Isolating the impacts of warming and acidification
Areas where projected coral reef habitat suitability is most critically degraded by ocean surface warming (Fig. 2) correspond to areas with the highest mean annual SSTs today (Fig. 1), and also map onto regions identified as being particularly susceptible to future coral bleaching (e.g., Guinotte et al., 2003; Donner et al., 2005; van Hooidonk et al., 2013). We find that the projected decrease in suitability is most strongly associated with the weekly maximum – and to a lesser degree, minimum – SST variable, and very high values (>32 °C) are linked to marked declines (see Figs. S1.7–S1.8 in Data S1 and Fig. S2.5 in Data S2). Both BRT and MaxEnt statistical modeling approaches assign a relatively high significance to maximum weekly SST; in 10 randomly generated versions of MaxEntOPT/BRTOPT, this parameter ranked 5th/8th of a total of 27 parameters in average relative contribution to model output (however, correlations between variables do make it difficult to clearly establish hierarchy; Couce et al., 2012). The importance placed by the Bioclimatic Envelope Models on SST variables in general, and the negative contribution to modeled suitability at the higher end of the SST range, indicates that the global data set used for model training contains sites where present-day coral reef distribution is already limited by an upper thermal threshold. Although equivalent maximum SSTs are tolerated by Red Sea and Arabian Gulf coral reefs, these sites could not be included in the training data set and analysis because it was not possible to simulate conditions using the UVic model in these enclosed seas. It has been suggested that Red Sea and Arabian Gulf coral are potentially conditioned to extreme SSTs by very high SST variability (Ateweberhan & McClanahan, 2010), however, this is not a characteristic shared with the equatorial regions projected to decline in habitat suitability (Lough, 2012). Adding support to our model projections of an equatorial retraction, in addition to a poleward expansion, is the matching distribution range shift identified in fossil records of coral distribution during the last interglacial period, when the global average temperature exceeded that of today (Kiessling et al., 2012).
The areas most affected by reductions in ΩArag levels correspond to off-equatorial latitudes, particularly in the western Pacific (Fig. 3 and 4b). Generally highly stratified, therefore with little buffering by mixing with subsurface waters, these regions will tend to experience the strongest degree of saturation decline. In addition, these are slightly cooler waters, where other environmental factors, including ΩArag, are likely to out-weigh temperature variables in terms of importance. For instance, Cooper et al. (2012) argued that, at present, SST was far more significant than ΩArag to explain long-term changes in calcification rates for massive Porites colonies off Western Australia, whereas the GBR has been identified, in modeling studies (McNeil et al., 2004) and skeletal measurements in massive Porites colonies (e.g., Cooper et al., 2008; De'ath et al., 2009), as a region already exceeding the thermal optima for coral calcification and sensitive to reduced ΩArag.
In contrast, our results suggest higher latitude sites (e.g., southern GBR, southern Japan) and upwelling regions (e.g., East Pacific) will be relatively little affected by the acidification resulting from the projected increases in atmospheric CO2. The suitability for reef presence at 1990 levels is maintained in these locations up until 2070 across the majority of the models (Fig. 3). These areas experience the lowest ΩArag levels at which reefs are found today and are considered marginal for reef formation (e.g., Smith, 1992; Kleypas et al., 1999a; Glynn et al., 2007; Manzello et al., 2008) due to low calcification and cementation rates, which are finely balanced with erosion processes (Glynn, 1976; Cortes, 1997). Previous studies (e.g., Hoegh-Guldberg et al., 2007; Cao & Caldeira, 2008; Manzello, 2010) have identified coral reefs from higher latitudes and upwelling regions as the most sensitive to future reductions in ΩArag. However, statistical modeling has indicated that SST variables and light availability may be stronger determinants of coral reef presence at these marginal sites than ΩArag (Couce et al., 2012). Our future projections reflect this balance of controls. In addition, upwelling of deeper waters, isolated from the atmosphere, leads to a substantial degree of pCO2 disequilibrium between ocean surface and atmosphere and may hence provide some buffering against ocean acidification in regions such as the Eastern Pacific. Fabricius et al. (2011) observed changes in coral reef communities along a natural pH gradient on volcanic seeps at Papua New Guinea, and found that despite significant losses in species composition and biodiversity, coral cover remained constant for acidification levels comparable to what is expected by 2070 under the A2 scenario. This was achieved through a transition from branching species such as Acropora sp. toward predominance of more massive Porites sp. corals, whose calcification rates were less affected. Such shifts in species composition could have significant implications for ecosystem services; however, this study models coral reef ecosystems as a single entity, and is not sensitive to changes in species composition or biodiversity losses.
Synergistic impacts of warming and acidification
Under the combined influences of ocean surface warming and acidification, habitat suitability for coral reef ecosystems declines across the latitudinal band between 20°N and 20°S, affecting significant reef areas such as the Caribbean, GBR, and Coral Triangle region (Fig. 4c and 5). Areas where shallow water substrate is available are particularly susceptible to the impacts of warming and acidification (e.g., in Fig. 4d, modeled suitability in shallow equatorial regions more than halves by 2070). Although a marginal improvement at higher latitudes is projected by 2070, range expansion is constrained by availability of shallow waters where benthic substrate is present within light penetration depths (Fig. 5). Range expansion is also projected onto areas at the extreme end of oceanographic circulation pathways along the western boundaries, where the most depauperate reefs are found today (e.g., Glynn et al., 2007; Macintyre, 2003; Veron & Minchin, 1992; but see Thomson & Frisch, 2010). Coral reef expansion to higher latitude sites with improved conditions will, therefore, depend on larval influx, introducing a likely limitation for sites situated away from current transport such as the south-eastern Pacific (Wood et al., 2013). The rate of projected range expansion of suitable environmental conditions for coral reef presence varies between models. The MaxEntOPT model predicts a moderate (ca. 5 km per decade) expansion at present, accelerating to 30–45 km per decade by 2070 (Fig. 4d; calculated as rate at which curve edges intersects a horizontal line at 0.2). In contrast, BRTOPT projects an initial global shrinkage under the effects of acidification (at ca. 25 km per decade), with increasingly rapid expansion from 2020–2030 as the effect of rising SST becomes dominant (up to 70–80 km per decade by 2070). For comparison, Burrows et al. (2011) found the average poleward speed of surface isotherm movement in the oceans (50°S to 80°N) over the last 50 years was 27.5 km per decade. As already stated, our coral reef range expansion rate estimates do not take into account connectivity limitations related to the dispersal of coral larvae by ocean currents, and thus follow the thermal shift tempered by ocean acidification.
The combined impacts of increasing SST and acidification have little impact on environmental suitability for coral reef presence in the eastern Pacific, south east Atlantic, and the north Brazilian coast over the coming decades. These are recognized as marginal reef sites (e.g., Smith, 1992; Kleypas et al., 1999a; Glynn et al., 2007; Manzello et al., 2008); however, the decline of ΩArag is relatively slow in these predominantly upwelling regions, average SST is within tolerance levels, and further increases in SST might actually be advantageous for reef formation. The low ΩArag values in the eastern Pacific are not a limiting factor according to any of our statistical models, which instead assign constant, or at times increasing, suitability values to areas with ΩArag values falling below 2.2–2.3. While these levels of ΩArag will obviously impact reef structure and the balance of carbonate accretion and dissolution (e.g., Kleypas et al., 1999b; Manzello et al., 2008), and is likely to affect species composition and biodiversity (e.g., Fabricius et al., 2011), the region remains suitable for reef presence in our models, potentially assisted by warming where temperatures were previously suboptimal. That conditions remain suitable does not necessarily imply that existing reefs will be able to cope with the changes anticipated by 2070. We have not considered future changes in SST variability in our analysis, and this factor could play a key role for coral reef survival in these marginal areas (Glynn & Colgan, 1992; Toth et al., 2012), which are ranked globally among the slowest in terms of recovering from disturbances (Graham et al., 2011b).
Model evaluation and limitations
Empirical evidence is important to increasing our confidence in Bioclimate Envelope Models and projected distribution changes (Araújo et al., 2005) and the fossil coral record can provide a test for the models’ predictive ability. If SST is the stronger limiting factor for the presence of shallow coral reefs at high latitudes, we would expect fossil evidence of such an expansion into higher latitudes during warmer geological periods. In Fig. 5, we compare our 2070 predictions with relevant fossil records, including reports of extensive relic reef formations in high-latitude sites in Japan (Veron, 1992), Florida (Lighty et al., 1978), and Lord Howe Island (Woodroffe et al., 2010) dating from the Holocene, and along the Western Australian coast dating back to the Late Pleistocene (Greenstein & Pandolfi, 2008). Kiessling et al. (2012) documented both a range expansion and an equatorial retraction of coral reef distribution during the last interglacial period of the Pleistocene (ca. 125 000 years ago; Fig. 5), when average SST might have been about 1 °C higher than today (McKay et al., 2011). In addition, there is evidence supporting present-day range expansion. Yamano et al. (2011) report several tropical coral species expanding their range into higher latitudes along the coast of Japan, among them two Acropora sp. key for reef formation in the Indo-Pacific region. Veron (1992) also describes a high latitude fossil reef in Japan being re-colonized as a response to increasing temperature and Marsh (1993) reported Acropora sp. growing at a high-latitude site off Perth, Western Australia. In the Atlantic region, Vargas-Ángel et al. (2003) describe thickets of Acropora cervicorvis growing off the coast of Florida at higher latitudes than found previously. All these reports of past and present range expansion are consistent with our projections under current environmental change (Fig. 5).
Bioclimatic Envelope Models trained with equilibrium data are not suitable tools for the prediction of transient coral bleaching episodes. In particular, the models employed here have not been trained with the most relevant environmental variables (e.g., Degrees Heating Weeks or other short-term measures of thermal stress), or the location, date, and severity of bleaching episodes. Our projections correspond, instead, to the expected equilibrium response if conditions were maintained constant for sufficient time. Within these limitations, we wanted to establish a comparison of our predictions for the present-day with current observations. Coral bleaching reflects a stress response (Glynn, 1996), and degrading environmental conditions may lead to an increased number of bleaching episodes. To test this hypothesis, we compared the distribution of over 3500 recent bleaching events (observed between 2008 and 2012) with projected changes in suitability between model training conditions (1990) and the present-day (2010). We found bleaching episodes were more frequent in areas to which both MaxEntSIM and BRTSIM's 2010 projections assign higher decreases in suitability, when compared to the average on presence cells (Fig. 7). This is particularly significant for the BRT ‘SIM’ model; however, no such correlation is present with projections by the ‘OPT’ models, developed from the full suite of 27 predictive variables. Models relying on a limited number of variables are more sensitive to any change in those variables, and thus the responses of the ‘SIM’ models generally amplify that of the ‘OPT’ models, which can help explain why the ‘SIM’ models correlate better to short-term variations. MaxEntSIM and BRTSIM do however differ substantially in their predictions, especially in the WPWP region and surrounding areas, where the highest decrease in suitability is to be found according to MaxEntSIM, but where BRTSIM actually forecasts improved conditions. Few bleaching events are reported for the WPWP, but as it contains some of the least monitored coral sites in the world, this may be due to underreporting. BRT methods tend to overfit to training data, despite careful model development designed to minimize this (e.g., Elith et al., 2008), and models relying on a limited number of variables are more affected. Despite acceptable performance by BRTSIM within training conditions (similar AUC scores to that of MaxEntSIM; Fig. S1.2 in Data S1), overfitting makes its projections unstable, creating an oscillatory response to rising temperatures (e.g., Fig. S1.7) and leading to unreliable projections for anything but the smallest changes in environmental variables, including out-of-range conditions (e.g., the WPWP and surrounding areas in Fig. 7). For this reason, we chose to exclude BRTSIM projections from the average 2070 change in suitability (Fig. 5). Interestingly, the same qualities that make BRTSIM a poor long-term predictor may determine its strength in identifying areas that are currently experiencing thermal stress (Fig. 7). The MaxEnt technique is not as prone to overfitting, and we find strong agreement between projections by the ‘SIM’ and ‘OPT’ model versions. In addition, this issue is not present with the full ‘OPT’ version of the BRT model, which has a similar performance to both MaxEnt models (Data S3).
Caveats in the use of Bioclimatic Envelope Models to predict the effect of climate change on the distribution of a species include uncertainties associated with the modelling technique used (e.g., Thuiller, 2004; Lawler et al., 2006), migration limitations (reviewed in Thuiller et al., 2008), the method employed for out-of-range extrapolation (e.g., Webber et al., 2011), the availability of absence data or background selection (e.g., Elith et al., 2010), the spatial scale of the analysis (e.g., Pearson & Dawson, 2003), and the equilibrium hypothesis (e.g., Hirzel et al., 2001). Some of these caveats can be addressed by understanding the limitations of the results, which represent changes in equilibrium habitat suitability. Projections of the fate of existing coral reef ecosystems require the consideration of additional factors, such as local anthropogenic pressures and management, stress responses and recovery, bleaching, mortality, rates of sea-level rise, potential for adaptation, and migration capabilities among others. Instead, the Bioclimatic Envelope Models used here simply identify areas with future conditions similar to those where present-day coral reefs are found. In addition, the 1° × 1° spatial scale of the study may be too coarse to resolve many of the subtleties of the changing environmental conditions; however, this scale is typical of global projections from current climate models. Within those limitations, our results are informative, representing an initial estimation of the magnitude of the impact and providing a basis for future modelling work.
We employed two Bioclimatic Envelope Model approaches that make different use of absence data. MaxEnt is a presence-only technique, meaning it does not assume absence in areas where presence has not been established, unlike BRT. These two approaches have divergent responses to some of the limitations listed above, including performance in out-of-equilibrium situations (Hirzel et al., 2001) and dealing with incompleteness of the training database and/or absences due to factors other than climatic unsuitability (e.g., Pulliam, 2000). In addition, the development of models with two different levels of complexity – the ‘OPT’ and ‘SIM’ versions – helps establish the impact of variable selection in the model output and illustrate model-related performance issues, such as BRTSIM's poor performance for long-term and out-of-equilibrium predictions. Overall, we find relatively high model agreement between presence-only MaxEnt and presence/absence BRT (for the ‘OPT’ versions) and between the ‘OPT’ and ‘SIM’ model versions in the case of MaxEnt, particularly in areas projected to experience worsening conditions (Fig. 5 and Data S3). This, together with the congruence with the distribution of fossils from warmer geological times, increases our confidence in the predictions.
Despite their limitations, we note that the use of Bioclimatic Envelope techniques together with climate models output remains among the best tools in the study of species or ecosystem responses to changing conditions (e.g., Pearson & Dawson, 2003; Wiens et al., 2009). This study represents a significant advance over previous studies discussing the conflicting effects of warming and acidification because our models do not rely on specified thresholds for SST and ΩArag variables, but instead make simultaneous use of all relevant variables in the definition of an optimal climatic ‘envelope’ based only on the statistical analysis of the current coral reef distribution. This ‘envelope’ allows for the synergistic or antagonistic responses between variables that complicate physiological experimental results of thermal tolerance and ocean acidification studies.
Our main findings for future environmental suitability of coral reef ecosystems in all three CO2 emission scenarios considered are as follows: (i) range expansion at the high-latitude boundaries; (ii) no decreased suitability in currently marginal eastern Equatorial Pacific locations as well as in the Atlantic generally; and (iii) severe temperature-driven impacts in the WPWP and surrounding regions. The potential range expansion at high latitudes, however, may in many places be severely constrained by a lack of suitable benthic environment available for colonization and could additionally be affected by dispersal limitations. Currently, reefs in the Eastern Tropical Pacific will remain marginal, with increased warming offsetting the negative impacts of ocean acidification, while impacts in the suitability of the Atlantic basin as a whole may be minor. However, our models also forecast a significant overall decline in coral reef habitat suitability, with a decrease in suitability for coral reef ecosystems by 2070 of up to 30% in shallow water areas. We find that the decline, driven mainly by short-term SST maxima, is greatest around the WPWP region, and therefore would affect some of the most biodiverse coral reef regions. These results present important implications for future coral reef management, as they suggest that more emphasis should be placed in conservation efforts on marginal reefs as they are not necessarily a ‘lost cause’. They also suggest that coral reef presence is more likely to be preserved throughout much of the central and western Indian Ocean as well as the Atlantic, assuming other anthropogenic stresses are minimized.