargassum Belt – Comparing quantitative assessments via manually processed OFOS-videos with AUV-images utilising deep learning models

doi:10.21203/rs.3.rs-4719438/v1

Download PDF

Research Article

Detecting abyssal megabenthos in the Great Atlantic Sargassum Belt – Comparing quantitative assessments via manually processed OFOS-videos with AUV-images utilising deep learning models

https://doi.org/10.21203/rs.3.rs-4719438/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

As the deep sea alone covers over 50 % of the Earth's surface and represents the largest benthic system, the assessment of deep-sea megafauna communities is of great importance. However, even today the number of studies investigating this system is small and is further hampered by the resource- and time-consuming nature of traditional manual surveys. Therefore, it is of great importance to find efficient sampling methods and optimize processing techniques. On the one hand, sampling can be done by a variety of non-invasive methods, such as video documentation using seabed observation systems (OFOS) or seabed mapping using autonomous underwater vehicles (AUV). On the other hand, automated data analysis could be a faster alternative to manual processing by incorporating deep learning models. Here, we compare both sampling and processing techniques by providing novel and unique data on epi-megafauna from two overlapping abyssal plains at 10°N in the Great Atlantic Sargassum Belt. The 2017 datasets provided by OFOS were processed manually, while a manually trained deep learning detection model based on the YOLO V5 algorithm was used to process the 2014/15 AUV dataset. Megafauna composition, total density and higher taxonomic densities as well as the occurrence of Sargassum patches were calculated and compared between the datasets. Significantly higher densities with additional megafauna forms were observed when the datasets were collected by an AUV and processed by the model. The results were discussed in terms of the advantages and disadvantages of the two approaches to sampling and processing.

Deep-sea ecology

abyssal plain

deep learning

megafauna abundance

methods

Sargassum

The abyssal deep sea, a realm covering over 50% of the Earth's surface, is a habitat characterized by extreme conditions and limited light penetration (Watling et al. 2013; Kennedy et al. 2020). Without primary production through photosynthesis, it was assumed that the fauna living there is primarily limited by slowly sinking organic material from the euphotic zone (e.g. Smith and Demopoulos 2003; Johnson et al. 2007; Smith et al. 2008). In recent years, however, several deep-sea studies have reported a surprising phenomenon observed in the North Atlantic - the sedimentation and accumulation of the macroalga Sargassum in abyssal areas (Smith et al. 2008; Krause-Jensen and Duarte 2016; Baker et al. 2018). This holopelagic brown alga is trapped in large quantities by the subtropical North Atlantic Gyre and forms floating mats that can cover large areas of the sea surface before naturally sinking to the seafloor (Fleury and Drazen 2013; Baker et al. 2018). Taking the organic carbon to the starved abyssal plains, Sargassum patches could be utilized there as an additional food source (Smith et al. 2008; Krause-Jensen and Duarte 2016; Baker et al. 2018), similar to the sinking of large pelagic vertebrates (Higgs et al. 2014) or other macroalgae (Krause-Jensen and Duarte, 2016). Despite its potential ecological importance, the number of studies investigating abyssal communities with regular Sargassum sedimentation events is still low today.

Therefore, the identification of efficient sampling methods and the optimization of processing techniques are of great value to increase the number of these studies (e.g. Beuchel et al. 2010; Gobi 2010; Cuvelier et al. 2024). Instead of traditional direct sampling, modern deep-sea studies are increasingly using video and photographic approaches such as Ocean-Floor Observation System (OFOS) or autonomous underwater vehicles (AUV), as they represent a non-destructive, ecosystem-preserving alternative (e.g. Beuchel et al. 2010; Schoening et al. 2012; Yang et al. 2022; Liu et al. 2024; Cuvelier et al. 2024). However, the manual processing of large data sets, which was a common practice until at least 2010 (e.g. Lebart et al. 2003; Chailloux et al. 2008; Gobi 2010), was and is an expensive and lengthy process (e.g. Lebart et al. 2003; Gobi 2010; Stanchev et al. 2020; Cuvelier et al. 2024). While the processing time can vary greatly depending on the number of scientists involved, their taxonomic knowledge and the habitat being studied, a recent study by Egbert et al. (2020) showed that a team of students needed around 600 hours to fully process 30 hours of deep-sea videos. However, as technology continues to improve, automated data analysis through the incorporation of deep learning models could offer a faster and more resource-efficient alternative (Beuchel et al. 2010; Durden et al. 2016b, 2021). Despite the advantages of OFOS and AUV documentation, as well as the potential of deep learning models for automated fauna detection and quantification, the number of studies investigating and comparing both sampling and processing techniques based on field studies is still hampered by the lack of sufficient quantitative and qualitative data.

Therefore, here we extend the resources both ecologically and methodologically by providing novel and unique data on the epimegafauna from two overlapping abyssal plains at 10°N in the Great Atlantic Sargassum Belt with documented Sargassum falls (Baker et al. 2018). Community structure and epimegafauna density were investigated using AUV images and OFOS video recordings acquired during the deep-sea expedition of R/V Sonne SO237 and R/V Meteor M139, respectively. In addition, the AUV images were processed by using and building a deep learning model for autonomous detection of abyssal megafauna based on the YOLOv5 algorithm. In addition to assessing the observed megabenthos community in an ecological context, both processing techniques and sampling devices were investigated and compared. Consequently, we investigated the following hypotheses: (i) Megafauna communities inhabiting the abyssal plain sampled by the AUV have a similar density and community composition to the section of the same plain sampled by the OFOS. The automated data analysis by the self-trained deep learning model is sufficiently effective because the megafauna density and community composition (ii) are similar for identical datasets when processed manually or by the model, and (iii) are the same for the different AUV subsections processed either manually or by the model.

Study areas and habitat description

Following the R/V Sonne expedition SO237, an abyssal plain with remarkable Sargassum falls near the Mid-Atlantic Ridge (MAR) was selected as a study area and investigated on January 1st, 2015 with an AUV (HYDROID Inc.) (Fig. 1; Baker et al. 2018). During the follow-up expedition of the R/V Meteor M139 in 2017, the same area was sampled with an OFOS camera dive on July 30, 2017 (Fig. 1; Arndt et al. 2017). As the stations near the MAR do not overlap perfectly and were sampled in different years, both are treated as separate sampling stations. The sampling stations investigated and compared included the OFOS station A5/6 of the M139 cruise and the AUV station A3 of the SO237 cruise (Supplementary Table 3).

Near the Mid-Atlantic Ridge, the first survey of the AUV station A3 by the R/V Sonne Expedition SO237 took place at N10°21.00', W036°57.62' (Devey et al. 2015). In addition to the start and end depths of 5116 m and 5131 m below the water surface, respectively, the AUV dive followed several serpentines until it covered approximately 354921 m² of the seafloor. Taking the R/V Meteor Expedition M139 into account, the abyssal plain investigated was between N10°20.37', W036°58.69' and N10°20.46', W036°56.65' (Fig. 1a & d; Arndt et al. 2017). With start and end depths of 5091 m and 5107 m, respectively, the eastward dive covered a distance of about 5235 m and 22094 m² of deep seafloor. In both cases, the stations were dominated by soft sediment and showed remarkable Sargassum accumulations (Supplementary Fig. 1–3).

Image and video sampling

The AUV Abyss (built by HYDROID Inc.) was used to conduct photo documentation (LED flash light system; 2048 × 2048 pixel images) during the R/V Sonne expedition SO237. This enabled continuous documentation through multiple, mostly overlapping images (Baker et al. 2018). Attached to the AUV was a Pike camera system equipped with a 15 mm Nikkor underwater lens with a focal length of 22 mm and providing a field of view of 41 degrees (Baker et al. 2018). Instruments built into the device also provided information on the depth, position and heading of the AUV, its distance from the seafloor and the approximate image pagination in meters for the majority of all images taken. Using the distance between the sampler and the seafloor (average values for AUV dive 163 equal 10.15 ± 0.42 m) together with the image side length (average values for AUV dive 163 equal 7.59 ± 0.31 m) and the number of images provides size measurements of the observed bottom area.

The video documentation of the R/V Meteor Expedition M139 followed the same procedure as described by Scepanski et al. (2024). An OFOS provided continuous documentation through several, automatically generated video files without temporal gaps. The equipment provided by the Helmholtz Center for Ocean Research GEOMAR consisted of a launcher, an HDCS camera system (Sea & Sun Technology GmbH, Trappenkamp, Germany) with an HD camcorder (CANON Legria, Tokyo, Japan, resolution 1920 × 1080 pixels) integrated into a titanium housing with a borosilicate dome aperture, LED spotlights (Bowtech Products Ltd., Aberdeen, UK) and three downward-facing laser pointers. The OFOS camera and laser system were towed behind the research vessel and moved at a speed of approximately 0.7 kn (3 hrs 52 min at 0.7 kn and 18 min at 0.4 kn), with the top and bottom lasers fixed vertically and the third laser tilted sideways downwards. Therefore, the upper and lower lasers projected stationary dots, while the third laser produced a horizontally moving dot between these dots depending on the distance to the ground. The constant distance between the upper and lower laser points allowed the size of both the animals and the observed area to be measured. When all laser points were aligned, the distance between OFOS and the seabed was exactly 1.5 m, with an average distance of 2 to 3 m.

Surveyed parameters and area assessment

During video and image processing, the quantity and quality of the megafauna epibenthos as well as the presence of organisms and large Sargassum patches were recorded. Only individuals of the epimegafauna were examined, including only taxa larger than 20 cm and visible on/above the deep-sea floor (epifauna) (Scepanski et al. 2024). The presence of buried organisms, such as polychaetes, was detected, but their numbers were not quantified here.

To quantitatively evaluate the R/V Sonne 237 data set, the extent of the seafloor examined was calculated for each replicate and the entire AUV dive. Using the mean image side length of 7.59 m (calculated on the basis of all AUV images provided), the corresponding area per image of 57.62 m² and the number of images examined (complete dive: 9229; see Supplementary Table 3), approximate values were calculated for the areas examined.

To enable further quantitative calculations for the R/V Meteor 139 data set, the area of the seabed was measured for all video files and the entire dive and processed according to Scepanski et al. (2024). Distance travelled was estimated based on speed and times between certain situations (e.g., bottom visibility, speed change, onset of surfacing). Width was calculated from three randomly selected moments of each replicate/video file (mean width 4.22 ± 1.17 m). We estimated megafaunal density for each replicate/video file by combining the length and width of the total dive observations with the duration of each video file showing the seafloor and summing for the entire dive.

Development and application of a deep learning detection model for megafauna and Sargassum identification

In order to investigate the occurrence of megafauna and Sargassum in the R/V Sonne 237 dataset, a deep learning recognition model was developed to speed up the processing. The images were therefore annotated with CVAT (Sekachev et. al. 2020) and used for model training utilising the YOLO V5 algorithm (Jocher et. al. 2022) on PyCharm (JetBrains s.r.o., Prague, Czech Republic). As the model was developed to reduce image processing time, the number of images used for training, validation and testing was continuously increased until an acceptable accuracy was achieved, which was then used to process all remaining images. The last two models created, hereafter referred to as the "older model" and "final model", were used to validate and process the dataset. Therefore, the older and final model were based on the first 2250 and 3105 images of the R/V Sonne 237 dive 163 datasets, respectively. While the older dataset was trained and tested on 2026 of the first 2250 processed images and validated on the next 855 images, the latest and final model was trained and tested on 2793 images and validated on 312 images to further increase accuracy. In both cases, images 0 to 500, 1000 to 1020 and 2000 to 2140, on which the model was based, were thoroughly tested (object search using an R tool to detect characteristic shapes and colours, followed by a manual search), while for images 500 to 999 and 2141 to 2870, an automatic annotation based on a model built on the previously mentioned images was tested and complemented by a quick manual search. For the remaining images, only the labelling of the final model was checked. With the exclusion of inanimate objects (e.g. shadows, rocks, indeterminate objects), the final model included large patches of Sargassum, characteristic Polychaeta holes often containing small pieces of the macroalgae, two forms of Porifera, Pennatulacea, Hydrozoa and undetermined Cnidaria, Crustacea, Crinoidea as well as an animal resembling both a Crinoidea and a brisingid Asteroidea, Holothuroidea, Asteroidea, which were intentionally combined with Ophiuroidea as the preliminary models could not distinguish between them, Echinoidea, Cephalopoda and Gastropoda, Teleostei, undetermined animals and potential Pennatulacea, as well as a completely unclassified megafaunal morphotaxon.

The final deep learning detection model was then used to automatically detect the organisms for all remaining images and replicates. However, due to the overlap between images, the number of detected tags did not correspond to the number of megafauna individuals present, as the same individual (as well as Sargassum patches and other objects) could be detected on consecutive images. Therefore, a script was written in PyCharm that uses the position of each detected organism to automatically subtract the animals that were counted multiple times due to overlap to determine the final number of megafauna individuals present. Our preliminary tests showed that the detection after automatic correction resulted in the correct number of animals in most cases. Accuracy was over 95%, with an error rate of less than 5% for missing an animal counted multiple times or misclassifying an individual as a single animal for each animal distinguished by the model.

Morphotaxa annotations and qualitative/quantitative estimates

To quantify and classify the studied epibenthic megafauna communities, we identified the observed animals based on their morphological characteristics down to the lowest possible taxonomic rank. For this purpose, we used the photographic identification catalogue of Vinha et al. (2022), the deep-sea species identification application Deep Sea ID v1.2 (Glover et al. 2015) and its underlying database, the World Register of Deep-Sea Species (WoRDSS) (Glover et al. 2023). Despite their distinctive morphological characteristics, some animals could not be determined and were treated as different morphotaxa. For better comparison, all morphotaxa of the megafauna were assigned to an artificially created group of higher taxa comprising Porifera, Pennatulacea, Hydrozoa, indeterminate Cnidaria, Crustacea, Crinoidea, an additional intermediate group related to either the Crinoidea or the brisingid Asteroidea, Holothuroidea, Ophiuroidea, Asterozoa, Echinoidea, Teleostei and undetermined Megafauna.

Sargassum quantification

Both manual and automated methods were used to quantify the submerged Sargassum patches. The quantitative parameters studied included algal occurrence, area covered by the algae, and Sargassum coverage per square meter of seafloor. In the case of the R/V Meteor 139 expedition, the area covered by Sargassum was measured manually by using the laser system connected to the OFOS and approximating the shape of each patch to simplified geometric shapes. During the R/V Sonne 237 expedition, all parameters were automatically recorded and calculated except for recording the occurrence of Sargassum patches during the manually processed section of the AUV dive. The area covered by Sargassum was measured using a PyCharm script that automatically extracted each Sargassum patch from each image where Sargassum was present, counted the number of pixels corresponding to the algae, and calculated the area based on the known dimensions of each image. All images were therefore standardized considering their lighting conditions. For each Sargassum patch, the corresponding bounding box (area covered by the algae) was extracted as a separate "bounding box image". Each bounding box image was then subjected to histogram equalization and adaptive thresholding to detect the pixels covered by Sargassum. The adaptive threshold calculation included contrast adjustments (alpha = 1. 5 and beta = -10), sharpness enhancement with a median filter (each pixel is replaced by the average of the surrounding 25 pixels), noise reduction with a median blur filter, simple binary thresholding (threshold = 250; maximum value = 255) as well as Adaptive Mean and Adaptive Gaussian thresholding (both: Maximum value = 255, block size = 111 and constant = 2) with "neighborhood mean" and "weighted sum of neighborhood values with Gaussian weights" as adaptive methods. Next, the number of Sargassum pixels in each bounding box image was counted and converted to the covered area. The area of each patch was summed for each replicate and divided by the corresponding area of the seafloor to obtain the Sargassum coverage per square meter seafloor.

Data analysis

Statistical tests and graphs were performed using RStudio version 4.2.2 (RStudio®, Boston, Massachusetts, USA;R Core Team 2022). The following R packages were used: “car” (Fox and Weisberg 2019), “colorspace” (Zeileis et al. 2020), “cowplot” (Wilke 2020), “dplyr” (Wickham et al. 2021), “forcats” (Wickham 2021), “ggplot2” (Wickham 2016), “grid” (R Core Team 2022), “gridExtra” (Auguie 2017), “hrbrthemes” (Rubis 2020), “indicspecies” (de Cáceres and Legendre 2009), “janitor” (Firke S 2023), “lawstat” (Gastwirth et al. 2020), “pacman” (Rinker and Kurkiewicz 2017), “PMCMRplus” (Pohlert 2022), “pvclust” (Suzuki et al. 2019), “RColorBrewer” (Neuwirth 2014), “Rcpp” (Eddelbuettel 2013), “reshape2” (Wickham 2007), “scales” (Wickham and Seidel 2020), “svglite” (Wickham et al. 2020), “tidyr”(Wickham 2020) and “tidyverse”(Wickham et al. 2019) as well as “vegan” (Oksanen et al. 2019). Furthermore, maps were created using the GEBCO Grid tool (© GEBCO 2020) together with the open-source software QGIS version 3.28.3.

As sampling techniques differed between expeditions and stations, two approaches were taken to obtain replicates for statistical testing and comparison between stations. The technical replicates were not perfectly independent, but the area sampled for the different technical replicates was similar, and the separation of replicates was independent of the observer and was separated before the data were analysed. For the AUV stations of expedition R/V Sonne SO237, along the serpentine dive, all images covering each of the 22 parallel straight segments and half of the neighbouring bays (see Fig. 1) were used as technical replicates, resulting in 22 technical replicates of similar length (see Supplementary Table 3). It should be noted that on the AUV dive, its direction over ground and its speed were chosen so that each straight section covered approximately the same area (mean number of images per replicate 406 ± 20). As the AUV dive was not originally deployed to investigate megafauna abundance but focused on Sargassum observations (Baker et al. 2018), all megafauna observations were random and not manipulated from onboard. For the remaining station of the R/V Meteor expedition M139, replicates were carried out according to the same principle as Scepanski et al. (2024). Thus, automatically generated video files based on time were used as technical replicas, with each file starting immediately after the end of the previous one. Overall, the amount of filming was relatively consistent between files (video recordings of mostly similar length with a mean duration of 37 minutes and 13 seconds ± 0 minutes and 2 seconds) (Supplementary Table 3). All OFOS dives, their corresponding speed and individual directions over the bottom could not be manipulated from onboard, resulting in variable range recording and random selection duration independent of the device operator. Ultimately, eight technical replicates were achieved at the OFOS station.

Accumulation curves were generated based on the richness of higher taxonomic groups to clarify whether enough area had been sampled to adequately represent the megafauna community at each station. We used the specaccum() function of the Vegan package with the random method for 100 permutations. Corresponding plots, all showing an acceptable pattern, are included in the online supplement (see Supplementary Figs. 4 & 5).

The homoscedasticity and the normal distribution within the technical replicates were tested using the Levene and Shapiro-Wilk test for all parameters. If the data showed homoscedasticity and normal distribution, comparisons were made using one-way ANOVAs and post-hoc Tukey tests. Otherwise, comparisons were made using a one-way Kruskal-Wallis analysis of variance and Dunn's post-hoc tests (pairwise comparisons with Dunn's test for multiple comparisons of independent samples). To avoid unnecessary repetition in the results section, information on tests, sample size, degrees of freedom and p-values were only provided at the relevant point in the text if no figure contained the corresponding data. Only differences with p-values below 0.05 were considered significant.

Possible differences in megafauna community composition between stations, sampling events and processing techniques were investigated by several non-metric multidimensional scaling diagrams (NMDS). For this purpose, densities (number of individuals per m²) for each higher taxonomic group at the corresponding level were used together with Bray-Curtis distance measures. With the exception of Supplementary Fig. 6a, in which three artificially generated replicates were used for extended validation of the old model, all replicates used were identical to those described above (further information in Supplementary Table 1). Stress values of 0.3 or more were interpreted as suspect, values below 0.2 as reliable, and values in between with caution. Analysis of similarities (ANOSIM) was performed for each NMDS representation, with ANOSIM applying 9999 permutations and Bray-Curtis distances for dissimilarity of the corresponding data sets. The function "r.g" (R Core Team 2022) was used to detect potential correlations between selected binary vectors. Indicator species analyses were performed to determine whether and - if so - which higher taxonomic groups were significantly associated with which station, sampling event and/or processing technique.

The differences between the overall densities of the individual stations, sampling methods and processing techniques and the specific densities of the higher taxonomic groups as well as Sargassum densities were investigated by comparing the corresponding data sets. The mean values were obtained from the aforementioned number of replicates. When variability is given in the text for megafauna density and Sargassum coverage it is provided as mean ± SD.

Overall megafauna composition, density and Sargassum occurrence

A total of 33655 different megafauna individuals were observed during the survey of the entire fauna, regardless of year, station, sampling method or processing technique (Fig. 2). Of these, 33460 individuals were observed during the AUV survey at station A3 and 195 individuals during the OFOS dive at station A 5/6. Including at least 33 different megafauna taxa, 15 higher taxonomic groups were distinguished, including Porifera, Pennatulacea, Hydrozoa, unidentified Cnidaria, Crustacea, Holothuroidea, Ophiuroidea, Asteroidea, Crinoidea, an additional intermediate group related to either the Crinoidea or the brisingid Asteroidea, Echinoidea, Gastropoda, Cephalopoda, Teleostei and an unclassified megafauna group (Fig. 2; Table 1). The AUV approach included all identified groups except the Crinoidea, whereas the OFOS approach included only the eight higher taxonomic groups of Porifera, Pennatulacea, Crustacea, Holothuroidea, Asteroidea, Crinoidea, Echinoidea and Teleostei. Overall, the Crustacea had the largest number of classified genera with Galathea and Benthesicymus (Fig. 2; Table 1). Other genera, species and families identified were the Pennatulacea Umbellula, the decapod family Pandalidae, the hydrozoan family Corymorphidae and the teleostean family Macrouridae (see Table 1 for more information).

When examining the individual occurrence of these higher taxonomic groups on each dive, starting with the AUV dive, it was found that all but a few higher taxonomic groups were continuously detected (Supplementary Fig. 2). These exceptions include Hydrozoa and other Cnidaria, Ophiuroidea, Cephallopoda, Gastropoda and the indeterminate megafauna group. While the latter group occurred relatively frequently in the sections processed with the deep learning model, the other groups were found relatively frequently in the manually processed sections. Hydrozoa, Ophiuroidea, Cephalopoda and Gastropoda could not be recognized by the model and appeared exclusively in the manually processed section. With the exception of the rounded Porifera, which had a single local maximum at the beginning of the manually processed section, no other maxima were visually observed. In addition, it was possible to continuously record potential Pennatulacea, Polychaeta and patches of sedimented Sargassum both manually and with the model. However, potential megafauna individuals were only recorded manually and later classified as "undetermined" during the learning process of the model. A relatively large number of individuals or objects in this category were measured in the section processed by the model. Looking at the occurrence of individuals during the OFOS dive, only Porifera and Teleostei were consistently present (Supplementary Fig. 1). With the exception of Crinoidea and Echinoidea, both of which were only observed three times and once respectively, the other higher taxonomic groups were relatively abundant in the second half of the dive. The occurrence of Sargassum was measured continuously during both sampling events.

Both the AUV approach (manual and deep learning model sections) and the OFOS approach showed large amounts of fauna in terms of megafauna density as well as large Sargassum accumulations (Table 2). Overall, the megafauna densities of the AUV approach were significantly higher than those of the OFOS survey (Fig. 3a). Corresponding overall mean densities of around 9.65 × 10^− 2 ± 1.91 × 10^− 2 ind./m² (mean ± SD) at the AUV station were more than ten times higher than those at the OFOS station with around 9.45 × 10^− 3 ± 4.97 × 10^− 3 ind./m² (mean ± SD Fig. 3a; Table 1). Sargassum densities were significantly greater at the AUV station with 1.16 × 10^− 3 ± 1.42.× 10^− 4 m² (mean ± SD) covered by the algae per m² ocean floor compared to 2.84 × 10^− 4 ± 4.78 × 10^− 4 m² (mean ± SD) covered by the algae per m² ocean floor at the OFOS station (Kruskal-Wallis chi-squared = 9.5806, df = 1, p-value = 1.966 × 10^− 3). Further comparisons between the OFOS approach and the differently processed AUV sections showed similar results (Fig. 3a-c; Table 1). For example, the manually processed AUV section 1, the total area processed by the deep learning model (AUV section 2 & 3) and the AUV subsection 3, which corresponds to the last seven automatically processed replicates, showed significantly higher megafauna densities compared to the OFOS approach (Fig. 3b & c). Only AUV subsection 2, which represents the first eight replicates processed with the deep learning model, did not differ from the OFOS approach concerning megafauna density (Fig. 3c). The manually processed AUV section also did not differ significantly from either the total area or the two subsections processed with the deep learning model with regard to megafauna density (Fig. 3b & c). Here, the mean densities corresponded to 1.07 × 10^− 1 ± 2.86 × 10^− 2 ind./m² (mean ± SD) processed manually, 9.16 × 10^− 2 ± 1.08 × 10^− 2 ind./m² (mean ± SD) processed with the deep learning model, where 8.72 × 10^− 2 ± 9.32 × 10^− 3 ind./m² and 9.67 × 10^− 2 ± 1.07 × 10^− 2 ind./m² (mean ± SD) were the mean values of the first eight and last seven replicates processed with the deep learning model, respectively (Fig. 3; Table 1).

In terms of Sargassum coverage per square meter seafloor, both the total area processed with the deep learning model (AUV section 2&3; Kruskal-Wallis chi-squared = 12.096, df = 1, p-value = 2.363 × 10^− 3) and each of these subsections had significantly higher Sargassum coverages per square meter seafloor compared to the OFOS approach (Kruskal-Wallis chi-squared = 12.101, df = 1, p-value = 7.045 × 10^− 3) (Table 1). Only the manually processed AUV subsection 1 did not differ from the OFOS approach, while it was neither significantly different from the total area nor from the two subsections processed with the deep learning model. The mean coverages per square meter seafloor corresponded to 1.06 × 10^− 3 ± 1.15 × 10^− 4 m² (mean ± SD) covered by Sargassum/m² when processed manually and 1.21 × 10^− 3 ± 1.31 × 10^− 4 m² (mean ± SD) covered by Sargassum/m² when processed with the deep learning model, where 1.21 × 10^− 3 ± 1.3 × 10^− 4 m²/m² and 1.21 × 10^− 3 ± 1.43 × 10^− 4 m²/m² (mean ± SD) were the mean values of the first eight and the last seven replicates processed with the deep learning model, respectively (Table 1).

Differences between manual processing and deep learning models

Validation and precision of an older and the final model

The two validation tables (Supplementary Tables 4 & 5), both based on the corresponding validation set, provide information on how effective an older model and the final model were in correctly recognizing each category. In terms of precision, several categories achieved a precision value above 0.5 for both models, with the old model achieving this threshold in six out of fourteen higher taxonomic groups and the final model in nine out of fourteen higher taxonomic groups. As for Sargassum patches, both models reached precision values around 0.8. The overall precision was 0.473 for the old model and 0.725 for the final model. In terms of recall, four and seven higher taxonomic groups, as well as the Sargassum category had a recall value equal to or above 0.5 for the old and final models, respectively, with an overall recall of 0.298 for the old and 0.346 for the final model. Similarly, together with Sargassum patches, five and six higher taxonomic groups achieved an F1 value of over 0.5 for the old and final models respectively, with total F1 values of 0.366 and 0.468, respectively.

Additional information on the effectiveness of both the older and the final model can be observed by comparing the detected community composition with that obtained manually by using non-metric multidimensional scaling based on Bray-Curtis distance measures (Supplementary Fig. 6). The older model, in comparison of the detected community from the extended validation dataset (images 2251 to 3105) with the identical manually processed section, revealed no significant differences in either the NMDS plot or the similarity analysis, with ANOSIM R and p values of 1 and 0.1, respectively. Subsequent analysis of indicator species revealed no indicators. Also, compering the community from the manually processed first AUV section with the identical one processed with the final model based on NMDS-plot and similarity analysis, revealed no significant differences (Supplementary Fig. 6) as represented by an ANOSIM R-value of -0.1127 and a p-value above 0.05 (0.9351). Subsequent analysis of indicator species revealed that three higher taxonomic groups were significantly associated with one of the two methods, with Hydrozoa and Ophiuroidea significantly associated with the manual method and undetermined megafauna significantly associated with the modelling approach.

Comparing the overall density provided by the old model from the extended validation dataset (images 2251 to 3105) with the same manually processed one, it is evident that the latter achieves significantly higher densities (Fig. 4). Regarding the density of the individual higher taxonomic groups and their percentage of the total density, four groups showed significant differences (Fig. 4). These included the Porifera, which had significantly higher densities in the manually processed section, as well as the Pennatulacea, the intermediate group related to either the Crinoidea or the brisingid Asteroidea, and the indeterminate megafauna group, which made a significantly higher contribution to the overall density when the section was processed with the model. Visually, the community composition was almost identical, with Porifera, Pennatulacea and Asteroidea representing the three largest relative proportions.

The comparison of the total density between the manually processed section and the same section processed with the final model showed no significant differences (Fig. 5). Also, when comparing the density of each higher taxonomic group and its percentage of the total density between the processing methods, only three groups showed significant differences (Fig. 5). These are Hydrozoa and Ophiuroidea, both of which are undetectable with the model, and the group of undetermined megafauna, which has significantly higher values when the section is processed with the model. Overall, the composition of the megafauna community was almost identical, with Porifera, Pennatulacea and Asteroidea accounting for the three largest relative proportions in both cases. When examining the individual occurrence of each higher taxonomic group as well as additional categories such as round Porifera, potential Pennatulacea or Polychaeta, the model correctly recognized the majority of all manually found groups (Supplementary Fig. 3). Exceptions included Hydrozoa, Ophiuroidea, Cephallopoda and Gastropoda, which could not be recognised by the model, as well as the group of “undetermined megafauna”, which was recognized more frequently than manually found. Overall, although some individuals were overlooked or only a few non-existent ones were recognized by the model, the overall occurrence patterns resulting from manual processing of the same section and processing with the model were similar (Supplementary Fig. 3).

The manually analysed AUV section compared with the remaining ones processed by the final model

A non-metric, multidimensional scaling based on Bray-Curtis distance measures together with a similarity analysis showed significant differences between the megafauna communities between the manually processed AUV section 1 and the other sections (AUV section 2 & 3) examined by the final deep learning model (Supplementary Fig. 7). However, in terms of community composition similarity, an ANOSIM R-value of 0.5832 indicated neither a completely identical (R-value = 0) nor a completely dissimilar (R-value = 1) composition of the megafauna communities. At a stress-value below 0.0001, no visible separation between manually and automatically processed sections could be detected. The minority of distinguished higher taxonomic groups clustered between/within both sections, with the majority located to the right of the automatically processed section. Subsequent analysis of the indicator species revealed that ten of the fourteen higher taxonomic groups were significantly associated with one of the two sections (Supplementary Table 1). While Porifera, Hydrozoa, the remaining Cnidaria, Holothuroidea and Ophiuroidea were significantly associated with the manually processed section, Crustacea, Asteroidea, Echinoidea, Teleostei and undetermined megafauna were associated with those examined by the deep learning model (Supplementary Table 1). Furthermore, splitting the section processed by the model into two sections of similar length (AUV section 2 and AUV section 3) to the manually processed section mirrored these results, showing significant differences and identical indicators along with a seemingly identical NMDS plot (Supplementary Fig. 7b; Supplementary Table 1).

In contrast, when comparing the density of each higher taxonomic group and their percentage share of the total density between the manually processed and the other, model-processed plots, quantitative differences in the identified communities become apparent (Fig. 6). While Porifera, Pennatulacea and Asteroidea consistently represent the three most important relative proportions, direct comparison reveals significant differences and variations. Porifera, the indicators significantly associated with the manually processed section, had significantly higher densities and a significantly greater proportion of the total density in the manually processed section than those processed with the model. The Pennatulacea did not differ in density, but showed a significantly different contribution to total density, with a significantly higher proportion measured in the area processed with the deep learning model. Hydrozoa and undetermined Cnidaria as well as Holothuroidea, Ophiuroidea and Cephalopoda reached significantly higher densities and percentage contributions to the total density in the manually processed section than in the model. In contrast, Crustacea as well as Asteroidea, Echinoidea and Teleostei reached significantly higher densities and percentage contributions to the total density in the model-processed section. The remaining groups showed no significant differences.

Differences between OFOS and AUV sampling

Community composition

The use of non-metric, multidimensional scaling based on Bray-Curtis distance measures together with similarity analysis revealed significant differences in megafauna community composition between the two sampling events (Fig. 7). With a stress value of approximately 0.0023, a clear separation was observed between the OFOS sampling event and the combined AUV sampling event, with the respective communities inhabiting the middle left and middle right areas of the plot, respectively. No overlap was observed between sampling events, and the majority of the distinguished higher taxonomic groups clustered in relative proximity to the AUV survey. The subsequent analysis of the indicator species revealed that ten of the originally fifteen distinguished higher taxonomic groups were significantly associated with one of the sampling events (Supplementary Table 2). Of these, nine (e.g. Pennatulacea, Porifera, Asteroidea, Teleostei and undetermined megafauna) were associated with the AUV survey and one (Crinoidea) with the OFOS. Additional NMDS-plots and similarity analysis comparing the OFOS dataset with the manually processed AUV dataset and either the entire or two subsections of the model-processed AUV dataset yielded nearly identical results, including significantly different megafauna communities (Supplementary Fig. 8; Supplementary Table 2). With identical stress values, sampling events and higher taxonomic groups, the clusters were almost identical to those shown in Fig. 7, with the AUV datasets overlapping each other. The corresponding analysis of indicator species showed all of the previously mentioned associations, with information on certain higher taxonomic groups being expanded. Thus, Cnidaria, Hydrozoa, Ophiuroidea, Holothuroidea and Cephalopoda were found to be significantly associated exclusively with the manually processed sections, while Echinoidea, Teleostei and undetermined megafauna were significantly associated exclusively with the automatically processed section. The remaining indicators were significantly associated with all AUV datasets.

Quantitative comparison

When comparing the density of each higher taxonomic group and its percentage of the total density between the OFOS approach and the combined AUV approach, quantitative differences in the identified communities become apparent (Fig. 8). While Porifera, Pennatulacea and Asteroidea accounted for the three largest relative proportions in the AUV study, the proportions were relatively balanced in the OFOS approach. Here, Pennatulacea and Crustacea dominated, followed by Porifera, Teleostei and Holothuroidea, respectively. The direct comparison between higher taxonomic groups also showed clear differences. Porifera, which were significantly associated with the AUV approach in this scenario, had significantly higher densities and a significantly greater contribution to total density during the AUV survey compared to the OFOS dive. Pennatulacea, while exhibiting significantly higher densities in the area surveyed by the AUV, did not differ in terms of percentage contribution to overall density. Instead, a relatively higher percentage contribution was found, using the OFOS approach. Although the remaining Cnidaria, more precisely Hydrozoa and undetermined Cnidaria, were only measured with the AUV, only the undetermined Cnidaria showed significant differences between sampling events, with significantly higher values obtained with the AUV approach. Both the specific density and the percentage of the total density were significantly higher for the Crustacea in the OFOS sampling. With the exception of the Ophiuroidea, all higher taxonomic groups belonging to the Echinodermata showed significant differences between the OFOS and AUV surveys. In terms of specific density, Holothuroidea, Asteroidea, the intermediate group of Crinoidea or brisingid Asteroidea and Echinoidea showed significantly higher densities in the AUV survey, while Crinoidea were significantly more abundant in the OFOS dive. In terms of percentage, Holothuroidea, Crinoidea and Echinoidea had a significantly greater proportion in the OFOS dive, while the remaining species did so in the AUV dive. While Gastropoda and Cephalopoda showed no significant differences, this was the case for Teleostei and undetermined megafauna. While Teleostei reached significantly higher densities following the AUV survey, their percentage was significantly higher during the OFOS dive. For the indeterminate form of megafauna, both the density and the percentage were significantly higher during the AUV survey.

Although many studies examined how different visual imaging and processing techniques affect investigations of abyssal megafauna communities (e.g. Gobalet 2001; Durden et al. 2016a, b), only a few evaluated both while providing a faunal quantification as well. Here, we provide one of the first quantitative assessments of epi-megabenthos inhabiting two intersecting abyssal plains within the Great Atlantic Sargassum Belt, expanding knowledge on habitats facing regular Sargassum falls. Further, by evaluating results provided by AUV and OFOS-sampling, as well as processing manually or by deep learning model, this study offers valuable insights into the efficiency of different visual imaging and processing practices.

Model evaluation

In general, the final model analysis of the epimegafauna performed well, with most of the distinguished model groups achieving precision, recall and F1 values above 0.5 and overall values of 0.725, 0.346 and 0.468, respectively (Supplementary Tables 4 & 5). For precision, defined as the ratio of correctly predicted positive instances to all positive instances predicted by the model, an overall value of about 72.5% indicates relatively few false positives (Durden et al. 2021; Cuvelier et al. 2024). For recall, which quantifies how many instances were identified by the model in relation to the total number of instances actually present, shows that only 34.6% of the manually annotated objects were also recognized by the final model (Durden et al. 2021; Cuvelier et al. 2024). Overall, with an F1 score of 0.468, the model appears to strike a relative balance between the two parameters, accurately quantifying the community while underestimating it to a certain extent (Durden et al. 2021; Cuvelier et al. 2024). Despite some bias from re-analyzing the images used for training, this overall effectiveness was supported by the final model accurately processing its training dataset. This was evidenced by the lack of significant differences in overall megafauna density (Fig. 5) and community composition (Fig. 6). As the accuracy, recall and corresponding F1-score of the final model are above 0.5 for most higher taxonomic groups as well as for Sargassum, it may also be remarkably effective at this level (Durden et al. 2021; Cuvelier et al. 2024). Categories below this threshold consisted mostly and in terms of precision exclusively of rarely observed model groups such as Hydrozoa, Cephalopoda or Gastropoda, for which were not enough images to adequately train the model (Durden et al. 2021; Cuvelier et al. 2024). As shown in studies by Durden et al. (2016b, 2021) or Liu et al. (2024), several hundred to thousand images per animal category are recommended for accurate recognition to cover a variety of animal poses, sizes, lighting conditions and environments (Gobi 2010; Cuvelier et al. 2024). Consequently, it fits the picture that the rarely observed model groups could not be recognized well by the final model, which also explains why significant and relative differences in both one-sided comparison (Fig. 5) and individual occurrence (Supplementary Fig. 3) were only measured for rare taxa. Considering the final model was, aside from the biased comparison only validated with 312 validation images, we investigated an additional “older” model trained and tested on fewer images (2026 of the first 2250 images) but validated with 855 manually processed images (see method section). Having been trained on fewer images, as expected, the overall precision of the older model decreased, as did the precision for certain categories. However, several categories still exceeded the precision threshold of 0.5 (Supplementary Tables 4 & 5). Together with the NMDS plot of the older model (Supplementary Fig. 6), which shows no significant differences, the statements made for the final model are already implied here. However, a comparison of the manually calculated densities with the densities derived from the older model based on the extended validation set revealed significant differences regarding the total megafauna and the common higher taxonomic groups with accuracies above 0.5. This shows that the older model is not as effective as it could be (Fig. 4, Supplementary Table 5). The lack of sufficient data in the datasets of the older models may therefore have prevented a definitive assessment of their ability to recognize and quantify the abyssal fauna of this area. Therefore, the lack of sufficient data in the older model's dataset may have prevented a definitive evaluation of their ability to detect and quantify the abyssal fauna of this area.

For this task, it is logical to compare the manually processed section with the areas processed by the final deep learning model. While the spatial distribution and occurrence of abyssal organisms is related to both biotic interactions (Milligan et al. 2016) and key environmental characteristics (Watling et al. 2013) including habitat heterogeneity (Buhl-Mortensen et al. 2010), our study could not observe any biological processes like e.g. predation. Considering only the relatively homogeneous abyssal plain dominated by soft sediments (Supplementary Fig. 2; Baker et al. 2018), we expected the megafauna community to be consistent across sections, as environmental characteristics should not differ (Watling et al. 2013; Schoenle et al. 2021; Scepanski et al. 2024). Indeed, the overall densities and, except for those groups for which the model's dataset was not large enough to make reliable predictions, the individual occurrences of most of the higher taxonomic groups over time provided by the model were consistent with manual processing, both significantly and relatively (Fig. 6; Supplementary Fig. 7). However, while Porifera, Pennatulacea and Asteroidea remained the most abundant groups, the community composition, specific density and percentage of higher taxonomic groups differed significantly between the analysis methods (Fig. 6; Supplementary Fig. 7; Supplementary Table 1& 2). As the abyssal plain is a relatively homogeneous area with assumed stable environmental factors (Watling et al. 2013; Baker et al. 2018), the influence could be due to possible changes in the spatial occurrence of sedimented Sargassum and it`s potential positive influence (see below) (Fleury and Drazen 2013; Molodtsova et al. 2017; Baker et al. 2018). However, their apparently constant presence undermines this explanation (Supplementary Figs. 2 & 3). Therefore, the differences at the higher taxonomic level could be related either to the effectiveness of the model used or to an unknown environmental or biotic factor, not assessed here, that differs between the sections compared. Considering the ability of the final deep learning model to effectively quantify the abyssal megafauna community, as well as sedimented Sargassum studied here, we conclude that it provides realistic overall measurements. However, its efficiency might decrease at a higher taxonomic level, as not enough training images were available for some groups. At the moment, this model does not yet allow fully unsupervised recognition, as some model groups are still underrepresented and need to be treated manually. However, since various results confirm the quality of the final model even at a higher taxonomic level, it could be used cautiously at this level of analysis as well.

Benefits and disadvantages of utilising (this) deep learning model

The assessment and quantification of deep-sea megafauna communities using deep learning models offers both advantages and disadvantages compared to manual assessments. Apart from the precision offered and the relative ease of application, which allows scientists outside of deep-sea ecology/zoology to use pre-trained models to study abyssal communities, the biggest advantage may be time. If the model is already trained on a robust database, it can process datasets much faster than manual observations (e.g. Schoening et al. 2012; Durden et al. 2016b, 2021; Liu et al. 2024; Cuvelier et al. 2024). This fact makes it possible to conduct studies on a larger scale (Beuchel et al. 2010; Stanchev et al. 2020). While the manual processing time in the different studies varies from seven (Cuvelier et al. 2024) to 20 minutes per image (this study) over 30–60 minutes per image (Schoening et al. 2012) up to 30 hours for 600 hours of video material (Stanchev et al. 2020), the model requires only a fraction of this. Even when multiple experts are involved to speed up processing, which increases costs, manual processing still takes significantly longer than using a model and can lead to inconsistent results (Gobalet 2001; Durden et al. 2016a, b; Cuvelier et al. 2024). Multiple experts may come to different taxonomic conclusions, changing the community assessed and producing different results (Gobalet 2001; Durden et al. 2016a, b). Should this happen during the training of a deep learning model, the inclusion of conflicting annotations could lead to a weakened model, as the quality of the dataset is very important. Even if it is necessary to train a model from scratch, implementing deep learning models from the beginning could significantly reduce the processing time. Studies could start with manual processing of the datasets to create an initial model that can then help process subsequent data sections. Through a continuous improvement cycle, the overall processing becomes faster and faster until the model can replace manual processing. Comparisons between the older and final model showed a significant increase in precision, meaning that newly processed images contain fewer errors, speeding up corrections. In general, using a model for abyssal data saves time and therefore resources and money. Another advantage of a well-trained model over manual processing is the consistent precision of the results compared to humans (Gobalet 2001; Durden et al. 2016a, b). Considering humans learn during manual processing, the reliability of labelling improves, and later images may be processed more accurately, which can lead to over- or underestimation of both early and later processed segments (Durden et al. 2016a). As human performance in visual tasks decreases by up to 70% after about 30 minutes due to fatigue and boredom, the reduced precision of manual processing could also be reflected in altered results over time (Colquhoun 1959; Durden et al. 2016a; Cuvelier et al. 2024). Considering the high time required for manual processing of abyssal datasets, both situations could occur frequently and drastically change the results provided.

However, deep learning models must also fulfil certain requirements in order to function as desired. Even if the annotation becomes faster with a further increase in model precision, it can take some time to optimize the annotation. In addition, sufficient computing power is required to train the model and make predictions. Furthermore, one should be careful to rely on the training dataset, as the performance of the model is limited by the quality, quantity and faunistic variation of the training dataset (e.g. Gobi 2010; Durden et al. 2016b, 2021; Liu et al. 2024; Cuvelier et al. 2024). This may also neglect one advantage of processing deep-sea data using models, namely the reduced influence of human error. While a deep learning model sufficiently trained on abyssal fauna can recognize and track multiple animals regardless of their number, humans may be overwhelmed (Colquhoun 1959; Durden et al. 2016a). Insufficient training based on a poor training dataset would undermine this advantage. Furthermore, the composition of the training dataset determines the applicability of the resulting model to other dataset types. Although the model can effectively handle datasets with similar composition, other datasets, such as those taken from a different angle, could pose a challenge (e.g., Gobi 2010; Durden et al. 2016b, 2021; Liu et al. 2024; Cuvelier et al. 2024). We checked this by applying our final model to another AUV dataset from a similar abyssal plain from the same expedition (Baker et al. 2018), which yielded a nearly identical dataset apart from slight angular and lightning variations. While parts of the megafauna community were correctly detected, the majority were missed or misclassified, indicating the lower effectiveness of the model on different datasets. Although training on different datasets could improve performance, this advanced training would need an even larger dataset to reliably detect the organisms, due to the higher variability.

Megafauna observed with OFOS and AUV investigations

For an ecological classification, it is essential to compare the investigated area with similar habitats in order to recognize conspicuous features. Unfortunately, there are no quantifications of the megafauna community in the immediate vicinity. However, according to Watling et al. (2013), the North Atlantic megafauna community in the abyssal does not vary greatly between habitats as it shares key environmental characteristics, including depth (here approximately 4672 m), average temperature (here 2.4°C) and salinity (here approximately 34.9 PSU), respectively, and approximately 2.1 g organic C/m² reaching the seafloor annually (Kürzel et al. 2023). Considering the separating effect of oceanic ridges on ecosystems (e.g.; Bober et al. 2018; Riehl et al. 2018a, b), several areas east of the MAR should therefore be suitable for the comparison of megafauna communities (Watling et al. 2013; Kürzel et al. 2023). While more remote areas such as the Porcupine Abyssal Plain near Great Britan were sampled extensively from 1988 to 2012 (e.g.; Thurston et al. 1994, 1998; Bett et al. 2001; Billett et al. 2001, 2010; Morris et al. 2014), the most comparable studies we found were conducted in the Madeira Abyssal Plain. There, Thurston et al. (1994, 1998) surveyed the megafauna community between 1990 and 1993 using epibenthic sleds and otter trawls and observed lower total densities of only 7.5 × 10^− 3 and 1.02 × 10^− 3 ind./m², respectively, compared to the two sampling events presented here (Table 1, Figs. 3 & 8). One explanation could be that Sargassum enriched the habitat of our study area by serving as an additional food source (e.g. Bett et al. 2001; Johnson et al. 2007; Baker et al. 2018). As Sargassum naturally accumulates in abyssal areas (e.g.; Smith et al. 2008; Krause-Jensen and Duarte 2016; Baker et al. 2018), it has long been suspected to positively impact abyssal communities (Fleury and Drazen 2013; Molodtsova et al. 2017; Baker et al. 2018). In fact, considering that the R/V Sonne 237 expedition observed significantly greater densities of megafauna and Sargassum accumulations compared to the R/V Meteor 139 expedition, this might be an additional explanation for the connection between abyssal megafauna and the quantity of available Sargassum patches (Fig. 3, Table 1). Several studies have already documented that benthic invertebrates such as Ophiuroidea (Bett et al. 2001), isopods and amphipods (e.g. Wolff 1976; Lawson et al. 1993; Kobayashi et al. 2012), molluscs, polychaetes and holothurians (Bett et al. 2001) are not only actively attracted (Fleury and Drazen 2013), but may even ingest Sargassum in the deep sea. Apart from filter feeders that may ingest the algae directly from the water column, aggregated Sargassum, similar to vertebrate carcasses (Smith and Baco 2003; Roman et al. 2014; Higgs et al. 2014), could provide a food source for various taxa during its decay (Baker et al. 2018; Kokubu et al. 2019). In early stages, microbes (Baker et al. 2018) or holothurians (Bett et al. 2001) feeding on surface sediments could feed directly on the algae, while larger taxa consume the fauna associated with the algae (Drazen et al. 2008; Fleury and Drazen 2013; Baker et al. 2018). Given the distance of the Madeira Abyssal Plain from the Great Atlantic Sargassum Belt and the lack of mention of Sargassum in the comparative study (Thurston et al. 1994, 1998; Wang et al. 2019), this could explain our elevated values. Interestingly, when comparing the community composition of this study with that of Thurston et al. (1994, 1998), there were differences in the Cnidaria community, as the Pennatulacea, which together with the Porifera and Asteroidea made up the largest proportion of the community in our study, were completely absent in addition to the Hydrozoa. The observation of Pennatulacea (Umbellula) along with other filter feeders growing directly on or near sedimented Sargassum (Fig. 2) could explain these differences, as they could be feeding on the potentially elevated microbe concentration (Baker et al. 2018). However, as recent stable isotope studies have not clearly clarified or refuted both direct and indirect consumption (Baker et al. 2018), the effects of Sargassum on deep-sea ecosystems are not yet clear. Given the implications provided by the dataset presented here, both the topic of Sargassum and the megafauna community of this area requires further investigation and comparison with other abyssal areas to clarify this question.

Quantitative differences between AUV- and OFOS-sampling of the (nearly) identical abyssal area

As location, depth, habitat heterogeneity and other environmental conditions are important factors influencing deep-sea habitats (e.g. Watling et al. 2013; Durden et al. 2015; Kürzel et al. 2023), it is reasonable to associate the differences found here with these aspects. Observed differences could also be related to seasonal variations within those environmental conditions, such as POC flux to the seafloor or Sargassum accumulation (Kokubu et al. 2019). However, aside from the prior discussed Sargassum influence, as location, depth, habitat structure and heterogeneity (see Supplementary Fig. 1–3), as well as environmental factors such as temperature (2–3°C), salinity (34.5–36 PSU) or POC flux to the seafloor, are almost identical between the two sampling events, the faunal differences could rather be related to the different sampling methods (Watling et al. 2013; Devey et al. 2015; Arndt et al. 2017; Schoenle et al. 2021). In fact, the datasets provided by OFOS may have underestimated the occurrence of megafauna in this region. While calculating densities allows quantitative comparisons between areas of different sizes by standardizing the number of individuals per m², sampling a larger area by AUV still increases the likelihood of encountering rare organisms (e.g. Morris et al. 2014; Cuvelier et al. 2024). Overall, the precision of abyssal biodiversity assessments is known to increase with the size of the seabed surveyed, settling at 1000–1500 m² (Cuvelier et al. 2024). Although this benchmark was exceeded, the area sampled by the OFOS represented less than 7% of the area covered by the AUV (Fig. 1; Supplementary Table 3), so significant differences in the megafauna composition are plausible (Figs. 3, 7 & 8). The much lower resolution of the OFOS videos may also have resulted in some hard-to-find animals being missed, such as certain Porifera or small shrimps that resemble rubble or small Sargassum patches, especially if they were not moving. In this case, the OFOS, which provided a colour observation, was better able to identify these animals than the here monochrome AUV (Durden et al. 2016b). Despite the absence of light and the fact that many animals are blind (Watling et al. 2013; Sumner-Rooney 2018), some animals, including shrimps, are easily distinguishable from the sediment due to their pigmentation, which ranges from pale yellow to bright red (Johnsen 2005; Durden et al. 2016b). This distinction most likely contributed to the fact that the OFOS observed significantly more Crustacea than the monochrome AUV camera (Figs. 2 & 8).

A further consideration could be whether the ecosystems studied were more affected by one sampling gear than the other, and whether the animals may have fled as a result (Richard 1968; Wynn et al. 2014; Morris et al. 2014). Considering that artificial light is known to alter deep-sea communities (Morris et al. 2014) and may attract some cephalopod and fish taxa while reducing crustacean numbers (Clarke and Pascoe 1985), the different amount of light provided by the two sampling devices may have led to the observed differences (Figs. 3, 7 & 8). As the AUV does not require constant light, the animals may have been less affected by it in general, which could lead to significant differences in community composition and quantity (Clarke and Pascoe 1985; Morris et al. 2014). On the other hand, the fish may have been attracted to the illuminated prey or the light itself, which could explain why their proportion was significantly higher during the OFOS-dive under constant illumination (Fig. 8). The proximity of the sampling device to the seafloor may also be a confounding factor, potentially leading to missed observations due to the smaller sight radius or the avoidance of mobile taxa. As the OFOS is much closer to the seafloor and some fishes were visibly reacting and fleeing from the OFOS, the detection of significantly higher megafauna densities with the AUV fits this reasoning (Fig. 3). In this context, it is interesting to note that teleosts actually reached significantly higher densities during the AUV dive, although, as previously mentioned, their proportion was much higher during the OFOS event (Fig. 8). As some deep-sea taxa such as fish are known to be deterred or attracted by certain sound waves, noise and vibration from sampling equipment could be the third important influencing factor explaining the observed results (Richard 1968; Wynn et al. 2014). Although motion is caused by noisy equipment such as propellers, AUVs are generally quieter than research vessels (Wynn et al. 2014) and both remotely operated vehicles (ROVs) and human-occupied vehicles (HOVs) (Morris et al. 2014). On the other hand, the cable connection of the OFOS to the research vessel can cause vibrations based on the vessel's movements, while the lack of propulsion of the OFOS reduces the emitted sound waves. Combined with the wave-induced up and down motion and the proximity of the device to the bottom, this may disturb abyssal ecosystems (Richard 1968; Wynn et al. 2014). However, as we did not take measurements of the noise and vibration generated in this study, we cannot say whether one method generates more or less noise/vibration than the other.

The density of megafauna was significantly greater when the area was sampled with the AUV, and a significantly different community was found with taxa reaching significantly different densities and percentages compared to the OFOS method. These results contradict our hypothesis that both density and community composition should not differ in the overlapping areas sampled by AUV and OFOS dives. Given similar environmental characteristics, including the accumulation of Sargassum, the differences were due to the sampling gear affecting both the ecosystem and the dataset differently. As the manual processing of the data takes a lot of time, the self-trained deep learning model helped a lot in speeding up the process. They are easy to set up and can provide quickly valuable help in annotating the organisms. The validation of the deep learning model showed that there were no significant differences in megafauna densities, community composition and percentages between the manually and automatically processed datasets for the abundant groups. This confirmed the hypothesis about the overall effectiveness of the model with identical datasets. The analysis revealed few significant differences in the less abundant groups, indicating that the accuracy at higher taxonomic levels was lower. This was due to the lack of sufficient training images for these groups. While the combined megafauna density remained consistent, the community composition, densities and percentages of specific higher taxonomic groups varied between differently processed AUV subsections. Therefore, the hypothesis of identical communities between AUV subsections could not be confirmed.

In conclusion, our study illustrates the influence of different visual sampling devices and processing techniques on the observed ecosystem and the collected dataset. Apart from some loss of accuracy at higher taxonomic levels due to underrepresentation, our deep learning model has shown promise as a time- and resource-saving technique to improve our understanding of abyssal communities. Furthermore, the gained quantitative insights on abyssal epimegabenthos facing regular Sargassum falls might be integrated into further research investigating the potential influence this macroalgae might exert on deep-sea ecosystems.

Funding

This work was supported by the German Ministry for Science and Education, grants 03G0237B to H.A and 03G0227A to A. Brandt & C. Devey, as well as by the German Research Foundation, grants MerMet 16-97 and AR 288/23,24 to H.A. and MerMet 17-82 to N.A. & C. Devey.

Conflicts of interest/Competing interests

The authors declare no conflicts of interest relevant to the content of this article.

Availability of data and material

The datasets generated during and/or analysed during the current study are available from the corresponding author on request.

Code availability

Not applicable.

Author contributions

Conceptualization: HA and DS; methodology: DS, JW, HA, NA, MR; field work: NA, MR, HA, JW; video/image analysis: DS; statistics: DS, JW; deep learning detection module: JW, DS; funding acquisition: HA and NA; supervision: HA; writing of the first draft of the manuscript: DS and HA; all authors contributed to the revision of the draft. All authors read and approved the final manuscript.

Ethics approval

No approval of research ethics committees was required to accomplish the goals of this study as fauna exploration was restricted to visual investigation through video recordings.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Compliance with Ethical Standards

All applicable international, national and/or institutional guidelines for sampling, care and experimental use of organisms for the study have been followed and all necessary approvals have been obtained. Documentary evidence is available on request.

Acknowledgements

We are grateful to the technical crews and the captains of RV Sonne (SO 237) and RV Meteor (M139), and to the scientific crews of both expeditions for valuable help and support. We thank Rosita Bieg, Brigitte Gräfe, Bärbel Jendral and Anke Pyschny for their valuable technical support. Special thanks go to the GEOMAR OFOS- and AUV teams. This work was supported by the German Ministry for Science and Education, grants 03G0237B to H.A and 03G0227A to A. Brandt & C. Devey, as well as by the German Research Foundation (DFG), grants MerMet 16-97 and AR 288/23,24 to H.A. and MerMet 17-82 to N.A. & C. Devey.

Arndt H, Ahlers J, Amano C, Augustin N, Dangl G, Feuling Y, Herrero T, Hohlfeld M, Jeuck A, Marx M, Mähnert B, Meißner R, Meyer C, Podobnik M, Palgan D, Paulmann C, Prausse D, Romankiewicz T, Schade M, Scherwaß A, Schiwitza S, Schoenle A, Schuffenhauer I, Sintes E, Stefanschitz J, Werner J, Wildermuth B, Živaljić S (2017) Cruise Report METEOR 139. https://oceanrep.geomar.de/id/eprint/50010/1/tagungsband_forschungsschiffe_2020_web.pdf
Auguie B (2017) gridExtra: Miscellaneous functions for “Grid” graphics. R package version 2.3. https://cran.r project.org/web/packages/gridExtra.
Baker P, Minzlaff U, Schoenle A, Schwabe E, Hohlfeld M, Jeuck A, Brenke N, Prausse D, Rothenbeck M, Brix S, Frutos I, Jörger KM, Neusser TP, Koppelmann R, Devey C, Brandt A, Arndt H (2018) Potential contribution of surface-dwelling Sargassum algae to deep-sea ecosystems in the southern North Atlantic. Deep Sea Research Part II: Topical Studies in Oceanography 148:21–34. doi: 10.1016/j.dsr2.2017.10.002
Bett BJ, Malzone MG, Narayanaswamy BE, Wigham BD (2001) Temporal variability in phytodetritus and megabenthic activity at the seabed in the deep Northeast Atlantic. Prog Oceanogr 50:349–368. doi: 10.1016/S0079-6611(01)00066-0
Beuchel F, Primicerio R, Lønne OJ, Gulliksen B, Birkely S (2010) Counting and measuring epibenthic organisms from digital photographs: A semiautomated approach. Limnol Oceanogr Methods 8:229–240. doi: 10.4319/lom.2010.8.229
Billett DSM, Bett BJ, Rice AL, Thurston MH, Galéron J, Sibuet M, Wolff GA (2001) Long-term change in the megabenthos of the Porcupine Abyssal Plain (NE Atlantic). Prog Oceanogr 50:325–348. doi: 10.1016/S0079 6611(01)00060-X
Billett DSM, Bett BJ, Reid WDK, Boorman B, Priede IG (2010) Long-term change in the abyssal NE Atlantic: The ‘Amperima Event’ revisited. Deep Sea Research Part II: Topical Studies in Oceanography 57:1406–1417. doi: 10.1016/j.dsr2.2009.02.001
Bober S, Brix S, Riehl T, Schwentner M, Brandt A (2018) Does the Mid-Atlantic Ridge affect the distribution of abyssal benthic crustaceans across the Atlantic Ocean? Deep Sea Res 2 Top Stud Oceanogr 148:91–104. doi: 10.1016/j.dsr2.2018.02.007
Buhl-Mortensen L, Vanreusel A, Gooday AJ, Levin LA, Priede IG, Buhl-Mortensen P, Gheerardyn H, King NJ, Raes M (2010) Bio- logical structures as a source of habitat heterogeneity and biodi- versity on the deep ocean margins. Mar Ecol 31:21–50. Doi: 10.1111/j.1439-0485.2010.00359.x
Cáceres M de, Legendre P (2009) Associations between species and groups of sites: indices and statistical inference. Ecology 90:3566–3574. doi: 10.1890/08-1823.1
Chailloux C, Allais AG, Simeoni P, Olu K (2008) Automatic classification of deep benthic habitats: Detection of microbial mats and siboglinid polychaete fields from optical images on the Håkon Mosby Mud Volcano. In: OCEANS 2008. IEEE, pp 1–7. doi: 10.1109/OCEANS.2008.5151812
Clarke MR, Pascoe PL (1985) The influence of an electric light on the capture of deep-sea animals by a midwater trawl. Journal of the Marine Biological Association of the United Kingdom 65:373–393. doi: 10.1017/S0025315400050499
Colquhoun WP (1959) The effect of a short rest-pause on inspection efficiency. Ergonomics 2:367–372. doi: 10.1080/00140135908930451
Cuvelier D, Zurowietz M, Nattkemper TW (2024) Deep learning–assisted biodiversity assessment in deep-sea benthic megafauna communities: a case study in the context of polymetallic nodule mining. Front Mar Sci. 11, doi: 10.3389/fmars.2024.1366078
de Cáceres M, Legendre P (2009) Associations between species and groups of sites: indices and statistical inference. Ecology 90:3566–3574. doi: 10.1890/08-1823.1
Devey C, Arndt H, Augustin N, Bober S, Borges V, Brandt A, Brenke N, Brix-Elsig S, Elsner N, Frutos I, Guggolz T, Heitland N, Jeuck A, Klischies M, Köhler J, Lejzerovicz F, Lins L, Linse K, Malytina M, Metz D, Minzlaff U, Prauße D, Palgan D, Riehl T, Rothenbeck M, Schmidt C, Schmidt C, Schoenle A, Schultze G, Schwabe E, Springer T, Steinführer A, Triebe L, Voltski I, Walter T, Wenzlaff E, Yeo I, Zinnkann A-C, Zoeller K (2015) Cruise Report SONNE 237.
Drazen JC, Popp BN, Choy CA, Clemente T, Forest L De, Smith KL (2008) Bypassing the abyssal benthic food web: Macrourid diet in the eastern North Pacific inferred from stomach content and stable isotopes analyses. Limnol Oceanogr 53:2644–2654. doi: 10.4319/lo.2008.53.6.2644
Durden J, Bett B, Schoening T, Morris K, Nattkemper T, Ruhl H (2016a) Comparison of image annotation data generated by multiple investigators for benthic ecology. Mar Ecol Prog Ser 552:61–70. doi: 10.3354/meps11775
Durden J, Bett B, Horton T, Serpell-Stevens A, Morris K, Billett D, Ruhl H (2016b) Improving the estimation of deep- sea megabenthos biomass: dimension to wet weight conversions for abyssal invertebrates. Mar Ecol Prog Ser 552:71–79. doi: 10.3354/meps11769
Durden JM, Bett BJ, Jones DOB, Huvenne VAI, Ruhl HA (2015) Abyssal hills – hidden source of increased habitat heterogeneity, benthic megafaunal biomass and diversity in the deep sea. Prog Oceanogr 137:209–218. doi: 10.1016/j.pocean.2015.06.006
Durden JM, Hosking B, Bett BJ, Cline D, Ruhl HA (2021) Automated classification of fauna in seabed photographs: The impact of training and validation dataset size, with considerations for the class imbalance. Prog Oceanogr 196:102612. doi: 10.1016/j.pocean.2021.102612
Eddelbuettel D (2013) Seamless R and C++ integration with Rcpp. Seamless R and C++ integration with Rcpp 40:1- 220. doi: 10.1007/978-1-4614-6868-4
Firke S (2023) janitor: Simple tools for examining and cleaning dirty data. R package version 2.2.0, https://CRAN.R-project.org/package=janitor.
Fleury AG, Drazen JC (2013) Abyssal scavenging communities attracted to Sargassum and fish in the Sargasso Sea. Deep Sea Research Part I: Oceanographic Research Papers 72:141–147. doi: 10.1016/j.dsr.2012.11.004
Fox J, Weisberg S (2019) An {R} Companion to applied regression, Third Edition. Thousand Oaks CA: Sage. URL: https://socialsciences.mcmaster.ca/jfox/Books/Companion/.
Gastwirth JL, Gel YR, Hui WLW, Lyubchich V, Miao W, Noguchi K (2020) lawstat: Tools for biostatistics, public policy, and law. R package version 3.4. https://CRAN.R-project.org/package=lawstat.
Gobalet KW (2001) A critique of faunal analysis; Inconsistency among experts in blind tests. J Archaeol Sci 28:377–386. doi: 10.1006/jasc.2000.0564
Gobi AF (2010) Towards generalized benthic species recognition and quantification using computer vision. In: 2010 Fourth Pacific-Rim Symposium on Image and Video Technology. IEEE, pp 94–100. doi: 10.1109/PSIVT.2010.23
Higgs ND, Gates AR, Jones DOB (2014) Fish food in the deep sea: Revisiting the role of large food-falls. PLoS One 9:e96016. doi: 10.1371/journal.pone.0096016
Jocher G, Chaurasia A, Stoken A, Borovec J, NanoCode012, Kwon Y, Michael K, TaoXie, Fang, J, Imyhxy, Lorna, Zeng Y, Wong C, Abhiram V, Montes D, Wang Z, Fati C, Nadar J, Laughing, UnglvKitDe, Sonck V, Tkianai, YxNONG, Skalski P, Hogan A, Nair D, Strobel M, Jain M (2022). ultralytics/yolov5: v7.0 - YOLOv5 SOTA Realtime Instance Segmentation. Zenodo. doi.org/10.5281/zenodo.3908559.
Johnsen S (2005) The red and the black: Bioluminescence and the color of animals in the deep sea. Integr Comp Biol 45:234–246. doi: 10.1093/icb/45.2.234
Johnson NA, Campbell JW, Moore TS, Rex MA, Etter RJ, McClain CR, Dowell MD (2007) The relationship between the standing stock of deep-sea macrobenthos and surface production in the western North Atlantic. Deep Sea Res 1 Oceanogr Res Pap 54:1350–1360. doi: 10.1016/j.dsr.2007.04.011
Kennedy BRC, Cantwell K, Malik M, Kelley C, Potter J, Elliott K, Lobecker E, Gray LM, Sowers D, White MP, France SC, Auscavitch S, Mah C, Moriwake V, Bingo SRD, Putts M, Rotjan RD (2020) Corrigendum: The unknown and the unexplored: Insights into the pacific deep-sea following NOAA CAPSTONE Expeditions. Front Mar Sci 6:2019–2020. doi: 10.3389/fmars.2019.00827
Kobayashi H, Hatada Y, Tsubouchi T, Nagahama T, Takami H (2012) The hadal amphipod Hirondellea gigas possessing a unique cellulase for digesting wooden debris buried in the deepest seafloor. PLoS One 7:e42727. doi: 10.1371/journal.pone.0042727
Kokubu Y, Rothäusler E, Filippi J-B, Durieux EDH, Komatsu T (2019) Revealing the deposition of macrophytes transported offshore: Evidence of their long-distance dispersal and seasonal aggregation to the deep sea. Sci Rep 9:4331. doi: 10.1038/s41598-019-39982-w
Krause-Jensen D, Duarte CM (2016) Substantial role of macroalgae in marine carbon sequestration. Nat Geosci 9:737- 742. doi: 10.1038/ngeo2790
Kürzel K, Brix S, Brandt A, Brenke N, Enderlein P, Griffiths HJ, Kaiser S, Svavarsson J, Lörz A-N, Frutos I, Taylor J, Linse K (2023) Pan-Atlantic comparison of deep-sea macro- and megabenthos. Diversity (Basel) 15:814. doi: 10.3390/d15070814
Lawson GS, Tyler PA, Young CM (1993) Attraction of deep-sea amphipods to macrophyte food falls. J Exp Mar Biol Ecol 169:33–39. doi: 10.1016/0022-0981(93)90041-L
Lebart K, Smith C, Trucco E, Lane DM (2003) Automatic indexing of underwater survey video: algorithm and benchmarking method. IEEE Journal of Oceanic Engineering 28:673–686. doi: 10.1109/JOE.2003.819314
Liu A, Liu Y, Xu K, Zhao F, Zhou Y, Li X (2024) DeepSeaNet: A bio-detection network enabling species identification in the deep sea imagery. IEEE Transactions on Geoscience and Remote Sensing 62:1–13. doi: 10.1109/TGRS.2024.3359350
Milligan RJ, Morris KJ, Bett BJ, Durden JM, Jones DOB, Robert K, Ruhl HA, Bailey DM (2016) High resolution study of the spatial distributions of abyssal fishes by autonomous underwater vehicle. Sci Rep 6: 26095. doi: 10.1038/srep26095
Molodtsova TN, Galkin S V., Kobyliansky SG, Simakova U V., Vedenin AA, Dobretsova IG, Gebruk A V. (2017) First data on benthic and fish communities from the Mid-Atlantic Ridge, 16°40′− 17°14′N. Deep Sea Research Part II: Topical Studies in Oceanography 137:69–77. doi: 10.1016/j.dsr2.2016.10.006
Morris KJ, Bett BJ, Durden JM, Huvenne VAI, Milligan R, Jones DOB, McPhail S, Robert K, Bailey DM, Ruhl HA (2014) A new method for ecological surveying of the abyss using autonomous underwater vehicle photography. Limnol Oceanogr Methods 12:795–809. doi: 10.4319/lom.2014.12.795
Neuwirth E (2014) RColorBrewer: ColorBrewer palettes. R package version 1.1-2. https://CRAN.R- project.org/package=RColorBrewer.
Oksanen J, Blanchet FG, Friendly M, Kindt R, Legendre P, Mcglinn D, Minchin PR, Hara RBO, Simpson GL, Solymos P, Stevens MHH, Szoecs E, Wagner H (2019) vegan: Community ecology package. R package version 2.5 6. https://CRAN.R- project.org/package=vegan.
Pohlert T (2022) PMCMRplus: Calculate pairwise multiple comparisons of mean rank sums extended. R package version 1.9.0. https://CRAN.R-project.org/package=PMCMRplus. 2022.
R Core Team (2022) R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URLhttps://www.R-project.org/.
Richard JD (1968) Fish attraction with pulsed low-frequency sound. Journal of the Fisheries Research Board of Canada 25:1441–1452. doi: 10.1139/f68-125
Riehl T, Lins L, Brandt A (2018a) The effects of depth, distance, and the Mid-Atlantic Ridge on genetic differentiation of abyssal and hadal isopods (Macrostylidae). Deep Sea Research Part II: Topical Studies in Oceanography 148:74–90. doi: 10.1016/j.dsr2.2017.10.005
Riehl T, Kaiser S, Brandt A (2018b) Vema-TRANSIT – An interdisciplinary study on the bathymetry of the Vema Fracture Zone and Puerto Rico Trench as well as abyssal Atlantic biodiversity. Deep Sea Research Part II: Topical Studies in Oceanography 148:1–6. doi: 10.1016/j.dsr2.2018.01.007
Rinker TW, Kurkiewicz D (2017) pacman: Package management for R. version 0.5. 0. Buffalo, New York http://github. com/trinker/pacman.
Roman J, Estes JA, Morissette L, Smith C, Costa D, McCarthy J, Nation JB, Nicol S, Pershing A, Smetacek V (2014) Whales as marine ecosystem engineers. Front Ecol Environ 12:377–385. doi: 10.1890/130220
Rubis B (2020) hrbrthemes: Additional themes, theme components and utilities for “ggplot2”. R package version 0.8.0. https://CRAN.R-project.org/package=hrbrthemes. URL http://sites.google.com/site/miqueldecaceres/.
Scepanski D, Augustin N, Dünn M, Scherwaß A, Xavier JR, Werner J, Waldvogel AM, Arndt H (2024) Vertical distribution of epibenthic megafauna of a large seamount west of Cape Verde islands (tropical North Atlantic). Marine Biodiversity. doi: 10.1007/s12526-023-01400-w
Schoening T, Bergmann M, Ontrup J, Taylor J, Dannheim J, Gutt J, Purser A, Nattkemper TW (2012) Semi-automated image analysis for the assessment of megafaunal densities at the arctic deep-sea observatory HAUSGARTEN. PLoS One 7:e38179. doi: 10.1371/journal.pone.0038179
Schoenle A, Hohlfeld M, Hermanns K, Mahé F, de Vargas C, Nitsche F, Arndt H (2021) High and specific diversity of protists in the deep-sea basins dominated by diplonemids, kinetoplastids, ciliates and foraminiferans. Commun Biol 4:501. doi: 10.1038/s42003-021-02012-5
Smith CR, Baco AR (2003) Ecology of whale falls at the deep-sea floor. In: Gibson RN, Atkinson RJA (eds) Oceanography and Marine Biology: An Annual Review. CRC Press, Boca Raton, pp 311–354
Smith CR, Demopoulos AWJ (2003) The deep Pacific Ocean floor. In: Tyler PA (ed) Ecosystems of the World, Volume 28: Ecosystems of the Deep Oceans. Elsevier, Amsterdam, pp 179–218
Smith CR, De Leo FC, Bernardino AF, Sweetman AK, Arbizu PM (2008) Abyssal food limitation, ecosystem structure and climate change. Trends Ecol Evol 23:518–528. doi: 10.1016/j.tree.2008.05.002
Stanchev L, Egbert H, Ruttenberg B (2020) Automating deep-sea video annotation using machine learning. In: 2020 IEEE 14th International Conference on Semantic Computing (ICSC). IEEE, pp 17–24. doi: 10.1109/ICSC.2020.00010.
Sumner-Rooney L (2018) The kingdom of the blind: Disentangling fundamental drivers in the evolution of eye loss. Integr Comp Biol 58:372–385. doi: 10.1093/icb/icy047
Suzuki R, Terada Y, Shimodaira H (2019) pvclust: Hierarchical clustering with P- Values via multiscale bootstrap resampling. R package version 2.2-0. https://CRAN.R- project.org/package=pvclust.
Thurston MH, Bett BJ, Rice AL, Jackson PAB (1994) Variations in the invertebrate abyssal megafauna in the North Atlantic Ocean. Deep Sea Research Part I: Oceanographic Research Papers 41:1321–1348. doi: 10.1016/0967-0637(94)90100-7
Thurston MH, Rice AL, Bett BJ (1998) Latitudinal variation in invertebrate megafaunal abundance and biomass in the North Atlantic Ocean Abyss. Deep Sea Research Part II: Topical Studies in Oceanography 45:203–224. doi: 10.1016/S0967-0645(97)00077-5
Wang M, Hu C, Barnes BB, Mitchum G, Lapointe B, Montoya JP (2019) The great Atlantic Sargassum belt. Science (1979) 365:83–87. doi: 10.1126/science.aaw7912
Watling L, Guinotte J, Clark MR, Smith CR (2013) A proposed biogeography of the deep ocean floor. Prog Oceanogr 111:91–112. doi: 10.1016/j.pocean.2012.11.003
Wickham H (2007) Reshaping data with the reshape package. Journal of Statistical Software, 21(12), 1-20. URL http://www.jstatsoft.org/v21/i12/./.
Wickham H (2016) ggplot2: Elegant graphics for data analysis. Springer International Publishing
Wickham H (2020) tidyr: Tidy Messy Data. R package version 1.1.2. https://CRAN.R- project.org/package=tidyr.
Wickham H (2021) forcats: Tools for working with categorical variables (Factors). R package version 0.5.1.https://CRAN.R-project.org/package=forcats.
Wickham H, Seidel D (2020) scales: Scale functions for visualization. R package version 1.1.1. https://CRAN.R- project.org/package=scales.
Wickham H, Averick M, Bryan J, Chang W, McGowan L, François R, Grolemund G, Hayes A, Henry L, Hester J, Kuhn M, Pedersen T, Miller E, Bache S, Müller K, Ooms J, Robinson D, Seidel D, Spinu V, Takahashi K, Vaughan D, Wilke C, Woo K, Yutani H (2019) Welcome to the Tidyverse. J Open Source Softw 4:1686. doi: 10.21105/joss.01686
Wickham H, Henry L, Pedersen TL, Luciani TJ, Decorde M, Lise V (2020) svglite: An “SVG” graphics device. R package version 1.2.3.2. https://CRAN.R- project.org/package=svglite.
Wickham H, François R, Henry L, Müller K (2021) dplyr: A grammar of data manipulation. R package version 1.0.7. https://CRAN.R-project.org/package=dplyr.
Wilke CO (2020) cowplot: Streamlined plot theme and plot annotations for “ggplot2”. R package version 1.1.1. https://CRAN.R-project.org/package=cowplot. 2020.
Wolff T (1976) Utilization of seagrass in the deep sea. Aquat Bot 2:161–174. doi: 10.1016/0304-3770(76)90017-6
Wynn RB, Huvenne VAI, Le Bas TP, Murton BJ, Connelly DP, Bett BJ, Ruhl HA, Morris KJ, Peakall J, Parsons DR, Sumner EJ, Darby SE, Dorrell RM, Hunt JE (2014) Autonomous Underwater Vehicles (AUVs): Their past, present and future contributions to the advancement of marine geoscience. Mar Geol 352:451–468. doi: 10.1016/j.margeo.2014.03.012
Yang L, Zhao S, Wang X, Shen P, Zhang T (2022) Deep-sea underwater cooperative operation of manned/unmanned submersible and surface vehicles for different application scenarios. J Mar Sci Eng 10:909. doi: 10.3390/jmse10070909
Zeileis A, Fisher JC, Hornik K, Ihaka R, McWhite CD, Murrell P, Stauffer R, Wilke CO (2020) colorspace : A toolbox for manipulating and assessing colors and palettes. J Stat Softw 96:1–49. doi: 10.18637/jss.v096.i01

Table 1 is available in the Supplementary Files section.

Download PDF

Reviewers agreed at journal
03 Sep, 2024
Reviewers invited by journal
21 Aug, 2024
Editor assigned by journal
21 Aug, 2024
First submitted to journal
19 Aug, 2024
Editorial decision: Revise and Resubmit
11 Aug, 2024

You are reading this latest preprint version

Detecting abyssal megabenthos in the Great Atlantic Sargassum Belt – Comparing quantitative assessments via manually processed OFOS-videos with AUV-images utilising deep learning models

Status:

Version 1

Abstract

Figures

Introduction

Material and methods

Study areas and habitat description

Image and video sampling

Surveyed parameters and area assessment

Morphotaxa annotations and qualitative/quantitative estimates

Data analysis

Results

Differences between manual processing and deep learning models

Validation and precision of an older and the final model

The manually analysed AUV section compared with the remaining ones processed by the final model

Differences between OFOS and AUV sampling

Community composition

Quantitative comparison

Discussion

Model evaluation

Benefits and disadvantages of utilising (this) deep learning model

Megafauna observed with OFOS and AUV investigations

Quantitative differences between AUV- and OFOS-sampling of the (nearly) identical abyssal area

Conclusion

Declarations

References

Table

Supplementary Files

Status:

Version 1