Degraded Coral Reefs are Becoming More Resistant to Hurricanes


 The frequency and intensity of Atlantic cyclonic storms are projected to increase as climate change warms the ocean 1,2. These changing disturbance dynamics, paired with simultaneous changes in the condition and composition of Caribbean coral reefs, could be altering reef resilience to storms in unexpected ways. For example, the observed decline of fast-growing, disturbance-sensitive species could promote resistance to and decrease recovery from storms3,4, increasingly locking reefs into a state dominated by weedy taxa. To test this hypothesis, we combined data from coral reef monitoring studies and historical hurricane records to develop a regional reef-storm interaction database. We found that as the living cover of Caribbean corals declined over the past 40 years, while resistance to storms increased, despite a concurrent increase in cyclonic storm frequency and intensity. Because storms selectively damaged branching coral species and had no measurable effect on the cover of “weedy” corals, reef composition shifted towards greater weedy dominance and reduced ecological function. Additionally, storms accelerated the loss rate of threatened acroporid corals, already in pre-storm decline, suggesting a worrisome synergism with other climate-related disturbances.

The purpose of this study was to measure the state-dependent effect of storms on Caribbean reef corals to better understand the causes and implications of changing coral reef resilience to storms. We compiled a coral reef-storm dataset that catalogues the historical record of storm crossings on coral reefs in the greater Caribbean and Gulf of Mexico (Fig. 1A). This database consists of 11,490 quantitative benthic surveys from 3,144 unique reef locations performed between 1970-2017. Survey data was compiled from various coral reef monitoring programs and the literature (Extended Data Table 1) and each survey reported the absolute living coral cover as a percentage of the benthic composition. A subset of the data sources also reported the percent cover of individual coral species (Extended Data Table 1), which we aggregated to determine the percent cover of three ecological functional groups (weedy, competitive, and stress tolerant, see Extended Data Table 2) across 672 unique reef locations. We then combined these coral community data with the tracks of Atlantic storms (tropical storm strength or stronger) sourced from the National Oceanic and Atmospheric Administration (NOAA)'s HURDAT2 dataset 12 . For the 47 years for which we had coral survey data there were 547 storms, 152 of which crossed at least one surveyed reef, resulting in 10,058 unique storm-reef intersections. When coral reef survey data was available for pre-and post-storm years at a site (1 year prior and 1 year post), we quanti ed the effect of storms on living coral cover and community structure, and their dependency on pre-impact reef state. Speci cally, we determined whether pre-storm living coral cover and community composition affected resistance to and recovery from storm damages.
We found that total living coral cover has steadily declined ( Fig. 1B) since prior regional assessments 5,6 that suggested coral loss may have plateaued by the late 1990s, at a regional mean of ~16%. However, our results indicate that from 1997 -2017, the median annual loss rate was ~0.25% per year, with a nal year regional mean of 9.5% +/-0.59% (SE). The more recent regional effects of stony coral tissue loss disease 11 , including reports of severe coral loss across the Caribbean 13 suggest further decline since the nal year of our study. Despite substantial sub-regional variation (Fig 1.B), coral cover continues to decline even on reefs with relatively high coral cover. The current consensus is that this pattern is driven largely by coral bleaching, disease, and other disturbances, with local stressors including shing and pollution playing a role at some locations 14 .
Our results also indicate that the frequency and intensity of Atlantic storms has increased (Figs. 1C & D, Extended Data Figure 1). Despite more, stronger storms, the immediate effect of storms on total living coral cover decreased markedly between the 1980s/1990s and the 2000s/2010s ( Fig. 2A). Storms now have no general effect on total living coral cover. In contrast, in the 1980s, the median short-term loss was -8.5% (for absolute total coral cover, n=10 impacted reefs, maximum loss = 27%). However, Gardner et al. 3 found that Caribbean hurricanes between 1980 and 2001 on average caused a 17% reduction in live coral cover (n=177 impacted reef sites).
One obvious difference is that pre-impact living coral cover has been substantially reduced. Our results indicate that reefs with greater pre-impact cover were less resistant (i.e., reefs lost more cover when other factors are controlled for Extended Data Figure 2, Extended Data Table 3). Numerous other studies have also reported this state-dependent resilience to storms 7,15 , and for other large-scale disturbances including disease outbreaks and bleaching events [16][17][18] . Although this phenomenon is often attributed to simply having more coral to lose, it could also be caused at least in part by the presence of disturbance sensitive species (which historically tended to dominate high-cover reefs 14,19,20 ); particularly coral species in the genus Acropora. Due to their fast growth and branching or plating morphology, these species are considered competitively dominant over other sessile reef organisms, including other coral species. Once established, acroporid corals can quickly monopolize incoming light and space on the sea oor -two critical resources for many sessile reef organisms.
Only four Caribbean coral species are considered "competitive dominants" 21 , three of which are in the genus Acropora (cervicornis, palmata, and their sterile hybrid prolifera). In U.S. waters, Acropora species are also listed under the Endangered Species Act as threatened. These species dominated Caribbean reefs until the early 1980s, when they were nearly extirpated by white band disease -a regional epizootic linked to ocean warming 7,8 . They are particularly sensitive not only to storms and other forms of physical disturbance (due to their morphology), but also to disease, predator outbreaks, and other environmental perturbations 4,8,21 . The result of these vulnerabilities means that Acropora-dominated reefs are less resistant to disturbance, despite greater apparent "health" 9,10 .
We found that storms drove a change in functional group composition of coral communities on Caribbean reefs (Fig. 2B). In general, acroporid and stress tolerant corals were negatively impacted by storms (average loss for competitive species -0.4% ± 0.12%, Wilcoxon p =0.009; average loss for stress tolerant species -0.66% ± 0.22%, Wilcoxon p = 0.007, also see Extended Data Table 4) causing the relative abundance of coral species categorized as "weedy" to increase.
The changing disturbance regime (e.g., more frequent marine heatwaves, bleaching events, storms, and disease outbreaks) on Caribbean reefs appears to have shifted reef composition into a lower cover but more resistant state. This shift has increased the relative dominance (and in some subregions the absolute cover) of coral species more able to tolerate disturbance, or at least able to quickly recolonize disturbed reefs due to the characteristics of their brooded larvae that stay relatively close to parental populations. Weedy corals grow slower than competitive species, thus post-disturbance recovery and reef accretion could also decline 22 . This compositional shift also degrades other important reef characteristics and functions including vertical reef accretion, surface complexity, and habitat provision 23,24 . In the Caribbean, competitive coral species, and some stress tolerant species in the Orbicella genera, disproportionately contribute to increased structural complexity and calci cation, whereas many weedy species do not because of their smaller colony sizes and atter morphologies. The loss of structural complexity is of unique importance due to the established links between complexity and multiple ecosystem services and because changes to complexity are highly in uenced by hurricane impacts, compared to other disturbances 25 .
Although the immediate effects of storms on coral communities have declined, the longer-term impacts linger for years. Like Gardner et al. 7 , we found little evidence of post-storm recovery. Instead, we found a brief period of stasis after the immediate impact of a storm, and then resumption of pre-storm decline (2005). This pattern could be due to delayed disease or coral predator outbreaks triggered by the storm For example, physical damage from tropical storms, such as coral fragmentation and increased sedimentation, may provide more opportunity for contact between pathogens and live coral tissue 26 .
When comparing the average rate of change in coral cover between storm-impacted and unimpacted sites, we found that storm-impacted sites still have a negative recovery rate (i.e., continued decline poststorm) in the 2010s while unimpacted sites have a slightly positive rate of change, indicating recovery ( Figure 4 below). We know that coral reefs experience a simultaneous myriad of disturbances and stressors, including storms, thermal stress (leading to coral bleaching), disease, and localized impacts from pollution, over shing, and development. Therefore, it is possible that concurrent or subsequent disturbances or stressors interact with storm impacts to further exacerbate coral cover decline. For example, synergistic interactions between hurricanes and coral disease, in which storm damaged sites have a higher prevalence of coral disease compared to unimpacted sites, has been documented in the Caribbean 27-29 . Thus, the effect of storms on cover, composition, and functioning often continue for a signi cant period of time after the event.
The contribution of hurricane impacts to coral decline in the Caribbean has decreased over the past few decades; however, localized immediate impacts from storms can still be substantial on reefs with high coral cover or longer time between storm events. Importantly, storms appear to be facilitating a shift towards coral assemblages dominated by weedy species. This shift has eroded reef functionality, but may also increase resistance to other disturbances.
Increasing trends in storm frequency and intensity have important regional implications, not only for natural marine and terrestrial ecosystems and wildlife, but also for people. More, stronger storms will result in increased damage to infrastructure and loss of ecosystem services that humans rely on. For example, recent reports quanti ed how damage to coral reefs from Hurricanes Irma and Maria in 2017 increased the ood risk to coastal communities in Florida and Puerto Rico. These risks translate to potentially hundreds of millions of dollars in direct and indirect damages to properties and buildings 30 .
Like coral communities, any socio-ecological community impacted by hurricanes will likely have decreased time for recovery between storm events under climate change.
Declarations Program (TCRMP). We thank the Reef Check Foundation, its Chapters and volunteers who have spent countless hours collecting these data. This publication would not have been possible without their contribution We thank Dr. James Umbanhowar for support in developing R code for creating the reefstorm intersection database and providing feedback on previous iterations of the analysis and manuscript. We also thank Dr. Peter White, Dr. Karl Castillo, and Dr. Allen Hurlbert for providing feedback on previous iterations of the analysis and manuscript.

Data Availability
The datasets analysed during the current study are available in Laura Mudge's GitHub repository: https://github.com/Lmudge13/Caribbean_Coral_Reef_Hurriane_Impacts. The compiled raw coral cover data set is available from the corresponding author on reasonable request; however, this dataset is not re-printed or publicly available due to license and user agreements from individual databases. All databases included in the coral cover dataset can be accessed using the references provided in Extended Data Table  1.

Coral Survey Data Acquisition
Coral reef benthic survey data was obtained from primary source databases, peer-reviewed literature, and grey literature. Databases included widely-used and publicly available coral reef monitoring programs, such as Reef Check, Reef Life Survey, and the Atlantic and Gulf Rapid Reef Assessment (AGRRA). Primary databases with download date are listed in Extended Data Table 1. This study combined a previous version of a Caribbean coral cover database used for analysis in Selig and Bruno 2010 27 and Schutte et al. 2010 28 , with more recent coral survey data from longitudinal studies and monitoring datasets. We relied heavily on data from monitoring programs because they provide large amounts of repeated measurements over long time frames and cover broad spatial scales, both of which are essential for making conclusions regarding regional trends and possibly help mitigate the effects of publication bias 3,29 . Most surveys were not conducted on permanent transects, and GPS coordinates were used to verify the location on the reef for resurveying over time. All monitoring programs collect data on characteristics of the habitat surveyed (e.g. reef zone, such as bank reef or patch reef), which aids in verifying that the same reef area is resurveyed.
Absolute living scleractinian coral cover was measured using quantitative techniques including line transect intercept (in-situ counts along a transect line) and point count (randomized points taken from video transects or photo quadrats). Despite differences in survey methodology and possibly precision and accuracy, the metric of percent cover of the benthos is recognized as a fairly coarse measurement, resulting in negligible differences that are not statistically affected by the method of collection 30,31,32,33,34 . The data collection methods for obtaining coral species data were similar (AGRRA or modi ed-AGRRA protocols) and conducted by trained scientists.
Studies were included if they reported a reef site location, survey date (or year), and a measure of absolute percent cover for scleractinian corals. When latitude and longitude coordinates of survey locations were not provided in the reference, we used site location descriptions and maps from the text to identify approximate coordinates using Google Earth, when possible. In addition to manual data entry from primary literature, three tools were used to extract data from pdf resources: the tabulizer package in R 35 was used to extract raw percent cover data from tables, and ImageJ (from previous Bruno lab database only, see Schutte et al., 2010) and/or Web Digitizer 36 was used to extract raw percent cover data from gures. If more than one survey was conducted at the same reef site on any given day, percent coral cover was averaged to produce one value per day/location combination. Coral reef survey locations were considered unique based on the latitude and longitude coordinates provided from the dataset or study. The resulting database includes survey data from 3,144 unique reef locations throughout the Caribbean with 11,490 measurements of coral cover between 1971 and 2017. Approximately 23% of the data came from peer-reviewed literature sources and 77% from coral reef monitoring programs.

Coral Cover by Life History Group
Absolute percent cover of distinct coral species was obtained from three sources, mostly focused in Florida and the US Virgin Islands (Extended Data Table 1, sources with **). Coral species were assigned a life history group (LHG) of either competitive, stress tolerant, or weedy based on classi cations made in Darling et al. 37 . These assignments are based on qualities related to species speci c growth and reproduction (Supplement Table 2). Coral species not yet assigned to a LHG were labeled as "unclassi ed". The relative cover for each life history group was calculated by site and year using the calculation: Building a hurricane and coral reef intersection database Historical storm track data was downloaded directly from the National Oceanic and Atmospheric Administration (NOAA) Atlantic Hurricane Database (HURDAT2) using the HURDAT package in R 38 . These historical records contain storm track location (latitudinal and longitudinal coordinates), wind speed (knots), low pressure (millibar), status (landfall, hurricane classi cation), date and time, with variables recorded every 6-hours. Historical track information from the earliest year (1851) to present was used to analyze overall storm patterns in the Atlantic basin. Linear models (ordinary least squares regression) were used to investigate changing trends in the frequency and intensity of tropical storms over time in the Atlantic.
Functional programming in R was used to catalog which hurricanes cross which reef sites in the coral reef survey dataset. Code for these procedures was adapted from Elsner and Jagger 39 . For each reef, I searched for all historical storms occurring within a 100km radius of the reef site coordinates. Storms of any strength were retained within a 35km radius of the reef coordinates, storms of category 3-5 on the Sa r-Simpson scale were further retained between 35-60km, and only category 4 and 5 storms retained between 65-100km. These buffers are based on previously published hurricane path impacts to coral reefs 40,34,6, . Each observation in the database is a unique reef-storm intersection. Therefore, reef locations appear multiple times in the database, if more than one storm has hit the reef since 1851, and individual storms appear multiple times if they struck multiple coral reef locations along their path.
Historically (1851-2017), approximately 32% of named storms in the Atlantic basin have hit a coral reef location (1,604 named storms, 521 hit a reef). Between 1970-2017, the time period of coral survey sampling, there were 547 storms total, 28% of which crossed over at least one coral reef site, for a total of 10,058 unique site-storm intersections. Out of 3,144 unique coral reef survey sites, 2,754 sites experienced at least one tropical storm since the beginning of storm records in 1851 (87.6% of reefs impacted, 12.4% of reefs unimpacted). Sites that were not impacted were located in the SW Caribbean, along the coast of Panama, Colombia, and Costa Rica.
For each unique reef site, we calculated several measurements pertaining to the disturbance regime of tropical storms, including the total number of storms to ever hit that reef, historical return time (average number of years between storm events), storm dispersion patterns, and the average historical maximum intensity of all storms, weighted by their distance to the reef. All of these variables were calculated from coral-storm intersections that occurred between 1851-2017. The dispersion statistic is used to assess the temporal clustering of hurricanes and has demonstrated ecological impacts on coral reef ecosystems 42,43 . Using previously described methods, we tabulated a count vector (Y) of storm events per reef for each year between 1851-2017. The dispersion statistics (Ψ ) is calculated as: Storm dispersion patterns were characterized as follows 42,43 : Stochastic (random): Ψ (Y) = 0 (i.e. variance = mean); Clustered (over-dispersed): Ψ (Y) > 0 (i.e. variance > mean); Regular (under-dispersed): Ψ (Y) < 0 (i.e. variance < mean).

Control Reefs
A subset of the larger coral cover database was identi ed to serve as a "control" dataset. This included coral cover data from sites that were either 1) never hit by a storm; 2) had a substantial amount of time (>10 years) between storm events. For sites that had been hit by a storm, coral cover data was only retained for a period of 10 years after a previous storm until the next storm hit. This is to ensure that we were not including potential storm recovery trajectories as part of a control condition. For each reef site, we calculated the annual rate of change in coral cover (CR) to use as a comparison against the rate of change in coral cover at storm-impacted sites 6 . The CR value was also calculated for each Caribbean subregion in order to account for anticipated spatial variation in coral cover and potential local conditions contributing to coral decline.

Quantifying Resilience
Resistance Coral resistance to tropical storm damage was measured as the absolute change in coral cover from initial conditions (one year prior to a storm) and one-year post-storm. Paired Wilcoxon tests were used to quantify differences in cover before and after each individual storm event at each reef (i.e. each sitestorm combination is one observation for this test). A Kruskal-Wallis test was used to test the hypothesis that coral resistance is greater (meaning less coral loss from storms) in more recent decades.

Recovery
Temporal patterns in coral recovery were quanti ed in two ways: as (1) the relative change in coral cover at any year pre-or post-storm, relative to coral cover in the year preceding a storm, here referred to as the initial conditions; and (2) as the annual rate of change in absolute coral cover (CR), post-storm 44,36,6 .
(1) Relative recovery = % cover at year relative to storm -% cover before impact (2) Annual rate of change in coral cover (CR) = (pca-pcb)/d Quantifying the relative change in coral cover for years both before and after a storm event allows us to compare the impact of storms on pre-disturbance trends. First, I used regression models to evaluate the trend in relative recovery for the time periods pre-and post-storm. Upon visual review of linear regression (using ordinary least squares models), it became apparent that one linear relationship did not persist throughout the time period of recovery, but rather multiple piecewise relationships might exist. We used the segmented package in R to estimate the appropriate breakpoints for the regressions 45 . We then compared the slopes in the piecewise regressions for several time periods pre-and post-storm to describe patterns of recovery. Next, we quanti ed the annual rate of change in coral cover (CR) after a storm event. CR is measured over each individual storm event time series, in which pca is the percent cover at the end of the time series, pcb is the coral cover immediately after a storm (post one-year), and d is the duration of the time series, calculated as the number of years between pcb and pca. If two or more storms occurred in the same year/site, the CR time series was kept for the stronger storm and/or later storm. Resistance and recovery were quanti ed for both absolute coral percent cover (all species) and the relative abundances of coral life history groups.
Linear mixed models were used to quantify the effects of a variety of disturbance characteristics on coral resistance and recovery. Predictors included a mix of event speci c characteristics and disturbance regime characteristics (Extended Data Table 3). All predictors were treated as xed effects, except for reef location, which was treated as a random effect to account for variation amongst individual reef sites. Prior to modeling, raw data were analyzed for normality, heteroscedasticity, outliers, and collinearity via pairs plots and variance in ation factors (VIF). Predictors with a VIF > 2 were removed from the model.
Historical return time and the historical number of storms were collinear and had high VIF and for each model, whichever variable had the higher VIF was removed. In the resistance models for coral life history groups, storm distance was also removed due to high collinearity with wind speed and high VIF. Response variables had a non-normal distribution and included both zeros and negative values, so a cube-root transformed was performed prior to modeling. All continuous xed effects were scaled prior to modeling. Model residuals were also evaluated to meet assumptions of normality and homoscedasticity.
Models were run using the lme4 package 46    Average yearly change in coral cover, relative to year of impact. The vertical black bar represents a storm event on a reef. Black dots are the Caribbean-wide average change in coral cover for any year pre/post storm (±bootstrapped con dence intervals), relative to initial cover (percent cover at one year prior to storm). Numbers represent sample sizes for averages. Lines represent the slope of change in coral cover pre-and post-storm, with the orange line representing immediate impact.