Aim: Forecasting the influence of climate change on coral biodiversity and reef functioning is important for informing policy decisions. Dominance shifts, tropicalisation and local extinctions are common responses of climate change, but uncertainty surrounds the reliability of predicted coral community transformations. Here, we use species distribution models (SDMs) to assess changes in suitable coral habitat and associated patterns in biodiversity across Western Australia (WA) under present-day and future climate scenarios (RCP 2.6 and RCP 8.5).

Location: Coral reef systems in WA.

Methods: We developed SDMs with model prediction uncertainty analyses, using specimen-based occurrence records of 188 hermatypic scleractinian coral species and seven variables to estimate present-day and future changes to coral species distribution and biodiversity patterns in WA under climate change conditions.

Results: We found that suitable habitat is predicted to increase across all regions in WA under RCP_2.6 ²⁰⁵⁰, RCP_8.5 ²⁰⁵⁰ and RCP_2.6 ²¹⁰⁰ scenarios with all tropical and subtropical regions remaining coral biodiversity strongholds. Under the extreme RCP_8.5 ²¹⁰⁰ scenario however, a clear tropicalisation trend could be observed with coral species expanding their range to mid-higher latitude regions, while a substantial drop in coral species richness was predicted at low latitude tropical coral reefs, such as the inshore Kimberley and offshore NW reefs. Despite the predicted expansion south, we identified a net decline in biodiversity across the WA coastline.

Main Conclusions: Results from the models predicted higher net biodiversity loss at low latitude tropical regions compared to net gains at mid-high latitude regions under RCP_8.5 ²¹⁰⁰. These results are likely to be representative of latitudinal trends across the southern hemisphere and highlight that increases in habitat suitability at higher latitudes may not lead to equivalent biodiversity benefits. Urgent action is needed to limit climate change to prevent spatial erosion of tropical coral communities, extinction events and loss of tropical ecosystem services.

Model fitting, variable selection and parsimonious model evaluation

In order to estimate the present-day and future suitable coral habitat along the WA coastline, we used the Maximum Entropy (MaxEnt) algorithm in the ‘dismo’ R package (Hijmans et al., 2017). MaxEnt is one of the most common modelling approaches for SDMs and was chosen due to its ability to handle presence-only data (Elith et al., 2006; Elith et al., 2011; Phillips & Dudík, 2008). To avoid overprediction, only species with > 10 spatially unique records were included (Wisz et al., 2008).

Prior to constructing the present-day and future coral suitable habitat, models were fitted and evaluated for these 205 species using background spatial point customization (Merow et al., 2013), data splitting into train/test datasets (Araújo et al., 2005) and variable selection to construct the most parsimonious model (Austin & Van Niel, 2011). For each species modelled, 1,000 background spatial points were randomly selected across the study area, bounded by the Australian EEZ depth of £ 40 m. Before the background points were generated, a 9 km radius mask around known locations was applied to exclude background points close to the occurrence locations and avoid variable replication. A radius of 9 km was used based on the coarsest resolution of raw environmental and geomorphological variable data.

Model performance for each species was evaluated with an independent test dataset, where the occurrence and background data were partitioned 75/25 into training/testing datasets (Araújo et al., 2005; Hastie et al., 2009; Merow et al., 2014; Vignali et al., 2020). Five replicate models were constructed on the training dataset with default model features and regularization settings, configured to select the highest contributing variables for the most parsimonious model (Elith et al., 2011). Average variable permutation importance across the five model replicates was used to exclude variables with a permutation importance >= 1% in the most parsimonious model (Li et al., 2020; Sobek-Swant et al., 2012; Williams et al., 2012). The threshold independent and dependent metrics, area under the curve (AUC) (Beaumont et al., 2019) and sensitivity, were used to evaluate model performance on the withheld test data.

Model predictions

For every species, continuous habitat suitability predictions were converted into binary values (occurrence [1] and absence [0]) using the maximum sensitivity and specificity threshold (Liu et al., 2013). Total area of occurrences was converted to km² to determine the present-day and future suitable habitat. Shifts in suitable habitat were estimated by overlaying the future binary habitat suitability predictions with present-day binary predictions using package ‘biomod2’ (Thuiller et al., 2009).

The R script details the model calibration, validation and prediction steps.

Sources of raw environmental and geomorphological data used to integrate into the models can be found in Table 1 in the main manuscript. Coral species occurrence museum records can be found in the Atlas of Living Australia (https://www.ala.org.au/) and upon request at the Western Australian Museum.

The R script details the model calibration, validation and prediction steps as described in the method section of the manuscript.