Understanding how emergent ecological assemblages have diverged from natural states is fundamental in predicting future functioning and services of ecosystems. Coral reefs are of particular concern due to their high susceptibility to anthropogenic stressors. Yet, little is known about their pre-disturbance ranges of natural states, and most reports of decline are based on a limited number of sites and high levels of uncertainty. Here, we used a novel approach to estimate the physical functionality of reefs across marine ecoregions based on habitat suitability and morpho-functional traits for coral species. We calibrated ecological niche models for 49 reef-building corals of the Greater Caribbean based on occurrence records and environmental predictors, which we combined with species-specific functional coefficients derived from morpho-functional traits reflecting their contribution to the reef three-dimensional structure to estimate the reef functional potential (RFP). We then assessed the degree of divergence of western Caribbean reefs by comparing our physical functionality estimates against recent field data evaluations. We found spatial variability in RFP across the Caribbean, with the highest mean value in the western Caribbean and the lowest in areas with marginal environmental conditions. Hotspots of RFP exist along the coast of Belize and the southeast of Cuba. Overall, 84% of sites along the western Caribbean showed a substantial reduction in their physical functioning, with the highest reductions occurring within hotspots, implying that reefs displaying the greatest changes have high initial RFP. We conclude that combining niche models with species morpho-functional traits is a valuable and promising approach to estimate the large-scale functional potential of communities and the degree of change in the absence of ecological baselines. These findings have important implications and could be used to guide efforts to preserve coral reefs functionality and define priority conservation areas in the Caribbean.

Species occurrence records

To create a geodatabase of occurrence records of Atlantic shallow-water reef-building corals we conducted a systematic search using the following sources: The Global Biodiversity Information Facilities (www.gbif.org/), the Ocean Biogeographic Information System (https://obis.org/), Biodiversity Information Serving Our Nation (https://bison.usgs.gov/#home), the Time Series Coral-cover from Florida and the Virgin Islands (Guest et al. 2018), the Caribbean Coastal Marine Productivity Program database (CARICOMP 2000), the Atlantic and Gulf Rapid Reef Assessment database (Marks and Lang 2018), and data from the laboratory of Biodiversity and Reef Conservation, which was curated, systemized, and is now included in Caribbean Reefs Information System database. Additionally, we performed a literature search to incorporate occurrence data for reefs areas that appeared to be underrepresented when the pooled occurrences were plotted in the geographic space (i.e. none or very few records). The compiled occurrences were reviewed to exclude duplicated coordinates per species, data coming from the fossil record and data with georeferencing errors (i.e. records on land).

We applied a spatial filter with a minimum nearest-neighbor distance between points in order to reduce spatial clumping of occurrences (Veloz 2009, Kramer-Schadt et al. 2013, Boria et al. 2014). To select the proper distance for the spatial thinning of occurrence data we tested models using ten coral species with different sample sizes and geographic range sizes as test examples. One group of models had no spatial thinning of the occurrence data; this yielded strong model overfitting coincident with a higher density of occurrences. A second group of models was run with a spatial filtering of 20 km applied to all data, and a third group was run with occurrences split in subsets according to different geographical clumping, where the data corresponding to the western Caribbean basin (higher clumping) was filtered with a distance of 40 km, and the rest of the data with a distance of 20 km. The resulting model outputs were converted to presence–absence maps using the minimum training presence threshold to cut off habitat suitability (Kramer-Schadt et al. 2013). We visually compared the two sets of presence–absence models with filtered occurrences against those generated with no thinning. As a result, we decided to use the 40–20 km filtering as these models showed a reduction in model overfitting, and better represented the known distribution of species. Once occurrence data were spatially filtered, the resulting subset was used as the training dataset, and the occurrences not included were used as testing dataset.

Divergence analysis

For this analysis we used published information on the estimation of the reef functional index (RFI; González-Barrios and Álvarez-Filip 2018), complemented with information obtained through the Caribbean Reefs Information System database. Data collection follows the AGRRA protocol point intercept (Lang et al. 2010) and ReefBudget (Perry et al. 2012) methodologies.

A total of 242 sites were included in this analysis covering a period from 2010 to 2017. The RFI obtained from field data was plotted against the modeled RFP to which we added an equivalence line. To construct this line, we used 0 as the minimum value for both axes and as maximum, the highest value that could potentially be achieved in each case. In the case of the FRI, the highest value would correspond to a hypothetical site with 100% of coral cover of the species with the highest functional coefficient, corresponding to a RFI of 0.977. In the case of modeled RFP the highest value would correspond to a cell in which the suitability value was 1 for all the 49 modelled species, in this case the RFP would be of 10.568. On the plot, a site with a coherent RFI and RFP values would be close to the line, sites with a higher RFI than RFP would be above the line, and sites with lower RFI than RFP would be below the line. As a measure of divergence from the RFP, we used the distance of each point to the equality line that we further transformed into percentage values.