Recurring bleaching events disrupt the spatial properties of coral reef benthic communities across scales

Marine heatwaves are causing recurring coral bleaching events on tropical reefs that are driving ecosystem change. Yet, little is known about how bleaching and subsequent coral mortality impacts the spatial properties of tropical seascapes, such as patterns of organism spatial clustering and heterogeneity across scales. Changes in these spatial properties can offer insight into ecosystem recovery potential following disturbance. Here we repeatedly quantified coral reef benthic spatial properties around the circumference of an uninhabited tropical island in the central Pacific over a 9‐year period that included a minor and subsequent severe marine heatwave. Benthic communities showed increased biotic homogenisation following both minor and mass bleaching, becoming more taxonomically similar with less diverse intra‐island community composition. Hard coral cover, which was highly spatially clustered around the island prior to bleaching, became less spatially clustered following minor bleaching and was indiscernible from a random distribution across all scales (100–2000 m) following mass bleaching. Interestingly, the reduced degree of hard coral cover spatial clustering was already evident by the onset of mass bleaching and before any dramatic wholesale loss in island‐mean coral cover occurred. Reductions in hard coral spatial clustering may therefore offer an early indication of the ecosystem becoming degraded prior to mass coral mortality. In contrast, the spatial clustering of competitive fleshy macroalgae remained unchanged through both bleaching events, while crustose coralline algae and fleshy turf algae became more spatially clustered at larger scales (200–700 m) following mass bleaching. Overall, benthic community spatial patterning became less predictable following bleaching and was no longer reflective of gradients in long‐term environmental drivers that typically structure these remote reefs. Our findings provide novel insights into how climate‐driven marine heatwaves can impact the spatial properties of coral reef communities over multiple scales.


Introduction
Disturbance can be a natural and beneficial ecosystem component, making resources such as light and space available for new recruits and enabling more species to coexist (Pickett & White, 1985).Yet, more frequent and intense disturbances can trigger abrupt system-wide changes (Burton et al., 2020;Hobbs & Huennneke, 1992).Widespread and persistent disturbance can promote the spatial expansion of "winners", often broadly adapted species that benefit from the resulting conditions, while "losers" can be lost from the area and native diversity reduced (McKinney & Lockwood, 1999).This can result in biotic homogenisation, and a decrease in the overall taxonomic diversity of ecological communities over time (Olden et al., 2004).Biotic homogenisation is documented across terrestrial and marine ecosystems, including forest plant communities following human land use change (Vellend et al., 2007), and fish assemblages following severe habitat loss (Richardson et al., 2018).
As the climate continues to change, ecosystem disturbance dynamics are also changing (Cheal et al., 2017;Coumou & Rahmstorf, 2012), including more frequent and longer-lasting marine heat waves that cause widespread coral bleaching on tropical reefs (Hughes, Kerry, et al., 2018).These extreme ocean temperatures break down the relationship between endosymbiotic algae and their coral host and can lead to reduced coral growth and mortality (Brown, 1997;Glynn, 1983).Significant coral cover reduction following disturbance can result in novel coral assemblages (Hughes, Anderson, et al., 2018) and dominance by non-reef-building benthic organisms such as fleshy macroalgae (Graham et al., 2015).Evidence suggests that coral reefs could persist into the future, but with novel compositions and ecosystem functions (Williams & Graham, 2019) and more homogenous benthic and fish communities (Bento et al., 2016;Darling et al., 2013;Edmunds, 2014;Hughes et al., 2007).
Changes to coral reef benthic communities following bleaching are well documented at organismal to ecosystem scales."Winners" and "losers" are driven by the biological characteristics of the coral hosts and their endosymbionts (Dilworth et al., 2021;van Woesik et al., 2011) and gradients in numerous abiotic factors (Evans et al., 2020;Robinson et al., 2019), including background temperature regimes (Safaie et al., 2018) and prior exposure to thermal stress (Ateweberhan & McClanahan, 2010).However, much less is known about how bleaching and subsequent coral mortality affect the spatial properties of tropical benthic communities across scales.We do know that the natural patterns of the spatial organisation of coral reef benthic communities are, in part, structured by long-term gradients in environmental drivers, including surface wave energy and upwelling (Aston et al., 2019;Ford, Gove, et al., 2021;Williams et al., 2013).However on many reefs, local and global human impacts such as climate change-induced ocean warming, coastal development, nutrient runoff and overfishing are now the dominant drivers of ecosystem state (Williams et al., 2019), making the influence of natural environmental factors hard to distinguish (Ford et al., 2020;Williams, Gove, et al., 2015).
Like many substrate-bound organisms including terrestrial plants (Plotkin et al., 2000;Réjou-Méchain et al., 2011), coral reef benthic communities are spatially heterogeneous.Their patterns of spatial organisation are often non-random across the seascape, with many individual benthic taxa showing evidence of spatial clustering at sub-metre to kilometre scales (Aston et al., 2019;Edwards et al., 2017;Ford, Gove, et al., 2021;Pedersen et al., 2019).Such heterogeneity is reputed to be an important driver of species diversity (Hautier et al., 2017), through the provision of prey refuges (Brusven & Rose, 1981) and increased available niche space (Stein et al., 2014).Quantifying and maintaining the natural spatial heterogeneity of an ecosystem is therefore often a target for conservation management (Wang et al., 2021).Importantly, disturbance events can temporarily disrupt patterns of biological clustering in natural ecosystems.For example, following severe wind disturbance, the degree of spatial clustering of trees in eastern North American forests was significantly increased at small scales (5-10 m), while clustering significantly decreased at larger scales (30-50 m) (Peterson, 2020).On tropical reefs, models suggest the degree of hard coral spatial clustering influences how they interact with neighbouring and competing benthic organisms, which in turn affects coral growth trajectories and ecosystem recovery potential following acute disturbance (Brito-Millán et al., 2019).
Here we quantified changes in the spatial properties of coral reef benthic taxa across scales over a 9-year period that included two marine heatwaves that resulted in bleaching and subsequent coral mortality.Our study system was Jarvis Island, an uninhabited tropical oceanic island and one of the world's most remote coral reef ecosystems.Our data were collected through towed-diver digital image surveys (Kenyon et al., 2006), yielding semicontinuous data across km-extents while preserving a high thematic resolution of the benthic communities (Ford, Gove, et al., 2021).These data allowed us to ask whether and how the spatial properties of coral reef benthic communities are impacted by coral bleaching across scales in the absence of confounding local human impacts.Specifically, we asked: (i) did benthic community composition and homogeneity change following bleaching?(ii) did the levels of spatial clustering of individual benthic organisms change across scales following bleaching?(iii) are the intra-island spatial patterns of coral reef benthic communities still reflective of the environmental factors that normally structure these remote reefs following bleaching?

Study site and disturbance history
Jarvis Island (0°22 0 S, 160°01 0 W) is a US-affiliated tropical island in the Line Islands Archipelago and forms part of the Pacific Islands Marine National Monument (Fig. 1).Jarvis has lacked a permanent human population throughout its geological history and is one of the most remote coral reef islands in the world.Previously labelled by the Ocean Health Index as the "healthiest" coral reef in the world (Halpern et al., 2012), Jarvis provides an opportunity to study the spatial ecology of coral reefs in the absence of confounding local human impacts (Ford et al., 2020;Williams, Gove, et al., 2015).Jarvis faces few threats aside from temperature stress; it has overall low coral disease prevalence (Brainard, Acoba, & Asher, 2019), low frequencies of extreme wave events caused by tropical storms (Gove et al., 2013), and a low abundance of coral predators such as the crown-of-thorns starfish (Acanthaster planci) (Brainard, Acoba, & Asher, 2019).
From late 2009 to early 2010, an El-Ni ño triggered ocean warming across the region (Vargas-Ángel et al., 2011;Williams, Knapp, et al., 2010) that resulted in minor coral bleaching prevalence ( $ 3%) at Jarvis (Barkley et al., 2018).According to the National Oceanographic and Atmospheric Administration's (NOAA) Coral Reef Watch data, this period reached a maximum of 11.7 Degree Heating Weeks (NOAA Coral Reef Watch, 2000).Degree Heating Week (DHW) is a common measure to indicate the accumulated thermal stress experienced by coral reefs, and this value exceeded the 8 DHW threshold expected to cause severe and widespread coral bleaching (Skirving et al., 2020).Jarvis then endured a second and more severe El-Ni ño triggered bleaching event that started in May 2015 and ended by May 2016, resulting in a 98% decline in mean live hard coral cover (Barkley et al., 2018;Brainard et al., 2018;Vargas-Ángel et al., 2019).NOAA's Coral Watch data showed temperature stress at a maximum of 29.1 DHW (NOAA Coral Reef Watch, 2000).

Benthic community digital surveys and spatial processing
Digital benthic images were taken around the circumference ( $ 13 km) of Jarvis' fore-reef habitat (reef slope facing the open ocean) by towed divers (following methods developed in Kenyon et al., 2006) between 2001 and 2017 as part of NOAA's Rapid Assessment and Monitoring Program (RAMP) and NOAA's National Coral Reef Monitoring Program (NCRMP).The diver, towed at $ 3 km h À1 along a targeted depth contour of 15 m, attempted to maintain a height of $ 1 m above the seafloor and captured images of the benthos at $ 15 m intervals using a downward-facing SLR camera (Canon EOS 10-D/50-D) fitted with strobes (Fig. 1).At this height, each image captured a mean area of 1.09 m 2 (Kenyon et al., 2006).The tow-board was fitted with a sub-surface temperature logger (SeaBird™ Electronics 39 subsurface pressure, temperature-depth recorder; 10 s sample rate, 0.002°C accuracy), and a GPS on the boat (combined with a layback algorithm) georeferenced each image.
From the available NOAA surveys, we post-processed images from 2008, 2010, 2015 and 2017 as they (i) provided the best spatial coverage around the island (each year had at least 70% coastline coverage with adequate digital image replication, see methods below), (ii) provided time points before and after minor and severe coral bleaching and (iii) included a 'baseline' survey year (2008) that represented $ 10 years since any previous mass bleaching events, allowing substantial time for benthic community recovery to a pre-disturbance state in the absence of local human impacts (Gilmour et al., 2013;Sheppard et al., 2008).The 2010 surveys occurred in April towards the end of the minor bleaching event.The surveys in 2015 also occurred in April, during the onset of mass coral bleaching but prior to any mass coral mortality (exact survey dates are detailed in Table S1).The 2017 survey provided a snapshot of benthic condition 11 months after the mass bleaching event in 2015-2016.
Although the benthic surveys targeted 15 m depth, some differences in depth did occur.We controlled for large differences in depth by selecting images within an 8-20 m depth range and every alternate benthic image was also selected to ensure continuity with previous work (Ford, Gove, et al., 2021).We used the analysis software CoralNet (Beijbom et al., 2015) to overlay 10 stratified random points over the entirety of each image (Fig. 1).The benthos under each point was then identified by a single analyst to ensure taxonomic consistency.Each hard coral, soft coral and macroalga was identified to a genus where possible, resulting in a total of 30 benthic groupings across all images (Table S2).These were later combined into eight higher-level benthic groups [hard coral, crustose coralline algae (CCA), calcifying macroalgae, soft coral, fleshy macroalgae, turf algae, other invertebrates and sand] (see Table S3 for group descriptions).Using ArcGIS and a custom Python script (Aston et al., 2019), we created 100-m wide grid cells around Jarvis' fore-reef and spatially joined them to the image-derived benthic cover data (Fig. 1).Where grid cells contained data from less than four images, we returned to the 'every alternate image' step in our workflow and, where possible, added additional data derived from images within the same grid cell.Following this, only grid cells containing at least four images in each of the survey years (70.5% of grid cells) were used in subsequent analyses (Ford, Gove, et al., 2021).

Quantifying benthic community change and heterogeneity over time
We tested for an effect of survey year (fixed factor and four levels) on benthic community structure, defined as the relative percentage cover of benthic organisms (n = 30, Table S2), using a permutational analysis of variance (PER-MANOVA) (Anderson, 2001) with the adonis function within the vegan package (Oksanen et al., 2019) for R version 4.0.2(R Core Team, 2020).We used island grid cells as replicates and analyses were based on square-root transformed data (to down-weight highly dominant benthic organisms), a Bray-Curtis similarity matrix, 999 permutations of the raw data, and type III (partial) sums-ofsquares.Where the global model was significant, we used the function adonis.pairwithin the EcolUtils package (Salazar, 2020) to identify differences between years.
To assess for changes in benthic community heterogeneity over time, we used multivariate dispersion (Anderson, 2006).Multivariate dispersion is a measure of group variance, here representing the relative variation in the cover of benthic genera (where possible) across island grid cells (individual replicates) within each survey year (group).We formally tested for differences in the mean distance to the centroid for each group using a Permutational Analysis of Multivariate Dispersions (PERMDISP) (Anderson et al., 2006) with the vegan functions betadisper and permutest.Relative changes in benthic communities over the 4 years were visualised using non-metric multidimensional scaling.To identify those benthic organisms most responsible for driving differences in community structure across years, we overlaid the correlations of the original benthic organisms with the ordination axes as a bi-plot using envfit (vectors scaled by square root of the r 2 ).
While each island grid cell contained benthic data from at least four images, the maximum number of images per grid cell differed within and between survey years.We, therefore, conducted a sensitivity analysis to assess what effect this might have on our benthic community analyses.For those grid cells containing more than four images, we randomly removed all but four of them and re-calculated the percentage cover of each benthic organism within each grid cell and repeated the above community analyses.The global effect of 'year' remained the same, regardless of the maximum number of images or permutation of four images, and we found no substantial changes in the relative patterns of biotic homogeneity (Table S4) or relative community dissimilarity between survey years (see Figure S1 for a more detailed description of the results from our sensitivity analysis).As a result, and to maximise the accuracy of the percentage cover estimates within any given grid cell, we did not constrain the upper limit of images within each grid cell during any of our subsequent analyses.

Quantifying benthic community spatial properties over time
We used the Moran's I statistic (Moran, 1950) to quantify the degree of spatial clustering across scales of the most abundant higher-level benthic groups (hard coral, turf algae, CCA and fleshy macroalgae; Tables S5-S8) using their percentage cover values over time.We did this in both directions around the island using a custom R function (Aston et al., 2019;Ford, Gove, et al., 2021).We defined the observed Moran's I value (OMI) as: where n is the number of observations, w i,j is the matrix of weights according to the inverse Euclidean distance between observations, x i is the observed value at location i, x j is the observed value at location j, x is the mean value and S 0 is the sum of spatial weights.The spatial weights are defined as the inverse of the minimum distance, d i,j , around the circumference of the island between locations i and j, as: for i≠j, 0 otherwise OMI values were calculated at the grid cell spatial resolution (100 m wide) and then again in a moving-window averaging process at increasing 100-m increments to a maximum of 2 km (limited by replication beyond that scale due to island size; Aston et al., 2019).At each scale we re-computed the Moran's I statistic and P-value for all possible 100 m grid cell starting locations of the moving window averaging process.This was iterated in both directions around the circumference of the island since spatial patterns in nature can be anisotropic.We report the range in OMI values for each scale from this process (shaded region in Figs.4A and 5) and the scale at which the upper bound of P exceeded 0.05 (i.e. did not significantly differ from a random spatial distribution).Significant OMI values away from zero indicated that the spatial pattern of the variable in question at that scale was highly organised in space.Significant positive OMI values indicated spatial clustering (positive autocorrelation).OMI values became negative (over-dispersed) at larger scales, however none of these values significantly differed from a random distribution for any of our higher-level benthic groups (indicated by the dashed horizontal lines in Figs.4A and 5) and were therefore not interpreted.

Quantifying environmental drivers
We quantified sub-surface seawater temperature and surface wave power at the grid cell resolution (100 m) to predict spatial variations in the cover of higher-level benthic groups over time.Previous work has shown these physical drivers to be spatially variable (Gove et al., 2006) and important in structuring Jarvis' benthic communities across scales (Aston et al., 2019;Ford, Gove, et al., 2021).
For example, sub-surface temperature gradients around Jarvis are a proxy for upwelling and seawater nutrient availability that positively correlate with hard coral cover (Aston et al., 2019).Furthermore, the upper spatial scales at which higher-level benthic groups show evidence of spatial clustering around Jarvis correlate with the upper scales at which surface wave power is spatially clustered (Ford, Gove, et al., 2021).We calculated mean subsurface temperature per grid cell using data from the tow-board temperature-depth recorders collected during that survey year, as well as the two prior survey years closest in the timeseries with adequate spatial coverage (to capture longer-term spatial patterning in temperature gradients) (Figure S2).Despite these temperature data representing a temporal snapshot, they capture the longer-term intra-island gradients in sub-surface temperature around Jarvis (Aston et al., 2019, see Figure S2).Furthermore, these data allowed for estimates of spatially explicit sub-surface temperature variation within each 100-m grid cell.The limited number (n = 5) of permanently deployed seabed-mounted temperature loggers around Jarvis within our study depth range lacked this spatial resolution (Aston et al., 2019).
To calculate surface wave power, data were extracted from NOAA's Wave Watch III 50-km resolution global model (http://polar.ncep.noaa.gov/waves).From the centre point of each 100-m grid cell, we created a 360°plot of lines (each line set to 100-km length) at 1°bearing intervals and used a vectorised shoreline of Jarvis and the surrounding islands to identify where each line intersected land, either on the island itself or a neighbouring island.Where this occurred, that degree bin was removed, leaving only bearings open to exposure along the incident wave swath.For each of these exposed degree bins, wave power and its corresponding direction were selected at each time-step for the closest Wave Watch III pixel to the centroid of Jarvis; only waves that travelled along an open bearing were retained.Wave power (WP in W m À1 ) was calculated from significant wave height (H s ) and peak period (T p ), which we extracted from Wave Watch III and defined as: where q is the density of seawater (1024 kg m À3 ) and g is the acceleration of gravity (9.8 m s À2 ).For high energy wave events, a key driver of benthic community structure (Gove et al., 2015), we calculated the mean of the top 2.5% of daily wave power values for each survey year (i.e., 9 per year) and the prior 3 years closest in the timeseries for each grid cell.We averaged across these 4 years to reflect the long-term spatial patterning in these anomalies around Jarvis for each survey year (Figure S2).

Quantifying the predictability of benthic communities following bleaching
To assess for changes in the ability to predict benthic community patterns from gradients in environmental drivers following bleaching, we modelled changes in CCA and turf algae cover.These benthic organisms were modelled (and not hard coral or fleshy macroalgae) due to their high abundance across the island in all survey years.This ensured that any changes in their predictability were a result of their driver-response relationship breaking down, rather than a loss of spatial replication leading to poor model performance.Here we focused on the changes in benthic organism spatial predictability over time rather than the nature and directionality of the biophysical relationships themselves, which have been extensively quantified for Jarvis in previous work (Aston et al., 2019).We used boosted regression trees (BRTs; Breiman, 1984;Elith et al., 2008) to model the relationships between gradients in CCA and turf algae cover and our two environmental drivers, building a separate model for each survey year.We also added a categorical predictor corresponding to the NOAA-defined geo-regions around Jarvis (Brainard, Acoba, Asher, et al., 2019) to account for any influence of other biophysical variables that we did not measure (Figure S2).BRTs are a non-linear regression technique that combine large numbers of simple regression trees by sequentially fitting each new tree to the residuals from the previous one to improve overall predictive performance (Buston & Elith, 2011).Prior to model building, we assessed the collinearity in our continuous predictors using Pearson's correlation coefficient (r).Since r ranged from 0.31 to 0.06 across the survey years, both environmental drivers were included in all models (Dormann et al., 2013).BRTs were constructed using the package gbm (Rigdeway, 2015) using a Gaussian distribution for the response variables.
BRT model performance was quantified using 10-fold cross-validation.This involved splitting the data into 10 training sets (nine unique subsets of the data used for training, one omitted and used for testing) (Elith et al., 2008).A BRT model is built on each training set and in each case model performance is tested with the remaining omitted data (Elith et al., 2008).Mean model performance is reported as the cross-validated percentage deviance explained [1 -(cross-validated deviance/mean total deviance)] and was calculated using the R package ggBRT (Jouffray et al., 2019).We optimised model performance by running combinations of the tree complexity (1, 2, 3, 4, 5), learning rate (0.001, 0.0001 and 0.00001) and bag fraction (0.1, 0.5 and 0.75) parameters, and identified the combination that resulted in the lowest cross-validation deviance for each individual BRT model (Williams, Aeby, et al., 2010) (see Table S9 for optimal settings used).
To further explore changes in the ability to predict benthic community patterns from gradients in environmental drivers following bleaching, we also modelled the relationships between the percentage cover of all 30 benthic variables, sub-surface seawater temperature and surface wave power using permutational distance-based multivariate multiple regression (McArdle & Anderson, 2001).Models were constructed using square-root transformed benthic cover data and a Bray-Curtis similarity matrix using the PERMANOVA+ add-on (Anderson et al., 2008) for PRIMER v.7 (Clarke & Gorley, 2015).The environmental predictors were normalised to account for their variations in units and ranges.All possible candidate models (unique combinations of the predictors) were computed, and we report the percentage variation in the response variable explained by the top candidate model for each survey year.All data and R code (where applicable) for the core analyses in this paper can be accessed from Dryad: https://doi.org/10.5061/dryad.xsj3tx9mt(Ford et al., 2023).

Changes in benthic community structure and heterogeneity over time
The relative cover of benthic organisms changed following both minor and mass bleaching (Fig. 2A).Island-mean hard coral cover (all hard coral genera combined) declined from 29.2% in 2008 (AE2.8SE) to 22.8% (AE2.2) in 2010 following minor bleaching.During the reprieve from disturbance between 2010 and 2015, island-mean hard coral cover increased to 27.0% (AE1.5) (Fig. 2A).During this entire period, hard coral communities were dominated by Montipora and Pocillopora (Fig. 2B; Tables S5-S7).CCA were also very abundant with an island-mean cover of 33.8% in 2008 that increased to 50.5% in 2010 following minor bleaching.Following mass bleaching, island-mean hard coral cover dropped to just 0.3% (AE0.1) by 2017, while there was a notable increase in turf algae and fleshy macroalgae (Fig. 2A).CCA also continued to dominate much of the seascape in 2017 (39.6%) following mass bleaching (Fig. 2A).Following mass bleaching there was a marked change in genera dominating the coral communities.The previously dominant Montipora gave way to Pocillopora and Porites and to a lesser extent Favites, Favia and Turbinaria dominating the surviving communities (Fig. 2B; Tables S5-S8).
Benthic community structure changed significantly over time following the two bleaching events (Pseudo-F 3,335 = 49.5, P = 0.001, Fig. 3A), with communities significantly different following minor bleaching  Prior to the bleaching events, the benthos was characterised by a higher cover of the hard corals Montipora, Pocillopora and Acropora and the soft coral Sinularia, than in the subsequent years (Fig. 3B).Following minor bleaching in 2010 and by the onset of mass bleaching in 2015, benthic communities were characterised by CCA and the fleshy macroalga Lobophora, the hard coral Favites and the fire coral Millepora.Benthic communities were markedly different again following mass bleaching (Fig. 3A) and were characterised by a higher cover of turf algae, the fleshy macroalga Galaxuara, Dictyosphaeria and Lobophora, cyanobacteria, and the encrusting calcified macroalga Peyssonelia (Fig. 3B).
Benthic communities also displayed increased homogeneity (had lower multivariate dispersion) over time (Fig. 3A).Multivariate dispersion was significantly lower following minor bleaching (2008 vs. 2010, P = 0.024) and then remained stable during the onset of mass bleaching (2010 vs. 2015, P = 0.883).The communities then became more homogenous again following mass bleaching (2015 vs. 2017, P = 0.044).Overall, the multivariate dispersion was significantly lower following the two bleaching events than at the start of the timeseries (2008 vs. 2017, P = 0.001).This pattern held when standardising the number of benthic images per grid cell (by limiting to four) around the island (Table S4).

Changes in benthic community spatial properties over time
Hard coral cover displayed similar patterns of spatial structure in 2008 and following minor bleaching in 2010, with cover significantly clustered up to 1.2 and 1.1 km scales, respectively (Fig. 4A).By the onset of mass bleaching in 2015, hard coral cover was only significantly clustered up to 0.5 km scales and the degree of clustering (mean OMI value) had almost halved (Fig. 4A).This appeared to be driven by a reduction in cover along the island's west coast, and localised increases in other locations, resulting in a more even distribution of coral cover around the island compared with earlier years (Fig. 4B).By 2017, and following mass bleaching, hard coral cover was no longer significantly spatially structured around the island and was indiscernible from a random distribution at all scales (Fig. 4A).
Like hard coral, CCA showed consistent patterns of spatial structure in 2008 and following minor bleaching in 2010 and was significantly clustered up to 0.9 and 1.0 km scales (Fig. 5A; Figure S3).By the onset of mass bleaching in 2015, the scale at which CCA cover was clustered dropped to 0.5 km and the overall degree of clustering declined (Fig. 5A).Similarly, the degree of spatial clustering of turf algae declined over time through to the onset of mass bleaching (Fig. 5B; Figure S4).Notably, following mass bleaching, both CCA and turf algae became more spatially clustered at larger scales (200-700 m) (Fig. 5A,B).In contrast, the spatial structure of fleshy macroalgae changed the least over time and was significantly clustered up to scales of $ 0.5 km across all survey years, before and after the bleaching events (Fig. 5C; Figure S5).

The predictability of benthic communities following bleaching
The ability of island geo-region, sub-surface seawater temperature and surface wave power to explain intra-island spatial variation in coral reef benthic communities decreased between the beginning (2008) and end of the timeseries (2017).In 2008 prior to bleaching, crossvalidated percentage deviance explained equalled 55.7% and 53.0% for CCA and turf algae cover, respectively.By 2017 and following the two bleaching events, it equalled 14.4% and 8.0% for CCA and turf algae, respectively (see Table S9 for detailed model summary outputs and relative influence of the predictor variables by survey year and see Figures S6 and S7 for the response-predictor relationships by survey year).This loss of predictability in turf algae and CCA cover around the island was mirrored when considering all benthic variables simultaneously.In 2008, 24.9% of the benthic community variation across grid cells was explained, but this dropped to 15.3% by 2017 (Table S10).

Discussion
Using a remote coral reef free from local human impacts, we show that repeated coral bleaching can dramatically reduce the cover of calcifying benthic organisms, homogenise benthic communities and disrupt their natural patterns of spatial organisation across the seascape.This results in a breakdown of the relationship between spatial patterns in the cover of benthic organisms and concurrent gradients in natural environmental drivers.Subsequently, the drivers no longer provide predictive power over the island's benthic seascape ecology, at least in the shortterm or perhaps under frequent bleaching disturbance.
Our results suggest that coral bleaching can disrupt the natural spatial properties of coral reefs, detected through expansive sampling that would otherwise be missed by more spatially limited observations.

Bleaching alters benthic community composition at a remote reef
Our study showed that even the most remote and protected coral reefs can be vulnerable to climate change.Despite being labelled as the world's healthiest marine ecosystem in 2012 (Halpern et al., 2012), mean hard coral cover around Jarvis Island dropped from 27.0% to 0.3% following mass coral bleaching.This exceptional amount of coral mortality was due to sustained high temperature stress and a temporary reduction in primary production, caused by reduced coastal upwelling, which otherwise might have provided corals with energetic subsidies (Brainard et al., 2018).Following extensive coral mortality, the benthos became characterised by non-reef-building benthic organisms in comparison with the pre-bleaching baseline, notably fleshy turf algae, the fleshy macroalgae Lobophora and Dictyosphaeria, the calcified macroalga Galaxaura, and cyanobacteria.Turf algal assemblages are highly competitive with corals (Barott et al., 2012), but also provide substrate for other macroalgae to grow and can inhibit coral recovery (Nieder et al., 2019).The fast-growing, mat-forming cyanobacteria can smother and outcompete other benthic organisms following acute disturbance, while their low palatability implies a lack of control by herbivory (Ford, Visser, et al., 2021).Benthic dominance by fleshy organisms is typically characteristic of reefs chronically impacted by local human impacts (Smith et al., 2016) but can also result from acute heat stress and mass bleaching (Graham et al., 2015) even in the absence of local human impacts (Gilmour et al., 2013).Mass bleaching at Jarvis also resulted in altered hard coral assemblages.Prior to mass bleaching and even following minor bleaching, coral assemblages were dominated by Montipora, with this single genus averaging 20.6% of the benthos around Jarvis between 2008 and 2015.Montipora subsequently suffered extreme (near 100%) mortality following mass bleaching, confirming previous observations at Jarvis over more spatially limited extents (Vargas-Ángel et al., 2019).Overall, our results and those of Vargas-Ángel et al. ( 2019) indicate that such high dominance by a single susceptible coral genus leaves a reef vulnerable to spatially extensive coral mortality following severe heat stress.Following mass bleaching, the surviving corals were from the genera Porites, Favia, Favites and Pocillopora.These corals likely represented survivors from the mass bleaching event since poritid and faviid corals are stress-tolerant genera (Donner et al., 2010;Kayal et al., 2015) and survived mass bleaching in the Persian Gulf and East China Sea (Bento et al., 2016;van Woesik et al., 2011).Different species of Pocillopora can vary in survival from bleaching events, but the genus as a whole shows resilience (Burgess et al., 2021).This altered hard coral composition at Jarvis could persist and cause the emergence of increasingly reported novel hard coral assemblages following mass bleaching (Hughes, Anderson, et al., 2018;Raj et al., 2021).

Bleaching homogenises coral reef benthic communities
Ecological communities can become more biotically heterogenous with increased species diversity following various types of disturbances (Gerwing et al., 2018;Odum, 1985).Here we found that benthic communities exhibited biotic homogenisation and became more spatially similar around Jarvis following both minor and mass bleaching disturbances.Similarly, models predicted the homogenisation of coral communities across the remote Chagos Archipelago in the Indian Ocean under realistic climate change disturbance scenarios (Riegl  et al., 2012).At Jarvis, the biotic homogenisation of benthic communities appears to be the result of low coral survivorship following mass bleaching and a spatially consistent increase in turf algae cover leading to a more homogenous seascape.Jarvis' abundant herbivorous fish populations (Williams, Baum, et al., 2015) could have maintained the spatial dominance of early successional turf algae through top-down control of fleshy macroalgae.Jarvis' remote location and paucity of larval supply (Maragos et al., 2008) could have inhibited coral recovery despite abundant CCA following mass bleaching that facilitates coral recruitment (Price, 2010).Our data do not span a sufficient post-disturbance period to know whether Jarvis' benthic community will remain in a homogenised state.However by 2018, less spatially expansive surveys showed the post-bleaching dominance by fleshy macroalgae and turf algae was only temporary; these were rapidly outcompeted by calcifying CCA and macroalgae from the genus Halimeda (Huntington et al., 2022).
It is therefore likely that the spatially expansive homogenisation of the benthic communities we document here is a temporary phenomenon at this remote reef.
Bleaching events disrupt the spatial properties of benthic communities across scales Increased community diversity and spatial heterogeneity following disturbance, even following severe disturbance, has generally been the rule rather than the exception in landscape ecology studies (Turner, 2010).However, there are some exceptions, like severe wind disturbance leading to decreased spatial clustering of trees across North American forest landscapes (Peterson, 2020).Here we find the same phenomenon in tropical marine benthic communitiesrecurrent bleaching resulted in the loss of significant spatial structure in hard coral cover around the island.The spatial clustering evident prior to mass bleaching broke down, resulting in a spatially random distribution of surviving coral cover across all scales examined.The reduced degree of hard coral cover spatial clustering was already evident during the onset of mass bleaching in 2015 and before any dramatic wholesale loss in island-mean coral cover at Jarvis.This may reflect residual impacts of the more minor marine heatwave that occurred 5 years earlier (Vargas-Ángel et al., 2011).Reductions in hard coral spatial clustering may therefore offer an indicator of the ecosystem already becoming degraded.Decreased numbers of large patches of vegetation and increased spatial homogeneity in habitat across landscapes are considered indicative of degraded systems within grasslands (Meyers et al., 2014), intertidal mudflats (Weerman et al., 2012) and arid ecosystems (Kefi et al., 2007).A more random distribution of hard coral cover following mass bleaching on coral reefs could have ecosystem recovery implications, since less aggregated coral communities recover to pre-disturbance states less effectively under modelled scenarios (Brito-Millán et al., 2019).
Like hard coral, the cover of CCA and turf algae generally became less spatially clustered following bleaching, particularly at smaller (100-m) scales.CCA and turf algae are naturally spatially clustered around Jarvis due to gradients in key environmental drivers that themselves exhibit strong degrees of spatial clustering (Ford, Gove, et al., 2021) (Figure S2).The reduction in the degree of spatial clustering and increasingly random distribution of CCA and turf algae following bleaching may reflect their opportunistic colonisation of newly available substrate in areas of localised high coral mortality.In contrast to hard coral cover, however, both CCA and turf algae became more spatially clustered at larger scales (200-700 m) following mass bleaching and this may reflect the occurrence of larger patches with a consistent amount of cover (Figures S3 and S4).In contrast to the other benthic groups, fleshy macroalgae around Jarvis remained spatially clustered up to $ 500-m scales following both bleaching events, despite an overall island-mean increase in cover following mass bleaching.The persistent spatial clustering of fleshy macroalgae both before and after bleaching disturbance is intriguing and warrants further investigation as to the driving forces behind their stable spatial properties.

Bleaching events disrupt the spatial predictability of coral reef benthic communities
The ability of long-term environmental drivers to effectively predict spatial variations in coral reef ecosystem state appears to be compromised under chronic local human impacts (Ford et al., 2020;Williams, Gove, et al., 2015).Here we show this concept extends to acute disturbance by anthropogenic climate change.Following bleaching, benthic organism spatial distributions around Jarvis became substantially less predictable using concurrent gradients in environmental drivers, despite the intraisland gradients in these environmental variables remaining relatively consistent across survey years (Figure S2).Bleaching-induced coral mortality, and the subsequent succession of newly available substrate, therefore, appears capable of disrupting the spatial structure of entire benthic communities across the seascape, meaning their distribution patterns no longer reflect long-term environmental forcing.In tropical rainforests, disturbance intensity predicted forest structure and diversity better than long-term environmental variables (Ding et al., 2012).Similarly, in grassland communities, stochasticity in community succession and loss in predictability by abiotic conditions was found after an extreme climatic event (Kreyling et al., 2011).Our findings also align with successional theory, that after large infrequent disturbances and when the recovering area is far from propagule sources, community composition is less predictable, owing to more stochastic processes driving their spatial structure (Turner et al., 1998).The degree to which a disturbance homogenises a community and reduces the predictability of communities is likely linked to the severity of the disturbance.As in terrestrial ecosystems, the loss in the spatial predictability of tropical benthic communities may only be temporary.As Jarvis' benthic community begins to recover (Huntington et al., 2022), niche-based processes may act to structure their spatial distributions (assuming differences among benthic organisms), becoming reflective once again of their surrounding long-term environmental setting (Mežaka et al., 2021).

Conclusion
Here we show that coral bleaching can disrupt the natural spatial properties of benthic communities on a tropical coral reef.Two marine heatwaves over a 9-year period resulted in a minor and a severe mass coral bleaching event at one of the world's most remote and protected reef ecosystems, the latter of which caused extensive coral mortality.Following each bleaching event, benthic communities became more taxonomically homogenous and key reefbuilding benthic organisms like hard corals lost their natural patterns of spatial clustering and became more randomly distributed across scales.As a result of these disruptions to their natural spatial ecology, benthic communities became less spatially predictable and were no longer reflective of gradients in long-term environmental drivers.Instead, their spatial patterns reflected intense and acute disturbance.Our data provide empirical evidence that global warming-induced marine heatwaves are capable of fundamentally re-structuring the spatial properties of tropical coral reef benthic communities across scales.set with a minimum of four images per grid cell but no upper limit (as in Fig. 3).Tables S5-S8.Summary statistics for all benthic organism genera (where possible) at the 100-m grid cell spatial resolution for 2008 (baseline year), 2010 (post-minor bleaching), 2015 (onset mass bleaching) and 2017 (post-mass bleaching).Table S9.Boosted regression tree model performance statistics from predicting the spatial distribution of turf algae and crustose coralline algae (CCA) cover around Jarvis Island based on concurrent gradients in sub-surface seawater temperature and surface wave power.Table S10.Results of permutational distance-based multivariate multiple regression models for each survey year.Island (by grid cell).Categorical 'geo-region' predictor included in our Boosted Regression Tree models also shown (remained constant across survey years).Figure S3.Spatial variation in crustose coralline algae (CCA) cover around the circumference of the fore reef habitat of Jarvis Island (central grey shape) over time.Figure S4.Spatial variation in turf algae cover around the circumference of the fore reef habitat of Jarvis Island (central grey shape) over time.Figure S5.Spatial variation in fleshy macroalgae cover around the circumference of the fore reef habitat of Jarvis Island (central grey shape) over time.Figure S6.The partial contribution of each predictor variable on the response variable (percentage cover of crustose coralline algae, CCA) with temperature shown in the left plot for each year and wave power on the right.Figure S7.The partial contribution of each predictor variable on the response variable (percentage cover of turf algae) with temperature shown in the left plot for each year and wave power on the right.

40 ª 2023
The Authors.Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.

Figure 1 .
Figure 1.Data collection, processing steps, and analytical pipeline for quantifying changes in coral reef benthic community composition and heterogeneity, spatial structure across scales, and spatial predictability following recurring coral bleaching events at Jarvis Island, central Pacific Ocean.

ª
2023 The Authors.Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.

44 ª 2023
The Authors.Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.

Figure 2 .
Figure 2. Variation in percentage cover values of the five most dominant benthic groups (A) and the five most dominant hard coral genera (B) before and after two coral bleaching events (one minor and one severe) at ∼ 15 m depth around the ∼ 13 linear km of Jarvis Island's fore-reef habitat at a spatial resolution of 100 m.Dots here represent outliers and boxes show the interquartile range and their middle lines represent median values.CCA, crustose coralline algae.

Figure 3 .Figure 4 .
Figure 3. Benthic community change over time around Jarvis Island.(A) Differences in community structure grouped and coloured by year with each point at the 100-m grid-cell resolution (n = 84 for each year).Shaded circles are the 95% confidence intervals around each year's group centroid (stress = 0.16).Multivariate dispersion (mean distance to centroid) for each year equals 0.2403 (2008), 0.2051 (2010), 0.2031 (2015) and 0.1809 (2017).(B) Benthic groups most responsible for driving differences among years.Vector lines in the bi-plot represent correlations (calculated by square root of r 2 ), with the direction indicating the relationship of each benthic group to the first two ordination axes.The length of each vector line is proportional to the strength of the correlation.Benthic organisms are resolved to genus level for some functional groups including calcifying macroalgae (e.g., Peyssonnelia and Galaxaura) fleshy macroalgae (e.g., Lobophora and Dictyosphaeria), hard corals (e.g., Pavona, Favites, Pocillopora and Montipora) and soft corals (e.g., Sinularia).Others are grouped to a less resolved taxonomic resolution, including fire corals (Millepora), Sand, Turf (turf algae) and CCA (crustose coralline algae).See TableS2for a detailed list of the benthic organisms included in the analysis.

Figure 5 .
Figure5.Changes in the spatial structure (solid horizontal lines) of (A) crustose coralline algae (CCA), (B) turf algae and (C) fleshy macroalgae cover around Jarvis Island over time at increasing 100-m scale increments (starting at 100-m).An Observed Moran's I value away from zero indicates that the spatial pattern of coral cover at that scale is highly organised in space, with positive values indicating a clustered distribution.At each scale we re-computed the range in Moran's I values for all possible 100-m grid cell starting locations iterating in both directions around the circumference of the island (shown as the shaded region around the solid horizontal lines).The vertical dotted lines indicate the upper scale at which there was no longer significant spatial structure in cover.Above this scale, the horizontal Moran's I values are dotted to indicate they do not differ from zero (i.e., a random distribution).Despite negative Moran's I values that would indicate overdispersion, none of these values differed from zero and are therefore not interpreted.

ª
2023 The Authors.Remote Sensing in Ecology and Conservation published by John Wiley & Sons Ltd on behalf of Zoological Society of London.

Figure S1 .
Differences between years in benthic community composition and biotic heterogeneity using only four images per grid cell across all survey years.
Figure S2.(a) Wave power (average of wave anomalies) and (b) Temperature (average of sub-surface seawater temperature from years) for each survey year (in bold) around Jarvis