Humans versus Machines: Artificial intelligence performs similarly to manual analysis in benthic reef monitoring

Humans versus Machines: Artificial intelligence performs similarly to manual analysis in benthic reef monitoring | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Humans versus Machines: Artificial intelligence performs similarly to manual analysis in benthic reef monitoring Marina Alves Méga de Andrade, Viviane Barroso, Carlos Eduardo Leite Ferreira, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7586184/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Tropical reefs are among the most biodiverse and threatened marine ecosystems and face both local and global stressors. Understanding community dynamics and ecosystem functioning under these pressures requires long-term monitoring. Image-based surveys are increasingly used in reef ecology because they enable rapid data acquisition and storage. However, manual identification of benthic organisms is time-consuming and potentially subjective, creating a bottleneck between image collection and ecological analysis. Automated annotation tools represent promising solutions, but few studies have directly compared their performance with that of manual approaches, especially under varying image quality conditions. Here, we assess the usefulness of the CoralNet automated tool for classifying benthic cover in tropical reef images and evaluate how image quality influences both manual and automated annotations. Using expert-validated images, we trained and tested both approaches, measuring performance with Cohen’s Kappa coefficient and modeling the effects of image imperfections. We found that manual and automated annotators achieved similar performance across most taxonomic and morphofunctional groups, and both were similarly affected by image quality. Furthermore, the most frequent imperfections were not the most influential. Our findings demonstrate that automated annotation is a reliable and efficient alternative to manual methods, with strong potential to enhance large-scale monitoring, biodiversity assessments, and conservation strategies in reef ecosystems. benthic community CoralNet image annotation long-term monitoring reef systems Figures Figure 1 Figure 2 Figure 3 Introduction Tropical reefs are among the most important ecosystems on Earth and are comparable to tropical rainforests in terms of high primary productivity (Hatcher 1997 ), complexity, and diversity (Reaka-Kudla et al. 1997). They also hold significant economic importance because of their fundamental ecosystem services, such as providing protein sources, supporting tourism, and protecting coastlines (Hughes et al. 2003 ; Moberg & Folke 1999 ). However, these ecosystems face global threats, such as ocean acidification and warming (Hoegh-Guldberg 2007; Hughes et al. 2018 ), and local impacts, such as overfishing and pollution (Zaneveld et al. 2016 ). Long-term monitoring is therefore crucial for understanding ecosystem functioning under these combined pressures (Quimbayo et al. 2018 ). Photoquadrat analysis is one of the most widely used methods for quantifying benthic community structure, with software such as CPC e (Kohler & Gill 2006 ) and PhotoQuad (Trygonis & Sini 2012 ) commonly applied. However, manual annotation of photoquadrats is a time-consuming and tedious process, creating a bottleneck between data collection and analysis (Stokes & Deane 2009 ). In addition, human annotation is subjective and inconsistent, leading to potential misclassifications, overestimations, or omissions on the basis of the experience of the individual annotator (MacLeod et al. 2010 ). Automated approaches based on artificial intelligence (AI) offer promising alternatives, reducing annotation subjectivity and shortening analysis time, allowing more images to be processed in less time (Shihavuddin et al. 2013 ). CoralNet is software similar to CPCe and PhotoQuad but includes an automated annotation tool using computer vision methods (Beijbom et al. 2015 ; Chen et al. 2021 ). After at least 20 images have been annotated by a researcher, CoralNet begins to automatically annotate the remaining images, improving its performance with more manually annotated images (Chen et al. 2021 ). The CoralNet AI tool operates in two stages: a convolutional neural network (CNN) and a simple classifier. CNNs are deep learning-based approaches designed to process large datasets that are structured in multiple arrays (LeCun et al. 2015 ). These deep neural networks have gained attention and are now being used in various fields, such as image recognition, speech recognition, and natural language processing (Ahmad et al. 2019 ). The typical architecture of CNNs consists of multiple layers that learn features hierarchically from the training dataset. Simpler features, such as edges and basic shapes, are learned in the initial layers, whereas deeper layers capture increasingly complex features as the data propagate through the network. During training, the network optimizes its weights via backpropagation, focusing on features that are more relevant to improve the classification task (Ahmad et al. 2019 ; González-Rivero et al. 2020). The second stage of the CoralNet AI tool is a simple classifier that receives a feature vector from the CNN and assigns a score to each label in the label set, selecting the highest score for annotation (Chen et al. 2021 ). Despite these advantages, challenges remain. A minimum number of images is required for accurate identification, making it more effective for common species than for rare ones. Moreover, incorrect annotations during training can lead to errors in automated analysis if not corrected by a human. Moreover, image quality also impacts analysis, and both manual and automated annotations are likely to be influenced by factors such as resolution, turbidity, and lighting (Beijbom et al. 2015 ). However, few studies have directly evaluated how image imperfections influence annotation outcomes or whether manual and automated approaches are similarly affected (Beijbom et al. 2015 ). The effectiveness of AI could help reduce the gap between collected and annotated images in long-term monitoring programs and enhance the reproducibility of analyses (Williams et al. 2019 ). Here, we evaluate the performance of an automated annotation tool for classifying benthic cover in tropical reefs, comparing its results with those of manual annotation. We further assess how image quality influences both approaches, providing insights into which imperfections are most detrimental to annotation accuracy. This study contributes to improving standards for image-based surveys, supporting large-scale monitoring and conservation of reef ecosystems. Materials and methods Study area and sampling We collected images in Rocas Atoll (03°50'S, 33°49'W), the first no-take, marine protected area in Brazil, established in 1978. Monitoring was conducted annually during May and June from 2014–2017 at four different sites (Âncoras, Cemitério, Falsa Barreta, and Podes Crer). At each site, three 20-meter transects were laid out following a systematic sampling method with 2-meter intervals. This approach resulted in a total of ten or eleven photoquadrats per site (Picolotto et al. 2021 , preprint ). In another study, all images were analyzed by an expert (Vitor André Passos Picolotto - Picolotto et al. 2021 , preprint ) via the CPCe program, with fifty random points per photoquadrat. Image quality and set separation We qualitatively classified the images (n = 499) based on the presence or absence of the imperfections listed in Table 1 . Inappropriate images were excluded from the dataset, resulting in 493 viable images, and we divided the transects equally (n = 24 each) into two different sets: i. the training set (248 images), which was used to train both human (manual) and AI (automated) annotators; and ii. evaluation set (245 images), in which images were reanalyzed by both manual and automated annotators using the same points as the expert but without knowledge of the previous annotations (Beijbom et al. 2015 ). The evaluation set was then used to compare and verify the agreement between the annotators (manual and automated) and the expert. Organisms were identified to the lowest taxonomic level possible (generally genera) and then grouped into morphofunctional groups (following Zamoner et al. 2021 ). The labelset used was developed by the Long Term Programs of Brazilian Oceanic Islands (PELD-ILOC) (Table S1 - Supplementary Material). Additionally, the effects of image quality on manual and automated annotations were analyzed via this set, allowing for the assessment of how each imperfection affects both analysis methods. Table 1 Image quality based on the presence or absence of imperfections. Imperfection Description Shadow Localized decrease in brightness (< 50% than surrounding) Suspension Sediment particles in front of the target or on the camera lens Sun Glint Reflection of sunlight on water (presence of “rainbow” beam) Overexposure Area with brightness excess and saturated pixels Distortion Deformation of image edges caused by lens curvature Unfocused Presence of unfocused area between two or more surfaces Parallax error Image not perpendicular to the target center Object inside area Presence of object inside the analysis area (e.g., tape) Loss of analysis area Square cut off the image, resulting in loss of analysis area Data analysis Performance of the annotators Traditionally, the agreement between observers has been assessed by the percentage of matches between observations (overall accuracy). However, this method does not consider agreements occurring by chance. To address this, Cohen's Kappa was introduced as a metric to estimate agreement while excluding random chance (Cohen, 1960 ). The Kappa statistic is calculated as follows: $$\:K\:=\frac{P\left(a\right)\:-\:P\left(e\right)}{1\:-\:P\left(e\right)}$$ where P(a) represents the observed proportion of agreement and P(e) denotes the proportion of agreement expected by chance (Carletta 2008). Cohen's Kappa ranges from − 1 to 1 and can be interpreted according to Landis & Koch ( 1977 ) (Table 2 ). This value is generally lower than the total accuracy because the total accuracy overestimates the agreement by ignoring chance factors. In this section, we use Cohen’s Kappa to compare each annotator with the expert (performance of the annotators) and to compare both annotators (manual and automated) with each other. We call Ka the performance of the automated annotator (AI), Km the performance of the manual annotator (human), and Kb the agreement between the annotators. The analysis was performed in two ways: i) evaluating the overall performance of both manual and automated annotators across all images (total Cohen’s Kappa), using annotations previously analyzed by the expert as a baseline for comparison; and ii) assessing the ability of the annotators to perform binary classification tasks (Cohen’s Kappa for each group), such as distinguishing Siderastrea from non- Siderastrea or corals from noncorals. Analyses were performed for both labels and morphofunctional groups (Beijbom et al. 2015 ). Cohen’s Kappa confidence intervals were set at 95%, and values were computed via the "Psych" package in R (Revelle, 2017 ). Table 2 Intensity of interobserver agreement according to Cohen’s Kappa values from Landis & Koch ( 1977 ) Cohen’s Kappa value Agreement intensity < 0.00 Poor 0.00–0.20 Slight 0.21–0.40 Fair 0.41–0.60 Moderate 0.61–0.80 Substantial 0.81–1.00 Almost perfect Effect of image quality on the performance of the annotators In this section, we used the Total Cohen’s Kappa for each annotator (manual and automated) on each image. Specifically, we assessed the performance of each annotator on individual images (e.g., the total Cohen’s Kappa of the automated annotator on image 1) to make the comparison. To reduce potential biases related to class abundance, we focused only on morphofunctional groups, as there are fewer rare morphofunctional groups than rare labels. We then ran a mixed model to identify which image quality factors influenced the annotations. In this analysis, the dependent variable was agreement (Cohen’s Kappa for each image), and the independent variables were image quality imperfections and the type of annotator (manual or automated), with images used as a random effect. The significance level (P value) was set at 5%. “Object within analysis area” and “loss of analysis area” were not included in the model because they do not lead to misclassification, although they reduce the sampled area and must be minimized during image acquisition. Results Overall performance At the label level, the overall performance of both manual and automated annotators was classified as moderate (Table 4), although the performance of manual annotators was higher than that of automated annotators. In addition, the overall agreement between annotators was greater than their individual performance (i.e., each annotator’s agreement with the expert) was (Fig. 1; Table 4). Total Cohen’s Kappa (overall performance) at the morphofunctional group level was greater than that at the label level for both annotators, although it remained moderate (Table 4). Nevertheless, the pattern was the same, with higher Kappa values for agreement between annotators, followed by the overall performance of the manual annotator and, finally, the overall performance of the automated annotator (Fig. 2; Table 4). Morphofunctional groups The morphofunctional groups best identified by the manual annotator were corals, leathery macroalgae, substrate, and filamentous macroalgae. For all of these groups, the manual annotator achieved the highest Kappa values and substantial performance (Table 4). For zoanthids, foliaceous macroalgae, and turf algae, the Kappa values decreased, and the performance was moderate. Finally, the morphofunctional group with the lowest Kappa values was crustose macroalgae, which were classified as having slight performance by the manual annotator (Table 4). For the automated annotator, the Kappa values were also higher for corals, leathery macroalgae, and substrate, all with substantial performance. However, the performance for filamentous macroalgae was moderate, with Kappa at the upper limit for this category (Fig. 1; Table 4). Turf algae and zoanthids also presented moderate performance but with relatively low Kappa values. Finally, foliaceous macroalgae and crustose macroalgae were classified as fair performers and represented the groups with the lowest Kappa values (Fig. 2; Table 4). When the annotation methods were compared, the performance did not differ significantly for corals, leathery macroalgae, or filamentous macroalgae. For the substrate, turf algae, zoanthids, and foliaceous macroalgae, the manual annotator outperformed the automated annotator, whereas the opposite was observed for crustose macroalgae. We found higher kappa values and agreement between annotators than for each individual performance metric. Agreement was greater for corals, leathery macroalgae, substrate, and filamentous macroalgae, whereas disagreement was greater for crustose macroalgae, zoanthids, and foliaceous macroalgae. For turf algae, the agreement between annotators was substantial (Fig. 2; Table 4). Labels For the labels, the manual annotator obtained the highest Cohen’s Kappa values for the coral Porites astreoides and for Caulerpa verticillata , both with almost perfect performance (Fig. 2; Table 3). Siderastrea spp., Sargassum spp., and the substrate also had high Kappa values and substantial performance. For Bryopsis spp., Zoanthus sociatus , filamentous turf, and Dictyota spp., the manual annotator showed moderate performance and intermediate Kappa values. Finally, the label with the lowest kappa values was crustose calcareous macroalgae, which was classified as having slight performance (Fig. 2; Table 3). For the automated annotator, the labels with the highest Kappa values were almost the same: Caulerpa verticillata , Siderastrea spp., Sargassum spp., and substrate. However, none of these labels achieved almost perfect performance, and only substantial performance was achieved (Fig. 2; Table 3). In contrast, the automated annotator identified Porites astreoides with moderate performance, which was lower than that of the manual annotator. Other labels with moderate performance were Bryopsis spp. and Zoanthus sociatus . Finally, the lowest performance was fair, observed for filamentous turf, Dictyota spp., and crustose calcareous macroalgae (Fig. 2; Table 3). Comparing the two annotation methods, Cohen’s Kappa values for Caulerpa verticillata , Siderastrea spp., and Sargassum spp. were similar for both manual and automated annotators. The same applied to Bryopsis spp. and Dictyota spp., where despite higher performance categories for the manual annotator, the confidence intervals overlapped (Fig. 2; Table 3). In this case, even for the labels that were grouped into different categories depending on the annotator (e.g., C. verticillata with almost perfect performance for the manual annotator and substantial performance for the automated annotator; Dictyota spp. with moderate performance for the manual annotator and fair performance for the automated annotator), no significant differences were found (Table 3). For substrate, Porites astreoides , Zoanthus sociatus , and filamentous turf, the manual annotator had higher Kappa values than did the automated annotator, whereas the opposite occurred for crustose calcareous macroalgae (Fig. 2; Table 3). Among the labels, those with the highest agreement between annotators were Siderastrea spp., Sargassum spp., Caulerpa verticillata , substrate, and Bryopsis spp., whereas the highest disagreement occurred for crustose calcareous macroalgae, Dictyota spp., Zoanthus sociatus , and Porites astreoides . For filamentous turf, the interannotator agreement was intermediate (Fig. 2; Table 3). Table 3: Classification of Cohen's Kappa values for each class (label or morphofunctional group) via the annotation method, according to Landis & Koch (1977). The performance can be ordered as follows: poor (up to 0.00), slight (0.00 - 0.20), fair (0.21 - 0.40), moderate (0.41 - 0.60), substantial (0.61 - 0.80), and almost perfect (0.81 - 1.00). We denoted Ka as the automated annotator performance (AI), Km as the manual annotator performance (human), and Kb as the agreement between the annotators. The values for the substrate are the same for both the labeled and morphofunctional groups, as it is the only label composing the morphofunctional group. Labels Km Intensity (Km) Ka Intensity (Ka) Kb Intensity (Kb) Substrate 0.770 (0.756 - 0.783) Substantial 0.740 (0.726 - 0.754) Substantial 0.816 (0.804 - 0.828) Almost perfect Bryopsis spp. 0.591 (0.559 - 0.622) Moderate 0.554 (0.523 - 0.586) Moderate 0.795 (0.773 - 0.817) Substantial Caulerpa verticillata 0.826 (0.804 - 0.847) Almost Perfect 0.799 (0.777 - 0.820) Substantial 0.819 (0.798 - 0.839) Almost perfect Crustose calcareous algae 0.184 (0.141 - 0.227) Slight 0.296 (0.251 - 0.340) Fair 0.234 (0.185 - 0.282) Fair Filamentous Turf 0.449 (0.434 - 0.464) Moderate 0.387 (0.370 - 0.404) Fair 0.561 (0.547 - 0.576) Moderate Zoanthus sociatus 0.574 (0.526 - 0.623) Moderate 0.462 (0.418 - 0.506) Moderate 0.397 (0.348 - 0.447) Fair Porites astreoides 0.952 (0.859 - 1.000) Almost Perfect 0.428 (0.103 - 0.753) Moderate 0.461 (0.125 - 0.797) Moderate Sargassum spp. 0.768 (0.714 - 0.822) Substantial 0.767 (0.712 - 0.821) Substantial 0.850 (0.804 - 0.896) Almost perfect Siderastrea spp. 0.771 (0.721 - 0.821) Substantial 0.773 (0.723 - 0.823) Substantial 0.881 (0.843 - 0.919) Almost perfect Dictyota spp. 0.439 (0.358 - 0.521) Moderate 0.371 (0.290 - 0.453) Fair 0.332 (0.243 - 0.420) Fair Total 0.526 (0.515 - 0.536) Moderate 0.497 (0.487 - 0.508) Moderate 0.636 (0.626 - 0.647) Substantial Morphofunctional Group Km Intensity (Km) Ka Intensity (Ka) Kb Intensity (Kb) Coral 0.779 (0.731 - 0.827) Substantial 0.765 (0.715 - 0.814) Substantial 0.868 (0.828 - 0.907) Almost perfect Leathery macroalgae 0.763 (0.708 - 0.817) Substantial 0.761 (0.253 - 0.341) Substantial 0.850 (0.804 - 0.896) Almost perfect Crustose macroalgae 0.187 (0.145 - 0.229) Slight 0.297 (0.726 - 0.754) Fair 0.224 (0.176 - 0.271) Fair Foliaceous macroalgae 0.580 (0.512 - 0.648) Moderate 0.400 (0.328 - 0.473) Fair 0.414 (0.338 - 0.490) Moderate Filamentous macroalgae 0.623 (0.604 - 0.642) Substantial 0.603 (0.584 - 0.622) Moderate 0.807 (0.792 - 0.823) Almost perfect Substrate 0.770 (0.756 - 0.783) Substantial 0.740 (0.726 - 0.754) Substantial 0.816 (0.804 - 0.828) Almost perfect Turf 0.523 (0.515 - 0.544) Moderate 0.497 (0.482 - 0.512) Moderate 0.674 (0.661 - 0.687) Substantial Zoanthids 0.573 (0.525 - 0.622) Moderate 0.462 (0.418 - 0.506) Moderate 0.397 (0.347 - 0.446) Fair Total 0.604 (0.593 - 0.615) Moderate 0.577 (0.566 - 0.588) Moderate 0.711 (0.701 - 0.721) Substantial Effect of image quality on annotator performance Some imperfections were more prevalent than others in the images obtained. The most frequent imperfections were shadow (95.6%) and distortion (89.2%) (Table 4). Unfocused patches (83.4%) and parallax errors (72.9%) were also common. In contrast, overexposure and loss of analysis area were less common, with 7.8% and 6.4%, respectively (Table 4). However, three of the most frequent imperfections did not show significant effects on image analysis (Fig. 3): shadow (p = 0.790), distortion (p = 0.167), and unfocused patches (p = 0.353). In contrast, sediment suspension (p = 0.003), parallax error (p = 0.004), sun glint (p = 0.01), and overexposure (p = 0.067) significantly reduced the annotators’ Cohen’s Kappa (Fig. 3). No significant differences were found between the Cohen’s Kappa values of the manual and automated annotators (p = 0.372) (Fig. 3). Table 4: Relative frequency of imperfections in the collected images. Note that an image may be classified in more than one imperfection. Imperfections Total Shadow (%) 95.6 Suspension (%) 33.7 Sun Glint (%) 26.3 Overexposure (%) 7.8 Unfocus (%) 83.4 Distortion (%) 89.2 Parallax error (%) 72.9 Object inside (%) 16.6 Loss of area (%) 6.4 Total number of images (n) 499 Discussion Humans versus machines Our results demonstrated that automated annotation with CoralNet performed comparably to manual annotation across most morphofunctional groups and labels, supporting the use of AI as an efficient and reliable tool for reef monitoring. The performance was generally greater at the morphofunctional group level than at the individual label level, which is consistent with the findings of previous studies (Beijbom et al. 2015 ; Bravo et al. 2021). Furthermore, for both manual and automated annotators, Cohen's Kappa values varied depending on the label or morphofunctional group, indicating that certain classes were easier to classify than others were. The high performance for corals and leathery macroalgae, without significant differences between manual and automated annotation, indicates that AI can reliably detect major habitat-forming organisms. Previous studies have investigated the accuracy of AI in identifying corals and macroalgae (Beijbom et al. 2015 ; González-Rivero et al. 2020) but have not used a morphofunctional approach. Here, the grouping of macroalgae revealed that while no significant differences were detected between Cohen's Kappa values for corals and leathery macroalgae for both annotators, other macroalgae groups, such as filamentous macroalgae and particularly foliaceous macroalgae, presented lower performance values. Previous studies that grouped macroalgae into one category likely masked the variation we observed between the different groups. Furthermore, González-Rivero et al. (2020) reported that the soft coral group presented intermediate error rates, which were higher than those of true corals but lower than those of algae, especially in Central America. Our study revealed lower annotator performance for zoanthids than for corals, with zoanthids falling between the algae morphofunctional groups. The specific label performance did not always match that of the morphofunctional groups. Among the filamentous macroalgae groups, Caulerpa verticillata had the highest Cohen's Kappa, but overall, the group performed worse than, for example, the leathery macroalgae group did. This lower performance for filamentous macroalgae was likely due to confusion of similar labels with "Filamentous Turf" from the morphofunctional group TURF. Additionally, the lack of training data for Cladophora spp. led to their misidentification as Bryopsis spp., lowering Cohen's Kappa for filamentous macroalgae for both annotators. In these situations, data augmentation could be particularly beneficial by increasing the representation of undersampled taxa. This technique is widely applied for CNN models that artificially enlarge datasets by creating new examples (synthetic data) from slight changes in samples of the original dataset and is a solution to the problem of limited or imbalanced datasets (Shorten and Khoshgoftaar, 2019 ; Stowell et al., 2019 ). The application of this technique would allow for better discrimination between similar genera by introducing greater morphological variation. Nevertheless, the high concordance for Bryopsis spp. between the two types of annotators indicates that, even for erroneously assigned points, the manual and the automated annotator agreed upon identifications, resulting in a substantial performance for both annotators. Both annotators struggled more with foliaceous macroalgae, probably because of confusion between Dictyota spp. and "Filamentous Turf". A turf, by definition, is a multispecific tangle that can be distinguished from macroalgae by size (Connell et al., 2014 ). Differentiating algae by size using photoquadrats is a challenging task, and in this study, these labels often overlapped. The low performance for turf is consistent with the findings of Beijbom et al. ( 2015 ) and González-Rivero et al. (2020), as it encompasses many organisms with different shapes, colors, and textures. This explains the low performance of the annotators for “ Dictyota spp.” and “Filamentous Turf”. However, for the morphofunctional groups (foliaceous macroalgae and TURF), the manual annotator outperformed the automated annotator. For crustose coralline macroalgae, the lowest-performing label, identification difficulties are due to their small size and lack of well-defined edges, which is consistent with the findings of Beijbom et al. ( 2015 ). Data augmentation techniques, including transformations in color (brightness, contrast, saturation) and scale (resizing, zooming), could help mitigate this by highlighting subtle features and boundaries that are critical for accurate identification for automated annotation. Notably, rare or underrepresented taxa, such as Porites astreoides , were better detected by human annotators. While the performance of the manual annotator was near perfect, the performance of the automated annotator was moderate. This reflects the current dependence of AI on large, balanced training sets. Nevertheless, previous studies have shown that once sufficient examples are included, CoralNet can reach human-level accuracy even for Porites astreoides (Beijbom et al. 2015 ; Williams et al. 2019 ). This suggests that expanding and diversifying image databases will directly enhance the automated recognition of rare but ecologically important taxa. For Siderastrea spp., this efficiency is confirmed by the substantial performance of both the automated and manual annotators. This is significant given the species abundance contribution to the coral group at Rocas Atol (Gherardi, 1995). When comparing annotation methods, it is crucial to consider that although manual annotators tend to align more with experts, AI tools such as CoralNet provide efficient annotations, with estimates close to those of humans (Bravo et al. 2021; González-Rivero et al. 2020; Williams et al. 2019 ). Furthermore, Beijibom et al. (2015) reported that even experts do not have 100% agreement when reannotating images (intraannotator variation) and that there is variability between different annotators (interannotator variation). A major advantage of AI is its consistency across time and observers. While human annotators may vary in accuracy due to fatigue or experience, CoralNet applies the same decision rules to all images, reducing intra- and interannotator variation (Williams et al. 2019 ). Given its high processing speed and lower costs (González-Rivero et al. 2020), AI is a powerful complement to human expertise in long-term ecological studies, particularly through the use of CoralNet ’s semiautomated tools. Ultimately, our findings support the integration of automated tools into reef monitoring programs, enabling broader spatial and temporal coverage without compromising accuracy. Effect of image quality on annotator performance This study is the first to explicitly test how different types of image imperfections affect the manual and automated annotation of reef photoquadrats. Surprisingly, the most common imperfections (shadow, distortion, blur, and parallax) did not significantly reduce performance, suggesting that both humans and AI are relatively robust to some level of image degradation. In contrast, sediment suspension, overexposure, sun glint, and severe parallax errors significantly reduced annotation accuracy. These imperfections obscure visual information and therefore represent critical challenges for both annotators. By demonstrating that only certain imperfections have substantial effects, our study advances upon previous works by establishing which imperfections are most detrimental for both humans and AI. The fact that shadow, distortion, and unfocus did not significantly affect annotation performance may be linked to the presence–absence methodology used here. Mild shadows, distortions, or a lack of focus may not impact annotations as much as intense ones. Similarly, the random points on the images may not coincidentally fall on the regions with imperfections. To better understand how image quality affects annotations, future studies should quantify imperfections such as saturation, brightness, and color to establish a minimum threshold for recording presence or create a ranking of the images analyzed to measure how different intensities of imperfections affect annotations. A novel contribution of our study is the identification of parallax error as critical, yet often overlooked, imperfection. While Bravo et al. (2021) discussed parallax in relation to in situ identification, they did not consider its impact on photoquadrats. Here, we show that improper alignment between the quadrat and the substrate can reduce the performance of both annotator types. Moreover, loss of analysis area and the presence of objects such as tape measures may not distort organism identification but reduce the effective sampling effort, leading to potential biases in community estimates. Although these two factors do not affect the quality of annotation, they should still be avoided, as valuable information could be lost. Importantly, our results demonstrate that humans and AI were similarly affected by image imperfections. This suggests that image acquisition is a stronger limiting factor than the annotator type itself. In other words, improving field protocols to reduce problematic imperfections through better lighting, careful positioning of quadrats, and optimized camera settings will simultaneously enhance the performance of both approaches. This highlights that ecological monitoring should focus not only on developing better analytical tools but also on refining data collection practices to ensure image quality. Conclusion The analysis of image quality and annotator performance revealed that image imperfections similarly affect both manual and automated annotations. Interestingly, the most common imperfections (shadow, distortion, blur, and parallax error) were not always the most influential in the annotation, with sediment suspension, parallax error, sunlight reflection, overexposure, and loss of analysis area having a more significant impact. To fully understand the influence of these factors, quantitative studies that account for the intensity of each imperfection are needed. Despite variations in image quality, both manual and automated annotators performed similarly across labels and morphofunctional groups. In particular, the performance was greater for morphofunctional groups than for individual labels. Groups with high contrast and less variation in shape and color, such as corals and leathery macroalgae, were easier to identify than those with subtle differences and smaller sizes, such as crustose algae and turf. Together, these results validate the use of AI-based annotation for monitoring reef communities, offering a fast, cost-effective, and consistent alternative to manual methods. By enabling the processing of larger datasets without loss of ecological accuracy, automated tools such as CoralNet can greatly enhance our ability to detect changes in benthic communities, strengthen biodiversity assessments, and inform conservation strategies in tropical reef ecosystems. Declarations Funding This study was funded in part by the Long-Term Ecological Research of Reef Communities at Oceanic Islands – PELD ILOC (CNPq 441327/2020-6; PI: C.E.L.F.) — Brazil. C.E.L.F. is also supported by grants from Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) and Fundação de Amparo à Pesquisa do Estado do Rio de Janeiro (FAPERJ E-26/200.215/2023). Competing interests The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Author contributions Marina A. Méga de A.: Data curation, Investigation, Methodology, Formal analysis, Visualization, Writing – original draft, Writing – review & editing. Viviane R. Barroso: Writing – original draft, Writing – review & editing. Carlos Eduardo L. Ferreira: Project administration, Funding acquisition, Resources, Writing – review & editing. Cesar A. M. M. Cordeiro: Conceptualization, Methodology, Project administration, Supervision, Writing – review & editing. Data availability The datasets generated and/or analyzed during the current study are available at https://github.com/marinamega/Humans-versus-Machines.git Ethics Approval This study did not involve human participants or animals that require formal ethical approval. All necessary permissions for fieldwork were obtained from the relevant authorities. References Ahmad J, Farman H, Jan Z (2019) Deep learning methods and applications. In: Khan M, Jan B, Farman H (eds) Deep learning: convergence to big data analytics. Springer, Gateway, pp 31–42. https://doi.org/10.1007/978-981-13-3459-7_3 Beijbom O, Edmunds PJ, Roelfsema C, Smith J, Kline DI, Neal BP, Dunlap MJ, Moriarty V, Fan TY, Tan CJ, Chan S, Treibitz T, Gamst A, Mitchell BG, Kriegman D (2015) Towards automated annotation of benthic survey images: variability of human experts and operational modes of automation. PLoS One 10:e0130312. https://doi.org/10.1371/journal.pone.0130312 Carletta J (1996) Assessing agreement on classification tasks: the kappa statistic. Comput Linguist 22:249–254. https://doi.org/10.48550/arXiv.cmp-lg/9602004 Chen Q, Beijbom O, Chan S, Bouwmeester J, Kriegman D (2021) A new deep learning engine for CoralNet. In: IEEE/CVF international conference on computer vision workshops (ICCVW), pp 3686–3695. https://doi.org/10.1109/ICCVW54120.2021.00412 Cohen J (1960) A coefficient of agreement for nominal scales. Educ Psychol Meas 20:37–46. https://doi.org/10.1177/001316446002000104 Connell SD, Foster MS, Airoldi L (2014) What are algal turfs? Towards a better description of turfs. Mar Ecol Prog Ser 495:299–307. https://doi.org/10.3354/meps10513 Gonzalez-Rivero M, Beijbom O, Rodriguez-Ramirez A, Bryant DE, Ganase A, Gonzalez-Marrero Y, Herrera-Reveles A, Kennedy EV, Kim CJS, Lopez-Marcano S, Markey K, Neal BP, Osborne K, Reyes-Nivia C, Sampayo EM, Stolberg K, Taylor A, Vercelloni J, Wyatt M, Hoegh-Guldberg O (2020) Monitoring of coral reefs using artificial intelligence: a feasible and cost-effective approach. Remote Sens 12:489. https://doi.org/10.3390/rs12030489 Hatcher BG (1997) Coral reef ecosystems: how much greater is the whole than the sum of the parts? Coral Reefs 16:S77–S91. https://doi.org/10.1007/s003380050244 Hoegh-Guldberg O, Mumby PJ, Hooten AJ, Steneck RS, Greenfield P, Gomez E, Harvell CD, Sale PF, Edwards AJ, Caldeira K, Knowlton N, Eakin CM (2007) Coral reefs under rapid climate change and ocean acidification. Science 318:1737–1742. https://doi.org/10.1126/science.1152509 Hughes TP, Baird AH, Bellwood DR, Card M, Connolly SR, Folke C, Grosberg R, Hoegh-Guldberg O, Jackson JBC, Kleypas J, Lough JM, Marshall P, Nyström M, Palumbi SR, Pandolfi JM, Rosen B, Roughgarden J (2003) Climate change, human impacts, and the resilience of coral reefs. Science 301:929–933. https://doi.org/10.1126/science.1085046 Hughes TP, Anderson KD, Connolly SR, Heron SF, Kerry JT, Lough JM, Baird AH, Baum JK, Berumen ML, Bridge TC, Claar DC, Eakin CM, Gilmour JP, Graham NAJ, Harrison H, Hobbs JA, Hoey AS, Hoogenboom M, Lowe RJ, McCulloch MT, Pandolfi JM, Pratchett M, Schoepf V, Torda G, Wilson SK (2018) Spatial and temporal patterns of mass bleaching of corals in the Anthropocene. Science 359:80–83. https://doi.org/10.1126/science.aan8048 Kohler KE, Gill SM (2006) Coral Point Count with Excel extensions (CPCe): a Visual Basic program for the determination of coral and substrate coverage using random point count methodology. Comput Geosci 32:1259–1269. https://doi.org/10.1016/j.cageo.2005.11.009 Landis JR, Koch GG (1977) The measurement of observer agreement for categorical data. Biometrics 33:159–174. https://doi.org/10.2307/2529310 LeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521:436–444. https://doi.org/10.1038/nature14539 Longo GO, Morais RA, Martins CDL, Mendes TC, Aued AW, Cândido DV, de Oliveira JC, Nunes LT, Fontoura L, Sissini MN, Teschima MM, Silva MB, Ramlov F, Gouvea LP, Ferreira CEL, Segal B, Horta PA, Floeter SR (2015) Between-habitat variation of benthic cover, reef fish assemblage and feeding pressure on the benthos at the only atoll in South Atlantic: Rocas Atoll, NE Brazil. PLoS One 10:e0127176. https://doi.org/10.1371/journal.pone.0127176 MacLeod N, Benfield M, Culverhouse P (2010) Time to automate identification. Nature 467:154. https://doi.org/10.1038/467154a Moberg F, Folke C (1999) Ecological goods and services of coral reef ecosystems. Ecol Econ 29:215–233. https://doi.org/10.1016/S0921-8009(99)00009-9 Picolotto VAP, Aued AW, Macedo-Soares LC, Zamoner JB, Segal B (2021) A functional perspective for reef benthic communities: temporal trends in the only atoll in Southern Atlantic. Res Sq [preprint]. https://doi.org/10.21203/rs.3.rs-683029/v1 Quimbayo JP, Mendes TC, Cordeiro CAMM, Longo GO, Giglio VJ, Vergara DC, Floeter SR, Villaça R, Freire AS, Fracini-Filho RB, Ferreira CEL (2018) Variação temporal das comunidades recifais no Arquipélago de São Pedro e São Paulo: ações de monitoramento de longa duração (PELD). In: Oliveira JE, Viana DL, Souza MAC (eds) Arquipélago de São Pedro e São Paulo: 20 anos de pesquisa. Design Publicações, pp 112–127 Reaka-Rudda ML (1997) The global biodiversity of coral reefs: a comparison with rainforests. In: Reaka-Rudda ML, Wilson DE, Wilson EO (eds) Biodiversity II: understanding and protecting our biological resources. Joseph Henry Press, Washington, pp 83–108 R Core Team (2016) R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna. https://www.R-project.org Revelle W (2017) psych: procedures for psychological, psychometric, and personality research. R package version 1.7.8. Northwestern University, Evanston, Illinois. https://CRAN.R-project.org/package=psych Shihavuddin ASM, Gracias N, Garcia R, Gleason AC, Gintert B (2013) Image-based coral reef classification and thematic mapping. Remote Sens 5:1809–1841. https://doi.org/10.3390/rs5041809 Shorten C, Khoshgoftaar TM (2019) A survey on image data augmentation for deep learning. J Big Data 6:60. https://doi.org/10.1186/s40537-019-0197-0 Stokes MD, Deane GB (2009) Automated processing of coral reef benthic images. Limnol Oceanogr Methods 7:157–168. https://doi.org/10.4319/lom.2009.7.157 Stowell D, Petrusková T, Šálek M, Linhart P (2019) Automatic acoustic identification of individuals in multiple species: improving identification across recording conditions. J R Soc Interface 16:20180940. https://doi.org/10.1098/rsif.2018.0940 Trygonis V, Sini M (2012) photoQuad: a dedicated seabed image processing software, and a comparative error analysis of four photoquadrat methods. J Exp Mar Biol Ecol 424:99–108. https://doi.org/10.1016/j.jembe.2012.04.018 Williams ID, Couch CS, Beijbom O, Oliver TA, Vargas-Ángel B, Schumacher BD, Brainard RE (2019) Leveraging automated image analysis tools to transform our capacity to assess status and trends of coral reefs. Front Mar Sci 6:222. https://doi.org/10.3389/fmars.2019.00222 Zamoner JB, Aued AW, Macedo-Soares LCP, Picolotto VAP, Eiras Garcia CA, Segal B (2021) Integrating oceanographic data and benthic community structure temporal series to assess the dynamics of a marginal reef. Front Mar Sci 8:762453. https://doi.org/10.3389/fmars.2021.762453 Zaneveld JR, Burkepile DE, Shantz AA, Pritchard CE, McMinds R, Payet JP, Welsh R, Correa AM, Lemoine NP, Rosales S, Fuchs C, Maynard JA, Thurber RV (2016) Overfishing and nutrient pollution interact with temperature to disrupt coral reefs down to microbial scales. Nat Commun 7:11833. https://doi.org/10.1038/ncomms11833 Supplementary Files renamed3783a.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revise and Resubmit 06 Jan, 2026 Reviewers agreed at journal 13 Sep, 2025 Reviewers invited by journal 13 Sep, 2025 Editor assigned by journal 11 Sep, 2025 First submitted to journal 10 Sep, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7586184","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":514596090,"identity":"fb5c0d50-0813-4fcd-b012-390e18360ded","order_by":0,"name":"Marina Alves Méga de Andrade","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA70lEQVRIiWNgGAWjYJCCA3DWByBmYydFC+MMkBZmUqxj5gGTBFTptp99eLigxi6an3/xw882v7bJ8zEzMH74mINbi9mZdIPDM44l586c8cxYOrfvtmEbMwOz5MxteLQcSGM4zMPGnLvhxgED6dye24xALWzMvPi0nH8G1PKvPnf/jeOff1v23LYnrOUG0BbetsO5G/h7zKQZftxOJEIL0BbevuO5M27wlFn2NtxObmNmbMbvl/NpzJ95vlXn9vcf33zjx5/btvPbmw9++IhHCwJIJADjsg3EYmwgRj0Q8B8AEn+IVDwKRsEoGAUjCgAA7R1V6m7AxPMAAAAASUVORK5CYII=","orcid":"https://orcid.org/0009-0008-9164-3226","institution":"Universidade Federal do Rio de Janeiro","correspondingAuthor":true,"prefix":"","firstName":"Marina","middleName":"Alves Méga","lastName":"de Andrade","suffix":""},{"id":514596091,"identity":"13c438f8-08e3-4eae-8a7a-cf574e43172e","order_by":1,"name":"Viviane Barroso","email":"","orcid":"","institution":"Instituto de Estudos do Mar Almirante Paulo Moreira","correspondingAuthor":false,"prefix":"","firstName":"Viviane","middleName":"","lastName":"Barroso","suffix":""},{"id":514596092,"identity":"870b9144-1d23-4e34-93d2-0ac00f9c6972","order_by":2,"name":"Carlos Eduardo Leite Ferreira","email":"","orcid":"","institution":"Universidade Federal Fluminense","correspondingAuthor":false,"prefix":"","firstName":"Carlos","middleName":"Eduardo Leite","lastName":"Ferreira","suffix":""},{"id":514596093,"identity":"60d2d04a-c213-4ed0-a597-24871545f09c","order_by":3,"name":"César Augusto Cordeiro","email":"","orcid":"","institution":"Universidade Estadual do Norte Fluminense: Universidade Estadual do Norte Fluminense Darcy Ribeiro","correspondingAuthor":false,"prefix":"","firstName":"César","middleName":"Augusto","lastName":"Cordeiro","suffix":""}],"badges":[],"createdAt":"2025-09-10 22:00:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7586184/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7586184/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":91952373,"identity":"6c6b6104-ff44-4e82-b093-2753182902f1","added_by":"auto","created_at":"2025-09-23 06:50:59","extension":"xml","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9296,"visible":true,"origin":"","legend":"","description":"","filename":"mabiMABID2500467.xml","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/7886b900e8e215567a89bbe4.xml"},{"id":91953620,"identity":"58a398d9-21f9-499e-949e-c9cb6c7b0720","added_by":"auto","created_at":"2025-09-23 06:58:59","extension":"xml","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1136,"visible":true,"origin":"","legend":"","description":"","filename":"MABID250046714254.go.xml","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/65657945da35d689f2fe46ee.xml"},{"id":91952376,"identity":"c90ce15e-ad6e-4520-b516-ac0d228acdf2","added_by":"auto","created_at":"2025-09-23 06:50:59","extension":"xml","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":798,"visible":true,"origin":"","legend":"","description":"","filename":"MABID2500467Import.xml","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/425da13bab76d16d2266ed8f.xml"},{"id":91953621,"identity":"d73d8c62-500c-46c5-89bc-75940c8edcee","added_by":"auto","created_at":"2025-09-23 06:58:59","extension":"xml","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":120806,"visible":true,"origin":"","legend":"","description":"","filename":"MABID25004670enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/eb2cf3fd5802180bf81b3ef2.xml"},{"id":91952380,"identity":"39a4ff78-ee20-45ee-b582-00b431d8d27f","added_by":"auto","created_at":"2025-09-23 06:50:59","extension":"eps","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":729736,"visible":true,"origin":"","legend":"","description":"","filename":"Fig1.eps","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/80645a95c24bf2418c2992a0.eps"},{"id":91953625,"identity":"220147f9-7e75-4d62-84c4-e0eefb5a287f","added_by":"auto","created_at":"2025-09-23 06:58:59","extension":"eps","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":982906,"visible":true,"origin":"","legend":"","description":"","filename":"Fig2.eps","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/1c73983d827c2dda53b5672e.eps"},{"id":91952382,"identity":"f8e9f1e5-aece-430a-bd81-058610424418","added_by":"auto","created_at":"2025-09-23 06:50:59","extension":"eps","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":470763,"visible":true,"origin":"","legend":"","description":"","filename":"Fig3.eps","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/7d64c7dfb5a14ee3842367a6.eps"},{"id":91954294,"identity":"7efa0857-7512-4640-bd73-785dc344faff","added_by":"auto","created_at":"2025-09-23 07:06:59","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":88124,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/7fe23a82f5d13ae88ed9db89.png"},{"id":91953623,"identity":"cb6fc20c-9a45-491f-86b3-ac64f8c8a939","added_by":"auto","created_at":"2025-09-23 06:58:59","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":111624,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/5fbf2b290d7c1267d6cf438f.png"},{"id":91952386,"identity":"f624c46f-554c-4cee-81fd-67c2474ab36c","added_by":"auto","created_at":"2025-09-23 06:50:59","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":147166,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/a20dcf30de3f86540b5acf3b.png"},{"id":91954293,"identity":"48922df1-c061-4395-8aef-3cc7cbe1751b","added_by":"auto","created_at":"2025-09-23 07:06:59","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28019,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/6963bfaf43ecb5014874fba1.png"},{"id":91953622,"identity":"92635aed-8ae3-4259-8075-23b26abcb3c3","added_by":"auto","created_at":"2025-09-23 06:58:59","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":35737,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/d8875cd937f1fd9b40a6f3bd.png"},{"id":91952381,"identity":"ff0d8860-5373-496b-9856-55887c7e0f9e","added_by":"auto","created_at":"2025-09-23 06:50:59","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":142845,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/df0e656a41a285047aa7845d.png"},{"id":91953627,"identity":"ec3c7f7b-f0be-443a-90fc-91ed2e6e6a68","added_by":"auto","created_at":"2025-09-23 06:58:59","extension":"xml","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":119454,"visible":true,"origin":"","legend":"","description":"","filename":"MABID25004670structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/5ef5f7ede1a08778e7fcced7.xml"},{"id":91952389,"identity":"77fcd9c7-d1e0-4a26-88b2-4faa5edb1f6e","added_by":"auto","created_at":"2025-09-23 06:50:59","extension":"html","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":127867,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/a8142b69978dfa1338e9e747.html"},{"id":91952374,"identity":"837ea2bd-1ab6-4444-a38c-c82a5fa22b3b","added_by":"auto","created_at":"2025-09-23 06:50:59","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":163727,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance of annotators and agreement between annotators, measured by Cohen's Kappa. The performance reflects the annotator's agreement with the expert, whereas the agreement between annotators is assessed by comparing manual and automated annotations. The x-axis shows Total Cohen’s Kappa and Cohen’s Kappa across morphofunctional groups.\u003c/p\u003e","description":"","filename":"fig1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/9e10e869045e695df7b19b6e.jpg"},{"id":91952372,"identity":"e8429381-2a6a-41ca-a8f2-53778132423a","added_by":"auto","created_at":"2025-09-23 06:50:59","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":187969,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance of annotators and agreement between annotators, measured by Cohen's Kappa. The performance reflects the annotator's agreement with the expert, whereas the agreement between annotators is assessed by comparing manual and automated annotations. The x-axis shows Total Cohen’s Kappa and Cohen’s Kappa across labels.\u003c/p\u003e","description":"","filename":"fig2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/02bbc96a712585d4b45899ac.jpg"},{"id":91952375,"identity":"e321440b-6e5b-453d-bcb7-71569d0f7167","added_by":"auto","created_at":"2025-09-23 06:50:59","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":256026,"visible":true,"origin":"","legend":"\u003cp\u003eEffects of image quality factors on the performance of the annotator (manual and automated). Dark gray areas are images where no imperfections occurred, and gray areas are images in which imperfection occurred. Significant factors are marked with an asterisk. The performance of the annotator was measured by Cohen’s Kappa as a morphofunctional group in each image.\u003c/p\u003e","description":"","filename":"fig3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/5e6403927916ff38a6dca363.jpg"},{"id":91956450,"identity":"8c8f1b69-b935-47d4-815c-79a913b08a76","added_by":"auto","created_at":"2025-09-23 07:15:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1378264,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/92206b97-2ab9-4711-b7f6-91abf88f64bc.pdf"},{"id":91952378,"identity":"205f8f26-ad15-45bd-9ef4-ef86905e89db","added_by":"auto","created_at":"2025-09-23 06:50:59","extension":"docx","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":9924,"visible":true,"origin":"","legend":"","description":"","filename":"renamed3783a.docx","url":"https://assets-eu.researchsquare.com/files/rs-7586184/v1/50b979693fb791037fe5f88d.docx"}],"financialInterests":"","formattedTitle":"Humans versus Machines: Artificial intelligence performs similarly to manual analysis in benthic reef monitoring","fulltext":[{"header":"Introduction","content":"\u003cp\u003eTropical reefs are among the most important ecosystems on Earth and are comparable to tropical rainforests in terms of high primary productivity (Hatcher \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e1997\u003c/span\u003e), complexity, and diversity (Reaka-Kudla et al. 1997). They also hold significant economic importance because of their fundamental ecosystem services, such as providing protein sources, supporting tourism, and protecting coastlines (Hughes et al. \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2003\u003c/span\u003e; Moberg \u0026amp; Folke \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e1999\u003c/span\u003e). However, these ecosystems face global threats, such as ocean acidification and warming (Hoegh-Guldberg 2007; Hughes et al. \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), and local impacts, such as overfishing and pollution (Zaneveld et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Long-term monitoring is therefore crucial for understanding ecosystem functioning under these combined pressures (Quimbayo et al. \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e\u003cp\u003ePhotoquadrat analysis is one of the most widely used methods for quantifying benthic community structure, with software such as \u003cem\u003eCPC\u003c/em\u003ee (Kohler \u0026amp; Gill \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2006\u003c/span\u003e) and \u003cem\u003ePhotoQuad\u003c/em\u003e (Trygonis \u0026amp; Sini \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2012\u003c/span\u003e) commonly applied. However, manual annotation of photoquadrats is a time-consuming and tedious process, creating a bottleneck between data collection and analysis (Stokes \u0026amp; Deane \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). In addition, human annotation is subjective and inconsistent, leading to potential misclassifications, overestimations, or omissions on the basis of the experience of the individual annotator (MacLeod et al. \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). Automated approaches based on artificial intelligence (AI) offer promising alternatives, reducing annotation subjectivity and shortening analysis time, allowing more images to be processed in less time (Shihavuddin et al. \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). \u003cem\u003eCoralNet\u003c/em\u003e is software similar to \u003cem\u003eCPCe\u003c/em\u003e and \u003cem\u003ePhotoQuad\u003c/em\u003e but includes an automated annotation tool using computer vision methods (Beijbom et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Chen et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). After at least 20 images have been annotated by a researcher, \u003cem\u003eCoralNet\u003c/em\u003e begins to automatically annotate the remaining images, improving its performance with more manually annotated images (Chen et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThe \u003cem\u003eCoralNet\u003c/em\u003e AI tool operates in two stages: a convolutional neural network (CNN) and a simple classifier. CNNs are deep learning-based approaches designed to process large datasets that are structured in multiple arrays (LeCun et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). These deep neural networks have gained attention and are now being used in various fields, such as image recognition, speech recognition, and natural language processing (Ahmad et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). The typical architecture of CNNs consists of multiple layers that learn features hierarchically from the training dataset. Simpler features, such as edges and basic shapes, are learned in the initial layers, whereas deeper layers capture increasingly complex features as the data propagate through the network. During training, the network optimizes its weights via backpropagation, focusing on features that are more relevant to improve the classification task (Ahmad et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Gonz\u0026aacute;lez-Rivero et al. 2020). The second stage of \u003cem\u003ethe CoralNet\u003c/em\u003e AI tool is a simple classifier that receives a feature vector from the CNN and assigns a score to each label in the label set, selecting the highest score for annotation (Chen et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eDespite these advantages, challenges remain. A minimum number of images is required for accurate identification, making it more effective for common species than for rare ones. Moreover, incorrect annotations during training can lead to errors in automated analysis if not corrected by a human. Moreover, image quality also impacts analysis, and both manual and automated annotations are likely to be influenced by factors such as resolution, turbidity, and lighting (Beijbom et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). However, few studies have directly evaluated how image imperfections influence annotation outcomes or whether manual and automated approaches are similarly affected (Beijbom et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). The effectiveness of AI could help reduce the gap between collected and annotated images in long-term monitoring programs and enhance the reproducibility of analyses (Williams et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eHere, we evaluate the performance of an automated annotation tool for classifying benthic cover in tropical reefs, comparing its results with those of manual annotation. We further assess how image quality influences both approaches, providing insights into which imperfections are most detrimental to annotation accuracy. This study contributes to improving standards for image-based surveys, supporting large-scale monitoring and conservation of reef ecosystems.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eStudy area and sampling\u003c/h2\u003e\u003cp\u003eWe collected images in Rocas Atoll (03\u0026deg;50'S, 33\u0026deg;49'W), the first no-take, marine protected area in Brazil, established in 1978. Monitoring was conducted annually during May and June from 2014\u0026ndash;2017 at four different sites (\u0026Acirc;ncoras, Cemit\u0026eacute;rio, Falsa Barreta, and Podes Crer). At each site, three 20-meter transects were laid out following a systematic sampling method with 2-meter intervals. This approach resulted in a total of ten or eleven photoquadrats per site (Picolotto et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, \u003cem\u003epreprint\u003c/em\u003e). In another study, all images were analyzed by an expert (Vitor Andr\u0026eacute; Passos Picolotto - Picolotto et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, \u003cem\u003epreprint\u003c/em\u003e) via the \u003cem\u003eCPCe\u003c/em\u003e program, with fifty random points per photoquadrat.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eImage quality and set separation\u003c/h3\u003e\n\u003cp\u003eWe qualitatively classified the images (n\u0026thinsp;=\u0026thinsp;499) based on the presence or absence of the imperfections listed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Inappropriate images were excluded from the dataset, resulting in 493 viable images, and we divided the transects equally (n\u0026thinsp;=\u0026thinsp;24 each) into two different sets: i. the training set (248 images), which was used to train both human (manual) and AI (automated) annotators; and ii. evaluation set (245 images), in which images were reanalyzed by both manual and automated annotators using the same points as the expert but without knowledge of the previous annotations (Beijbom et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). The evaluation set was then used to compare and verify the agreement between the annotators (manual and automated) and the expert. Organisms were identified to the lowest taxonomic level possible (generally genera) and then grouped into morphofunctional groups (following Zamoner et al. \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The labelset used was developed by the Long Term Programs of Brazilian Oceanic Islands (PELD-ILOC) (Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e - Supplementary Material). Additionally, the effects of image quality on manual and automated annotations were analyzed via this set, allowing for the assessment of how each imperfection affects both analysis methods.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eImage quality based on the presence or absence of imperfections.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eImperfection\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDescription\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eShadow\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLocalized decrease in brightness (\u0026lt;\u0026thinsp;50% than surrounding)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSuspension\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSediment particles in front of the target or on the camera lens\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSun Glint\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eReflection of sunlight on water (presence of \u0026ldquo;rainbow\u0026rdquo; beam)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOverexposure\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eArea with brightness excess and saturated pixels\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDistortion\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDeformation of image edges caused by lens curvature\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eUnfocused\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePresence of unfocused area between two or more surfaces\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eParallax error\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eImage not perpendicular to the target center\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eObject inside area\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePresence of object inside the analysis area (e.g., tape)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLoss of analysis area\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSquare cut off the image, resulting in loss of analysis area\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003eData analysis\u003c/h2\u003e\u003cdiv id=\"Sec6\" class=\"Section3\"\u003e\u003ch2\u003ePerformance of the annotators\u003c/h2\u003e\u003cp\u003eTraditionally, the agreement between observers has been assessed by the percentage of matches between observations (overall accuracy). However, this method does not consider agreements occurring by chance. To address this, Cohen's Kappa was introduced as a metric to estimate agreement while excluding random chance (Cohen, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e1960\u003c/span\u003e). The Kappa statistic is calculated as follows:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:K\\:=\\frac{P\\left(a\\right)\\:-\\:P\\left(e\\right)}{1\\:-\\:P\\left(e\\right)}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003ewhere \u003cem\u003eP(a)\u003c/em\u003e represents the observed proportion of agreement and \u003cem\u003eP(e)\u003c/em\u003e denotes the proportion of agreement expected by chance (Carletta 2008). Cohen's Kappa ranges from \u0026minus;\u0026thinsp;1 to 1 and can be interpreted according to Landis \u0026amp; Koch (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e1977\u003c/span\u003e) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). This value is generally lower than the total accuracy because the total accuracy overestimates the agreement by ignoring chance factors.\u003c/p\u003e\u003cp\u003eIn this section, we use Cohen\u0026rsquo;s Kappa to compare each annotator with the expert (performance of the annotators) and to compare both annotators (manual and automated) with each other. We call \u003cem\u003eKa\u003c/em\u003e the performance of the automated annotator (AI), \u003cem\u003eKm\u003c/em\u003e the performance of the manual annotator (human), and \u003cem\u003eKb\u003c/em\u003e the agreement between the annotators. The analysis was performed in two ways: i) evaluating the overall performance of both manual and automated annotators across all images (total Cohen\u0026rsquo;s Kappa), using annotations previously analyzed by the expert as a baseline for comparison; and ii) assessing the ability of the annotators to perform binary classification tasks (Cohen\u0026rsquo;s Kappa for each group), such as distinguishing \u003cem\u003eSiderastrea\u003c/em\u003e from non-\u003cem\u003eSiderastrea\u003c/em\u003e or corals from noncorals. Analyses were performed for both labels and morphofunctional groups (Beijbom et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). Cohen\u0026rsquo;s Kappa confidence intervals were set at 95%, and values were computed via the \"Psych\" package in R (Revelle, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eIntensity of interobserver agreement according to Cohen\u0026rsquo;s Kappa values from Landis \u0026amp; Koch (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e1977\u003c/span\u003e)\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCohen\u0026rsquo;s Kappa value\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAgreement intensity\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePoor\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e0.00\u0026ndash;0.20\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSlight\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e0.21\u0026ndash;0.40\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFair\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e0.41\u0026ndash;0.60\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eModerate\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e0.61\u0026ndash;0.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSubstantial\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e0.81\u0026ndash;1.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAlmost perfect\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\n\u003ch3\u003eEffect of image quality on the performance of the annotators\u003c/h3\u003e\n\u003cp\u003eIn this section, we used the Total Cohen\u0026rsquo;s Kappa for each annotator (manual and automated) on each image. Specifically, we assessed the performance of each annotator on individual images (e.g., the total Cohen\u0026rsquo;s Kappa of the automated annotator on image 1) to make the comparison. To reduce potential biases related to class abundance, we focused only on morphofunctional groups, as there are fewer rare morphofunctional groups than rare labels. We then ran a mixed model to identify which image quality factors influenced the annotations. In this analysis, the dependent variable was agreement (Cohen\u0026rsquo;s Kappa for each image), and the independent variables were image quality imperfections and the type of annotator (manual or automated), with images used as a random effect. The significance level (P value) was set at 5%. \u0026ldquo;Object within analysis area\u0026rdquo; and \u0026ldquo;loss of analysis area\u0026rdquo; were not included in the model because they do not lead to misclassification, although they reduce the sampled area and must be minimized during image acquisition.\u003c/p\u003e"},{"header":"Results","content":"\u003ch2\u003e\u003cem\u003eOverall performance\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eAt the label level, the overall performance of both manual and automated annotators was classified as moderate (Table 4), although the performance of manual annotators was higher than that of automated annotators. In addition, the overall agreement between annotators was greater than their individual performance (i.e., each annotator\u0026rsquo;s agreement with the expert) was (Fig. 1; Table 4). Total Cohen\u0026rsquo;s Kappa (overall performance) at the morphofunctional group level was greater than that at the label level for both annotators, although it remained moderate (Table 4). Nevertheless, the pattern was the same, with higher Kappa values for agreement between annotators, followed by the overall performance of the manual annotator and, finally, the overall performance of the automated annotator (Fig. 2; Table 4).\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eMorphofunctional groups\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eThe morphofunctional groups best identified by the manual annotator were corals, leathery macroalgae, substrate, and filamentous macroalgae. For all of these groups, the manual annotator achieved the highest Kappa values and substantial performance (Table 4). For zoanthids, foliaceous macroalgae, and turf algae, the Kappa values decreased, and the performance was moderate. Finally, the morphofunctional group with the lowest Kappa values was crustose macroalgae, which were classified as having slight performance by the manual annotator (Table 4).\u003c/p\u003e\n\u003cp\u003eFor the automated annotator, the Kappa values were also higher for corals, leathery macroalgae, and substrate, all with substantial performance. However, the performance for filamentous macroalgae was moderate, with Kappa at the upper limit for this category (Fig. 1; Table 4). Turf algae and zoanthids also presented moderate performance but with relatively low Kappa values. Finally, foliaceous macroalgae and crustose macroalgae were classified as fair performers and represented the groups with the lowest Kappa values (Fig. 2; Table 4).\u003c/p\u003e\n\u003cp\u003eWhen the annotation methods were compared, the performance did not differ significantly for corals, leathery macroalgae, or filamentous macroalgae. For the substrate, turf algae, zoanthids, and foliaceous macroalgae, the manual annotator outperformed the automated annotator, whereas the opposite was observed for crustose macroalgae. We found higher kappa values and agreement between annotators than for each individual performance metric. Agreement was greater for corals, leathery macroalgae, substrate, and filamentous macroalgae, whereas disagreement was greater for crustose macroalgae, zoanthids, and foliaceous macroalgae. For turf algae, the agreement between annotators was substantial (Fig. 2; Table 4).\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eLabels\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eFor the labels, the manual annotator obtained the highest Cohen\u0026rsquo;s Kappa values for the coral \u003cem\u003ePorites astreoides\u003c/em\u003e and for \u003cem\u003eCaulerpa verticillata\u003c/em\u003e, both with almost perfect performance (Fig. 2; Table 3). \u003cem\u003eSiderastrea\u0026nbsp;\u003c/em\u003espp., \u003cem\u003eSargassum\u0026nbsp;\u003c/em\u003espp., and the substrate also had high Kappa values and substantial performance. For \u003cem\u003eBryopsis\u0026nbsp;\u003c/em\u003espp., \u003cem\u003eZoanthus sociatus\u003c/em\u003e, filamentous turf, and \u003cem\u003eDictyota\u0026nbsp;\u003c/em\u003espp., the manual annotator showed moderate performance and intermediate Kappa values. Finally, the label with the lowest kappa values was crustose calcareous macroalgae, which was classified as having slight performance (Fig. 2; Table 3).\u003c/p\u003e\n\u003cp\u003eFor the automated annotator, the labels with the highest Kappa values were almost the same: \u003cem\u003eCaulerpa verticillata\u003c/em\u003e, \u003cem\u003eSiderastrea\u0026nbsp;\u003c/em\u003espp., \u003cem\u003eSargassum\u0026nbsp;\u003c/em\u003espp., and substrate. However, none of these labels achieved almost perfect performance, and only substantial performance was achieved (Fig. 2; Table 3). In contrast, the automated annotator identified\u003cem\u003e\u0026nbsp;Porites astreoides\u003c/em\u003e with moderate performance, which was lower than that of the manual annotator. Other labels with moderate performance were \u003cem\u003eBryopsis\u0026nbsp;\u003c/em\u003espp. and \u003cem\u003eZoanthus sociatus\u003c/em\u003e. Finally, the lowest performance was fair, observed for filamentous turf, \u003cem\u003eDictyota\u0026nbsp;\u003c/em\u003espp., and crustose calcareous macroalgae (Fig. 2; Table 3).\u003c/p\u003e\n\u003cp\u003eComparing the two annotation methods, Cohen\u0026rsquo;s Kappa values for \u003cem\u003eCaulerpa verticillata\u003c/em\u003e, \u003cem\u003eSiderastrea\u0026nbsp;\u003c/em\u003espp., and \u003cem\u003eSargassum\u0026nbsp;\u003c/em\u003espp. were similar for both manual and automated annotators. The same applied to \u003cem\u003eBryopsis\u0026nbsp;\u003c/em\u003espp. and \u003cem\u003eDictyota\u0026nbsp;\u003c/em\u003espp., where despite higher performance categories for the manual annotator, the confidence intervals overlapped (Fig. 2; Table 3). In this case, even for the labels that were grouped into different categories depending on the annotator (e.g., \u003cem\u003eC. verticillata\u003c/em\u003e with almost perfect performance for the manual annotator and substantial performance for the automated annotator; \u003cem\u003eDictyota\u0026nbsp;\u003c/em\u003espp. with moderate performance for the manual annotator and fair performance for the automated annotator), no significant differences were found (Table 3). For substrate, \u003cem\u003ePorites astreoides\u003c/em\u003e, \u003cem\u003eZoanthus sociatus\u003c/em\u003e, and filamentous turf, the manual annotator had higher Kappa values than did the automated annotator, whereas the opposite occurred for crustose calcareous macroalgae (Fig. 2; Table 3). Among the labels, those with the highest agreement between annotators were \u003cem\u003eSiderastrea\u0026nbsp;\u003c/em\u003espp., \u003cem\u003eSargassum\u0026nbsp;\u003c/em\u003espp., \u003cem\u003eCaulerpa verticillata\u003c/em\u003e, substrate, and \u003cem\u003eBryopsis\u0026nbsp;\u003c/em\u003espp., whereas the highest disagreement occurred for crustose calcareous macroalgae, \u003cem\u003eDictyota\u0026nbsp;\u003c/em\u003espp., \u003cem\u003eZoanthus sociatus\u003c/em\u003e, and \u003cem\u003ePorites astreoides\u003c/em\u003e. For filamentous turf, the interannotator agreement was intermediate (Fig. 2; Table 3).\u003c/p\u003e\n\u003cp\u003eTable 3:\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eClassification of Cohen\u0026apos;s Kappa values for each class (label or morphofunctional group) via the annotation method, according to Landis \u0026amp; Koch (1977). The performance can be ordered as follows: poor (up to 0.00), slight (0.00 - 0.20), fair (0.21 - 0.40), moderate (0.41 - 0.60), substantial (0.61 - 0.80), and almost perfect (0.81 - 1.00). We denoted Ka as the automated annotator performance (AI), Km as the manual annotator performance (human), and Kb as the agreement between the annotators. The values for the substrate are the same for both the labeled and morphofunctional groups, as it is the only label composing the morphofunctional group.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"588\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"7\" valign=\"top\" style=\"width: 100%;\"\u003e\n \u003cp\u003eLabels\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003eKm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eIntensity (Km)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003eKa\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eIntensity (Ka)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003eKb\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eIntensity (Kb)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eSubstrate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.770 (0.756 - 0.783)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.740 (0.726 - 0.754)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.816 (0.804 - 0.828)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eAlmost perfect\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003e\u003cem\u003eBryopsis\u003c/em\u003e spp.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.591 (0.559 - 0.622)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.554 (0.523 - 0.586)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.795 (0.773 - 0.817)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003e\u003cem\u003eCaulerpa verticillata\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.826 (0.804 - 0.847)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eAlmost Perfect\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.799 (0.777 - 0.820)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.819 (0.798 - 0.839)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eAlmost perfect\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eCrustose calcareous algae\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.184 (0.141 - 0.227)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSlight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.296 (0.251 - 0.340)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.234 (0.185 - 0.282)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eFilamentous Turf\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.449 \u0026nbsp;(0.434 - 0.464)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.387 (0.370 - 0.404)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e\u0026nbsp;0.561 (0.547 - 0.576)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003e\u003cem\u003eZoanthus sociatus\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.574 (0.526 - 0.623)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.462 (0.418 - 0.506)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e\u0026nbsp;0.397 (0.348 - 0.447)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003e\u003cem\u003ePorites astreoides\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.952 (0.859 - 1.000)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eAlmost Perfect\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.428 (0.103 - 0.753)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.461 (0.125 - 0.797)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003e\u003cem\u003eSargassum\u003c/em\u003e spp.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.768 (0.714 - 0.822)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.767 (0.712 - 0.821)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.850 (0.804 - 0.896)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eAlmost perfect\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003e\u003cem\u003eSiderastrea\u003c/em\u003e spp.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.771 (0.721 - 0.821)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.773 (0.723 - 0.823)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.881 (0.843 - 0.919)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eAlmost perfect\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003e\u003cem\u003eDictyota\u003c/em\u003e spp.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.439 (0.358 - 0.521)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.371 (0.290 - 0.453)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.332 (0.243 - 0.420)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.526 (0.515 - 0.536)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.497 (0.487 - 0.508)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.636 (0.626 - 0.647)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"7\" valign=\"top\" style=\"width: 100%;\"\u003e\n \u003cp\u003eMorphofunctional Group\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003eKm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eIntensity (Km)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003eKa\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eIntensity (Ka)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003eKb\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eIntensity (Kb)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eCoral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.779 (0.731 - 0.827)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.765 (0.715 - 0.814)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.868 (0.828 - 0.907)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eAlmost perfect\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eLeathery macroalgae\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.763 (0.708 - 0.817)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.761 (0.253 - 0.341)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.850 (0.804 - 0.896)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eAlmost perfect\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eCrustose macroalgae\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.187 (0.145 - 0.229)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSlight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.297 (0.726 - 0.754)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.224 (0.176 - 0.271)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eFoliaceous macroalgae\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.580 (0.512 - 0.648)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.400 (0.328 - 0.473)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.414 (0.338 - 0.490)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eFilamentous macroalgae\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.623 (0.604 - 0.642)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.603 (0.584 - 0.622)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.807 (0.792 - 0.823)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eAlmost perfect\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eSubstrate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.770 (0.756 - 0.783)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.740 (0.726 - 0.754)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.816 \u0026nbsp;(0.804 - 0.828)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eAlmost perfect\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003e\u003cem\u003eTurf\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.523 (0.515 - 0.544)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.497 (0.482 - 0.512)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.674 (0.661 - 0.687)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eZoanthids\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.573 (0.525 - 0.622)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.462 (0.418 - 0.506)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.397 (0.347 - 0.446)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 15.9864%;\"\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.4354%;\"\u003e\n \u003cp\u003e0.604 (0.593 - 0.615)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.577 (0.566 - 0.588)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.0952%;\"\u003e\n \u003cp\u003e0.711 (0.701 - 0.721)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.7959%;\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003e\u003cem\u003eEffect of image quality on annotator performance\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eSome imperfections were more prevalent than others in the images obtained. The most frequent imperfections were shadow (95.6%) and distortion (89.2%) (Table 4). Unfocused patches (83.4%) and parallax errors (72.9%) were also common. In contrast, overexposure and loss of analysis area were less common, with 7.8% and 6.4%, respectively (Table 4).\u003c/p\u003e\n\u003cp\u003eHowever, three of the most frequent imperfections did not show significant effects on image analysis (Fig. 3): shadow (p = 0.790), distortion (p = 0.167), and unfocused patches (p = 0.353). In contrast, sediment suspension (p = 0.003), parallax error (p = 0.004), sun glint (p = 0.01), and overexposure (p = 0.067) significantly reduced the annotators\u0026rsquo; Cohen\u0026rsquo;s Kappa (Fig. 3). No significant differences were found between the Cohen\u0026rsquo;s Kappa values of the manual and automated annotators (p = 0.372) (Fig. 3).\u003c/p\u003e\n\u003cp\u003eTable 4:\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eRelative frequency of imperfections in the collected images. Note that an image may be classified in more than one imperfection.\u003c/p\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"487\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78.0287%;\"\u003e\n \u003cp\u003eImperfections\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9713%;\"\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78.0287%;\"\u003e\n \u003cp\u003eShadow (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9713%;\"\u003e\n \u003cp\u003e95.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78.0287%;\"\u003e\n \u003cp\u003eSuspension (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9713%;\"\u003e\n \u003cp\u003e33.7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78.0287%;\"\u003e\n \u003cp\u003eSun Glint (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9713%;\"\u003e\n \u003cp\u003e26.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78.0287%;\"\u003e\n \u003cp\u003eOverexposure (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9713%;\"\u003e\n \u003cp\u003e7.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78.0287%;\"\u003e\n \u003cp\u003eUnfocus (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9713%;\"\u003e\n \u003cp\u003e83.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78.0287%;\"\u003e\n \u003cp\u003eDistortion (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9713%;\"\u003e\n \u003cp\u003e89.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78.0287%;\"\u003e\n \u003cp\u003eParallax error (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9713%;\"\u003e\n \u003cp\u003e72.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78.0287%;\"\u003e\n \u003cp\u003eObject inside (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9713%;\"\u003e\n \u003cp\u003e16.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78.0287%;\"\u003e\n \u003cp\u003eLoss of area (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9713%;\"\u003e\n \u003cp\u003e6.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 78.0287%;\"\u003e\n \u003cp\u003eTotal number of images (n)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9713%;\"\u003e\n \u003cp\u003e499\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e"},{"header":"Discussion","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003eHumans versus machines\u003c/h2\u003e\u003cp\u003eOur results demonstrated that automated annotation with \u003cem\u003eCoralNet\u003c/em\u003e performed comparably to manual annotation across most morphofunctional groups and labels, supporting the use of AI as an efficient and reliable tool for reef monitoring. The performance was generally greater at the morphofunctional group level than at the individual label level, which is consistent with the findings of previous studies (Beijbom et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Bravo et al. 2021). Furthermore, for both manual and automated annotators, Cohen's Kappa values varied depending on the label or morphofunctional group, indicating that certain classes were easier to classify than others were.\u003c/p\u003e\u003cp\u003eThe high performance for corals and leathery macroalgae, without significant differences between manual and automated annotation, indicates that AI can reliably detect major habitat-forming organisms. Previous studies have investigated the accuracy of AI in identifying corals and macroalgae (Beijbom et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Gonz\u0026aacute;lez-Rivero et al. 2020) but have not used a morphofunctional approach. Here, the grouping of macroalgae revealed that while no significant differences were detected between Cohen's Kappa values for corals and leathery macroalgae for both annotators, other macroalgae groups, such as filamentous macroalgae and particularly foliaceous macroalgae, presented lower performance values. Previous studies that grouped macroalgae into one category likely masked the variation we observed between the different groups. Furthermore, Gonz\u0026aacute;lez-Rivero et al. (2020) reported that the soft coral group presented intermediate error rates, which were higher than those of true corals but lower than those of algae, especially in Central America. Our study revealed lower annotator performance for zoanthids than for corals, with zoanthids falling between the algae morphofunctional groups.\u003c/p\u003e\u003cp\u003eThe specific label performance did not always match that of the morphofunctional groups. Among the filamentous macroalgae groups, \u003cem\u003eCaulerpa verticillata\u003c/em\u003e had the highest Cohen's Kappa, but overall, the group performed worse than, for example, the leathery macroalgae group did. This lower performance for filamentous macroalgae was likely due to confusion of similar labels with \"Filamentous Turf\" from the morphofunctional group TURF. Additionally, the lack of training data for \u003cem\u003eCladophora\u003c/em\u003e spp. led to their misidentification as \u003cem\u003eBryopsis\u003c/em\u003e spp., lowering Cohen's Kappa for filamentous macroalgae for both annotators. In these situations, data augmentation could be particularly beneficial by increasing the representation of undersampled taxa. This technique is widely applied for CNN models that artificially enlarge datasets by creating new examples (synthetic data) from slight changes in samples of the original dataset and is a solution to the problem of limited or imbalanced datasets (Shorten and Khoshgoftaar, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Stowell et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). The application of this technique would allow for better discrimination between similar genera by introducing greater morphological variation.\u003c/p\u003e\u003cp\u003eNevertheless, the high concordance for \u003cem\u003eBryopsis\u003c/em\u003e spp. between the two types of annotators indicates that, even for erroneously assigned points, the manual and the automated annotator agreed upon identifications, resulting in a substantial performance for both annotators. Both annotators struggled more with foliaceous macroalgae, probably because of confusion between \u003cem\u003eDictyota\u003c/em\u003e spp. and \"Filamentous Turf\". A turf, by definition, is a multispecific tangle that can be distinguished from macroalgae by size (Connell et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Differentiating algae by size using photoquadrats is a challenging task, and in this study, these labels often overlapped. The low performance for turf is consistent with the findings of Beijbom et al. (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) and Gonz\u0026aacute;lez-Rivero et al. (2020), as it encompasses many organisms with different shapes, colors, and textures. This explains the low performance of the annotators for \u0026ldquo;\u003cem\u003eDictyota\u003c/em\u003e spp.\u0026rdquo; and \u0026ldquo;Filamentous Turf\u0026rdquo;. However, for the morphofunctional groups (foliaceous macroalgae and TURF), the manual annotator outperformed the automated annotator. For crustose coralline macroalgae, the lowest-performing label, identification difficulties are due to their small size and lack of well-defined edges, which is consistent with the findings of Beijbom et al. (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). Data augmentation techniques, including transformations in color (brightness, contrast, saturation) and scale (resizing, zooming), could help mitigate this by highlighting subtle features and boundaries that are critical for accurate identification for automated annotation.\u003c/p\u003e\u003cp\u003eNotably, rare or underrepresented taxa, such as \u003cem\u003ePorites astreoides\u003c/em\u003e, were better detected by human annotators. While the performance of the manual annotator was near perfect, the performance of the automated annotator was moderate. This reflects the current dependence of AI on large, balanced training sets. Nevertheless, previous studies have shown that once sufficient examples are included, \u003cem\u003eCoralNet\u003c/em\u003e can reach human-level accuracy even for Porites astreoides (Beijbom et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Williams et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). This suggests that expanding and diversifying image databases will directly enhance the automated recognition of rare but ecologically important taxa. For \u003cem\u003eSiderastrea\u003c/em\u003e spp., this efficiency is confirmed by the substantial performance of both the automated and manual annotators. This is significant given the species abundance contribution to the coral group at Rocas Atol (Gherardi, 1995).\u003c/p\u003e\u003cp\u003eWhen comparing annotation methods, it is crucial to consider that although manual annotators tend to align more with experts, AI tools such as CoralNet provide efficient annotations, with estimates close to those of humans (Bravo et al. 2021; Gonz\u0026aacute;lez-Rivero et al. 2020; Williams et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Furthermore, Beijibom et al. (2015) reported that even experts do not have 100% agreement when reannotating images (intraannotator variation) and that there is variability between different annotators (interannotator variation). A major advantage of AI is its consistency across time and observers. While human annotators may vary in accuracy due to fatigue or experience, \u003cem\u003eCoralNet\u003c/em\u003e applies the same decision rules to all images, reducing intra- and interannotator variation (Williams et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Given its high processing speed and lower costs (Gonz\u0026aacute;lez-Rivero et al. 2020), AI is a powerful complement to human expertise in long-term ecological studies, particularly through the use of \u003cem\u003eCoralNet\u003c/em\u003e\u0026rsquo;s semiautomated tools. Ultimately, our findings support the integration of automated tools into reef monitoring programs, enabling broader spatial and temporal coverage without compromising accuracy.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003eEffect of image quality on annotator performance\u003c/h2\u003e\u003cp\u003eThis study is the first to explicitly test how different types of image imperfections affect the manual and automated annotation of reef photoquadrats. Surprisingly, the most common imperfections (shadow, distortion, blur, and parallax) did not significantly reduce performance, suggesting that both humans and AI are relatively robust to some level of image degradation. In contrast, sediment suspension, overexposure, sun glint, and severe parallax errors significantly reduced annotation accuracy. These imperfections obscure visual information and therefore represent critical challenges for both annotators. By demonstrating that only certain imperfections have substantial effects, our study advances upon previous works by establishing which imperfections are most detrimental for both humans and AI.\u003c/p\u003e\u003cp\u003eThe fact that shadow, distortion, and unfocus did not significantly affect annotation performance may be linked to the presence\u0026ndash;absence methodology used here. Mild shadows, distortions, or a lack of focus may not impact annotations as much as intense ones. Similarly, the random points on the images may not coincidentally fall on the regions with imperfections. To better understand how image quality affects annotations, future studies should quantify imperfections such as saturation, brightness, and color to establish a minimum threshold for recording presence or create a ranking of the images analyzed to measure how different intensities of imperfections affect annotations.\u003c/p\u003e\u003cp\u003eA novel contribution of our study is the identification of parallax error as critical, yet often overlooked, imperfection. While Bravo et al. (2021) discussed parallax in relation to \u003cem\u003ein situ\u003c/em\u003e identification, they did not consider its impact on photoquadrats. Here, we show that improper alignment between the quadrat and the substrate can reduce the performance of both annotator types. Moreover, loss of analysis area and the presence of objects such as tape measures may not distort organism identification but reduce the effective sampling effort, leading to potential biases in community estimates. Although these two factors do not affect the quality of annotation, they should still be avoided, as valuable information could be lost.\u003c/p\u003e\u003cp\u003eImportantly, our results demonstrate that humans and AI were similarly affected by image imperfections. This suggests that image acquisition is a stronger limiting factor than the annotator type itself. In other words, improving field protocols to reduce problematic imperfections through better lighting, careful positioning of quadrats, and optimized camera settings will simultaneously enhance the performance of both approaches. This highlights that ecological monitoring should focus not only on developing better analytical tools but also on refining data collection practices to ensure image quality.\u003c/p\u003e\u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThe analysis of image quality and annotator performance revealed that image imperfections similarly affect both manual and automated annotations. Interestingly, the most common imperfections (shadow, distortion, blur, and parallax error) were not always the most influential in the annotation, with sediment suspension, parallax error, sunlight reflection, overexposure, and loss of analysis area having a more significant impact. To fully understand the influence of these factors, quantitative studies that account for the intensity of each imperfection are needed. Despite variations in image quality, both manual and automated annotators performed similarly across labels and morphofunctional groups. In particular, the performance was greater for morphofunctional groups than for individual labels. Groups with high contrast and less variation in shape and color, such as corals and leathery macroalgae, were easier to identify than those with subtle differences and smaller sizes, such as crustose algae and turf. Together, these results validate the use of AI-based annotation for monitoring reef communities, offering a fast, cost-effective, and consistent alternative to manual methods. By enabling the processing of larger datasets without loss of ecological accuracy, automated tools such as \u003cem\u003eCoralNet\u003c/em\u003e can greatly enhance our ability to detect changes in benthic communities, strengthen biodiversity assessments, and inform conservation strategies in tropical reef ecosystems.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding\u003cbr\u003e\u003c/strong\u003eThis study was funded in part by the Long-Term Ecological Research of Reef Communities at Oceanic Islands \u0026ndash; PELD ILOC (CNPq 441327/2020-6; PI: C.E.L.F.) \u0026mdash; Brazil. C.E.L.F. is also supported by grants from Conselho Nacional de Desenvolvimento Cient\u0026iacute;fico e Tecnol\u0026oacute;gico (CNPq) and Funda\u0026ccedil;\u0026atilde;o de Amparo \u0026agrave; Pesquisa do Estado do Rio de Janeiro (FAPERJ E-26/200.215/2023).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003cbr\u003e\u003c/strong\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003cbr\u003e\u003c/strong\u003eMarina A. M\u0026eacute;ga de A.: Data curation, Investigation, Methodology, Formal analysis, Visualization, Writing \u0026ndash; original draft, Writing \u0026ndash; review \u0026amp; editing.\u003cbr\u003e\u0026nbsp;Viviane R. Barroso: Writing \u0026ndash; original draft, Writing \u0026ndash; review \u0026amp; editing.\u003cbr\u003e\u0026nbsp;Carlos Eduardo L. Ferreira: Project administration, Funding acquisition, Resources, Writing \u0026ndash; review \u0026amp; editing.\u003cbr\u003e\u0026nbsp;Cesar A. M. M. Cordeiro: Conceptualization, Methodology, Project administration, Supervision, Writing \u0026ndash; review \u0026amp; editing.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003cbr\u003e\u003c/strong\u003eThe datasets generated and/or analyzed during the current study are available at https://github.com/marinamega/Humans-versus-Machines.git\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study did not involve human participants or animals that require formal ethical approval. All necessary permissions for fieldwork were obtained from the relevant authorities.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAhmad J, Farman H, Jan Z (2019) Deep learning methods and applications. In: Khan M, Jan B, Farman H (eds) Deep learning: convergence to big data analytics. Springer, Gateway, pp 31\u0026ndash;42. https://doi.org/10.1007/978-981-13-3459-7_3\u003c/li\u003e\n\u003cli\u003eBeijbom O, Edmunds PJ, Roelfsema C, Smith J, Kline DI, Neal BP, Dunlap MJ, Moriarty V, Fan TY, Tan CJ, Chan S, Treibitz T, Gamst A, Mitchell BG, Kriegman D (2015) Towards automated annotation of benthic survey images: variability of human experts and operational modes of automation. PLoS One 10:e0130312. https://doi.org/10.1371/journal.pone.0130312\u003c/li\u003e\n\u003cli\u003eCarletta J (1996) Assessing agreement on classification tasks: the kappa statistic. Comput Linguist 22:249\u0026ndash;254. https://doi.org/10.48550/arXiv.cmp-lg/9602004\u003c/li\u003e\n\u003cli\u003eChen Q, Beijbom O, Chan S, Bouwmeester J, Kriegman D (2021) A new deep learning engine for CoralNet. In: IEEE/CVF international conference on computer vision workshops (ICCVW), pp 3686\u0026ndash;3695. https://doi.org/10.1109/ICCVW54120.2021.00412\u003c/li\u003e\n\u003cli\u003eCohen J (1960) A coefficient of agreement for nominal scales. Educ Psychol Meas 20:37\u0026ndash;46. https://doi.org/10.1177/001316446002000104\u003c/li\u003e\n\u003cli\u003eConnell SD, Foster MS, Airoldi L (2014) What are algal turfs? Towards a better description of turfs. Mar Ecol Prog Ser 495:299\u0026ndash;307. https://doi.org/10.3354/meps10513\u003c/li\u003e\n\u003cli\u003eGonzalez-Rivero M, Beijbom O, Rodriguez-Ramirez A, Bryant DE, Ganase A, Gonzalez-Marrero Y, Herrera-Reveles A, Kennedy EV, Kim CJS, Lopez-Marcano S, Markey K, Neal BP, Osborne K, Reyes-Nivia C, Sampayo EM, Stolberg K, Taylor A, Vercelloni J, Wyatt M, Hoegh-Guldberg O (2020) Monitoring of coral reefs using artificial intelligence: a feasible and cost-effective approach. Remote Sens 12:489. https://doi.org/10.3390/rs12030489\u003c/li\u003e\n\u003cli\u003eHatcher BG (1997) Coral reef ecosystems: how much greater is the whole than the sum of the parts? Coral Reefs 16:S77\u0026ndash;S91. https://doi.org/10.1007/s003380050244\u003c/li\u003e\n\u003cli\u003eHoegh-Guldberg O, Mumby PJ, Hooten AJ, Steneck RS, Greenfield P, Gomez E, Harvell CD, Sale PF, Edwards AJ, Caldeira K, Knowlton N, Eakin CM (2007) Coral reefs under rapid climate change and ocean acidification. Science 318:1737\u0026ndash;1742. https://doi.org/10.1126/science.1152509\u003c/li\u003e\n\u003cli\u003eHughes TP, Baird AH, Bellwood DR, Card M, Connolly SR, Folke C, Grosberg R, Hoegh-Guldberg O, Jackson JBC, Kleypas J, Lough JM, Marshall P, Nystr\u0026ouml;m M, Palumbi SR, Pandolfi JM, Rosen B, Roughgarden J (2003) Climate change, human impacts, and the resilience of coral reefs. Science 301:929\u0026ndash;933. https://doi.org/10.1126/science.1085046\u003c/li\u003e\n\u003cli\u003eHughes TP, Anderson KD, Connolly SR, Heron SF, Kerry JT, Lough JM, Baird AH, Baum JK, Berumen ML, Bridge TC, Claar DC, Eakin CM, Gilmour JP, Graham NAJ, Harrison H, Hobbs JA, Hoey AS, Hoogenboom M, Lowe RJ, McCulloch MT, Pandolfi JM, Pratchett M, Schoepf V, Torda G, Wilson SK (2018) Spatial and temporal patterns of mass bleaching of corals in the Anthropocene. Science 359:80\u0026ndash;83. https://doi.org/10.1126/science.aan8048\u003c/li\u003e\n\u003cli\u003eKohler KE, Gill SM (2006) Coral Point Count with Excel extensions (CPCe): a Visual Basic program for the determination of coral and substrate coverage using random point count methodology. Comput Geosci 32:1259\u0026ndash;1269. https://doi.org/10.1016/j.cageo.2005.11.009\u003c/li\u003e\n\u003cli\u003eLandis JR, Koch GG (1977) The measurement of observer agreement for categorical data. Biometrics 33:159\u0026ndash;174. https://doi.org/10.2307/2529310\u003c/li\u003e\n\u003cli\u003eLeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521:436\u0026ndash;444. https://doi.org/10.1038/nature14539\u003c/li\u003e\n\u003cli\u003eLongo GO, Morais RA, Martins CDL, Mendes TC, Aued AW, C\u0026acirc;ndido DV, de Oliveira JC, Nunes LT, Fontoura L, Sissini MN, Teschima MM, Silva MB, Ramlov F, Gouvea LP, Ferreira CEL, Segal B, Horta PA, Floeter SR (2015) Between-habitat variation of benthic cover, reef fish assemblage and feeding pressure on the benthos at the only atoll in South Atlantic: Rocas Atoll, NE Brazil. PLoS One 10:e0127176. https://doi.org/10.1371/journal.pone.0127176\u003c/li\u003e\n\u003cli\u003eMacLeod N, Benfield M, Culverhouse P (2010) Time to automate identification. Nature 467:154. https://doi.org/10.1038/467154a\u003c/li\u003e\n\u003cli\u003eMoberg F, Folke C (1999) Ecological goods and services of coral reef ecosystems. Ecol Econ 29:215\u0026ndash;233. https://doi.org/10.1016/S0921-8009(99)00009-9\u003c/li\u003e\n\u003cli\u003ePicolotto VAP, Aued AW, Macedo-Soares LC, Zamoner JB, Segal B (2021) A functional perspective for reef benthic communities: temporal trends in the only atoll in Southern Atlantic. Res Sq [preprint]. https://doi.org/10.21203/rs.3.rs-683029/v1\u003c/li\u003e\n\u003cli\u003eQuimbayo JP, Mendes TC, Cordeiro CAMM, Longo GO, Giglio VJ, Vergara DC, Floeter SR, Villa\u0026ccedil;a R, Freire AS, Fracini-Filho RB, Ferreira CEL (2018) Varia\u0026ccedil;\u0026atilde;o temporal das comunidades recifais no Arquip\u0026eacute;lago de S\u0026atilde;o Pedro e S\u0026atilde;o Paulo: a\u0026ccedil;\u0026otilde;es de monitoramento de longa dura\u0026ccedil;\u0026atilde;o (PELD). In: Oliveira JE, Viana DL, Souza MAC (eds) Arquip\u0026eacute;lago de S\u0026atilde;o Pedro e S\u0026atilde;o Paulo: 20 anos de pesquisa. Design Publica\u0026ccedil;\u0026otilde;es, pp 112\u0026ndash;127\u003c/li\u003e\n\u003cli\u003eReaka-Rudda ML (1997) The global biodiversity of coral reefs: a comparison with rainforests. In: Reaka-Rudda ML, Wilson DE, Wilson EO (eds) Biodiversity II: understanding and protecting our biological resources. Joseph Henry Press, Washington, pp 83\u0026ndash;108\u003c/li\u003e\n\u003cli\u003eR Core Team (2016) R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna. https://www.R-project.org\u003c/li\u003e\n\u003cli\u003eRevelle W (2017) psych: procedures for psychological, psychometric, and personality research. R package version 1.7.8. Northwestern University, Evanston, Illinois. https://CRAN.R-project.org/package=psych\u003c/li\u003e\n\u003cli\u003eShihavuddin ASM, Gracias N, Garcia R, Gleason AC, Gintert B (2013) Image-based coral reef classification and thematic mapping. Remote Sens 5:1809\u0026ndash;1841. https://doi.org/10.3390/rs5041809\u003c/li\u003e\n\u003cli\u003eShorten C, Khoshgoftaar TM (2019) A survey on image data augmentation for deep learning. J Big Data 6:60. https://doi.org/10.1186/s40537-019-0197-0\u003c/li\u003e\n\u003cli\u003eStokes MD, Deane GB (2009) Automated processing of coral reef benthic images. Limnol Oceanogr Methods 7:157\u0026ndash;168. https://doi.org/10.4319/lom.2009.7.157\u003c/li\u003e\n\u003cli\u003eStowell D, Petruskov\u0026aacute; T, \u0026Scaron;\u0026aacute;lek M, Linhart P (2019) Automatic acoustic identification of individuals in multiple species: improving identification across recording conditions. J R Soc Interface 16:20180940. https://doi.org/10.1098/rsif.2018.0940\u003c/li\u003e\n\u003cli\u003eTrygonis V, Sini M (2012) photoQuad: a dedicated seabed image processing software, and a comparative error analysis of four photoquadrat methods. J Exp Mar Biol Ecol 424:99\u0026ndash;108. https://doi.org/10.1016/j.jembe.2012.04.018\u003c/li\u003e\n\u003cli\u003eWilliams ID, Couch CS, Beijbom O, Oliver TA, Vargas-\u0026Aacute;ngel B, Schumacher BD, Brainard RE (2019) Leveraging automated image analysis tools to transform our capacity to assess status and trends of coral reefs. Front Mar Sci 6:222. https://doi.org/10.3389/fmars.2019.00222\u003c/li\u003e\n\u003cli\u003eZamoner JB, Aued AW, Macedo-Soares LCP, Picolotto VAP, Eiras Garcia CA, Segal B (2021) Integrating oceanographic data and benthic community structure temporal series to assess the dynamics of a marginal reef. Front Mar Sci 8:762453. https://doi.org/10.3389/fmars.2021.762453\u003c/li\u003e\n\u003cli\u003eZaneveld JR, Burkepile DE, Shantz AA, Pritchard CE, McMinds R, Payet JP, Welsh R, Correa AM, Lemoine NP, Rosales S, Fuchs C, Maynard JA, Thurber RV (2016) Overfishing and nutrient pollution interact with temperature to disrupt coral reefs down to microbial scales. Nat Commun 7:11833. https://doi.org/10.1038/ncomms11833\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"marine-biology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"mabi","sideBox":"Learn more about [Marine Biology](https://www.springer.com/journal/227)","snPcode":"227","submissionUrl":"https://submission.nature.com/new-submission/227/3","title":"Marine Biology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"benthic community, CoralNet, image annotation, long-term monitoring, reef systems","lastPublishedDoi":"10.21203/rs.3.rs-7586184/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7586184/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eTropical reefs are among the most biodiverse and threatened marine ecosystems and face both local and global stressors. Understanding community dynamics and ecosystem functioning under these pressures requires long-term monitoring. Image-based surveys are increasingly used in reef ecology because they enable rapid data acquisition and storage. However, manual identification of benthic organisms is time-consuming and potentially subjective, creating a bottleneck between image collection and ecological analysis. Automated annotation tools represent promising solutions, but few studies have directly compared their performance with that of manual approaches, especially under varying image quality conditions. Here, we assess the usefulness of the \u003cem\u003eCoralNet\u003c/em\u003e automated tool for classifying benthic cover in tropical reef images and evaluate how image quality influences both manual and automated annotations. Using expert-validated images, we trained and tested both approaches, measuring performance with Cohen\u0026rsquo;s Kappa coefficient and modeling the effects of image imperfections. We found that manual and automated annotators achieved similar performance across most taxonomic and morphofunctional groups, and both were similarly affected by image quality. Furthermore, the most frequent imperfections were not the most influential. Our findings demonstrate that automated annotation is a reliable and efficient alternative to manual methods, with strong potential to enhance large-scale monitoring, biodiversity assessments, and conservation strategies in reef ecosystems.\u003c/p\u003e","manuscriptTitle":"Humans versus Machines: Artificial intelligence performs similarly to manual analysis in benthic reef monitoring","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-23 06:50:54","doi":"10.21203/rs.3.rs-7586184/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revise and Resubmit","date":"2026-01-06T11:53:21+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"","date":"2025-09-13T14:57:23+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-09-13T13:46:59+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-09-11T13:29:58+00:00","index":"","fulltext":""},{"type":"submitted","content":"Marine Biology","date":"2025-09-10T17:59:42+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"marine-biology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"mabi","sideBox":"Learn more about [Marine Biology](https://www.springer.com/journal/227)","snPcode":"227","submissionUrl":"https://submission.nature.com/new-submission/227/3","title":"Marine Biology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"61ff0d9e-c897-41e5-84a3-615f415e3470","owner":[],"postedDate":"September 23rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-03-09T16:14:59+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-23 06:50:54","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7586184","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7586184","identity":"rs-7586184","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}