Niche space of corals along the Florida reef tract

Over the last three decades corals have declined precipitously in the Florida Keys. Their population decline has prompted restoration effort. Yet, little effort has been invested in understanding the contemporary niche spaces of coral species, which could assist in prioritizing conservation habitats. We sought to predict the probability of occurrence of 23 coral species, including the critically endangered Acropora cervicornis, using observations at 985 sites from 2011–2015. We ran boosted regression trees to evaluate the relationship between the presence of these corals and eight potential environmental predictors: (i) bathymetry (m), (ii) mean of daily sea surface temperature (SST) (°C), (iii) variance of SST (°C), (iv) range of SST (°C), (v) chlorophyll-a concentration (mg m3), (vi) turbidity (m-1), (vii) wave energy (kJ m-2), and (viii) distance from coast (km). The Marquesas and the lower and upper Florida Keys were predicted to support the most suitable habitats for the 23 coral species examined. A. cervicornis had one of the smallest areas of suitable habitat, which was limited to the lower and upper Florida Keys, the Dry Tortugas, and nearshore Broward-Miami reefs. The best environmental predictors of site occupancy of A. cervicornis were SST range (4–5°C) and turbidity (K490 between 0.15–0.25 m-1). Historically A. cervicornis was reported in clear oligotrophic waters, although the present results find the coral species surviving in nearshore turbid conditions. Nearshore, turbid reefs may shade corals during high-temperature events, and therefore nearshore reefs in south Florida may become important refuges for corals as the ocean temperatures continue to increase.


Introduction
Since the late 1970s, there has been a steady decline in live coral cover throughout the Caribbean [1,2]. This decline has included unprecedented mortality of two of the Caribbean's most historically important reef-building coral species, A. cervicornis and Acropora palmata [1,3]. Most of the acroporid mortality in the Caribbean was caused by disease and thermal-stress events [4][5][6]. In 2006 this decline prompted the listing of both acroporids as 'threatened' under the U.S. Endangered Species Act [7], and in 2008 they were listed as 'critically endangered' on the International Union for Conservation of Nature Red List. Decades after the initial a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 2011-2015. Note that the models only predicted niche spaces for 23 of the 32 coral species because nine of the species were reported in less than 10% of the sites, making predictions uncertain. The FRRP data were collected at 1028 sites using 10 m x 1 m (10 m 2 ) belt transect surveys (Fig 1). The niche models were trained on a random subset (80%) of the sites and were tested against the remaining 20% of the sites (Table A in S1 File).
Environmental data. Eight environmental predictors, which have been previously shown to influence coral physiology and survival [35][36][37], were initially examined for incorporation in the niche model: (i) bathymetry (m), (ii) mean of daily sea surface temperature (SST) (˚C),  Table 1). A 1-km resolution depth (m) raster file was developed by Ames [38], which was a combination of satellite data and in-situ field measurements. Daily measurements of SST (˚C) and chlorophyll-a concentration (mg m -3 ), collated as raster data, with a 0.25 km resolution, were obtained from the University of Southern Florida's Optical Oceanography laboratory from 2011-2015 (https://optics.marine.usf.edu/). The average SST (˚C) was calculated as the mean of daily SST from 2011-2015. The variance of SST (˚C), a measurement of thermal variability, was calculated as the variance of daily SST from 2011-2015. The range of SST was calculated by subtracting the minimum temperature from the maximum temperature at every pixel. The average daily chlorophyll-a concentration was calculated as the mean chlorophyll-a concentration from 2011-2015. Turbidity was quantified by the diffuse light attenuation coefficient K at 490 nm (m -1 ) from NOAA CoastWatch, averaged daily from 2013-2015. Although pre-2013 K 490 imagery exists, it does not include the region of interest and so pre-2013 data were not included in the analysis. Wave energy was calculated using inputs of fetch (i.e., the distance of open ocean over which winds travel unobstructed) calculated using the 'fetchR' package in R [39], and daily wind speed and wind direction raster data were obtained from Remote Sensing Solutions [40] from 2011-2015. Wave energy was calculated using equations in Chollett and Mumby [41] adapted from [42], where each cell's fetch was evaluated in the dominant wind direction. If fetch was less than 38 km, then the seas were considered 'fetch-limited', whereas if the fetch was greater than 38 km then the seas were considered 'fully developed.' A complete explanation of the wave energy calculations is available in the online S1 File. Distance from coast (km) was calculated at the 1-km resolution using the distance function in the 'raster' package [43] and the coastline polygon from the high-resolution map in the package 'RWorld-Map' [44] in R [45]. For spatial consistency, the final input raster files were resampled to a 1-km resolution and masked to a 1km buffer of the Fish and Wildlife Conservation Unified Florida Reef Tract Map (http://geodata.myfwc.com/datasets/unified-florida-reef-tract-map).
Coral niche model. A niche model was developed initially for 32 coral species along the Florida reef tract at 1028 sites from 2011-2015. However, 9 coral species were found in less than 10% of the sites making those models unstable, therefore the results are not included here in the 23 coral species niche model ( Table 2). In addition, because A. cervicornis is listed as 'critically endangered' and is of special interest in this study, an exception to the 10% rule was made for this species. A. palmata is also of special interest but was only recorded in < 0.5% of the sites (Table 2), which made modeling problematic. We used boosted regression trees (BRTs) [46] to fit the presence and absence of the coral species data, at 1028 sites, to seven of the eight potential environmental predictors (Table 1, Table A in S1 File). Variance of SST was excluded as a potential environmental predictor because there was a strong positive correlation between variance of SST and range of SST (0.71, Fig A in S1 File). Data for each environmental factor were then extracted for each site. Any sites that did not have values for all 7 environmental factors were removed. Of the 1028 total sites, 43 were removed: 22 sites had no SST data, 20 sites had no turbidity data, and 7 sites Table 2. In-situ presence of coral species at sites (%); the modeled area under the receiver operating curve (AUC), which is a diagnostic for model performance; and the percent suitable habitat area (%) predicted by the niche model along the Florida reef tract using data from 985 sites from 2011-2015.

Species Name Presence at Sites (%) AUC Suitable Habitat Area (%)
Siderastrea radians 47 Where � indicates coral species that were not present at >10% of the survey sites and were therefore excluded from the model. �� indicates an exception for the 10% rule, because the critically endangered A. cervicornis was a species of special interest in this study. The dashes indicate that the coral species were found in less than 10% of the sites, therefore the results are not included.
https://doi.org/10.1371/journal.pone.0231104.t002 had no wave energy data (6 of which also had no SST data). In total, 985 sites were used in the analysis (Fig 1). We used k-fold partitioning to randomly divide the data into five sections. We used data from four of those sections (i.e., 80% of the data) to train the model and data from one of the sections (i.e., 20%) to test the model. We used a machine-learning algorithm in the form of BRTs to evaluate the relationship between the presence of each species and potential environmental predictors. BRTs fit data by recursively adding 'trees' (n-branching nodes) at each iteration-bagged trees take a new bootstrap sample from the training data and choose the next tree that minimizes the 'loss' function. We set the bag fraction to 0.8 to introduce some stochasticity into the niche model, which indicates that 80% of the training data were used to fit each individual tree. Additionally, we weighted all sites to generate an equal weight of presences and absences [47]. We built the model using the 'gbm' R package [46] and code adapted from [48]. The niche model was set to a tree complexity dependent on number of environmental factors being tested, a learning rate of 0.0015, and an initial condition of 30 trees.
A 1-km buffer of the coral-reef polygons, found within Florida Fish and Wildlife Conservation Commission-Fish and Wildlife Research Institute's Unified Florida Reef Tract spatial layer, was used to clip the geographical extent of the model results (Fig 1). The amount of suitable habitat was then calculated as the combined area of cells (km 2 ), above a 50% threshold value. The stepwise iteration process of the BRTs cross-validates at each iteration using data that were not in the immediate bootstrapped training set. The model was run multiple times using different environmental combinations, based on their performance in previous runs, to optimize the model (i.e., the best area under the receiver operating curve) and remove unpredictive environmental variables. The model was then tested for accuracy against the remaining 20% of the dataset, to produce diagnostics of model performance which included constructing confusion matrices (Tables B and C in S4 File). Confusion matrices were computed using the 'caret' package [49] in R [45].
We were also interested in the effects of disturbances on the niche space of Acropora cervicornis. Several disturbances occurred either side of our field-sampling window (2011-2015), including a major cold-snap in 2010 [5], Hurricane Irma in 2017, and a disease outbreak (stony-coral-tissue-loss disease) in 2016. Therefore, we took a landscape-metrics approach to simulate the influence of a 'generic' disturbance on the predicted patch attributes of Acropora cervicornis along the Florida reef tract. Because Acropora cervicornis colonies are spread across the Florida reef tract as viable patches of subpopulations, disturbances are likely to remove viable patches from the metapopulation, increasing the average distance among the patches. To test this concept and calculate the relative distances between predicted patches of Acropora cervicornis subpopulations on modern reefs along the Florida reef tract, we computed the Euclidean nearest-neighbor distance among the predicted niche patches to examine the average distances between the patches using a series of probability-of-occurrence thresholds. These thresholds simulate different intensities of disturbance, with increasing values simulating increasing intensities of disturbance. All data and R code are available at https://github.com/ rvanwoesik/Florida-Niche, and kmz (Google Earth) files of the predicted probability of occurrence of each coral species is available in S4 File.

Ethics statement
The field data were collected by participants in the Florida Reef Resilience Program Disturbance Response Monitoring (DRM) https://myfwc.com/research/habitat/coral/drm/. Permission to visit the study sites was granted by the Florida Fish and Wildlife Conservation Commission and the National Oceanic and Atmospheric Administration. The field studies did not involve the collection of endangered or protected species.

Probability of occurrence
The niche model was run on 23 coral species ( Table 2). The coral species Siderastrea radians and S. siderea were predicted to have the largest area of suitable habitat along the Florida reef tract ( Table 2, Table 3). S. intersepta, S. bournoni, M. alcicornis, P. astreoides, and P. strigosa were also predicted to have large areas of suitable habitat, whereas A. cervicornis was among the species with the smallest area of suitable habitat along the Florida reef tract ( Table 2, Table 3).
The Marquesas, the lower Florida Keys, and the upper Florida Keys were predicted to support the most suitable habitats for the coral species examined (  (Fig 2) we focus on 4 contrasting coral species: A. cervicornis, Mycetophyllia sp., O. franksi, and S. siderea, although maps showing the predicted probability of occurrences of all 23 coral species are presented in the online supporting document (Figs B-X in S1-S4 Files). The geographical subregions with the highest probability of occurrence of the critically endangered species Acropora cervicornis included the lower Florida Keys (195 km 2 ), the Dry Tortugas (156 km 2 ), the upper Florida Keys (112 km 2 ), and nearshore Broward-Miami reefs ( Table 3, Fig 2). The probability of occurrence was lower at Deerfield to South Palm Beach, and the reefs north of South Palm Beach were predicted as unsuitable for Acropora cervicornis at the time of surveys (Table 3, Fig 2). Mycetophyllia sp. was predicted to occur from the Dry Tortugas through to the middle Florida Keys, whereas the probability of occurrence of O. franksi was more restricted, and only included the upper Florida Keys and Biscayne. S. siderea had a wide geographic extent and had a particularly high probability of occurrence from the Dry Tortugas through to Miami (Fig 2, Table 3).

Environmental predictors
Distance from the coast, range of SST, bathymetry, and wave energy were the 4 most consistent predictors of the 23 coral species examined (Table 4). Again, for illustrative purposes, we focus on four contrasting coral species (A. cervicornis, Mycetophyllia sp., O. franksi, and S. siderea) (Fig 3), although the partial dependency plots that outline the best environmental predictors of the 23 coral species are presented in the online supporting document (Figs B-X in S1-S4 Files).
The best environmental predictors of site occupancy of A. cervicornis were SST, moderate turbidity (K 490 0.15-0.25 m -1 ), and moderate wave energy (>0.5-1.5 kJ m -2 ) ( Table 4, Fig 3). The probability of occurrence of A. cervicornis was lower where wave energy was > 1.5 kJ m -2 ( Table 4, Fig 3). Mycetophyllia sp., had the highest probability of occurrence when the temperature range was between 4-6˚C, the bathymetry was deeper than 5 m, the chlorophyll-a concentrations were below 1.5 mg m 3 , and the mean SST was 24 o C ( Table 4, Fig 3). The highest probability of occurrence for O. franksi occurred in the Dry Tortugas, > 60 km from shore, where wave energy was relatively high (1.2-1.5 kJ m -2 ), and where mean SST was around 25 o C (Table 4, Fig 3). Note the flat line in Fig 3B represents a lack of sampling sites between 10 km and 70 km from shore. S. siderea had the highest probability of occurrence in habitats that had low chlorophyll-a concentrations (< 0.3 mg m 3 ), mean SST around 26 o C, and wave energy between 1.2-1.5 kJ m -2 (Table 4, Fig 3).

Discussion
In the last four decades the Florida reef tract has lost a significant proportion of coral populations, particularly the reef-building corals Orbicella and Acropora species [1,2,50]. The niche models predicted that 13 species of coral were likely to have favorable habitats from the Dry Tortugas to Miami. These species included: D. stokesii, D. labyrinthiformis, E. fastigiata, M. alcicornis, O. faveolata, P. astreoides, P. divaricata, P. furcata, P. clivosa, S. radians, S. siderea, S. bournoni, S. intersepta (S3 File for Google Earth kmz files). Seven coral species appeared more sensitive to the environmental variables examined, and their probability of occurrence was The coral species are ranked according to their total area of suitable habitat space (km 2 )-depicted in the right-hand column of the  (Table 3). Ginsburg and Shinn [51] first reported on the negative influence of Florida Bay on the middle Florida Keys, and recently Toth et al. [52] showed that reef accretion terminated significantly earlier in the middle Florida Keys than elsewhere, which they suggested was most likely because of negative influences from Florida Bay. It is likely that Florida Bay will continue to influence reefs in the middle Florida Keys, which may prove a disadvantage for many coral species along those reefs.

PLOS ONE
The best environmental predictors of site occupancy of the 'critically endangered' A. cervicornis were moderate turbidity, SST, and wave energy. Although historically Acropora corals are known to survive best in oligotrophic waters [53][54][55], the present results suggest that on modern reefs, near zero turbidity was not optimal for A. cervicornis [56]. Rather, a moderate turbidity value K 490 of between 0.15-0.25 m -1 showed the highest probability of occurrence. These results agree with physiological studies, which show that reducing light by shading can effectively reduce the influence of temperature anomalies [57][58][59]. Indeed, thermal-stress events may be shifting the optimal niche space of corals toward more turbid habitats, for example to nearshore reefs of Broward-Miami subregion, as ocean temperatures increase [56]. The present study also suggests that moderate wave energy (>0.5-1.5 kJ m -2 ) is favorable for A. cervicornis. In support, D'Antonio et al. [27] showed that colonies of A. cervicornis were most common close to reef edges, where water-flow rates were high. Physiological experiments also show Acropora colonies are particularly intolerant to stagnant waters, with low rates of mass transfer [60].
Although the niche models were 87% accurate at predicting localities for restoration for A. cervicornis, and 88% accurate for O. franksi and Mycetophyllia sp. (Table C in S4 File), there are some caveats that need consideration. Firstly, these types of models suffer from incomplete geographic sampling and mismatches of scale between the organism and environmental covariates. For example, an observed absence of a coral in a 10 m 2 belt transect in the field does not necessarily imply a complete absence throughout a 1-km 2 grid cell, at which the environmental variables are considered. Therefore, an absence might not be considered a 'true absence' [61] and would reduce the model's predictive capacity. Secondly, while a dominant species might occupy most of its fundamental niche space, rare species might occupy only a small proportion of their fundamental niche [62]. Dispersion limitation may further prevent the rare species from occupying all the potential niche space, and therefore predicting the probability of occupancy may be over-estimated (i.e., with high false positives).
Since niche models are known to be prevalence-dependent [63], low in-situ occurrences will translate to low accuracies. Indeed, the niche models had high specificity and low sensitivity (see S4 File for full specificity and sensitivity results). Specificity is an indicator of how good the model is at detecting true negatives, whereas sensitivity is an indicator of how good the model is at detecting true positives. In other words, the niche models were good at predicting habitats in which a particular species was unlikely to be present, but less accurate at predicting habitats in which a species could occur. This strong specificity and low sensitivity are expected in localities such as the Florida reef tract, where the organisms do not occupy the entire niche space because the system has undergone recent disturbances. Such disturbed environments reflect data that are unbalanced toward absences, although we did compensate for this issue by generating an equal weight of presence and absence sites [47]. The simulations that examined disturbances to the predicted patches of Acropora cervicornis, using a series of probability-ofoccurrence thresholds, found that patch distances were on average 2-3 km (Fig Y in S4 File) on modern reefs along the Florida reef tract, and that distances between the patches are likely to increase with an increase in disturbances. These results are troublesome considering the fragile nature of the modern metapopulation of Acropora cervicornis. The niche models can be improved in the future by hierarchically adding spatial and temporal complexity, although we expect that the general geographic patterns will hold up. The models can also be improved by adding more local information, such as macroalgal cover, since low macroalgae coverage has been shown to increase the survival of A. cervicornis [64].

Summary and conclusions
Recent changes in environmental conditions along the Florida reef tract may have shifted, and even narrowed, the niche space of some sensitive coral species [8], and consequently information on the distribution of coral species from decades past may no longer provide information for present niche space. Previously, the most optimal purported niche space for A. cervicornis was in clear oligotrophic waters, although the present results suggest that turbid conditions are more optimal on modern reefs that frequently experience high heat stress. Therefore, the nearshore reefs along the Florida reef tract may become important refuges for corals as the ocean temperatures continue to increase.