High-resolution modeling of thermal thresholds and environmental influences on coral bleaching for local and regional reef management

Observational records of coral bleaching

We used observations from the Japanese coasts submitted to the Sango Map Project (http://www.sangomap.jp/) up to March 2017. Observations were composed of the following information: (1) presence or absence of corals; (2) longitude and latitude of the location, searchable through the Google Maps API (https://developers.google.com/maps/); (3) name of the location; (4) date, month, and year of the observation; (5) method of survey (scuba diving, snorkeling, glass boat, walking, or other); (6) water depth in meters; (7) observer’s professional background (professional scientist, nonprofit or nongovernmental organization, tourism, or other); (8) level of severity of coral bleaching (high, medium, low, nonbleaching, or not available) derived from the bleaching dataset in ReefBase. We confirmed or rejected questionable records, such as observations made on land or in the open ocean and observations of doubtful coral species.

After quality control and exclusion of records lacking information on bleaching, we obtained 668 independent records between July 2004 and October 2016. Of these observations, 52 were submitted by professional scientists, 152 by nonprofit or nongovernmental organizations, and 134 by tourists. Fifty-nine observations were conducted as part of CoralWatch (http://www.coralwatch.org/) and 63 as part of ReefCheck Japan (http://www.reefcheck.jp/). The records provided good spatial coverage of coral reefs in Japan (Fig. 1). Most of the records were obtained in the first three years following the launch of the Sango Map Project, including 449 records from 2008 to 2010 alone. In addition, 82 and 111 records were reported in 2013 and 2016, respectively, when mass bleaching events were observed throughout Japan (Kayanne, 2017; Kayanne, Suzuki & Liu, 2017).

Records of bleaching not induced by thermal stress were regarded as nonbleaching observations for the purpose of this study. We therefore reclassified 107 bleaching observations as nonbleaching observations (Step 1 in Table 1). Following screening, the prevalence of records was more biased than prior to screening, with 228 bleaching and 440 nonbleaching observations collated (Fig. 1). However, the risk of biased predictions was still deemed low (Step 2 in Table 1). Annual and spatial patterns of bleaching occurrences were consistent with those reported previously in Japan (Kayanne, 2017; Kayanne, Suzuki & Liu, 2017). We assessed spatial autocorrelation in the residuals of the prediction model of the NOAA CRW DHW (Step 2 in Table 1), using the spatial autocorrelation coefficient (Moran’s I). We confirmed that there was no significant autocorrelation in the residuals, indicating no significant spatial bias in the data (Dormann et al., 2007).

Thermal indices

To calculate thermal indices, we used daily data at a spatial resolution of 1 km (0.01°) from the Multi-scale Ultra-high Resolution Sea Surface Temperature (MUR SST) Analysis version 4.1 (JPL MUR MEaSUREs Project, 2015) (http://dx.doi.org/10.5067/GHGMR-4FJ04). The MUR SST product is a blend of SST from six satellites and thus provides higher accuracy than single-satellite products. MUR SST data are only available from 2002 to May 2017, an insufficient period of time to calculate maximum monthly mean (MMM) climatology (Table 2). In addition to data from the MUR SST, we also used data for the years 1985–2002 from the Optimum Interpolation SST (OI SST) version 2 (Reynolds et al., 2007) (http://www.esrl.noaa.gov/psd/data/gridded/data.noaa.oisst.v2.highres.html). To correct the SST bias between MUR SST and OI SST, we added the bias in the 2002–2017 monthly climatology to the OI SST data, after down-scaling to 0.01° using inverse distance weighting interpolation (Tabor & Williams, 2010; Yara et al., 2011).

Using the monthly mean SST from 1985 to 2015, we obtained two types of MMM climatologies. The first MMM climatology follows the NOAA CRW version 3 protocol (Liu et al., 2014, 2017). The temporal midpoint was recentered to that of the heritage 50 km MMM (1985–1990, 1993) using the approach of Heron et al. (2014) as follows: ${\text{SST}}_{recenteredi}={\text{SST}}_i-{\text{slope}}_i\times \left(T_{1985-2015}-T_{1985-1993}\right),$where SST_i is the SST climatology as obtained above and SST_recenteredi is the recentered SST climatology at cell i. The linear trend of monthly mean SST between the center times of the two-time durations (T_1985–2015, T_1985–1993) is represented by slope_i at cell i. The down-scaled and recentered MMM correlated significantly with the CRW MMM version 3 (Figs. S1 and S2).

The second climatology, known as MMM_max climatology (Donner, 2009, 2011), uses the mean of the warmest month of each year instead of the mean of the warmest month in the climatological years in the MMM climatology (Table 2). The warmest month is not always the same among years and, therefore, MMM_max is larger than MMM (Table 2; Figs. S3A–S3C) and represents the seasonal peak in SST more accurately than MMM climatology. This method is particularly effective in tropical zones with reduced seasonality (Donner, 2011).

We calculated eight types of thermal and cooling indices for each grid cell and observation day, including mean weekly and monthly SST, DHM, DHW (MMM + α °C), DHW (MMM + α °C) using SST variation as the bleaching alert threshold, DHW (MMM_max + α °C), DHW (MMM + βσ_m °C), and degree cooling weeks (DCW) (see Table 2 for a detailed derivation of the indices). Historical SST variability (σ_m) (Table 2) was calculated with the monthly mean SST from 1985 to 2015, and ranged from 0.36 to 0.71 with a median of 0.57 (Fig. S2D). Although DCW is calculated with a similar algorithm to that used for DHW, DCW was not significantly correlated with DHW. We therefore included DCW as a covariate in our models. The filtering thresholds (α and β) were fixed to 1 in the standard thermal indices and optimized in our indices.

Additional environmental variables

Monthly UV-B and PAR data were obtained from the Japan Aerospace eXploration Agency Satellite Monitoring for Environmental Studies (JASMES; http://kuroshio.eorc.jaxa.jp/JASMES/; accessed 25 June 2017) and derived from the average of data extracted from the Aqua and Terra sensors of moderate resolution imaging spectroradiometer (MODIS; http://modis.gsfc.nasa.gov/data/dataprod/). Although both UV-B and PAR may affect coral bleaching (Hoegh-Guldberg, 1999), the variables were significantly correlated (r = 0.79) (Yee, Santavy & Barron, 2008). We excluded PAR from our analysis (Step 3 in Table 1) as parameters may be misestimated in statistical modelings and machine learnings under multicollinearity (Dormann et al., 2013).

To quantify the speed of surface currents, we extracted data from the HYCOM+NCODA Global 1/12° Analysis GLBu0.08 from 1997 to 2017 (https://hycom.org/dataserver/gofs-3pt0/analysis/; accessed 22 June 2017). We obtained climatological median from July to September, the months during which most of the recorded bleaching events occurred. To quantify wind speed, typhoon tracking data were obtained from the Regional Specialized Meteorological Center Tokyo (http://www.jma.go.jp/jma/jma-eng/jma-center/rsmc-hp-pub-eg/trackarchives.html; accessed 22 June 2017). We calculated the wind speed index for each grid cell as the length of time without typhoons, with typhoons defined as wind speeds over 15 ms⁻¹. However, the wind speed index was strongly correlated with DHW (r = 0.86) and therefore was excluded from the analysis.

We used the diffuse attenuation coefficient (K₄₉₀) as an index of water turbidity, which can reduce light radiation stress involved in bleaching (Table 2). A monthly composite of K₄₉₀ (4 km, Level-3 binned MODIS AQUA products) was obtained from the NOAA OceanColor database (https://oceancolor.gsfc.nasa.gov; accessed 8 September 2017) for the months July–September.

Data on current speed and diffuse attenuation were down-scaled to 1 km using bilinear interpolation. When environmental variables were not available for coastal cells, we used inverse distance weighting interpolation to estimate coastal values.

Model evaluation and optimization

We evaluated coral bleaching models based on the accuracy of both positive (bleaching) and negative (nonbleaching) predictions. Most studies have evaluated models of coral bleaching based only on overall accuracy, such as the proportion of correct predictions and AIC (Maina et al., 2008, 2011; McClanahan, Maina & Ateweberhan, 2015; Kayanne, 2017; Welle et al., 2017), whereas a few studies have differentiated the accuracy of positive and negative predictions (Yee, Santavy & Barron, 2008; van Hooidonk & Huber, 2009; Donner, 2011). When the number of bleaching and nonbleaching observations is unequal (i.e., under class imbalance), model predictions can be biased. We therefore used four evaluation metrics: overall accuracy, true positive rate (TPR), true negative rate (TNR), and true skill statistic (TSS) (defined in Table 2). TSS quantifies prediction skill as the index weighs positive and negative predictions equally (Allouche, Tsoar & Kadmon, 2006).

To assess the combined effects of thermal stress and multiple environmental influences on coral bleaching, we constructed prediction models of bleaching with two approaches: generalized linear model (GLM) with a binomial error distribution and a logit link function, and random forests (RFs; Breiman, 2001). Although both models compute predictions in the form of probabilities, the models are based on different algorithms. GLM is an extension of regression models, whereas RF is a machine learning method that uses randomly repeating classifications to capture complex interactions among explanatory variables (Table 2). Therefore, GLM has an advantage of which fitted model can be written as a formula that is easy to be used subsequently.

We confirmed that the data met the assumption of binomial GLM (i.e., the residual deviance per degree of freedom was less than 1.5; Zuur et al., 2009). GLM was applied with the “glm” function in base R (R Core Team, 2017), and RF was applied with the “randomForest” function of the randomForest R package. RF was used under standard settings to avoid overfitting the training data. However, we followed the recommendation of Hijmans et al. (2017), by specifying the model as a “regression model,” even though the response variable was categorical. The relative importance of explanatory variables was calculated with the “importance” function of the MuMIn package (Bartoń, 2015) for GLM and of the randomForest package for RF (Liaw & Wiener, 2002).

Predicted probabilities were transformed into bleaching and nonbleaching categories with the threshold that maximized the sum of TPR and TNR (Liu et al., 2005). The midpoint (i.e., 0.5) is often used as a threshold (Fig. S1C), although it is sensitive to class imbalance in the training data and may, therefore, lead to inaccurate predictions (Liu et al., 2005). This issue is addressed in studies of species distribution modeling, although it remains poorly addressed in studies of mass bleaching (van Hooidonk & Huber, 2009). We used the “evaluation” function of the dismo R package (Hijmans et al., 2017) for evaluation and optimization of the threshold.

Models were evaluated by 10-fold cross-validations using TSS as the evaluation index. In each repeat, we separated 30% of the data as testing data and used the remaining 70% for constructing GLM and RF (Step 4 in Table 1). We optimized the filtering thresholds (Fig. S1A) for DHM and DHWs by cross-validation, while the filtering thresholds were fixed at 1.0 °C in the standard indices (Step 4 in Table 1). We selected the optimum filtering threshold between 0 and 1.5 °C for indices using a constant threshold (α), whereas we examined the coefficient of σ_m (β) between 0.1 and 2.5 for indices based on historical variability (Table 1). We conducted optimizations with 0.01 precision for both types of indices, i.e., with 151 and 241 submodels, respectively.

For models with multiple explanatory variables, we considered DCW, historical SST variability, UV-B, water turbidity, water depth, and current speed, in addition to the thermal index. The two most influential variables (DCW and UV-B) were always included in models with multiple explanatory variables (Step 4 in Table 1). The optimum set of explanatory variables was specified through cross-validation. The set of variables that best explain variation in the testing data was selected among all 15 possible combinations. In total, we evaluated 22,650 and 36,150 models (10 cross-validations × 15 variable combinations × 151 or 241 submodels) for each GLM and RF model, respectively.

Finally, we predicted coral bleaching in the warmest month of the main coral habitat in the study area using the best predictive model (Step 5 in Table 1). We also assessed reduction in UV-B as a possible adaptive measure by reducing UV-B radiation by 40% and increasing water turbidity by 40%, thereby simulating the effects of screening with fishnets (Step 5 in Table 1). These levels of changes were consistent with in situ examination in Onna Village in the Ryukyu Islands (Okinawa Prefecture, 2017). Predictions were obtained from each model built in the 10 cross-validations, and subsequently averaged among the models. Spatial data were obtained from the Global Map Japan version 2.1 Vector data, provided by the Geospatial Information Authority of Japan (2015).

All analytical codes (available in Supplemental Information 1) were written in R version 3.4.1 (R Core Team, 2017).

Effects of environmental variables

Predicted probability of bleaching increased with increasing values of thermal indices, including SST, DHM, DHW, and UV-B (Fig. 2). Predicted probability of bleaching decreased with DCW, water turbidity, and water depth (Fig. 2). Relationships between bleaching and historical SST variability and current speed were not significant, with 95% confidence intervals (CIs) ranging from negative to positive. Relationships between bleaching and monthly and weekly SST were positive, although the widths of the 95% CIs suggest these variables are not reliable indices of coral bleaching. Alert thresholds for predicted bleaching were found to be lower than standard thresholds, except for DHM (Table 3).

Ranking of important variables was similar between GLM and RF (Fig. 3): the best explanatory variable was DHW (100% in both of GLM and RF), followed by DCW. UV-B, water turbidity, and historical SST variability also explained substantial variation in coral bleaching. The explanatory powers of historical SST variability and current speed were high, despite inconsistent relationships with coral bleaching (Fig. 2). Absolute variable importance differed between GLM and RF. In GLM, most variables explained more than 25% of variation in bleaching each, while in RF, only DHW, DCW, and UV-B explained more than 25% of variation in bleaching each.

Optimization and assessment of filtering thresholds

Optimization of filtering thresholds improved the predictive performance of DHM and DHWs, although the improvement was small for GLM (Fig. 4). Improvement by the optimization was around 0.01 in TSS in GLM, while the improvement of DHW using the historical SST variability (σ_m) as the filtering threshold was 0.02–0.03 in TSS in RF.

We compared the predictive performance of all models including standard and optimized thermal indices, and the optimized set of explanatory variables (Fig. 5). In models based only on a thermal index with the standard threshold, TNR was larger than TPR, indicating that high overall accuracy can result from effective identification of nonbleaching occurrences, despite ineffective identification of bleaching occurrences. TSS was indicative of both TPR and TNR. Among thermal indices with the standard threshold, DHW using historical SST variability (σ_m) as the bleaching alert threshold and DHW using σ_m as the filtering threshold scored the best performances in TSS (0.60 and 0.65, respectively) (Fig. 5). Weekly SST had no prediction skill (TSS = 0) and the skill of monthly SST was low (0.20). TPR of DHW using historical SST variability as the bleaching alert threshold was the highest (0.85) among models, although the false-positive rate (1–TNR = 0.20) was also the highest. Using historical SST variability as the filtering threshold for DHW decreased the false-positive rate (0.09), but also decreased TPR (0.69). DHW using MMM_max as the baseline climatology showed the lowest predictive performance (TSS = 0.37) among all DHW indices, with the lowest TPR (0.38).

Predictive performance was improved by optimizing the evaluation threshold (Fig. 5; Table 4). Under optimized evaluation thresholds, DHW with historical SST variability as the filtering threshold performed best (TSS = 0.72 for GLM) among the three types of DHW, although differences were small. Predictive performance of DHM was the lowest (TSS = 0.40). When both evaluation and filtering thresholds were optimized (Table 5), improvements in evaluation thresholds were only small. DHW with historical SST variability as the filtering threshold remained the highest performing index under optimization (TSS = 0.73 in both GLM and RF). Optimized filtering thresholds were less than 1 °C in all the models using one thermal index, using the optimized DHM or DHW other than DHW using historical SST variability (σ_m) as the filtering threshold (see also Fig. S3D for the distribution of σ_m).

The predictive skill of models with multiple explanatory variables was negligibly higher than that of models including only one thermal index. TSS increased by 0.1 at most in all models, except for the DHM model where TSS increased by 0.3 in both GLM and RF. Increases in predictive performance were mainly due to increases in TNR (i.e., reductions in false-positive rates; Fig. 5). Models including no thermal indices showed high prediction skill (TSS = 0.65 in GLM; TSS = 0.72 in RF). In models with multiple explanatory variables, differences in predictive performance between models with optimized evaluation thresholds and models with optimized evaluation and filtering thresholds were smaller than differences between GLM and RF models for most thermal indices. RF always performed better than GLM, with differences in TSS of 0.04 to 0.05. Although the TPR of GLM exceeded that of RF in most cases, the TNR of GLM was lower than those of RF, i.e., the risk of false positives was higher in GLM.

The RF model based on DHW with MMM + 0.97 °C filtering threshold, DCW, UV-B, water turbidity, historical SST variability, and current speed (Table 6) showed the best predictive performance (TSS = 0.79; TPR = 0.90; TNR = 0.89). Among the GLM, the model consisting of DHW with MMM + 1.83·σ_m °C filtering threshold, DCW, UV-B, and turbidity showed the best predictive skill.

Predictions of coral bleaching

We predicted probabilities of coral bleaching in the main coral-habitable areas of Japan with the optimized best multivariate model of RF using DHW with MMM + 0.97 °C filtering threshold (Fig. 6). The mean predicted probability of bleaching ranged from 0.46 to 0.74 among areas. Spatial variation in the probability of bleaching was found in both the eastern (Fig. 6A) and the western (Fig. 6C) Ryukyu Islands. Hotspots with higher bleaching probabilities were found in the southeastern part of Okinawa Island, the eastern part of the Kerama Islands (Fig. 6A), and the northern part of Ishigaki Island (Fig. 6C). This resulted in bleaching in 2.5–5 (44–100%) of five years (2008–2010, 2013, and 2016; Figs. 6B and 6D).

Under reduced UV-B radiation, coral bleaching decreased in most areas, with particularly large decreases in bleaching hotspots (Figs. 7A and 7C). Predicted probabilities of bleaching ranged from 0.34 to 0.63 (Figs. 7A and 7C), and bleaching frequencies ranged from 28% to 92% (Figs. 7B and 7D). Decreases in probabilities of bleaching of up to 0.24 were observed, resulting in a significant decrease in bleaching frequency of up to 56% (Figs. 7B and 7D). Bleaching in fewer than three out of five years occurred in most areas.

Optimizing coral bleaching models

The ability to predict coral bleaching was improved by optimizing thermal indices, particularly SST and DHWs. As in a previous study (Donner, 2011), we generally found lower TPR than TNR, but both TPR and TNR were improved by optimization. The sole contribution of optimizing the filtering threshold was small. However, optimizing the filtering threshold and the evaluation threshold while combining multiple environmental variables achieved large improvements in TPR and TNR (reaching ∼0.9). We also found that cooling (DCW), UV-B, and screening (water turbidity) were important predictors of bleaching, particularly in RF models.

Our results are mostly consistent with those of Donner (2011), who showed that DHW using historical SST variability as bleaching alert threshold had a higher TPR but a higher false-positive rate than NOAA CRW DHW, and that DHW using MMM_max did not predict bleaching accurately but was suitable in equatorial zones. Because our study was conducted at a higher latitude, using MMM_max resulted in lowest performance among the DHW indices, as expected. We also used historical SST variability as a filtering threshold, and this method showed the highest prediction skill among the models with only one thermal index. Historical SST variation (σ_m) may be a particularly effective predictor of bleaching in Japan, as variation was larger in our study area (ca. 0.56) than in the study area (ca. 0.25) of Donner (2011). Indeed, southern islands in Japan are encircled by the strong Kuroshio boundary current flowing poleward, and tropical waters brought by the current can cause faster warming than the global average (Wu et al., 2012), leading to high levels of historical SST variation in the area.

Optimized filtering thresholds were smaller than 1 °C in GLM and RF models using only one thermal index. Previous studies have suggested that thermal stress not exceeding 1 °C can induce coral bleaching (Brown, 1997; McWilliams et al., 2005; Kleypas, Danabasoglu & Lough, 2008). Nevertheless, bleaching thresholds had not been statistically optimized before our study.

Random forest was an excellent method for predicting coral bleaching and could be used more widely in studies of coral ecology. The use of RF has much increased in ecological studies in the 10 years since the introduction of Cutler et al. (2007). RF shows considerable potential for ecological analyses including classification, regression, and survival, due to its high predictive accuracy and its ability to model complex interactions among explanatory variables (Cutler et al., 2007).

Models with multiple environmental variables are becoming more popular, and show high explanatory power when modeling bleaching (Maina et al., 2008; Yee, Santavy & Barron, 2008; McClanahan, Maina & Ateweberhan, 2015; Welle et al., 2017). We found UV radiation to be an important explanatory factor for coral bleaching, consistent with previous studies (Hoegh-Guldberg, 1999; Maina et al., 2008, 2011; McClanahan, Maina & Ateweberhan, 2015). Other variables related to cooling (DCW; Jones et al., 2017) and screening (water turbidity; West & Salm, 2003; Oliver, Berkelmans & Eakin, 2009; Maina et al., 2011; Oxenford & Vallés, 2016) also explained variation in the occurrence of coral bleaching. Small-scale topographic variables, including water depth, are known to reduce thermal stress on corals (West & Salm, 2003; Oliver, Berkelmans & Eakin, 2009). Strong winds may also reduce bleaching risk (Maina et al., 2008, 2011; Yee, Santavy & Barron, 2008; McClanahan, Maina & Ateweberhan, 2015; Welle et al., 2017), but the importance of this variable could not be evaluated in our study due to high correlations with thermal indices. Our predictions may be negatively affected by environmental variation at small temporal and spatial scales that has not been adequately included in our study. For example, small-scale water flow may improve the resistance of corals to bleaching (Nakamura & van Woesik, 2001), but the 8 km resolution of current speed in our study is too coarse to represent such effects. The microstructure of the sea floor at the meter scale may also be related to local water flow or shading of corals (West & Salm, 2003; Oliver, Berkelmans & Eakin, 2009), but was not incorporated in our study.

Bleaching responses and thermal thresholds vary among coral species (Maynard et al., 2008; Yee, Santavy & Barron, 2008; Guest et al., 2012; Harii et al., 2014; McClanahan, 2014). Branching corals of Acropora and Pocillopora spp. are more susceptible to thermal stress than massive corals such as Porites spp. (Maynard et al., 2008; Yee, Santavy & Barron, 2008; Guest et al., 2012; Harii et al., 2014; McClanahan, 2014). Thermal tolerance increases with repeated excessive thermal stress (Brown et al., 2002; Maynard et al., 2008; Guest et al., 2012), highlighting the potential of corals to adapt to thermal stress (Brown et al., 2002; Maynard et al., 2008; Guest et al., 2012). However, the effects of past thermal conditions have not been fully explained with historical SST variability in our analysis or in previous studies (Donner, 2011; McClanahan, Maina & Ateweberhan, 2015). Variation in thermal tolerance can result from interspecific differences (Maynard et al., 2008; Guest et al., 2012; Harii et al., 2014; McClanahan, 2014) in acclimation, genotypes, and epigenetics of host corals and symbiotic algae (Palumbi et al., 2014; Torda et al., 2017). The effects of such differences remain poorly known and should be prioritized for further research.

Coral bleaching in Japan and reef management

Four studies in the Japanese region have analyzed coral bleaching occurrences with temperature anomalies and DHW at coarse resolutions (>50 km; Strong et al., 2002; Harii et al., 2014; Kayanne, 2017; Kayanne, Suzuki & Liu, 2017). Coarse DHW captures regional trends in the onset of coral bleaching, although it fails to predict bleaching within smaller reefs (Strong et al., 2002; Harii et al., 2014). Indeed, our bleaching predictions for the Ryukyu Islands with DHW at 1 km resolution exhibited a TPR of 0.58, compared to a value of 0.44 calculated from the tables of Kayanne (2017). However, the improvement in predictive performance may result from the use of high-resolution temperature data, or the optimization of thermal thresholds of DHW. Further studies are required to establish the importance of high-resolution temperature data for predicting coral bleaching in Japan.

The high performance of our bleaching model at 1 km resolution has practical implications for the local and regional management of coral reefs. Our predictions revealed high frequencies of coral bleaching in many parts of the Ryukyu Islands. However, predictions of bleaching frequency were based on the lowest levels of bleaching severity; hence, further analyses may be required to establish full distributions of bleaching frequencies according to levels of severity.

Practical management to reduce the risk of coral bleaching should include control of coastal water turbidity (Fabricius, 2005). Increases in water turbidity by terrestrial runoff may decrease the resistance (Wooldridge & Done, 2009) or resilience (Hongo & Yamano, 2013) of corals to bleaching. Turbid coastal regions may provide refuges from climate warming due to limited increases in temperature and solar radiation (Cacciapaglia & van Woesik, 2016). However, coastal turbidity may increase the incidence of coral diseases and promote the growth of competing algae (Fabricius, 2005). Consequently, coastal turbidity should be carefully managed.

Reducing UV radiation may reduce bleaching risk and may constitute a powerful adaptive measure against climate warming. In the Onna Village of the Ryukyu Islands, in situ reduction of UV radiation with no increase in water turbidity has already been tested (Okinawa Prefecture, 2017). Reduction of UV radiation by 30–44% with large fishery nets resulted in a survival rate of 80% in cultured coral colonies in the summer of 2016, when the most severe thermal stress was recorded in the 2004–2016 study period (Kayanne, Suzuki & Liu, 2017). Reduction in UV radiation was similar to that used in our study (40%), so our predictions could provide a quantitative basis for future reef management in this area.