Aim

Spatially explicit protections of coastal habitats determined on the current distribution of species and ecosystems risk becoming obsolete in 100 years if the movement of species ranges outpaces management action. Hence, a critical step of conservation is predicting the efficacy of management actions in future. We aimed to determine how foundational, habitat‐building species will respond to climate change in Fiji.

Location

The Republic of Fiji.

Methods

We develop species distribution models (SDMs) using MaxEnt, General Additive Models and Boosted Regression Trees and publicly available data from the Global Biodiversity Information Facility to predict changes in distribution of suitable habitat for mangrove forests, coral habitat, seagrass meadows and critical fisheries invertebrates under several IPCC climate change scenarios in 2070 or 2100. We then overlay predicted distribution models onto existing Fijian protected area network to assess whether today's conservation measures will afford protection to tomorrow's distributions.

Results

We develop species distribution models (SDMs) using MaxEnt, General Additive Models and Boosted Regression Trees and publicly available data from the Global Biodiversity Information Facility to predict changes in distribution of suitable habitat for mangrove forests, coral habitat, seagrass meadows and critical fisheries invertebrates under several IPCC climate change scenarios in 2070 or 2100. We then overlay predicted distribution models onto existing Fijian protected area network to assess whether today's conservation measures will afford protection to tomorrow's distributions.

Main conclusions

Species distribution models are a critical tool for conservation managers, as linking spatial distribution data with future climate change scenarios can aid in the creation and resiliency of protected area programmes. New protected area designations should consider the future distribution of species to maximize benefits to those taxa.

We sampled mangrove presence points from mangrove global distribution polygons created by Hamilton and Casey (2016) based on estimates from 2000.  Seagrass presence points were downloaded from McKenzie et al. (2021) who thoroughly reviewed the known distribution of seagrass meadows in the Pacific Islands. Coral presence points were sampled from coral reef global distribution polygons created by UNEP-WCMC (2021). Presence points for each taxa were reviewed to eliminate points without a minimum of three decimal places as well as any outliers. Publicly available distribution data on mud crabs (Scylla serrata), ark shells (Anadara antiquata), triton shells (Charonia tritonis), sea cucumbers (Holothuria atra, Bohadschia argus, and Holothuria edulis), and giant clams (including Tridacna gigas, T. maxima, and T. squamosa) were downloaded from the Global Biodiversity Information Facility. Presences reported on GBIF went through a cleaning process to remove points without latitude and longitude, coordinates with reported uncertainty larger than 10 km², duplicates, points without a minimum of three decimal places, and points without enough identifying information using the R package scrubr (Chamberlain 2020). Geographic distribution of points were reviewed per taxa to determine if spatial sampling bias occurred, and model study areas reflect the closest and most thorough sampling effort to Fiji (Table 1). Finally, to address pseudoreplication and bias, presence points were gridded over the entire study area and reduced to only one presence point per 5‐arcminute cell. Background points were identified from nonoccupied cells of this grid, masked for land area. Biophysical variables were selected for inclusion in analysis based on life history traits of each taxa type, using available literature and prior SDM studies on similar species to limit the amount of irrelevant predictors (Santini et al. 2020). Climatic variables were sourced from WorldClim, and oceanographic variables were sourced from Marspec and Bio-Oracle at a resolution of 5 arcminutes.

We compared the performance of MaxEnt, BRT, and GAM using area under the receiver operating characteristic curves (AUC) , the true skill statistic (TSS), and the presence-only metric Boyce Index. We assumed that AUC = 1 was overfit, AUC values < 0.8 were poorly fit, and TSS values < 0.6 were poorly fit (see Mainella 2016; Shabani et al. 2016). Boyce Index values vary from -1 to 1 and values >0 indicate that presence points and model predictions are consistent and better than random, while values <0 indivate poor performance. Validation, modelling and projections were completed in the SDMTune package for MaxEnt and BRT (Vignali et al. 2020), and in Biomod2 for GAM models (Thuiller et al. 2021).