Coral reefs are threatened by climate change, thus effective policies and management require spatial prioritization for conservation investments. Our aim was to develop a spatially explicit ecological model to predict current (2020) and future (2050) numbers of coral taxa at moderate scales (i.e., ~6 km²). Our machine learning predictive models of coral community attributes in 7039 mapped reef cells in the western Indian Ocean were based on 35 spatially complete influential environmental proxies trained with ~1000 field surveys. Four models explored: influences of climate change, water quality, direct human-resource extraction, and variable selection processes on numbers of coral taxa. Two predictive models examined the predictions of all variables and compared them to a variable-restricted climate change (8 commonly used variables) and human influence model (9 variables). The most frequently selected temperature variables in all models were the median, skewness, excess heat, rate of temperature rise, and kurtosis. However, non-temperature variables of observer, depth, wave energy, dissolved oxygen, salinity, chlorophyll-a, calcite concentrations, sunlight, and net primary productivity were frequently as important or stronger. Human influences of national jurisdiction, distance to people, sediments, and nutrients were selected but less influential when compared to the climate or the full variable models. Comparing models indicated the importance of variable pre-selection processes and variable interactions in predicting climate change and human influences on coral diversity. Comparing climate scenarios in the moderate RCP2.6 and extreme RCP8.5 emission scenarios indicated fewer losses in coral taxa (RCP2.6 = 5.2%, RCP8.5 = 8.1% respectively) relative to cover (RCP2.6 = 14%, RCP8.5 = 34%) over the 30 years. Excess heat and rate of temperature rise variables used in the Intergovernmental Panel on Climate Change (IPCC) forecasts predict more negative effects on corals than our four models but shown here to have low to modest effects.

Corals were sampled haphazardly while snorkeling or scuba diving in either visually estimated or measured quadrats of ~2 m² where all corals >5-cm were identified and counted in ~15-20 replicates (McClanahan et al. 2007). We also recorded the depth and habitats of the sites as reef edge, reef crest, reef flat, or reef lagoon. Thus, the values used here were the total number of taxa in ~40 m². Taxa identification was to the genus level, but Porites colonies were distinguished further as massive, branching, or Porites rus and Galaxea as either G. astreata or G. fascicularis. A total of 67 taxa in 1001 sites were sampled in 6 of the ecoregions. Sixteen observers contributed to the database, but most contributed few sites and three observers combined sampled 939 of the 1001 sites and had non-significant differences between them.

We used one of the latest iterations of coral reef maps to establish reef distribution patterns (https://data.unep-wcmc.org/datasets/1). Specifically, we used the map of the Western Indian Ocean O composed of ~7039 6-km² cells in 9 ecoregions, namely the Northern Monsoon Current Coast, East African Coral Coast, Seychelles, Cargados Carajos/Tromelin Island, Mascarene Islands, Southeast Madagascar, Western and Northern Madagascar, Bight of Sofala/Swamp Coast, and Delagoa. Empirical data were pooled to the 6-km² reef cells over the entire sampling period (1998 – 2022).

Environmental data compilations accessed several sources from satellite and shipboard measurements. Environmental layers included those expected to influence marine organisms including oceanographic layers such as wave energy, photosynthetic active radiation (PAR), light diffusion attenuation, pH, calcite, dissolved oxygen, salinity, net primary productivity, chlorophyll-a, and phytoplankton carbon. Additionally, several water temperature or thermal stress metrics known to influence chronic and acute stress on marine organisms were calculated including SST mean, median, range, standard deviation, skewness, kurtosis, rate of rise, and cumulative degree-heating weeks. Two multivariate integrated metrics of thermal stress and water quality were included. Finally, we used several layers that measure connectivity, including average net flow, indegree, outdegree, and retention metrics for each reef cell.

Human influence variables included several geographic variables including nation, wilderness (>4 hours travel time from human population), travel distance to people, shore and ports, and market gravity or the number of people living on the shore or cities as divided by the square of the distance or travel time. The ecoregion was also included. Reefs were assigned 4 fisheries management categories including unrestricted fishing, restricted fishing, low compliance closures, and high compliance closures.  These classifications were based on information in published literature, the experience of the data providers, and discussions with knowledgeable observers.