Modelling present and future distribution of intertidal macroalgae to evaluate the efficacy of Western Australia’s marine protected areas

The coasts of Western Australia are highly diverse in their characteristics. The coastline extends 20,781 km (~ 34.8% of the total coastline of Australia) (Edyvane 2005) from tropical to temperate regions with a difference in sea surface temperature of ~ 10 °C (Wijffels et al. 2018). In the north, the coast is macrotidal (a tidal range of about 8 m) with low significant wave height (SWH) (~ 1 m), while in the south, it tends to be microtidal (a tidal range of less than 1 m) with high SWH (4 m) (Bosserelle et al. 2012; Harker et al. 2019). The coast is also oligotrophic as the Leeuwin Current delivers low nutrient waters from the tropics, with little riverine inflow (Hanson et al. 2007; Molony et al. 2011; McLaughlin et al. 2019). Rocky intertidal shores constitute about 19% of the Western Australian coastline (Edyvane 2005) that is interrupted by sandy beaches or mangroves (Wilson 2013). Those shores are karstified Pleistocene limestone with different profiles and erosional features depending on the coastal orientation to wind and waves, climates, and riverine influx (Semeniuk and Johnson 1985).

Macroalgal samples were collected from 14 localities, between 18°S and 34°S, keeping 1° latitudinal distance between localities to represent the coastline of Western Australia and changes in sea surface temperature, from September 2020 (spring) to January 2021 (summer) during low tide (Fig. 1, Table S1). At each locality, one to three sites with horizontal limestone rock platforms were selected as these habitats are often dominated by macroalgae (Bessey et al. 2019). The number of sites and distance between them in each locality differed depending on availability and accessibility of rock platforms. We could not access rocky intertidal shores between 23°S and 27°S due to high cliffs. In total, there were 38 sites across 14 localities.

At each site, three transects separated by 25–50 m were set out perpendicularly to the shoreline. Along each transect, three quadrats of 0.2 m x 0.2 m were haphazardly placed at the inner, middle, and outer of the rock platform to capture macroalgal variations toward offshore. The length of the transects differed among sites following the length of the rock platform at those sites. Previous studies have also used the same size, number, and placement of quadrats to analyse diversity and density patterns of macroalgae at regional scales (Pereira et al. 2006; Wieters et al. 2012; Lathlean et al. 2015). Macroalgae inside those quadrats were scraped and preserved in 95% ethanol in the field. In the laboratory, we identified whole-plant macroalgae to species level and followed Open Nomenclature qualifiers (e.g., sp. 1, sp. 2, etc.) if species names were unknown (Sigovini et al. 2016), then the wet biomass of each species was measured. Macroalgal data were collated at the transect level to account for the variation of macroalgae along the rock platform. Species found in more than 10% of total transects (n = 11) were then selected. This number falls between the minimum required number of occurrences for narrow-ranged species (n = 3) and widespread species (n = 13) to develop accurate species distribution modelling (van Proosdij et al. 2016).

A set of nineteen environmental variables, including physical factors (cloud cover, significant wave height, tidal amplitude, surface current velocity, and photosynthetically available radiation), sea surface temperature derivatives (mean, minimum, maximum, mean in the coldest month, mean in the warmest month, and range), and chemical factors (salinity, dissolved oxygen, pH, nitrate, phosphate, particulate organic carbon, particulate inorganic carbon, and total suspended matter), was selected as those variables often explain the biogeographical distribution of rocky intertidal assemblages (Fenberg et al. 2015; Ibanez-Erquiaga et al. 2018) and have been used as environmental predictors to model present and future distribution of intertidal macroalgae (Jueterbock et al. 2013). Cloud cover, surface current velocity, minimum and maximum sea surface temperature, dissolved oxygen, pH, nitrate, and phosphate data at the resolution of 5 arc-minutes were downloaded from Bio-ORACLE (Tyberghein et al. 2012); significant wave height data at the resolution of 5 arc-minutes were downloaded from Copernicus Marine Service (https://​marine.​copernicus.​eu/​); tidal amplitude, particulate inorganic carbon, and total suspended matter data at the resolution of 5 arc-minutes were downloaded from Global Marine Environment Datasets (http://​gmed.​auckland.​ac.​nz); mean sea surface temperature, mean sea surface temperature in the coldest month, mean sea surface temperature in the warmest month, and range of sea surface temperature data at the resolution of 30 arc-seconds were downloaded from MARSPEC (Sbrocco and Barber 2013); photosynthetically available radiation and particulate organic carbon data for 2020 at the resolution of ~ 2.40 arc-minutes were downloaded from the Aqua MODIS satellite (http://​oceancolor.​gsfc.​nasa.​gov). Environmental data with 30 arc-second and ~ 2.40 arc-minute resolutions were adjusted into 5-arc minute resolution using bilinear interpolation to equalise the spatial resolution of all variables.

The collinearity (non-linear correlations) among environmental variables was assessed using Spearman’s rank-order correlation analysis. Some environmental variables were used as surrogates for other variables that were highly correlated (|r_s| > 0.90). Maximum sea surface temperature, mean sea surface temperature in the warmest month, dissolved oxygen, and phosphate were excluded as these variables were highly correlated with mean sea surface temperature. Mean sea surface temperature in the coldest month and salinity were also excluded as these variables showed high correlations with minimum sea surface temperature and photosynthetically available radiation. We retained mean and minimum sea surface temperature and also photosynthetically available radiation as surrogates for maximum sea surface temperature, mean sea surface temperature in the coldest and warmest months, dissolved oxygen, and phosphate, and salinity (Table S2) because the former parameters often explain a high proportion of the distribution of rocky intertidal assemblages (Fenberg et al. 2015; Ibanez-Erquiaga et al. 2018; Hadiyanto et al. 2023). Thus, 16 environmental variables were used as predictors.

We developed species distribution models using occurrence and biomass data. Biomass data were logarithmic transformed (log(x + 1) to meet the assumption of data normality before the model was constructed. Models were built based on hierarchical modelling of species communities (HMSC) using probit regression for occurrence data due to binomial (presence-absence) data and linear regression for biomass data due to continuous data (Ovaskainen et al. 2017). The transect was included in those models as a spatial random factor within sites. We developed occurrence and biomass models using R-package Hmsc 3.0 (Tikhonov et al. 2020). HMSC was selected because this algorithm can model multiple species distributions simultaneously to assess assemblage-level patterns, allow spatial or temporal random factors, and provide the evaluation of species associations to capture biotic interactions and the effect of missing environmental covariates. Thus, HMSC recognises the multivariate nature of assemblages by allowing species to respond the environment jointly as well as to associate each other (Ovaskainen et al. 2017; Ovaskainen and Abrego 2020; Tikhonov et al. 2020). This algorithm has also been used to model the distribution patterns of terrestrial and aquatic assemblages (Deflem et al. 2021; Antão et al. 2022; Stark and Fridley 2022).

We ran two Monte Carlo Markov chains (MCMC) for 150,000 iterations each to achieve the model convergence. The first 50,000 samples were discarded as transient. The remaining samples (100,000 samples) were thinned by 100, which means discarding all samples but saving every 100th sample, to yield 1,000 posterior samples per chain (Tikhonov et al. 2020). The convergence of models was evaluated by computing Gelman-Rubin diagnostic. This measure compares the estimated between- and within-chain variances for each model parameter, with a small difference between those variances indicating convergence (Gelman and Rubin 1992).

The predictive power (accuracy and discrimination) of models was assessed based on the five-fold cross-validation (here referred to as internal cross-validation). In this validation, original samples were randomly divided into five equal sized subsamples, with one subsample retained for testing the model and the remaining four subsamples used for training the model. This process was repeated five times, with each subsample used once for testing the model, and then the five results were averaged to yield a single estimation. The accuracy of occurrence and biomass models (how far the data are from the best predictions) was measured using root mean square error (RMSE), with a lower RMSE indicating a more accurate model. The discrimination of occurrence model was measured using the area under the receiver operating characteristic curve (AUC) and the coefficient of discrimination (Tjur R²) due to presence-absence data with probit regression. Despite different units, AUC and Tjur R² indicate how well the occurrence probabilities differentiate whether sampling units are occupied or not, with higher AUC (more than 0.5) and Tjur R² indicating that the model discriminates better empty and occupied sampling units. The discrimination of biomass model was measured using the coefficient of determination (R²) due to continuous data with linear regression. R² compares the correlation between predicted and observed values of linear regression, with a higher R² indicating that the model predicts better (Ovaskainen and Abrego 2020).

As a supplementary analysis, the predictive power of the occurrence model for species that could be assigned a Linnean name was tested using external datasets (here referred to as external cross-validation). HMSC requires both presence and absence data. In this validation, presence data of those species were downloaded and compiled from Ocean Biodiversity Information System (OBIS 2021), Global Biodiversity Information Facility (GBIF.org 2021), and Atlas of Living Australia (ALA 2021) on 6 November 2021. Duplicate species within location records were removed. We used absence data from fieldwork and 1,000 pseudo-absences that were generated randomly within the study area (13–35°S and 112–125°E) for each species to account false absence data.

The relative importance of environmental groups, i.e., physical factors (cloud cover, current velocity, tidal range, and significant wave height), sea surface temperature derivatives (mean, minimum, and range), and chemical factors (nitrate, photosynthetically available radiation, pH, particulate inorganic carbon, and total suspended matter), and also transect (as a random effect) in explaining species occurrences and biomass was evaluated by measuring variation partitioning for each species. The variation partitioning was the coefficient determination (R²) of linear regression on each group of explanatory variables in the model, ignoring the covariances among those groups. Residual correlations between species (i.e., the amount of covariation among species, after accounting for the effect of explanatory variables) were also measured to assess the influence of species associations on occurrence and biomass distributions.

Models were applied to predict occurrence probability and biomass of macroalgae along the coast of Western Australia from 18°S to 34°S using 16 environmental values (at the resolution of 5 arc minutes or ~ 9.2 km in tropics) as inputs. We only selected environmental values from coastal cells to represent intertidal habitats, thus there were 324 new cells to be predicted. The linear correlation between predicted occurrence probability and biomass for each species was analysed using the Pearson correlation test to determine whether those parameters showed similar pattern or not.

The future distribution of macroalgae with respect to two climate change scenarios (2050 and 2100) was predicted using the same set of environmental variables to model the present distribution of macroalgae, but the value of sea surface temperature derivatives (mean, minimum, and range) was changed following those climates. The future scenarios were based on the fifth phase of the Coupled Model Intercomparison Project (CMIP5) (Taylor et al. 2012) and two representative concentration pathway scenarios (RCP) for each climate: peak and decline (RCP2.6) and increase (RCP8.5) in emissions (Moss et al. 2010). Future sea surface temperature derivatives were downloaded from Bio-ORACLE (Assis et al. 2018).

Observed prevalence (the proportion of presences relative to the total number of observations) was used as a threshold to transform the probability of occurrence into binary presence-absence data. The prevalence approach is the best threshold of occurrence in terms of sensitivity, specificity, overall prediction success, and Kappa (a measure to correct the overall accuracy of model predictions by the accuracy expected to occur by chance) (Liu et al. 2005). We intersected the map of presence-absence and biomass to define the distribution area of each species. All sites of a given species that were predicted to contain species presence and positive biomass data were assigned to be the distribution area of that species. All species distribution areas were stacked to analyse species richness, total biomass, species composition, and bioregions of macroalgae.

Changes in the species composition of macroalgae between present and future climate within cells were calculated based on the Bray-Curtis dissimilarity index using the unweighted pair-group method with arithmetic averages for logarithmic-transformed (log(x + 1)) species biomass data. The dissimilarity between cell-based macroalgal composition and latitudinal species pool (i.e., all species within latitudes) was also quantified based on that index to determine the spatial homogenisation of macroalgae. We calculated differences between dissimilarities in the present and future climate to estimate changes in spatial homogenisation over future climate scenarios. Reduction in dissimilarity index means that the spatial homogenisation of macroalgae increases in the future climate (that is, the composition of macroalgae gets similar in the future climate) and vice versa.

We also determined whether a species moves poleward or equatorward in the future climate by comparing the latitude of distribution centroid (DC) (that is the mean latitude of a species within the study area) between present and future climates. The latitude of distribution centroid for each species was calculated based on the approach of Cheung et al. (2012):

where Li is the latitudinal coordinates of a cell (decimal degree), Bi is the species’ biomass of a cell (g), and n is the total number of cells within the study area (324 cells).

We evaluated the efficacy of Western Australia’s MPAs by analysing two parameters: (1) the effectiveness of the MPAs in protecting macroalgal distribution and (2) the representativeness of the MPAs in describing macroalgal bioregions. Boundary coordinates of MPAs (marine parks) were obtained from the Department of Biodiversity, Conservation and Attractions (DBCA) (https://​www.​dpaw.​wa.​gov.​au). To assess the effectiveness of the MPAs, we overlaid the map of species distributions with the MPA boundaries, then species distribution areas inside and outside those MPAs were calculated. To evaluate the representativeness of the MPAs, we overlaid the map of macroalgal bioregions with the MPA boundaries, then the proportion of those MPAs within the bioregions was calculated. We also calculated the proportion of total species inside MPAs within the bioregions to examine the adequacy of those MPAs to protect macroalgal diversity. In this analysis, macroalgal bioregions were delineated by a hierarchical cluster analysis based on the Bray-Curtis dissimilarity index using the unweighted pair-group method with arithmetic averages method for logarithmic-transformed (log(x + 1) species biomass data to meet the assumption of data normality. We used biomass data in the delineation of bioregions to account commonness and rarity patterns. The number of clusters was determined using the silhouettes approach. This approach measures the distance from each point in one cluster to points in the adjacent clusters, with a high value indicates that the object is matched better to its own cluster than to the adjacent clusters (Rousseeuw 1987). The separated clusters were then assigned as distinct bioregions. Bioregions that comprised less than five cells were joined to adjacent bioregions.

In total, we collected 34,384.59 g macroalgae belonging to 187 species (Chlorophyta: 43 species, Ochrophyta: 61 species, Rhodophyta: 83 species). Thirty six species (Chlorophyta: 4 species, Ochrophyta: 17 species, Rhodophyta: 15 species) (Fig. S1) were found in more than 10% of total transects, with the highest occurrence (50% of total transects) observed for Jania micrarthrodia, followed by Lobophora variegata (48% of total transects) and Sirophysalis trinodis (45% of total transects). Based on the selected species, the number of species per transect ranged from 1 species to 22 species and the total biomass per transect ranged between 3.67 g and 266.74 g. Further, we used these 36 species to develop occurrence and biomass models of macroalgae.

Occurrence and biomass models achieved the MCMC convergence, with the potential scale reduction factors close to the ideal value of one (Fig. S2). Accuracy and discrimination of occurrence model were higher than those of biomass model. Based on internal cross-validation, occurrence model showed a RMSE of 0.31 ± 0.01, an AUC of 0.86 ± 0.02, and a Tjur R² of 0.38 ± 0.03, while biomass model showed a higher RMSE (0.54 ± 0.05) and a lower R² (0.28 ± 0.03) (Table S3). Supplementary analyses showed that occurrence model had a RMSE of 0.48 ± 0.04 for external cross-validation with absence data from fieldwork and 0.46 ± 0.03 for external cross-validation with 1,000 pseudo-absences, an AUC of 0.71 ± 0.03 for external cross-validation with both absence data from fieldwork and 1,000 pseudo-absences, and a Tjur R² of 0.22 ± 0.05 for external cross-validation with absence data from fieldwork and 0.05 ± 0.06 for external cross-validation with 1,000 pseudo-absences. The coefficient of correlations between predicted occurrence probability and biomass of macroalgae was less than 0.90, except for Caulocystis cephalornithos (r = 0.94) (Table S4).

Environmental variables explained approximately 91.93% variation in species occurrences and 81.11% variation in species biomass (Fig. 2a). Sea surface temperature derivatives explained the largest proportion of the variation in species occurrences (34.19%) and chemical factors were the most important predictor in explaining the variation in species biomass (31.48%). However, environmental variables that explained the largest proportion of the variation in macroalgal distribution differed between species. The best environmental variables in explaining species occurrences were sea surface temperature derivatives for 22 species, physical factor for 7 species, chemical factor for 5 species, and transect random factor for 2 species. The best environmental variables in describing species biomass were chemical factor for 17 species, physical factor for 6 species, transect random factor for 6 species, and sea surface temperature derivatives for 5 species. Occurrence and biomass in some species were explained best by same environmental variables, namely physical factor for Dictyota sp1; sea surface temperature derivatives for Padina sp1, Sporochnaceae 1, Laurencia sp1, and Laurencia sp2; chemical factor for Halimeda macroloba and Sargassum sp10; and transect random factor for Lobophora variegata.

Strong relationships between macroalgae and environmental variables (posterior probability of more than 0.9) were detected more often in occurrence model (Fig. 2b). Further, strong environmental relationships were observed more often in Amphiroa sp1 and Jania rosea based on occurrence model (12 variables) and Gelidiella acerosa based on biomass model (8 variables). Nevertheless, species occurrences had different responses to environmental variables, with most species showing negative relationships with mean sea surface temperature (15 species) and particulate organic carbon (15 species), and positive relationships with tidal amplitude (17 species). Similarly, species biomass had different responses to environmental variables, with most species showing negative relationships with particulate organic carbon (11 species) and particulate inorganic carbon (11 species).

Residual species correlations with posterior probability of more than 0.9 were detected more often in biomass model than in occurrence model (Fig. 3). Based on occurrence model, Laurencia sp2, Champia sp1, Sarconema sp1, Dictyota sp2, Sporochnaceae 1, Sargassum sp20, Hydroclathrus clathratus, Hypnea cornuta, Padina sp1, L. variegata, Sargassum sp16, Sirophysalis trinodis, Jania micrarthrodia, and Leveillea sp, positively co-occurred each other. Based on biomass model, those species also positively co-occurred each other but negatively co-occurred with filamentous Chlorophyta 1; Laurencia sp1; Caulocystis cephalornithos; Jania verrucosa, and Ulva sp1.

In the present climate, species richness and biomass of macroalgae were higher in temperate regions (23.5°S to 35°S) than in tropical regions (18°S to 23.5°S). Species richness was 7.93 ± 0.25 species/transect in tropical regions and 14. 73 ± 0.30 species/transect in temperate regions (Fig. 4a, Table S5). Total biomass was 40.38 ± 4.66 g/transect in tropical regions and 52.82 ± 3.59 g/transect in temperate regions (Fig. 4b, Table S5).

Species richness and biomass of macroalgae were projected to decrease under all future climate change scenarios. Species richness was predicted to decline by about 19% (RCP2.6) to 26% (RCP8.5) by 2050 and about 18% (RCP2.6) to 49% (RCP8.5) by 2100 (Fig. 5a, Table S6). Reduction in species richness will be 2–4 times faster in temperate regions (23.5°S to 35°S). Biomass was expected to increase by approximately 6% (RCP8.5) to 7% (RCP2.6) by 2050 and about 5% (RCP2.6) to 21% (RCP8.5) by 2100 in tropical regions (18°S to 22°S) but decrease by approximately 6% (RCP2.6) to 11% (RCP8.5) by 2050 and about 3% (RCP2.6) to 48% (RCP8.5) by 2100 in temperate regions (23.5°S to 35°S) (Fig. 5b). As a result, total biomass for the Western Australian coastline was projected to remain stable under RCP2.6 scenario but reduce by about 4% by 2050 and 20% by 2100 under RCP8.5 scenario (Table S6).

The species composition of macroalgae was projected to change in future climates. The average dissimilarity index of macroalgae between present and future climates was predicted to be 0.21 (RCP2.6) to 0.22 (RCP8.5) by 2050 and 0.20 (RCP2.6) to 0.43 (RCP8.5) by 2100 (Fig. 6a, Table S5). Reduction in dissimilarity index (that is, increased spatial homogenisation) was predicted to occur in 46–56% of total cells, especially within the latitude of 18°S to 21°S and the latitude of 25°S to 34°S (Fig. 6b, Table S5). However, the latter latitudinal range will show increases in dissimilarity index (that is, decreased spatial homogenisation) under 2100 RCP8.5 scenario. Increases in dissimilarity index will also be found within the latitude of 22°S to 24°S under 2050 RCP2.6, 2050 RCP8.5, and 2100 RCP2.6 scenarios but not under the 2100 RCP8.5 scenario.

Species biomass, distribution area, and distribution centroid of macroalgae were all projected to change in the future climate (Fig. 7, Table S7). More than 60% of species were expected to decrease in biomass by about 40% (RCP2.6) to 45% (RCP8.5) by 2050 and 40% (RCP2.6) to 65% (RCP8.5) by 2100. The distribution area of more than 60% of species was also predicted to reduce by about 29% (RCP2.6) to 33% (RCP8.5) by 2050 and 26% (RCP2.6) to 70% (RCP8.5) by 2100. Poleward shifts will occur in more than 60% of species, with the change in distribution centroid of about 0.32° (RCP2.6) to 0.42° (RCP8.5) by 2050 and 0.32° (RCP2.6) to 0.80° (RCP8.5) by 2100. Amphiroa gracilis, Canistrocarpus cervicornis, Hypnea cornuta, and Sarconema sp1 were expected to go extinct by 2100 under RCP8.5 scenario. In contrast, Caulerpa racemosa, Eucheuma sp., Gelidiella acerosa, and Sargassum sp10 will move poleward and show increases in biomass and distribution area.

In the present climate, Western Australia’s MPAs protect about 43% of the distribution area of macroalgae, with nine species occurring more often inside MPAs (i.e., more than 50% distribution area was inside MPAs). The distribution area of macroalgae inside MPAs was projected to be around 46–47% by 2050 and 47–52% by 2100, with 13–15 species present more often inside MPAs than outside (Fig. 8).

The distribution of macroalgae could be delineated into nine bioregions under the present climate scenario (Fig. 9). The coverage of MPAs was high (50–100%) within bioregions from 22°S to 26°S and from 33°S to 34°S but low (12–43%) within bioregions from 18°S to 21°S and from 25°S to 31°S. More than 65% of total species within those bioregions were inside MPAs. However, bioregions were projected to be fewer (five to seven) in future climates. In addition, a bioregion from 25°S to 31°S will be separated into two distinct regions (25°S to 29°S and 30°S to 31°S) under 2100 RCP8.5 scenario. There was no MPA within the northernmost bioregion, and all studied species inside the MPA within the southernmost bioregion will be extinct under that scenario.