Predicting climate change impacts on critical fisheries species in Fijian marine systems and its implications for protected area spatial planning

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.


| INTRODUC TI ON
Predicting ecological disturbance in the face of impending climate change and sea level rise has become a priority for many marine conservation biologists as the ocean is predicted to warm between 0.73°C and 2.58°C and sea levels to increase 0.43-0.84m by 2100 (IPCC, 2019 Report).Because of these projected climatic changes, entire marine systems will likely become disrupted through range shifts (Bates et al., 2014), productivity loss (Moore et al., 2018) and functional redundancy loss (Teixidó et al., 2018).Preserving biodiversity requires an understanding of the current distribution of species and the ecological niches they inhabit.In aid of this, species distribution modelling has become a frequent tool for conservators and managers (Guisan et al., 2013;Jiménez-Valverde et al., 2011;Leidenberger et al., 2015;Pearson & Dawson, 2003).
These tools allow managers to predictively assess both the areas that species may occur in and the threats they may face.Together this allows scientists to forecast mismatches between species' distributions and their threats and therefore best allocate conservation reserves.
Marine habitats on small island states in the Pacific are already impacted by sea level rise (Church et al., 2006), degradation from industrial shipping and pollution, invasive species and overexploitation (Hunt, 2003;Johnson et al., 2020;Zann, 1994).Many coastal communities in the Pacific are intricately tied to marine habitats and utilize their marine resources for livelihoods, sustenance and cultural maintenance (Lai, 1984;Thaman et al., 2017 but see Turner et al., 2007 on declining reliance).Local stakeholders in Fiji, who often have a long history of marine conservation through cultural practices like tabu (a culturally defined limited access fishery) and local governments are increasingly looking to codify protections as marine protected areas (Robertson et al., 2020;Weeks & Jupiter, 2013).However, in exploring the 'velocity of climate change', Loarie et al. (2009) suggest that only 8% of established protected areas have a residency time of over 100 years.Coupling parks together in a dynamic management model could provide 'safe landing zones' for communities forced to move because of climate change (D'Aloia et al., 2019).This requires an understanding of the potential distribution of habitat in 2100 or beyond.
Thus, the information provided by species distribution models may help organize current efforts to restore ecologically and culturally important species in Fiji and to improve future ecosystem functioning throughout the region.
In addition to concerns about focus species being displaced from protected areas, managers need to simultaneously consider the distribution and ecological interaction among foundational taxa.
Foundational taxa, like seagrass, mangroves and coral, are especially important in the Pacific as they build and structure entire marine biomes.Though made up of a suite of species (e.g.Acropora spp. and Porites spp.) that may respond differently to climate change, predicting where habitats may exist in future independent of individual species is critically important.While several studies have looked at individual species or a limited number of keystone species, few have taken comprehensive views of entire interconnected systems (see Adams et al., 2016).Although single taxa may act as umbrella or keystone species, species inhabiting the same habitat type may experience wildly different futures based on their environmental niche limits, and co-evolved communities may suffer from 'climate reshuffling' (sensu Ordonez & Williams, 2013).Therefore, comprehensive, ecosystem-wide studies are needed to better predict widespread change.
Mangroves are threatened worldwide from development and the pressures of growing populations (Sandilyan & Kathiresan, 2012).Because they are immobile, mangrove forests are limited in responding to a rapidly changing climate (Gilman et al., 2008).With a lifespan of hundreds of years, there is potential for a disconnect between the climate within which the mangrove seed landed and the climate where the tree will live.In Florida, there is already a northern migration of mangroves and their replacement of salt marshes as forests track the minimum freeze zone (Cavanaugh et al., 2015).Mangroves in Oceania are particularly at threat due to loss of habitat from sea level rise and climate change (Alongi, 2008;Lovelock et al., 2015).As managers consider protecting mangrove forests (Cameron et al., 2021), a better assessment of their future distributions on land-limited island coasts in the face of rising global temperatures and sea level is needed.
In the Pacific, seagrass habitats are an important food resource for local communities (Cullen-Unsworth et al., 2014;Thaman et al., 2017) as well as supporting culturally important megafauna (Craig et al., 2004), providing critical nursery habitat for reef fisheries, nutrient cycling and erosion control.Seagrass meadows are in global decline (Orth et al., 2006), and in Fiji, they are threatened by pollution and environmental degradation (Cullen-Unsworth et al., 2014).Seagrasses have been previously mapped using SDMs at the functional level (Adams et al., 2016) and species level (Bittner et al., 2020).Approximately 16% of Fiji's seagrass habitat is estimated to occur within an MPA (Allen Coral Atlas, 2020;McKenzie et al., 2021), but climate change is likely to negatively impact the distribution of seagrass meadows due to increased wave action and rising temperatures (Brodie & de Ramon N'Yeurt, 2018).
Boosted regression trees, Climate change, Generalized additive models, MaxEnt, Protected areas, Species distribution models Bleaching events are a major threat to warm-water reefs, though reefs in Fiji have shown some resilience to bleaching events and other sources of disturbance (Harding et al., 2003;Sykes & Lovell, 2009).In addition to their ecological role, reefs provide local human communities with many well-studied ecosystem services like fisheries, tourism, cultural significance and building materials (Golden et al., 2021;Mangubhai et al., 2020;Woodhead et al., 2019).Current coral distribution has been extensively mapped (Li et al., 2020) and projected worldwide distribution in future (Freeman et al., 2013), but detailed maps of potential local change in Fiji are still needed.
As habitat moves, so too do interstitial sessile invertebrate species.While many studies have used species distribution models to predict how warm-water fishes will be impacted by climate change (e.g.Beger & Possingham, 2008;Knudby et al., 2010;Pittman & Brown, 2011), less attention has been focussed on tropical invertebrates (but see Moya et al., 2017;Young et al., 2020), due in part to a lack of available distribution data.Invertebrates often have limited movement once settled and can have thin occupancy bands limited by temperature, depth, air exposure, sea level and competition.Many invertebrates are critically important to local communities for cultural uses (Thaman et al., 2017), sustenance (Kuster et al., 2005) and income (Fay-Sauni et al., 2008).Conservation of Scylla serrata (mangrove mud crab) is one of the main priorities of conservation managers and is intricately tied to the well-being of fisherwomen in the Pacific (Thomas et al., 2020).Ark shells, or Anadara antiquata, are also critical components of women's subsistence fishing in Fiji (Fay-Sauni et al., 2008), often gleaned from mud flats.Ceremonially important species like Charonia tritonis, a conch shell blown at the start of important meetings (Thaman et al., 2017), may also play important roles in preventing Crown-of-Thorns outbreaks (Hall et al., 2017).Sea cucumbers are one of the most lucrative fisheries in small island states, and their removal can have a dramatic impact on nutrient cycling in inshore areas (Lee et al., 2018;Purcell et al., 2018).Giant clams (Tridacna spp.) are the focus of a concentrated reseeding effort in the Pacific and have strong economic and cultural ties to coastal Fijian communities (Bao & Drew, 2017;Thaman et al., 2017).
Here, we develop species distribution models (SDMs) to predict changes in distribution of suitable habitat for mangrove forests, coral reefs, seagrass meadows and culturally and commercially important invertebrates in Fiji under several IPCC (Intergovernmental Panel on Climate Change) scenarios.We then overlay these predicted distribution models onto existing Fijian protected area network to assess whether today's conservation measures will afford protection to future distributions.Knowledge of marine species distributions is limited due to the lack of biological census data, but biophysical data are widely available and SDMs can greatly increase the impact of limited observational datasets.Our objectives are to (1) build appropriate SDMs for biome and invertebrate distribution in Fiji, (2) improve knowledge of the change in habitat in Fiji in the face of sea level rise and (3) provide information on important fisheries species to managers.Understanding how future climate change might impact the distribution of species is vitally important to conservation managers who are working to define protected area boundaries.

| Species occurrence data
Our work draws on existing datasets that are publicly available for Fijian coasts.We focussed on the islands of Vanua Levu and Viti Levu and immediately surrounding area, where we expect protected areas are likely to be concentrated in future.Due to the rarity of species-specific distribution data for dominant biome taxa in Fiji (seagrass, mangrove and coral reefs), we took a holistic approach and built models for each habitat.Marine biomes, large geographic areas dominated by one type of plant or animal life, can be made up of a number of species but data on these critical habitats are typically collected on the functional level or higher.A previous study on seagrass modelling highlighted that, in the face of limited speciesspecific data, biome data are more spatially accurate than using species data (McKenzie et al., 2021).
We sampled mangrove presence points from mangrove global distribution polygons created by Hamilton and Casey (2016)  Presence points for each taxon were reviewed to eliminate points without a minimum of three decimal places as well as any outliers (e.g.inland points; Table 1).Although sampling from expert-derived polygons may cause false positives, these remote sensing-based expert distribution maps are highly accurate and likely to produce similar results to publicly available databases (Fourcade, 2016).All data analysis was performed in the statistical computing language R 4.0.2(R Core Team, 2020).
Marine invertebrates of interest were selected based on a literature review of Fijian fisheries interests and ecosystem services, reports of important or endangered species from the wider region, consultation with local conservation organizations and discussion with relevant agency officials.Publicly available distribution data on mud crabs (Scylla serrata), ark shells (Anadara antiquata), triton shells (Charonia tritonis), sea cucumbers (Bohadschia argus, Holothuria atra, and Holothuria edulis) and giant clams (including Tridacna gigas, T. maxima, and T. squamosa) were downloaded from the Global Biodiversity Information Facility (GBIF, 2022) using the R package rgbif version 3.5.2(Chamberlain et al., 2023).Scylla and Tridacna gigas models were built using GBIF data available from Australia, due to an extremely low number of online observations in Fiji.Presences reported on GBIF went through a cleaning process to remove points without latitude and longitude, coordinates with reported uncertainty larger than 10 km 2 , duplicates, points without a minimum of three decimal places, and points without enough identifying information using the TA B L E 1 Variables included in species distribution models.R package scrubr version 032 (Chamberlain, 2020).Geographic distribution of points was reviewed per taxa to determine whether 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.This greatly reduced presence points in some cases (Table S1).

| Environmental and climate data
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 number 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 (Table 1) using R package sdmpredictors version 0.2.9 (Bosch, 2020).
Human influence has had a profound impact on mangrove, seagrass and coral habitat (Halpern et al., 2008), so therefore we also incorporated anthropogenic variables in our models.Distance to the nearest city centre was calculated using rgeos version 0.5-5 and raster version 3.4-10 packages (Bivand & Rundel, 2020;Hijmans, 2021).
Finally, we incorporated a level of protection, using IUCN categories for polygons acquired from IUCN-UNEP (2021).Polygons lacking an IUCN designation were reviewed by the authors and categorized manually using available information and authors' knowledge of the field site.Polygons were then dissolved into a raster layer such that the highest level of protection (lowest IUCN category) in each cell was retained.
Biophysical and anthropogenic data were evaluated for collinearity (>0.8) using the R function corSelect, which builds a bivariate model of each variable against the response and excludes one with the higher AIC value (package fuzzySim version 3.0; Barbosa, 2015).
Models were built using only the remaining layers (Table 1).

| Species distribution modelling
An ODMAP (Overview, Data, Model, Assessment and Prediction; Zurell et al., 2020) protocol describing the modelling pipeline is available in the Supplementary material (Table S2).To compare the predictive capability of different environmental variables, we ran the models in three iterative stages.First, we built SDMs using all environmental variables identified as ecologically important and projected these across Fiji.Then, using a smaller pool of variables that also had available future projections, (see To minimize methodological biases and to provide a comparative framework, we used three commonly used species distribution modelling methods to predict taxa distribution under future climate scenarios (Elith et al., 2006): Boosted Regression Trees (BRT; Friedman et al., 2000), MaxEnt (Phillips et al., 2006) and General Additive Models (GAM; Hastie & Tibshirani, 2017).Before running the final models, we tuned model hyperparameters for MaxEnt and BRT using package SDMtune version 1.1.4,which optimizes model performance using a genetic algorithm implementation (Vignali et al., 2020) and package Biomod2 version 3.4.6 for GAM (Thuiller et al., 2021).Models were divided into three divisions, and a training set was used to run iteratively (k = 5 folds) through the manipulated hyperparameters and a testing set used to determine the best performance of each model using AUC.Hyperparameters used in the best performing model for each species distribution modelling technique were then input into the final species distribution models.
Once environmental variables and hyperparameters were optimized, full models were created and then evaluated for model fit.
We compared the performance of MaxEnt, BRT and GAM using area under the receiver operating characteristic curves (AUC), 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 indicate poor performance (Boyce et al., 2002;Hirzel et al., 2006).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).Boyce Index values were determined using ecospat package version 3.2 using spearman rank method (Broennimann et al., 2021).Models were further evaluated using fivefold cross-validation and comparing average AUC and TSS scores from each fold.We included TSS scores in our assessment of model performance as AUC scores alone can be overconfident when many background points are used.Cross-validation was performed with random k-folds, as limited distribution of presence points prevented spatial block cross-validation for some taxa, which may result in low error estimates (Roberts et al., 2017).

| Model comparison
After evaluating each model for fit and performance, we compared projections of species habitat suitability visually and quantitatively.
Maps were projected to our study region; however, due to the growing concern of probabilities presented as only binaries (e.g. an area is either suitable for a species or not) which reduces biological nuances in the outputs (Santini et al., 2020), we developed two levels of habitat suitability to compare.We generated habitat suitability thresholds using the maximizing the sum of sensitivity and specificity method (MaxSSS) and minimum training presence (MTP).MaxSSS had been shown to be a highly objective threshold and to perform better than many other threshold metrics even with presence-only datasets (Liu et al., 2013(Liu et al., , 2016)).The minimum training presence finds the lowest predicted suitability for any one occurrence point and aims to minimize omission errors (Cord et al., 2014) and was evaluated using a sdm_threshold function (Morrow, 2019).MaxSSS was evaluated using the dismo package version 1.3-3 (spec_sens; Hijmans et al., 2020).

| Habitat suitability
Distribution of habitat suitability was highly variable for each biome type (Mangrove, Seagrass and Coral) and invertebrate species (Figure 1).Details of the best-fit models (Table S1) and in some cases additional well-fit models (Table S2) are presented for each taxon, and detailed information on model fit is available in Supporting Information.We found MaxEnt to more reliably model invertebrates overall, likely due to the ability to compensate for a low initial sample size (Wisz et al., 2008).BRT and GAM models were better fit for habitat analyses, which had higher input points and a wider initial distribution.

| Fijian habitats
The most important environmental variable for models of mangrove distribution under all time frames and fit models was elevation (average importance 40.07%± 0.43 SE; Table S4), followed by temperature seasonality (average importance 30.63% ± 0.07 SE), and precipitation of driest month (average importance 17.96% ± 0.17 SE).The relationship between mangroves and precipitation of the driest month shows a highly variable relationship but that precipitation levels of <30 mm, as seen in the Yasawas, are related to low mangrove presence, and an increase to levels between 30 and 80 mm will likely increase the presence of mangroves.Since the mangrove biome includes many species of coastal mangrove in Fiji, and mangrove trees grow in speciesspecific zones each with different relationship to salinity and soil water content (Ball, 1998), there is a complex relationship between precipitation and mangrove presence, as each species will have different precipitation niches.
The most important variable for seagrass distribution across all reduced variable models was distance to shore (76.2% importance in BRT).GAM and BRT models placed current velocity and max bathymetry depth as second and third most important variables.
For coral reefs, distance to shore and max bathymetry depth were the most important environmental variables for all reduced variable models, together accounting for 95.4% on average (0.86% SE), though rank of importance differed between modelling methods.

| Interstitial invertebrates
For Scylla serrata, all three MaxEnt models showed that max temperature was the most important variable to explain Scylla's distribution in Oceania (average importance 94.9% ± 4.0 SE).
For all MaxEnt models of Anadara antiquata, distance to shore was the most important variable (average importance 93.2% ± 0.65 SE), explaining almost all of the species' habitat suitability within the models.

| Sea cucumbers
There was a large difference between present-day and future climate models for sea cucumber Bohadschia argus.In present-day models, when chlorophyll was included as an environmental variable, it occurred in the top two most important variables.Calcite was also important in predicting habitat suitability for the species in the present-day MaxEnt models.In the future models, bathymetry and distance to shore were the top two most important variables; however, calcite and chlorophyll were not included as environmental predictors in these models due to a lack of future projections.
For sea cucumber Holothuria atra, all models found current velocity to be an important environmental variable for predicting habitat suitability and distance to shore was also important in most models.
All models found distance to shore to be an important variable in predicting habitat suitability for Holothuria edulis, and future and sea level rise models also listed temperature range.

| Giant clams
Distance to shore and bathymetry were the two most important variables for predicting habitat suitability of Tridacna maxima.MaxEnt models of future climate and sea level rise reported slope as third most important (average 2.1%) variable.BRT models did not include bathymetry as an important variable but did have protected areas as the second most important predictor (average 27.9% important) and current velocity as the third most important (average 13.3%).
Boyce Index values (−0.25) indicate that despite well-fitting AUC and TSS scores for Tridacna gigas MaxEnt models (Table S3), model predictions and species presence points are not well aligned.For GAM models for future climates and sea level rise, bathymetry was most the important predictor of habitat suitability, with distance to shore and slope making up the remaining contributions.
For Tridacna squamosa, there was moderate agreement among models of the most important environmental variables.MaxEnt present-day models identified bathymetry (78.7%) and distance to shore (4%) as the top two most important predictors of habitat suitability, while future climate and sea level regime models found distance to shore to be most important, and current velocity (15.05% ± 1.45 SE) or protected area (15.4% ± 2.2 SE) second or third.BRT models found protected areas most important in presentday models and future models (64.4% ± 2.98 SE) with current velocity also important (20% ± 1.89 SE).Chlorophyll (19.5%) was also found to be important in predicting the present-day habitat suitability; however, it was not included in future models due to a lack of future projections.

F I G U R E 1
Present-day probability of habitat suitability of invertebrate species in Fiji based on MaxEnt models built with a reduced number of climate variables.One indicates suitable habitat, while 0 indicates unsuitable habitat.

| Habitat response to climate change
There is a predicted increase in suitable habitat over MaxSSS and MTP thresholds habitats for mangroves under RCP 2.6 and 8.5.
Mangrove models project a decrease in suitable habitat during RCP 4.5, likely due to the variability of change in rainfall, which was one of the most important environmental variables.Maps of projected change (Figure 2) show that despite overall projected increase in suitable habitat, the change is spatially dependent.Major expansions in suitable habitat are predicted to occur in the Yasawa Islands, where precipitation is likely to increase, while the island of Kadavu projects losses in suitable habitat (Figure 2).
The addition of sea level rise to models projecting mangrove distribution slightly increased the amount of projected increased suitable habitat in the RCP 2.6 and 8.5 (Table 2) scenarios, likely due to the relative change in amount of land area projected to be within the preferred elevation above sea level.
The best-fit models of seagrass showed the greatest disagreement between the change in suitable habitat predicted by MaxSSS and MTP.
Suitable habitat in MaxSSS was projected to be just 3.82% of the overall area, with small but variable change between RCP scenarios.For MTP-based values, 8.13% of the area was projected to be suitable habitat, with loss projected in each RCP scenario.The addition of sea level rise as a variable in models slightly decreases the amount of suitable habitat under MaxSSS thresholds but increases for MTP.
Coral reefs showed little overall response to climate change in our distribution models (Figure 3).BRT and GAM models project slight increases in suitable habitat under future climate scenarios, though model map projections do show a redistribution of suitable habitat.Sea level rise had a very low impact on projected coral distribution in our models.Suitable coral habitat was predicted to remain at a higher percentage of the overall study area than any other biome type for future projections (Table 2).

| Interstitial invertebrates
Scylla serrata, which is often associated with mangrove forests, is predicted to undergo range expansion under future climate scenarios.Under all three modelling methodologies, Scylla was projected to increase in suitable habitat under the RCP 4.5 and 8.5 scenarios.
MaxEnt models, which showed that temperature changes play an important role in the projected distribution, predict that suitable habitat for Scylla will increase by nearly 100% in the RCP 2.6 and 4.5 scenarios and by over 200% in the RCP 8.5 scenario (Figure 4).
Habitat availability for Anadara antiquata showed a variable response to climate change (Figure 4).Anadara habitat is projected to have a variable response to climate projections, with RCP scenarios 2.6 and 8.6 projected to increase suitable habitat but RCP 4.5 decreasing by nearly 20%, potentially due to the nonlinear pattern of salinity change projected in the RCP 4.5 scenario.
Charonia tritonis showed little change in suitable habitat distribution similar to coral reefs.As such, habitat for Charonia is not projected to change under future climate scenarios, and MaxEnt models found it to have an extremely high amount of suitable habitat, with at least 10.6% of the entire study area qualifying as suitable habitat.

| Giant clams response to climate change
For Tridacna maxima, models reported a projected decrease of suitable habitat, especially in the RCP 4.5 scenario.Overall, T. squamosa saw only moderate change in the projected climate change and sea level rise scenarios, though RCP 4.5 did predict loss.Tridacna gigas models were not well fit (Boyce Index −0.25), so these projections are not meaningful.

| Protected area impact
We Protected areas are overlaid in dashed lines (Sykes et al., 2018).Blue ( 1) is an increase from present-day projected suitability, red (−1) is a loss of projected suitability, and 0 is no change.
TA B L E 2 Table of percent of habitat predictions higher than maximizing the sum of specificity and sensitivity (MaxSSS) thresholds and minimum training presence (MTP) thresholds for each of the three habitat-building taxa, using results from Boosted Regression Tree (BRT) models.Note: Model scenarios were Current (present-day projections using an extended list of environmental variables), Current restricted variables (present-day projections using only variables with corresponding future predictions), Projections 2.6, 4.5 and 8.5, and the last four repeated but with the addition of a sea level variable.Values with * designate that for this time period the model was overfit.
F I G U R E 3 Predicted percent change in suitable habitat above maximizing the sum of specificity and sensitivity (MaxSSS) and minimum training presence (MTP) thresholds from present to the climate regimes representative concentration pathway (RCP) 2.6, 4.5 and 8.5 for the years 2070 (mangroves) and 2100 without sea level rise included for habitat models.
the century under RCP 2.6 and 8.5.Precipitation changes in the currently arid Yasawa archipelago and the nearby Mamanuca islands were predicted to drive mangrove expansion to the north.
The Coral Coast, however, decreased in habitat suitability for mangroves in all three scenarios.The mangrove-associated crab Scylla, driven by a preference for warmer temperatures and the possibility of mangrove expansion, will also likely have more suitable habitat in the coming decades.Other studies have noted that mangroves may be able to keep up with rising sea levels (Ellison, 2000) and that mangrove expansion has already begun along the Florida coast.Mangroves offer critical shoreline protection and fishing grounds in Fiji (Brander et al., 2012;Lai, 1984;Walters et al., 2008) and are already threatened by development, deforestation and increasing storm intensity (Alongi, 2008;Giri et al., 2011;Lovelock et al., 2015;Sandilyan & Kathiresan, 2012).In areas where mangroves are projected to decrease, alternative income sources and enhanced storm protection should be focussed to prevent harm to local populations.& Baird, 2000;van Woesik et al., 2011).Our results, which suggest that there will be suitable habitat for at least some of the species that make up Fijian reefs, do not negate the potential negative impact of climate change on the biodiversity of that biome.Additionally, research has shown that while some coral species in Japan have shown poleward range expansion, even up to 14 km per year (Yamano et al., 2011), movement was species-specific.
Climate change will likely have a negative impact on the sea cucumber and giant clam fisheries in Fiji, but response will be speciesspecific.Tridacna maxima and Holothuria atra had the most severe decreases in habitat suitability during RCP 4.5.Efforts to raise Tridacna in aquaculture facility and reseed local habitats should focus species-specific efforts to locations where the clams will continue to thrive far into the future, especially on those species that may be robust to climate change, like T. squamosa.We also expect that many invertebrate species' distributions will be limited more strongly by habitat availability and degradation rates than the environmental abiotic factors such as those included as predictive variables in our models.
For some sessile species, physiological limits on development may prevent further reproduction in changing habitats, or for mobile species like Scylla, direct migration into new habitats may occur.
Mangroves, which reproduce via viviparity and can rapidly spread into newly suitable territories, may naturally spread quickly compared with other species.Coral reefs that experience stress-related bleaching events from rising temperatures may be able to adapt quickly, but phenological changes in reproductive spawning schedules, responding to temperature and seasonal cues (as in plankton, Edwards & Richardson, 2004), interrupt reproductive success.
Moreover, isolated coral reefs, with low Symbiodinium diversity, may not be able to adapt quickly to climate-related stressors (Donner et al., 2005).Rising temperatures may also stress some sea cucumber species, disrupting righting times and other physiological functions (see Holothuria atra Buccheri et al., 2019).
It is unlikely that all species tested in this study will be able to disperse naturally to all future potentially suitable habitats.Despite this, maps of potentially viable habitat are useful for conservation managers looking to prioritize areas for future conservation interventions.Careful consideration should be taken when introducing species into a habitat where they do not naturally disperse, as there may still be unknown barriers to success.However, species with human-assisted migration actions, such as mangrove reseeding efforts and restocking efforts for Tridacna clams, may not need to consider dispersal likelihood in planting maps.
Current protected area designations did not have a strong influence on projected species distributions when included as a model variable.This could be due to the size of each cell projected (5 arcminute), which likely includes some unprotected and some protected waters.Moreover, while they were not the most important factors influencing distribution, protected areas were positively related to distribution for most species.We observed that within current protected areas, many taxa have remarkable predicted change in habitat suitability.However, these losses or increases in suitability may not immediately result in direct changes in distribution.Marine protected areas may remain a vital tool for conservation managers even after suitability declines if there is a lag between habitat suitability changes and population response.Additionally, species distribution and potential distribution is only one factor of protected area delimitation.Habitat condition, fishing pressure, political will, source-sink dynamics and many other considerations may prove more important to species conservation than habitat suitability, and by controlling those factors, species may be able to more quickly adapt to changing environmental conditions.
Our results indicate that there are few areas within Fiji that will serve as climate refugia across all species.In fact, the majority of the areas for which we have data show loss of habitat suitability, especially for foundational species, with the exception of mangroves.
These results therefore suggest that a plan of spatially explicit marine protected areas by themselves will be insufficient on their own to forestall climate-mediated biodiversity losses in future.This is not to say however that marine protected areas have no value, as they can alleviate numerous stressors that exacerbate species' declines, and in doing so may provide a more 'fair fight' for those systems against climate change (Williams et al., 2019).Ultimately, our work suggests that the multitude of ecosystem services, including cultural provisioning, may change in a warming world and that we need to consider these losses through the lens of a more integrated and dynamic human/natural system (Woodhead et al., 2019).

| Data limitations and challenges
We relied on the accuracy of GBIF contributors for our invertebrate distributions, which may not be completely accurate despite our cleaning process.Many invertebrate species do not have robust abundance data records for Fiji and Oceania broadly, despite their relative importance to local communities.Two of our models (Scylla and T. gigas) were built using species occurrence data from eastern Australia, where a higher species richness may alter biological interactions between species and therefore distribution.This dearth of research highlights the need for invertebrate censuses in the area to better understand how biodiversity will be impacted by changing climates in the tropics.In this study, presence points from Australia for T. gigas did not align with model predictions (Boyce Index −0.25), and predictions for Fiji could be made more trustworthy with increased biodiversity surveys there.
Similarly, there may be issues related to sampling of presence points from distribution polygons which may result in false positives, as was executed for coral (Hamilton & Casey, 2016) and mangrove (UNEP-WCMC, 2021) biomes.However, as these categories mapped the entire biome regardless of species-specific distributions, expert-derived distributions of entire habitats are more likely to be accurate as they depict reasonable continuous coverage.
Moreover, range maps used in this study were based on Landsat and global remote sensing data, rather than expert delineations around known occurrence points akin to IUCN maps (Herkt et al., 2017).
Other research has shown that presence points derived from polygons perform similarly well to other publicly available point databases (Fourcade, 2016).
There are many potential issues to be aware of when interpret- there is no single answer to protecting complex systems and increasing our understanding of each single taxa is vital to its protection.

ACK N O WLE D G E M ENTS
We would like to thank the staff of WCS Fiji, especially Sangeeta Mangubhai for their direction during the planning stages of this process and help identifying important species of interest.We would also like to thank Eric Treml for advice during the analysis stage and Márcia Barbosa for training on the implementation of SDMs in R.
Many other people gave advice during the planning and analysis stage, especially Kate Henderson and Leah Rubin.

FU N D I N G I N FO R M ATI O N
We also recognize our funding sources at SUNY-ESF that supported the project, including an Independent Exploration Grant.
This material is based upon work supported by the National Science Foundation 2020 Graduate Research Fellowship under Grant No. 85697.
based on global remote sensing 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).

Future
predictions under climate change were compared with the location of key protected areas to investigate how protection coverage of each habitat and species range would change under future climate scenarios.Fiji has a series of locally managed marine areas (LMMAs) and traditional fishery management units qoliqoli, which are used to temporarily restrict access to threatened marine environments and implement a wide range of habitat and population protections.Based on the patterns of our results and local importance, we focus on four major areas: Namena Marine Reserve, Qoliqoli Cokovata, the Yasawa Islands and the Coral Coast.The Namena MarineReserve primarily encompasses a barrier reef and is Fiji's largest notake reserve(Goetze & Fullwood, 2013).Qoliqoli Cokovata protects the northern coast of Vanua Levu and a major section of the Great Sea Reef, as well as extensive seagrass and mangrove habitat.It was designated as Ramsar site in 2018(Madigibuli, 2020).The Yasawa Islands span northward from the western shore of Viti Levu and are an important centre for tourism(Comley et al., 2005;Murphy et al., 2018).The Coral Coast is protected by a series of LMMAs that span most of the southwestern coast of Viti Levu and encompass an extensive coral reef and includes Shark Reef marine protected area (Ward-Paige et al., 2020).
further visually examined projected change in a few critically important habitats and places which saw extensive projected change from our models.Areas in the Coral Coast, Yasawas, and in Qoliqoli Cokovata will likely see some change in habitat suitability for important fisheries species.Under future climate change scenarios, the Coral Coast was predicted to undergo a decrease in habitat suitability for mangroves (Figure 5) and seagrass towards the westernmost edge.Qoliqoli Cokovata was predicted to have lower habitat suitability for corals under future climate change scenarios (Figure 5), as well as lower habitat suitability for Tridacna maxima.The mud crab Scylla serrata is projected to increase in suitability widely, particularly in the Yasawas.In Namena, coral decreases in viability slightly, though not as extreme as in other areas of Fiji, and Bohadschia argus and T. maxima are projected to decrease in the RCP 4.5 scenario.4 | DISCUSS ION All species and biomes we analysed in this study had remarkably different predicted responses to climate change projections.Current important marine protected areas are likely to experience highly localized reactions to change; for example, Qoliqoli Cokovata is projected to see major decreases in the distribution of suitable coral habitat due to rising temperatures while mangrove suitability is projected to decrease along the Coral Coast and move instead into the Yasawa Islands.Seagrass meadows had variable projections, with RCP 4.5 MaxSSS values predicting a decrease in suitable habitat, but low-to-moderate rise in other scenarios.The focal species in this study are vital parts of the marine coastal system, and predicted shifts in habitat suitability under climate change can be used to better implement efficient monitoring networks.Our results suggest that climate change may cause a substantial expansion in Fiji in suitable habitat of mangroves by the end of F I G U R E 2 Change maps showing gains and loss in predicted habitat suitability in representative concentration pathway (RCP) scenarios 2.6, 4.6 and 8.5 for Boosted Regression Tree (BRT) models built using future climatic variables in years 2070 (mangroves) and 2100.
Seagrass, on the contrary, is predicted to have reduced habitat suitability under RCP 4.5, especially along the Coral Coast of Fiji, in the Mamanucas, and on the west of Viti Levu near Nadi, as current velocities are altered by changing environments.All models estimated that seagrass distribution is strongly related to distance to shore and max depth, likely due to light attenuation, with seagrass appearing in a small halo of distribution in shallow, nearshore habitats.A narrow band of suitable current velocities also controls the prevalence of seagrass, and if sea level rise changes the location of this band, it will be important to ensure that seagrass meadows are able to keep up with the rate of change.Seagrass meadows in Fiji have been relatively quick to respond to disturbance, suggesting potential for range movement (McKenzie et al., 2021).Seagrass biomes are crucial parts of the life cycle for many fisheries species in Fiji (e.g.emperors Lethrinidae and snappers Lutjanidae, Eggertsen et al., 2022), and local governments should manage run-off in places where seagrass will be most threatened, to prevent turbid and polluted waters from further impacting seagrass meadows.F I G U R E 4 Predicted percent change in suitable habitat above maximizing the sum of specificity and sensitivity (MaxSSS) and minimum training presence (MTP) thresholds from present to the climate regimes representative concentration pathway (RCP) 2.6, 4.5 and 8.5 for the year 2100, without sea level rise included, for invertebrate taxa.Original illustrations by the authors.F I G U R E 5 Change in projected suitability from the present day for Scylla serrata in the Yasawa Islands for 2100 (1st row); mangroves along the coral coast for 2070 (2nd row); coral change in Qoliqoli Cokovata for 2100 (above Vanua Levu) and Namena (below; 3rd row); and Tridacna maxima (4th row) for 2100 under the three projected climate regimes.Our models did not predict a net loss of suitable coral habitat due to climate change or sea level rise; however, the location of the most suitable habitat will likely change, and coral habitat in Qoliqoli Cokovata may be especially threatened.Additionally, many studies have shown that different coral species will likely react very differently to warming temperatures and acidifying oceans (e.g.Marshall ing our model results.The number of input environmental variables did significantly change the projected suitable habitat in many cases, whether by lacking an important variable (like chlorophyll and calcite for Bohadschia argus, and calcite for Charonia tritonis) or by reducing the redundancy and preventing overfitting.Moreover, the range of possible values of the input environmental variables may not initially include values that RCP scenarios project may be possible in future, such that the model is extrapolating to values not yet observed in the study space.This can result in the model missing known biological limits, but the initial selection of a spatially limited study space is critical to ensuring background absences do not result in overfit and overconfident models.Finally, the resolution of our environmental variables was limited to 5 arcminutes, as no finer resolution data were available for some of the datasets (i.e. for future models of marine environmental variables under climate change scenarios).The accuracy and specificity of species distribution studies would be substantially improved by using finer resolution maps, and the creation and dissemination of such should therefore be a high priority for Pacific data managers.5| CON CLUS IONSSpecies 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.For species with little initial distribution data, like understudied marine invertebrates, models can greatly improve our current understanding of their potential range.Habitat suitability in many of the strongest marine protected areas in Fiji is likely to radically change for many species in the upcoming decades and tracking the possible range shifts or population declines will be important for many species studied here.Direct action conservation efforts, like mangrove reseeding and giant clam nurseries, should use model results to focus geographic efforts, while industries like sea cucumber harvesters may preferentially focus on species that may be resilient to climate change.New protected area designations should consider the future distribution of species in Fiji based on climate change to best maximize future benefits to those taxa, and managers of current protected areas, like Qoliqoli Cokovata, should heighten efforts to be prepared for potential coral losses due to climate change.The variability between species and among climate scenarios indicates