When is enough enough? Utilizing bootstrap resampling to estimate benthic cover

Abstract

Reliable estimates of benthic metrics are fundamental for monitoring coral-reef ecosystems and informing management decisions. Here, we present an analytical approach to optimize the sampling design of benthic photoquadrats derived from spatially rectified orthophotomosaics using photogrammetric techniques. Specifically, the numbers of photoquadrats and annotation points per photoquadrat required to achieve the accuracy and precision required for a specific research question can be assessed through simulations and confidence intervals established by bootstrap resampling. Using live coral cover as an example, we demonstrate that both accuracy and precision improve with increasing annotation points and photoquadrat counts, and that the point of diminishing returns beyond which additional annotation points and photoquadrats yield minimal benefits can be identified. The use of bootstrap resampling to quantify precision also provides analytical flexibility by establishing confidence intervals around the point estimates of benthic metrics, which enables robust comparisons of the benthic metrics across time or space. In our live coral cover example, both the level of coral cover and its variability across photoquadrats within each plot affected precision, highlighting the importance of considering benthic heterogeneity when designing sampling strategies. Balancing annotation efforts with the desired level of precision can enable researchers to determine an optimal sampling design for reef monitoring studies.

1 Introduction

Quantitative characterization of coral-reef habitat is an important process in ecological studies, particularly under the increasing levels of threats that coral reefs worldwide are currently experiencing due to environmental stressors (Hughes et al., 2017). Monitoring studies designed to detect spatial and/or temporal changes in the abundances and distributions of reef organisms are critical in management decisions, thus reliable estimates of these metrics are fundamental in coral-reef monitoring. While various survey methods for estimating benthic cover, such as belt transect, line transect and linear point intercept methods, have been proposed and evaluated (Nadon and Stirling, 2006; Vallès et al., 2019; Carneiro et al., 2024), the use of underwater photogrammetric techniques has become relatively popular in recent years due to the capacity to reconstruct coral reefs in high resolutions and extract quantitative data later in the laboratory (Burns et al., 2015; Fukunaga et al., 2019). Resulting reconstructions of survey areas offer a wide range of options to characterize benthic habitat through quantifications of benthic organisms and architectural complexity (Bryson et al., 2017; Fukunaga et al., 2019; Fukunaga and Burns, 2020; Fukunaga et al., 2020; Urbina-Barreto et al., 2022) and provide long-term records with the potential for re-analyses under different approaches in the future.

Proportion of live coral cover in a given habitat is widely used as an ecological descriptor in coral-reef studies (e.g. Rogers and Miller, 2006; Pratchett et al., 2014). Live coral cover, or cover of any benthic organisms, is often estimated from photographs of benthic quadrats (photoquadrats) or from still frames extracted from video recordings of the benthos by randomly overlaying points on the images and digitally annotating the benthic organisms or features beneath each point (Trygonis and Sini, 2012; Molloy et al., 2013). This approach of image annotation has an advantage over in-situ methods as the images can be archived as permanent records, but it can be labor intensive as the workload is dependent on the numbers of images and points that need to be annotated per image. While some studies have investigated the number of images and annotation points per image required for reliable benthic cover estimation (Pante and Dustan, 2012; Perkins et al., 2016; Taormina et al., 2020), these requirements vary with the abundance of benthic categories, with low-cover categories generally requiring greater sampling efforts (Pante and Dustan, 2012). More importantly, the process of determining sampling designs for benthic analyses, including the numbers of photoquadrats and annotation points per photoquadrat, is often unclear or underreported in coral-reef studies and frequently appears to be influenced by logistical constraints of SCUBA-based fieldwork (e.g. limited bottom time) and the time required for image annotation. In many cases, particularly in monitoring programs, survey designs are based on established or “standard” protocols that have been successfully applied across a range of reef systems. These approaches have provided useful consistency and enabled long-term comparisons across studies. However, they are not always optimized for the specific environmental conditions or research questions of individual studies, and formal evaluation of the accuracy and precision of benthic cover estimates remains limited. Given that estimates of benthic cover are widely used to assess reef condition, detect changes and inform management decisions, even modest levels of uncertainties can have important implications. This highlights the need to more rigorously evaluate and optimize benthic annotation and sampling effort and raises a fundamental question: when is enough enough?

Benthic surveys utilizing underwater photogrammetric techniques allow researchers to capture a large study area in a single high-resolution orthophotomosaic (Fukunaga et al., 2019). This approach can potentially reduce the taxonomic expertise of divers in the field required to conduct in-situ identifications and eliminate the need to carry a quadrat framer or quadrapod to photograph individual quadrats, as random sampling of virtual photoquadrats can be completed in the laboratory from the orthophotomosaic. Benthic annotation of the virtual photoquadrats can now also be sped up greatly through a semi-automated image annotation assisted by artificial intelligence (AI; e.g. CoralNet, Beijbom et al., 2015). Given the technological advancements, we revisited the question of how many photoquadrats and annotation points per photoquadrat are required for accurate habitat characterization, emphasizing that there is no “magic number” or one-size-fits-all sampling design for benthic cover analyses. Here, we describe an analytical approach that takes advantage of the ever-improving data processing speed and utilizes empirical data (i.e. orthophotomosaics) to determine an appropriate sampling design based on the level of accuracy and precision required to answer a specific research question. Our approach involves intensive benthic annotation of a small number of orthophotomosaics, which are then used to simulate different combinations of photoquadrat counts and annotation points. We further perform bootstrap resampling for each of the simulated combinations to assess the accuracy and precision of the estimates of benthic metrics. This approach allows for direct comparisons across different sampling strategies and supports data-driven decisions when designing or refining coral-reef monitoring protocols, providing a quantitative framework to determine when sampling effort is sufficient to characterize the benthos reliably.

This study provides valuable insights into sampling designs for benthic characterization using orthophotomosaics and virtual photoquadrats. We encourage scientists and managers to determine the optimal combination of the numbers of photoquadrats and annotation points based on the magnitude of change they aim to detect in their benthic metric of interest, rather than relying on historical conventions. We also demonstrate how bootstrapping techniques can be used to establish confidence intervals around estimates of benthic metrics, which is particularly useful for monitoring designs based on repeated surveys of fixed transects. Together, these approaches provide a practical framework for evaluating sampling efforts using empirical data and contribute to improving the consistency and effectiveness of benthic monitoring. By tailoring a sampling design to the specific goals and conditions of a study, researchers and managers can develop more efficient survey techniques that are both statistically robust and adaptable to resource limitations. Importantly, this work underscores that benthic cover estimates are statistical approximations, and that explicitly quantifying their uncertainty is essential for accurately interpreting ecological patterns and trends.

2 Materials and methods

2.1 General workflow to obtain annotation data from orthophotomosaics

Underwater photogrammetry reconstructs benthic habitat in three dimensions by processing sequential, overlapping photographs taken throughout a study area. The overlapping photographs are processed in photogrammetry software (e.g. Metashape Professional) to generate a spatially orthorectified photomosaic of the study area (survey plot). The survey plot is then divided into virtual photoquadrats using a grid system (Figure 1). To estimate the benthic metrics of interest (e.g. proportion of live coral cover, other invertebrates, macroalgae, etc.), annotation points are randomly placed on each photoquadrat, and organisms and abiotic features under the annotation points are identified and recorded to obtain the estimates of benthic metrics of interest per photoquadrat. Benthic annotation can be performed manually using image analysis or GIS software, or semi-automatically with AI-assisted tools such as CoralNet (https://coralnet.ucsd.edu). For the simulation and bootstrapping procedures described in the following section, we recommend maximizing the numbers of photoquadrats and annotation points to a reasonable extent. In our case study, we utilized all photoquadrats generated using a 1-m² grid system and placed 1,000 random annotation points per 1-m² virtual photoquadrat (Figure 1).

Figure 1

2.2 Simulation and bootstrapping

Our proposed approach simulates combinations of different numbers of photoquadrats and annotation points per photoquadrat to assess an optimal sampling design for a survey plot. For each plot, we randomly subset photoquadrats and annotation points within the photoquadrats (i.e. sampling without replacement), repeating this process many times for each combination of photoquadrat/annotation point numbers. For example, if the combination of 20 photoquadrats and 40 annotation points per photoquadrat is one of the potential sampling designs, we randomly subset 20 photoquadrats out of all available photoquadrats and randomly subset 40 annotation points from all available annotation points for each of the 20 photoquadrats to obtain a simulated dataset. Then we repeat the whole subsetting process, for example, 1,000 times to obtain 1,000 simulated datasets for the sampling design of 20 photoquadrats and 40 annotation points per photoquadrat.

For each simulated dataset, we use bootstrap resampling methods to obtain 95% confidence intervals for the benthic metric of interest. First, estimates of the benthic metric are obtained from all photoquadrats in the simulated dataset, then 10,000 bootstrap resampling iterations (i.e. sampling with replacement) are performed using those estimates.

For each bootstrap resampling, the bootstrap distribution of a t statistic is.

where is the mean of the benthic metric of interest under bootstrap resampling, is the mean of the observed benthic metric of interest for the simulated dataset (i.e. original data before bootstrapping), and is the standard error under bootstrap resampling (Hesterberg, 2015). After 10,000 bootstrap resampling of the simulated dataset, bootstrap t interval is obtained by calculating the 2.5% and 97.5% quantiles of the bootstrap t-distribution, and the 95% confidence interval of the benthic metric is given by.

where and are the 2.5% and 97.5% quantiles of the bootstrap t-distribution and is the standard error for the simulated data (Hesterberg, 2015).

After 10,000 bootstrap resampling of all simulated datasets per sampling design, we calculate the number of times that the 95% confidence intervals successfully included the best estimate of the benthic metric obtained from the maximum (all available) annotation points. We propose that the metric of accuracy is then calculated by dividing the number of successes by the number of simulated datasets. We also calculate the range of 95% confidence interval by subtracting the lower confidence limit from the upper confidence limit for each of the simulated datasets. The metric of precision is then obtained by taking the median range from all simulated datasets per sampling design.

2.3 Case study

2.3.1 Data collection

Images of coral reefs were collected in August 2021 from Kapou (Lisianski Island) in Papahānaumokuākea (the Northwestern Hawaiian Islands) and in April 2022 from Hawai‘i Island in the main Hawaiian Islands as part of larger coral-reef monitoring efforts during expeditions aboard M/V Imua and NOAA ship Oscar Elton Sette, respectively. In both surveys, a transect tape was laid at each survey site and overlapping photographs were taken in a lawn-mower pattern along the tape with 70-80% overlap between the photographs. The imagery was collected using a Sony α7R III mirrorless camera with a 24-mm lens in an underwater housing equipped with an 8-inch dome port. Multiple scale bars were placed around the transect tape and included during the image-collection processes to scale the resulting data products accurately for geospatial analyses. The overlapping photographs were then processed in the software Metashape Professional (Agisoft LLC., St. Petersburg, Russia) to generate orthorectified photomosaics of the survey plots using photogrammetric techniques. The resulting photomosaics (plots) for data analysis were approximately 10 m × 5 m for the ones from Kapou and 25 m × 5 m for those from Hawai‘i Island.

For the case study, we selected five plots from Kapou and six plots from Hawai‘i Island to represent a gradient of coral cover, ranging from relatively low to high, based on visual approximation of coral cover at each site. Each plot was then divided into 1×1 m virtual photoquadrats (Figure 1), and each photoquadrat was processed using the online benthic annotation application CoralNet (https://coralnet.ucsd.edu) by randomly placing 1,000 points per photoquadrat and annotating each point. The automatic annotation function of CoralNet was used to speed up the overall processing time, but each annotation point was manually confirmed to ensure the accuracy. There were 50 and 125 1-m² virtual photoquadrats per plot for the plots from Kapou and Hawai‘i Island, respectively, covering the entire plots for annotation. As part of the larger monitoring efforts, the annotation process labeled all coral to the species level. For the case study, however, we re-labeled them to live coral and others that included non-coral organisms and abiotic features and focused on live coral cover. We obtained the best estimate of live coral cover for each of the 11 plots using all available annotation points (i.e. 50,000 annotation points for 50-m² plots from Kapou and 125,000 annotation points for 125-m² plots from Hawai‘i Island).

2.3.2 Simulation and bootstrap resampling

The simulation and bootstrap resampling procedure were completed using the software R 4.3.0 (R Core Team, 2023) with the package collection tidyverse (Wickham et al., 2019). The numbers of photoquadrats used for the simulations were 10, 20, 30 and 40 photoquadrats per plot for Kapou and 10, 20, 40, 60, 80, 100 and 120 photoquadrats per plot for Hawai‘i Island. The numbers of annotations points were 10, 20, 50, 100, 200, 400, 600, 800 and 1,000 points per photoquadrat, for a total of 36 combinations of the numbers of photoquadrats and annotation points for Kapou and 63 combinations for Hawai‘i Island. We then repeated the subsetting process to generate 1,000 simulated datasets for each combination of the sampling design per plot (i.e. 36 × 1,000 simulated datasets for Kapou and 63 × 1,000 simulated datasets for Hawai‘i Island).

The metrics of accuracy and precision were calculated for each combination of the sampling designs for each of the 11 plots based on the 1,000 simulated datasets using 10,000 bootstrap resampling as described in Section 2.2. The metric of accuracy for a given sampling design of a given plot was the proportion of the 1,000 simulated datasets whose 95% bootstrap confidence intervals included the best estimate of live coral cover for the plot. The metric of precision was calculated as the median value of the ranges of the 1,000 bootstrap confidence intervals.

3 Results

The best estimate of live coral cover (%) for each plot based on all of the 1×1 m virtual photoquadrats (50 for Kapou and 125 for Hawai‘i Island) with 1,000 annotation points per photoquadrat ranged from 18% to 38% for Kapou and from 5% to 22% for Hawai‘i Island (Table 1). Variability in live coral cover among the photoquadrats within each plot varied across different plots, with one plot from Kapou (LIS_CW1) having a particularly large range from 0.8% to 84% (Supplementary Figure 1). Coefficient of variation (CV) was overall higher for plots from Hawai‘i Island ranging from 0.56 to 0.82, while those from Kapou ranged from 0.23 to 0.52 except for LIS_CW1 with the high among-photoquadrat variability, which had the value of 1.06 (Table 1).

For the simulations using plots from Kapou, a 95% bootstrap confidence interval could not be reliably established in some simulations for LIS_CW1 due to an excessive number of zero live coral cover estimates per photoquadrat for those simulating 10 photoquadrats with 10 or 20 annotation points per plot (Table 2). This issue was not observed with any other plots from Kapou. For the simulations with plots from Hawai‘i Island, the same issue was observed with some simulations mostly for scenarios simulating 10 photoquadrats with 10 annotation points, and they were more pronounced for plots with lower live coral cover (Table 2). The plots with >15% live coral cover had less than 15 simulations (out of 1,000) with the issue, while the plot with the lowest coral cover (5%, WH_S11_169) had more than half of the simulations (522) with the issue for 10 photoquadrats with 10 annotation points and 65 simulations even for 20 photoquadrats with 10 annotation points.

Accuracy of live coral cover estimate for Kapou was relatively high for any combinations of the numbers of photoquadrats and annotation points, being above ≈0.95 even when using only 10 photoquadrats to estimate coral cover (Figure 2a). This relatively high accuracy when using low numbers of photoquadrats was reflective of low (wide) precision obtained by establishing bootstrap confidence intervals. Overall, increasing the number of annotation points above 50 per photoquadrat had small effects on the improvement of both accuracy (Figure 2a) and precision (Figure 2b). Both accuracy and precision also improved as the number of photoquadrats increased. Using 30 photoquadrats, which corresponds to coverage of >50% of the total plot area, resulted in accuracy greater than 0.99 and precision (ranges of 95% confidence intervals for coral cover estimates) of <0.2.

Figure 2

Accuracy of live coral cover estimate for Hawai‘i Island showed patterns similar to those observed with plots from Kapou. It was relatively high for any combinations of the numbers of photoquadrats and annotation points, being above 0.94 even when using only 10 photoquadrats (Figure 3a). The high accuracy with 10 photoquadrats with 10 annotation points for the lowest coral cover plot (Figure 3a) was, however, due to the artifact from the extremely low precision (i.e. large 95% confidence interval, Figure 3b). Overall, increasing the number of annotation points above 50 per photoquadrat had small effects on the improvement of both accuracy (Figure 3a) and precision (Figure 3b). For the number of photoquadrats, there was a relatively small change in accuracy or precision with ≈60 photoquadrats, which corresponds to coverage of approximately 50% of the total plot area.

Figure 3

4 Discussion

Estimates of benthic cover are fundamental to assessing the condition of coral reefs, and the capacity to detect either spatial or temporal changes depends on the accuracy and precision of those estimates. Photogrammetry-based benthic surveys have been gaining popularity in the field of coral-reef ecology in the past decade due to their ability to create high-resolution and spatially accurate three-dimensional (3D) reconstructions of survey plots (e.g. Burns et al., 2015; Fukunaga et al., 2019, Pascoe et al., 2021, Urbina-Barreto et al., 2022, Ferreira et al., 2023). The use of resulting orthophotomosaics to assess benthic cover offers flexibility in the data extraction pipeline after field surveys as shown in the present study; virtual photoquadrats can be generated from an orthophotomosaic using a grid and randomly selected for benthic annotation. However, one of the findings of the present study highlights an important limiting factor that needs to be considered prior to a field survey. While the accuracy and precision of benthic estimates improve with an increasing number of annotation points per photoquadrat, these gains reach an asymptote relatively quickly (between 50 to 100 points for 1-m² photoquadrat; Figures 2, 3). Thus, the number of photoquadrats, and ultimately, the total area of study plot from which they are randomly sampled, are primary drivers of the maximum accuracy and precision that can be achieved through benthic annotation.

For the numbers of photoquadrats and annotation points per photoquadrat, low numbers lead to erroneous estimates of 0% live coral cover and failure to establish confidence intervals in some simulations (Table 2). This was particularly true for the combination of 10 photoquadrats and 10 annotation points per photoquadrat with low live coral cover (Table 2). This is consistent with previous studies demonstrating more sampling efforts being required for habitats with low benthic cover (Pante and Dustan, 2012). This pattern is also intuitive; when coral cover is low (e.g. under 10%), having only 10 annotation points increases the likelihood that all points fall on non-coral features and produce an estimate of 0% coral cover for that photoquadrat. Having fewer photoquadrats also increases the likelihood that all or nearly all photoquadrats yield coral cover estimates of 0%. However, while this seems to be largely dependent on the level of live coral cover, the plot LIS_CW1 from Kapou with 18% live coral cover collectively had more failed simulations than the WH_S02_026 plot from Hawai‘i Island with 15% live coral cover. More importantly, one simulation failed for LIS_CW1 even with 20 annotation points (Table 2), while none failed with 20 annotation points for the two plots from Hawai‘i Island with 14% and 15% live coral cover. This is likely due to the high variability in live coral cover in the LIS_CW1 plot as demonstrated by the high CV values (Table 1) and confirms the importance of considering the benthic heterogeneity where reef organisms are non-randomly distributed (e.g. crumped distribution). These findings highlight that benthic heterogeneity can be as influential as overall coral cover in determining the reliability of sampling-based estimates, underscoring the need to account for spatial variability when optimizing sampling strategies.

Levels of precision for live coral cover estimates established through bootstrap resampling were generally wider for those with higher live coral cover for the plots from Hawai‘i Island (Figure 3), but this was not the case for the plots from Kapou with the plot with the lowest coral cover (LIS_CW1) having the widest range (Figure 2). As scientists and managers are often interested in detecting percent changes in live coral cover, we further calculated, as a post-hoc exploratory analysis, the ratio of precision to mean coral cover estimates for Kapou (Figure 4) and Hawai‘i Island (Figure 5). Once the ratios were plotted, there were no clear associations between coral cover and precision for either the Kapou or Hawai‘i Island plots. However, CV values seem to offer some guidance (Figures 4, 5), with sites exhibiting lower CV values generally showing smaller ratios (e.g. LIS_R7 with 35% coral cover and CV of 0.23 and WH_S07_260 with 9% coral cover and CV of 0.56). Here, the conversion to the ratio also offers direct estimations of percent change in coral cover detectable given the number of photoquadrats from one year to another. For example, if a 25% increase or decrease needs to be detected, the ratio should ideally be below 0.5 (25% below and 25% above around the coral cover estimate). However, this would be a rough estimate since coral cover estimates from both years would have their own confidence intervals, and true detectability depends on the overlap between them. In our case, achieving this level of detectability required 30 and 60 photoquadrats for the Kapou and Hawai‘i Island plots, respectively, representing at least 50% of the total area of each orthophotomosaic with the exception of LIS_CW1, which had a high CV value of 1.06 (Figure 4, 18% coral cover). The lower number of photoquadrats required for Kapou likely reflect the overall lower variability in live coral cover across photoquadrats as observed by their CV values (except for LIS_CW1), again highlighting the importance of considering the benthic heterogeneity when developing a sampling design. These results demonstrate that the sampling effort required to detect meaningful changes in coral cover is strongly context dependent, reinforcing that sampling sufficiency cannot be assumed and must be explicitly evaluated based on site-specific variability. Unlike traditional power analyses, bootstrap resampling provides a non-parametric framework that derives power calculations directly from empirical data, thereby avoiding stringent distributional assumptions that often fail to characterize complex reef communities.

Figure 4

Figure 5

The bootstrapping approach presented here allows researchers to determine an optimal sampling design by selecting the number of photoquadrats based on their desired level of precision. For a plot-level estimate, photoquadrats are random samples from the population (i.e. plot), thus benthic metrics should be calculated for each of the photoquadrats, which in turn are used for bootstrapping. When the number of annotation points per photoquadrat is consistent, however, the point estimate of benthic metric (e.g. coral cover in our case study) would be the same by either calculating the benthic metric for each photoquadrat first then averaging them for the plot or pooling annotation points from all photoquadrats first then calculating the benthic metric using all annotation points for the plot (see Supplementary Material 2 for details). Incorrectly using the latter “pooled” approach in bootstrapping (i.e. randomly resampling pooled annotation points and calculating bootstrap estimates) results in considerably different accuracy and precision metrics (Supplementary Figure 2.1 & S2.2); the bootstrap resampling procedure produces much narrower confidence intervals than those obtained through the appropriate procedure and often fail to include the best estimates of coral cover, resulting in precise yet inaccurate confidence intervals. This highlights the importance of correctly identifying the resampling unit for bootstrapping to obtain appropriate representations of uncertainty in estimates of benthic metrics. More broadly, it demonstrates that resampling at the photoquadrat level versus pooling annotation points can yield substantially different estimates of accuracy and precision, with pooled approaches producing misleadingly precise but inaccurate results.

While we strongly encourage researchers to utilize the bootstrap resampling methods described here on a small subset of their own data to determine their optimal sampling designs, our results seem to support a simple “50–50 Rule” for reliably estimating benthic cover from orthophotomosaics. This approach recommends sampling at least 50% of the total plot area using 1-m² photoquadrats, each annotated with at least 50 points. The 50–50 Rule provides a practical balance between sampling effort and data accuracy, offering a starting point for designing efficient and statistically robust sampling strategies. Notably, our findings on the number of annotation points were consistent with a previous study showing no significant difference in benthic cover estimates when comparing 50 points to 450 points per photoquadrat (Carneiro et al., 2024). However, the total number of photoquadrats required remain sensitive to site-specific conditions; a previous study have found that while 7–10 quadrats were sufficient for some reefs (Carneiro et al., 2024), another study required as many as 38 to detect a 20% relative change in coral cover (Leujak and Ormond, 2007). Considering the variability in the distributions of organisms across sampling units, increasing the photoquadrat size may be beneficial in areas with high benthic heterogeneity, although this reduces the possible number of photoquadrats that can be placed within a plot of fixed size, making it important to consider plot dimensions carefully during the planning phase of photogrammetry surveys. The number of annotation points per photoquadrat also need to be considered in relation to photoquadrat size. An overall cost-benefit analysis that balances the total number of annotation points against the resulting precision should guide the determination of an optimal sampling design. Future research that systematically explores these trade-offs will further enhance the development of efficient, scalable and statistically robust sampling strategies for benthic habitat monitoring. Together, these considerations highlight that determining when enough is enough depends on balancing sampling effort, spatial scale and desired precision, with the 50–50 Rule serving as a practical starting point for achieving reliable estimates of benthic cover.

We employed the bootstrap-t method due to its accuracy and robustness with small sample sizes compared to the standard quantile method (Hesterberg, 2015). However, the bootstrap-t method can encounter issues with zero variance in resamples, as observed here, leading to a failure to establish a confidence interval (see Table 2). When optimizing sampling designs, such a failure serves as a diagnostic indicator that the number of photoquadrats or annotation points is insufficient for the specific benthic metric. Once an optimal sampling design is determined, bootstrap resampling remains a versatile tool for tracking and comparing benthic metrics, allowing researchers to generate plot-level confidence intervals. Non-overlapping confidence intervals between plots suggest statistically significant differences, and the Bias-Corrected and Accelerated (BCa) bootstrap method (Efron, 1987) offers a robust alternative should the bootstrap-t method fail to establish confidence intervals. The BCa method, implemented in R packages such as resample (Hesterberg, 2022), is particularly effective as it adjusts for both bias and skewness in small datasets. Furthermore, bootstrap resampling is easily adapted to simple random sampling designs. When multiple plots are randomly sampled within a site, the plots can serve as the resampling unit to derive site-level estimates. This flexibility underscores the utility of bootstrapping in ecological monitoring, provided resampling unit is correctly identified.

One of the advantages of photogrammetric benthic surveys is the capacity to utilize spatially rectified orthophotomosaics. Here, we generated photoquadrats for point annotation, but an alternative approach is to digitize benthic organisms by tracing polygons in geographic information system (GIS) software. Fully digitizing an orthophotomosaic allows for accurate quantification of planar benthic cover within a plot, provided that boundaries are delineated correctly, and can yield exact benthic cover values for temporal comparisons. Partial digitization can also be used to estimate benthic cover (Lechene et al., 2019), however, digitization is more labor intensive than point annotation and can be challenging when boundaries between benthic organisms are unclear due to overgrowth. In contrast, point annotation can be substantially accelerated using AI assisted platforms, particularly as training datasets expand and model performance improves over time (Beijbom et al., 2015, Williams et al., 2019). Continued advances in AI may further streamline or even automate digitization and identification, expanding the range of practical approaches for quantifying benthic cover. It is important for researchers and managers to leverage advances in photogrammetry, computational methods and AI to evaluate and refine sampling designs critically. Given that benthic cover estimates underpin assessments of reef conditions and inform management decisions, ensuring their accuracy and precision remains essential.

References

• 1

  BeijbomO.EdmundsP. J.RoelfsemaC.SmithJ.KlineD. I.NealB. P.et al. (2015). Towards automated annotation of benthic survey images: variability of human experts and operational modes of automation. PloS One10, e0130312. doi: 10.1371/journal.pone.0130312. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 2

  BrysonM.FerrariR.FigueiraW.PizarroO.MadinJ.WilliamsS.et al. (2017). Characterization of measurement errors using structure‐from‐motion and photogrammetry to measure marine habitat structural complexity. Ecol. Evol.7, 5669–5681. doi: 10.1002/ece3.3127. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 3

  BurnsJ. H. R.DelparteD.GatesR. D.TakabayashiM. (2015). Integrating structure-from-motion photogrammetry with geospatial software as a novel technique for quantifying 3D ecological characteristics of coral reefs. PeerJ3, e1077. doi: 10.7717/peerj.1077. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 4

  CarneiroI. M.SáJ. A.Chiroque-SolanoP. M.CardosoF. C.CastroG. M.SalomonP. S.et al. (2024). Precision and accuracy of common coral reef sampling protocols revisited with photogrammetry. Mar. Environ. Res.194, 106304. doi: 10.1016/j.marenvres.2023.106304. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 5

  EfronB. (1987). Better bootstrap confidence intervals. J. Am. Stat. Assoc.82, 171–185. doi: 10.1080/01621459.1987.10478410. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 6

  FerreiraS. B.BurnsJ. H. R.PascoeK. H.KaponoC. A.ReyesA. J.FukunagaA. (2023). Prediction of habitat complexity using a trait-based approach on coral reefs in Guam. Sci Rep. 13, 11095. doi: 10.1038/s41598-023-38138-1

  • CrossRef
  • Google Scholar

• 7

  FukunagaA.BurnsJ. H. R. (2020). Metrics of coral reef structural complexity extracted from 3D mesh models and Digital Elevation Models. Remote Sens.12, 2676. doi: 10.3390/rs12172676. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 8

  FukunagaA.BurnsJ. H. R.CraigB. K.KosakiR. K. (2019). Integrating three-dimensional benthic habitat characterization techniques into ecological monitoring of coral reefs. J. Mar. Sci. Eng.7, 27. doi: 10.3390/jmse7020027. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 9

  FukunagaA.BurnsJ. H. R.PascoeK. H.KosakiR. K. (2020). Associations between benthic cover and habitat complexity metrics obtained from 3D reconstruction of coral reefs at different resolutions. Remote Sens.12, 1011. doi: 10.3390/rs12061011. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 10

  HesterbergT. (2022). “ resample: resampling functions,” in R package version 0.6. Available online at: https://CRAN.R-project.org/package=resample (Accepted April 1, 2026).

  • Google Scholar

• 11

  HesterbergT. C. (2015). What teachers should know about the bootstrap: resampling in the undergraduate statistics curriculum. Am. Stat.69, 371–386. doi: 10.1080/00031305.2015.1089789. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 12

  HughesT. P.KerryJ. T.Álvarez-NoriegaM.Álvarez-RomeroJ. G.AndersonK. D.BairdA. H.et al. (2017). Global warming and recurrent mass bleaching of corals. Nature543, 373–377. doi: 10.1038/nature21707. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 13

  LecheneM. A. A.HaberstrohA. J.ByrneM.FigueiraW.FerrariR. (2019). Optimising sampling strategies in coral reefs using large-area mosaics. Remote Sens.11, 2907. doi: 10.3390/rs11242907. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 14

  LeujakW.OrmondR. F. G. (2007). Comparative accuracy and efficiency of six coral community survey methods. J. Exp. Mar. Biol. Ecol.351, 168–187. doi: 10.1016/j.jembe.2007.06.028. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 15

  MolloyP. P.EvansonM.NellasA. C.RistJ. L.MarcusJ. E.KoldeweyH. J.et al. (2013). How much sampling does it take to detect trends in coral-reef habitat using photoquadrat surveys? Aquat. Conserv.: Mar. Freshw. Ecosyst.23, 820–837. doi: 10.1002/aqc.2372. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 16

  NadonM.StirlingG. (2006). Field and simulation analyses of visual methods for sampling coral cover. Coral. Reefs.25, 177–185. doi: 10.1007/s00338-005-0074-5. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17

  PanteE.DustanP. (2012). Getting to the point: accuracy of point count in monitoring ecosystem change. J. Mar. Sci.2012, 802875. doi: 10.1155/2012/802875

  • CrossRef
  • Google Scholar

• 18

  PascoeK. H.FukunagaA.KosakiR. K.BurnsJ. H. R. (2021). 3D assessment of a coral reef at Lalo Atoll reveals varying responses of habitat metrics following a catastrophic hurricane. Sci Rep11, 12050. doi: 10.1038/s41598-021-91509-4

  • CrossRef
  • Google Scholar

• 19

  PerkinsN. R.FosterS. D.HillN. A.BarrettN. S. (2016). Image subsampling and point scoring approaches for large-scale marine benthic monitoring programs. Estuar. Coast. Shelf. Sci.176, 36–46. doi: 10.1016/j.ecss.2016.04.005. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 20

  PratchettM. S.HoeyA. S.WilsonS. K. (2014). Reef degradation and the loss of critical ecosystem goods and services provided by coral reef fishes. Curr. Opin. Environ. Sustain.7, 37–43. doi: 10.1016/j.cosust.2013.11.022. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 21

  R Core Team (2023). R: A language and environment for statistical computing (Vienna, Austria: R Foundation for Statistical Computing).

  • Google Scholar

• 22

  RogersC. S.MillerJ. (2006). Permanent 'phase shifts' or reversible declines in coral cover? Lack of recovery of two coral reefs in St. John, US Virgin Islands. Mar. Ecol. Prog. Ser.306, 103–114. doi: 10.3354/meps306103. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 23

  TaorminaB.MarzloffM. P.DesroyN.CaiseyX.DugornayO.Metral ThiesseE.et al. (2020). Optimizing image-based protocol to monitor macroepibenthic communities colonizing artificial structures. ICES. J. Mar. Sci.77, 835–845. doi: 10.1093/icesjms/fsz249. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 24

  TrygonisV.SiniM. (2012). photoQuad: A dedicated seabed image processing software, and a comparative error analysis of four photoquadrat methods. J. Exp. Mar. Biol. Ecol.424-425, 99–108. doi: 10.1016/j.jembe.2012.04.018. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 25

  Urbina-BarretoI.EliseS.GuilhaumonF.BruggemannJ. H.PinelR.KulbickiM.et al. (2022). Underwater photogrammetry reveals new links between coral reefscape traits and fishes that ensure key functions. Ecosphere13, e3934. doi: 10.1002/ecs2.3934. PMID:

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 26

  VallèsH.OxenfordH. A.HendersonA. (2019). Switching between standard coral reef benthic monitoring protocols is complicated: proof of concept. PeerJ7, e8167. doi: 10.7717/peerj.8167

  • CrossRef
  • Google Scholar

• 27

  WickhamH.AverickM.BryanJ.ChangW.McGowanL. D. A.FrançoisR.et al. (2019). Welcome to the tidyverse. J. Open Source Softw.4, 1686. doi: 10.21105/joss.01686

  • CrossRef
  • Google Scholar

• 28

  WilliamsI. D.CouchC. S.BeijbomO.OliverT. A.Vargas-AngelB.SchumacherB. D.et al. (2019). Leveraging automated image analysis tools to transform our capacity to assess status and trends of coral reefs. Front Mar Sci6, 222. doi: 10.3389/fmars.2019.00222

  • CrossRef
  • Google Scholar