Predicting Coral Offspring Survival From Reef Thermal Histories in Ningaloo World Heritage Reef

ABSTRACT

Identifying coral populations with heritable thermal tolerance is critical for developing climate‐resilient reef conservation practices. We quantified thermal tolerance in larvae from 51 selectively bred Acropora families using acute heat stress assays (35.5°C) then modelled larval survival against 20 satellite‐derived SST predictors (1985–2022) using boosted regression trees (BRT) to predict the spatial distribution of thermally resilient coral offspring along Ningaloo World Heritage Reef (East Indian Ocean). BRT models had an AUC of 0.67 (±0.01 SE) and revealed that thermal anomaly frequency, cumulative heat stress and temperature variability, particularly diurnal temperature range, most strongly predicted larval survival (37.2%, 22.7% and 21.7%, respectively). Survival peaked at thermal stress of 1.14°C –weeks (Degree Heating Weeks) and variability of daily temperature range of ~1°C day⁻¹ highlighting potential ecological thresholds in larval survival in this area. Spatial outputs identified six reefs (17% of surveyed sites), all north of Point Cloates, as having the highest predicted larval survival (> 45%), contrasting with consistently lower survival probabilities across most southern sites. By linking experimental physiology with spatial modelling, this study delivers the first spatially explicit predictions of thermally resilient larval supply along Ningaloo Reef, providing a scalable framework for guiding marine spatial planning, broodstock selection and restoration prioritisation across a remote World Heritage coral reef.

Coral collection, spawning, selective breeding, larval rearing, heat stress test, using survival data and satellite‐derived sea surface temperature data to carry out a boosted regression tree analysis to predict areas of expected higher larval survival.

1. Introduction

Marine heatwaves, intensified by global warming, have increased in duration and frequency, posing a growing threat to coral reefs worldwide (van Hooidonk et al. ²⁰¹⁶; Frölicher et al. ²⁰¹⁸; Spady et al. ²⁰²⁶). The fourth global coral bleaching event (2023–2025), the most spatially extensive on record, impacted approximately 84% of the world's coral reef area across at least 83 countries and territories (NOAA Coral Reef Watch ²⁰²⁵). Coral bleaching is the disruption of the symbiotic relationship between the coral animal and its endosymbiotic algae (Symbiodiniaceae) and generally occurs when temperatures exceed the Maximum of the Monthly Mean SST (MMM) by approximately 1°C–2°C (Berkelmans and Willis 1999), with extended exposure leading to significant mortality (Hughes et al. 2018). Corals experiencing thermal stress exhibit increased vulnerability to diseases (Miller et al. 2009), diminished energy reserves and reproductive output (Johnston et al. 2020). Corals from locations with higher MMM and highly variable temperature regimes that have withstood past bleaching events tend to be more resistant to thermal stress (Fine et al. 2013; Guest et al. ²⁰¹²; Safaie et al. ²⁰¹⁸; Quigley and van Oppen ²⁰²²).

Adaptation is characterised by changes in genotypes in populations towards phenotypes with higher fitness in the present environment. One way in which this occurs is through adaptive genetic variation, that is, allelic variation accumulated over time and maintained in a population through natural selection or other evolutionary mechanisms. This contributes to a population's ability to adapt in response to selective pressures (Orr 2005). Given the widespread impact of climate warming on populations and species, identifying whether and where adaptive genetic variation exists within species is critical as this variation underpins the capacity of populations to respond to environmental change and helps predict the likelihood and rate of adaptation to warming (Willi et al. 2022). On coral reefs, local adaptive variation in heat tolerance has been detected and is often evidenced by differential bleaching between colonies (Brown, Eyal, et al. ²⁰²³; Brown, Lenz, et al. ²⁰²³; Langlais et al. ²⁰¹⁷; Selig et al. ²⁰¹⁰). Thermal tolerance has been demonstrated to be heritable in corals, including in Acropora, with quantitative genetic studies reporting moderate to high narrow‐sense heritability for survival and bleaching response under heat stress (Bairos‐Novak et al. ²⁰²¹; Dziedzic et al. ²⁰¹⁹; Kirk et al. ²⁰¹⁸). It has been posited that, given enough time, this natural variation may provide the necessary genetic variation to spur adaptation to further warming (Urban et al. 2024). However, limited reef connectivity and the spatial structure of gene flow in some locations may mean that the spread of heat‐tolerant alleles into regions with less underlying tolerance to heat will be slow, hindering natural adaptation across the broader reef system. For example, in the Great Barrier Reef, it would take over 30 generations for adaptive gene variants relating to thermal resilience to spread beyond northern reefs to the central and southern GBR, even under multiple release scenarios (Quigley et al. 2019). This suggests that without active intervention, heat tolerance is unlikely to spread beyond the northern GBR reefs in time, given current warming rates (Quigley et al. 2019; Torda and Quigley ²⁰²²).

Phenotypic plasticity may also drive adaptive responses to environmental change, enhance the probability of population persistence and potentially support genetic rescue (Reed et al. 2011). Phenotypic plasticity has been broadly defined as the property of a given genotype to produce different phenotypes in response to distinct environmental conditions (see Pigliucci ²⁰⁰¹) and may include transgenerational effects (Mousseau and Fox 1998), in which offspring performance is heavily influenced by the environment experienced by the parents (Shama et al. 2014). Corals may achieve this through epigenetic mechanisms such as DNA methylation (Liew et al. 2020), or through rapid restructuring of the coral‐symbiont relationship (i.e., symbiont shuffling or switching; Berkelmans and van Oppen ²⁰⁰⁶; Howells et al. ²⁰¹³). Nevertheless, evidence suggests that corals may possess limited capacity to acclimatise or adapt beyond the thermal tolerance thresholds of one or both, partners in the coral‐symbiont relationship (Howells et al. 2013). Hence research efforts increasingly aim to enhance the adaptive capabilities of reef‐forming corals to improve their probability of survival (Voolstra et al. 2021).

One method to enhance the potential for adaptation to future warming is through selective breeding, by which genetic and phenotypic variability are leveraged to improve thermal resilience (Dixon et al. 2015; Drury et al. ²⁰²²). A handful of studies have intentionally manipulated coral reproduction to investigate phenotypic outcomes, finding that offspring from selectively bred parents show markedly improved survival under heat stress after just one generation (Quigley 2023). In this work, identifying naturally occurring, thermally tolerant parental genotypes often require the screening of thousands of colonies through acute heat stress assays, at large cost and challenging logistical constraints (Guest et al. 2014; Humanes et al. ²⁰²¹). Hence there is a need to accurately predict the location of heat‐resistant adult corals (which can serve as brood stock) using environmental predictors and machine learning. This machine learning framework has been developed on the Great Barrier Reef and identified hundreds of reefs as having heat tolerant colonies (Quigley and van Oppen ²⁰²²). This has yet to be done in other critical reef regions, like Ningaloo World Heritage Reef.

Ningaloo is a fringing coral reef located along the northwestern coast of Western Australia, extending over 200 km. It supports diverse coral assemblages (van Keulen and Langdon ²⁰¹¹) and experiences strong environmental gradients due to its latitudinal span. The reef is influenced by the warm southward‐flowing Leeuwin Current, against the climatological mean equatorward winds, contributing to relatively high sea surface temperatures compared to other reefs at similar latitudes (Church et al. 1989). Most sea surface temperature variability in the ocean just offshore of the reef is due to wind‐driven fluctuations (Lowe et al. 2012). Notably, Ningaloo has experienced both acute and chronic stress events. This includes widespread coral bleaching being observed in both 2011 and 2016, though mortality varied considerably among sites and across years (Depczynski et al. 2013; Le Nohaïc et al. ²⁰¹⁷). Extensive bleaching also occurred in 2025 (Quigley 2025). Generally, northern Ningaloo typically experiences greater thermal variability and higher maximum temperatures, while the southern reefs are generally cooler and more stable (NOAA Coral Reef Watch ²⁰¹⁸). This thermal heterogeneity may create natural gradients in coral exposure to heat stress, making the region a valuable system for studying local adaptation and resilience to climate change.

To this end, we performed multiple years of selective breeding in two reef‐building hard coral species to produce 51 larval families. Survival data was obtained from these quantitative genetic breeding experiments under control and heat stress. A boosted regression trees (BRT) framework was applied to estimate locations of high heat tolerance in relation to environmental predictors. This framework enabled forecasting the conditions required for adult colonies with high thermal tolerance and heritable resilience to persist, thus identifying relative resistance (i.e., regions supporting corals capable of re‐populating reefs with offspring tolerant of sustained temperature increases up to 3°C). These results are critical to inform management efforts and indicate important targets for conservation planning in Ningaloo Reef, one of the world's longest fringing reefs.

2. Methods

2.1. Remotely Sensed Data

Remote‐sensing tools allow for the evaluation of the spatial distribution of environmental conditions related to heat tolerance, which can then be related back to phenotypes. The large size of Ningaloo Reef, spanning approximately 260 km, poses challenges to the use of in situ loggers commonly used in ecological studies. Remotely sensed datasets were therefore sourced from three publicly available repositories that varied in spatial and temporal resolution and in the period of record; (i) Daily Global 5 km Satellite Sea Surface Temperature (CoralTemp v3.1) (NOAA Coral Reef Watch ²⁰¹⁸), (ii) Himawari‐8 L3C hourly high‐resolution geostationary satellite interpreted by the Australian Bureau of Meteorology (IMOS ²⁰²⁴) and (iii) Coral Reef Temperature Anomaly Database (CoRTAD) (v.6), a global database of weekly, 4 km resolution Sea Surface Temperature (SST) and related Thermal Stress Metrics (Saha et al. 2018). CoRTAD was selected for thermal stress metrics (SSTA frequency and TSA_DHW) because it provides pre‐calculated coral reef indices widely used in previous global bleaching studies (e.g., van Woesik and Kratochwill ²⁰²⁴). Candidate predictors were selected a priori to represent key components of the thermal environment affecting coral physiology. An initial set of 22 variables (Table 1) was compiled within this framework. To reduce redundancy and overfitting, final predictor retention was guided by relative contribution to model performance within the boosted regression tree framework using cross‐validation (Section 2.6), ensuring a parsimonious, hypothesis‐driven model structure.

TABLE 1.

Sea surface temperature (SST) metrics derived from remotely sensed satellite data products and included as environmental predictors in the boosted regression tree models. Variables are grouped into five categories representing ecologically relevant aspects of the thermal environment. Seasons were defined such that each season spanned three complete months, with austral spring and summer defined as September through February.

Variable [units]	Description	Cut‐offs	Range [frequency]	Source
Baseline thermal environment
SSTmean [°C]	Mean measurement of the nighttime ocean temperature at the sea surface	01/01/1985 – 29/12/2022	5 km [nightly]	CoralTemp
SSTstdev	Standard deviation of mean SST
SSTmix	Minimum SST
SSTmax	Maximum SST
SSTannual	Mean yearly SST
Min_temp_range	Minimum SST in that yearday	01/09/2015 – 13/12/2022	2 km [hourly]	Himawari‐8
Max_temp_range	Maximum SST in that yearday
Cumulative thermal stress
TSA_DHW [°C‐weeks]	Mean degree heating weeks, DHW being the sum of previous (12 weeks when thermal stress anomaly is ≥ 1°C).	01/01/1985 – 19/12/2022	4 km [weekly]	CoRTAD
TSA_DHWstdev	Standard deviation of DHW
TSA_DHWmin	Minimum TSA_DHW
TSA_DHWmax	Maximum TSA_DHW
TSA_DHWpost2011	Mean degree heating weeks from 2011 to 2022	01/01/2011 – 19/12/2022
Thermal anomaly frequency
SSTA_Freq [count (events/year)]	Number of times over previous 52 weeks when SSTA (weekly SST minus weekly climatological SST) is ≥ 1°C	01/01/1985 – 19/12/2022	4 km [weekly]	CoRTAD
SSTA_Freqstdev	SD of temp. anomalies
High‐frequency temperature variability
Daily temperature range (DTR) [°C]	Difference between maximum and minimum temperatures for each day, excluding the calculation on days that have any NA values	01/09/2015 – 13/12/2022	2 km [hourly]	Himawari‐8
DTR_stdev	Standard deviation of DTR
DTRss	Mean DTR spring and summer
DTRss_stdev	Standard deviation of DTR in spring and summer
DTRaw	Mean DTR in autumn and winter
DTRaw_stdev	Standard deviation of DTR in autumn and winter
DTR distribution shape
Kurtosis [—]	Kurtosis of full time series of DTR values	01/09/2015 – 13/12/2022	2 km [hourly]	Himawari‐8
Skewness [—]	Skewness of full time series of DTR values

SST metrics were extracted from a single pixel per site, consistent with the spatial resolution of the satellite products used. For each site, the nearest pixel to the GPS coordinates was identified and extracted; in cases where the nearest pixel overlapped with land, the adjacent open‐water pixel was selected to ensure data quality. Within‐site temporal variability metrics were calculated from the full time series of the extracted pixel. To assess diurnal heating and cooling patterns, the daily temperature range (DTR) was calculated as the difference between the daily maximum and minimum temperatures for each time series, excluding days with any values flagged as poor quality. Temporal variations and ranges in DTR were analysed by creating composite averages of mean, min and max temperatures across yeardays 1–366 (i.e., the sequential day numbers in a calendar year). To assess the symmetry and tail‐density of DTR distributions, skewness and kurtosis values were calculated for each time series. Skewness was calculated to determine whether temperature ranges were predominantly low or high, indicating potential seasonal patterns or anomalies. Kurtosis was used to assess the frequency of extreme temperature ranges; high kurtosis suggests that while most DTRs were moderate, occasional extreme values (e.g., unusually high or low temperatures) occurred. The CoRTAD dataset contains Sea Surface Temperature Anomaly (SSTA) Frequency (number of times over previous 52 weeks that SSTA is ≥ 1°C; SSTA being weekly SST minus weekly climatological SST) and Thermal Stress Anomaly Degree Heating Weeks (TSA_DHW, sum of previous 12 weeks when TSA is ≥ 1°C). To capture recent ecologically significant thermal stress, we calculated mean DHW cumulatively from 2011 to 2022. The 2011 event was selected as the starting point because it represented the most severe marine heatwave recorded in Western Australia, causing widespread bleaching and coral mortality (see Pearce et al. ²⁰¹¹). The cumulative 2011–2022 metric therefore incorporates subsequent heatwaves, including the 2013 and 2015/16 events, and contrasts with the longer‐term DHW metric calculated across the full satellite record (1985–2022). Statistical differences in environmental SST metrics were performed using non‐parametric Kruskal–Wallis tests followed by the post hoc Dunn's test.

2.2. Spatial Data Visualisation

Shapefiles for the land mass (grey) and extent of coral reefs (light blue) represented in Figures 1a and 4 were obtained from Open Street Map (https://osmdata.openstreetmap.de) and Allen Coral Atlas, respectively. The map was generated in R (R Core Team ²⁰²⁴) using the ggplot2 (Wickham 2024), sf (Pebesma 2025) and patchwork (Pedersen 2024) packages.

FIGURE 1.

Spatial and temporal patterns of sea surface temperature (SST) and thermal stress indicators across four locations in Ningaloo Reef: Tantabiddi, Oyster Stacks, Bruboodjoo and Pelican Point. (a) Map showing sampling locations of gravid Acropora tenuis and Acropora millepora colonies. Site colours are burgundy (Tantabiddi), blue (Oyster Stacks), brown (Bruboodjoo) and orange (Pelican Point). Inset map of indicates the location of the Ningaloo Reef region within Australia (see Section 2.2). (b) Mean annual SST (°C) from 1985 to 2022, illustrating long‐term temperature trends. (c) Frequency distribution of SST anomalies relative to the historical mean, highlighting the prevalence of extreme temperature events. Violin shading represents the underlying density distribution of SSTa values. Boxplots include the median values (centre lines), means (diamonds), upper and lower quartiles (box limits) and 1.5× interquartile range (whiskers). (d) Yearly composites of mean daily temperatures (line) and interannual temperature ranges (minimum–maximum) for each day of the year (shading). (e) Histograms of daily temperature range with frequency measured in number of days (between 2015 and 2022). (f) Distribution of degree heating weeks (DHW) across reef sites during the 2011 thermal anomaly, shown for illustrative purposes only; DHW in 2011 alone was not included as a BRT predictor variable (see Section 2.1). Solid vertical lines indicate the mean; shaded regions and dashed lines indicate mean ± SE. Annotations indicate statistical significance of pairwise contrasts at ɑ = 0.05: n.s.= not significant, **p < 0.01, ***p < 0.001.

FIGURE 4.

Predicted percent larval survival across Ninglaoo World Heritage Reef derived from boosted regression tree models. Prediction sites were restricted to 5 km from each other. A continuous colour gradient connects sites, with colours interpolated between adjacent locations to represent the spatial gradient in predicted survival, ranging from blue–grey (low survival) to dark red (high survival). Red points indicate predicted high‐survival reefs while black points denote named reference locations included for geographic context.

2.3. Coral Spawning and Larval Rearing

The horizontally transmitting broadcast‐spawning coral species Acropora tenuis (Dana, 1846) and Acropora millepora (Ehrenberg, 1834) were chosen as they are known to produce substantial larval output, necessary for large scale experimentation. In 2023, 10 gravid colonies of each species were collected from two sites on the Ningaloo Reef: Oyster Stacks in the north and Pelican Point in the south. In 2024, only A. tenuis colonies were collected, from two new sampling sites (Tantabiddi and Bruboodjoo; Figure 1a). These sites were selected based on species presence and to capture as much of the full thermal gradient along the reef as possible, spanning approximately 1.4° of latitude (~155 km) from Tantabiddi in the north to Pelican Point in the south. In both years, colonies were transported to flow‐through holding tanks set to the ambient temperature of the respective reefs (Pelican Point: 27.6°C; Oyster Stacks: 27.9°C; Tantabiddi and Bruboodjoo: 28°C). Following the appearance of egg–sperm bundles under the oral disc and polyp (setting behaviour), colonies were isolated into individual containers. Gamete collection, fertilisation and rearing followed established methodologies (Quigley et al. 2016, 2020; Weeriyanun et al. ²⁰²²). Briefly, gametes were collected, washed and filtered, before the sperm from one parental colony was added at a concentration of approximately 10⁵ mL to eggs from another parental colony to create distinct coral families. Of the colonies that spawned, a total of 51 families (40 in 2023 and 11 in 2024) were selectively bred and used in the subsequent heat stress experiment (see Kler Lago et al. ²⁰²⁵). Family pedigree and the total number of maternal and paternal colonies used in this study are listed in Tables S2–S4. The families produced (hereafter termed crosses) are considered biological replicates and are either pure‐bred (intrapopulation cross) or hybrid (interpopulation cross). Crosses are denoted using a two‐letter code, where the first letter represents the maternal colony's reef of origin, and the second letter represents the paternal colony's reef of origin. For example, ‘P × O’ indicates eggs from a Pelican Point colony fertilised with sperm from an Oyster Stacks colony. Each cross was maintained at a density of 0.5 to 1 planula larvae per ml in flow‐through 80 L conical larval tanks at 27.5°C with gentle aeration until the start of the heat stress experiment.

2.4. Larval Heat Stress Experiment

By day five post‐fertilisation, larvae were ciliated and motile, consistent with the 96‐h stage of larval development reported for A. millepora (Ball et al. 2002). To quantify the variation in coral heat tolerance, the aposymbiotic larvae were placed in net‐wells on floating holders, allowing water and waste exchange with the large volume (1200 L) of filtered seawater in the water bath. Approximately 20 larvae were aliquoted into each floating net‐well (three replicate wells per population cross, n≈60 per cross per temperature treatment). In the elevated temperature treatment, the temperature was ramped from 27.1°C to 35.5°C over 16 h (or 0.5°C/h until temperatures reached 35.5°C). The larvae that were alive in each net‐well were counted approximately every 11 h over the following 143 h (92 h for A. millepora , 88 h in 2024) to estimate survival rates per replicate. The elevated temperature treatment was chosen to ensure comparability with previous selective breeding experiments where aposymbiotic larvae were exposed to an equivalent level of thermal stress (35.5°C) (Dixon et al. 2015; Quigley and van Oppen ²⁰²²; Weeriyanun et al. ²⁰²²). The final survival counts occurred when ∼50% of the larval crosses reached 50% survival in the heat treatment.

2.5. Survival Statistical Analyses

Summary statistics were computed at two hierarchical levels: by cross type and by family. By cross type: summary statistics were calculated for each cross type, species and temperature. By family: summary statistics were computed at the family level, grouped by species, temperature and cross type. Assumption testing for parametric analyses included a Shapiro–Wilk test for normality, conducted within each species and temperature group for families exhibiting variability in percent survival. Homogeneity of variance was confirmed using Levene's test across temperature, species and family. Paired t‐tests were conducted for each species and family, comparing survival at 35.5°C to 27.1°C, using a one‐sided alternative hypothesis that survival was lower at 35.5°C. To account for multiple comparisons and correct for false discovery rate, p‐values were adjusted using the Benjamini‐Hochberg (BH) method. To assess whether survival differed significantly between cross types within each species at 35.5°C, one‐way ANOVA were performed separately for each species. Tukey's HSD test was conducted post hoc to determine pairwise differences between cross types.

2.6. Boosted Regression Trees

BRT modelling was performed to assess the relative importance of different environmental predictors to larval survival under heat stress using stochastic gradient boosting. This approach was developed previously for the Great Barrier Reef and outlined in (Quigley and van Oppen ²⁰²²). This approach was favoured as BRT models use a limited number of user‐defined parameters and allow the modelling of non‐linear relationships whilst handling both qualitative and quantitative variables, and being robust to outliers, missing data and collinearity among variables (Buston and Elith 2011; De'ath ²⁰⁰⁷; Elith and Leathwick ²⁰¹⁷). BRT enhance predictive performance by combining multiple regression trees with a boosting algorithm, creating an ensemble model that aggregates predictions from simpler models, unlike traditional regression methods that rely on a single predictive model. All BRTs were fitted in R (version 4.3.3) (R Core Team ²⁰²⁴) using the ‘gbm’ package version 2.2.2 (Ridgeway ^2024a), including code provided by Elith et al. (2008).

Whilst there are different ways to approach the fitting of a boosted model, here we used a stochastic gradient descent, a stepwise process wherein each subsequent tree is fitted to the residuals of the set so far selected. By reducing the influence of each individual tree and averaging across the final ensemble, boosting produces a model that is more robust than a single regression tree and can capture complex, non‐linear relationships (Elith et al. 2008).

The binary response variable (individual larva survival) was coded as 1 = survived, 0 = dead. An initial BRT model was fitted using the gbm.step function of the ‘dismo’ package version 1.3–14 (Hijmans et al. 2024) with a Bernoulli error distribution. To prevent overfitting, which is particularly concerning for small datasets, a 10‐fold cross‐validation was performed using a ‘random’ search of the hyper‐parameter space to identify the optimal number of trees. Various parameter combinations and diagnostics were tested on our data, to assess if CV adequately balanced the squared error loss versus the number of trees. Model parameters were set to a tree complexity of 5, a learning rate of 0.001 and a bag fraction of 0.5. Therefore, the learning rate was slow enough to build models that estimate responses reliably whilst tree complexity was high enough to build models with complex interactions. More information about user defined parameters can be found in Elith et al. (2008) and Ridgeway (2024b).

The relative importance of the environmental predictors was evaluated based on the frequency with which each predictor was selected during tree construction and its contribution to overall model performance. The relative influence values were then normalised to sum to 100%. The function gbm.simplify was used to reduce model complexity by iteratively removing predictors with minimal influence, resulting in a simplified model containing only predictors with a relative influence of more than 5%. Model performance was evaluated using total mean deviance, residual deviance, cross‐validation deviance, training data correlation, cross‐validation correlation and receiver operating characteristic (ROC) scores. This sequential ensemble approach maximised predictive power by constructing, selecting and pruning models to identify the most influential predictors of heat stress survival in larvae. The function predict.gbm was used to predict values to new sites based on the generalised boosted model object. We applied a modelling approach using a Bernoulli error distribution on a dataset of 3035 larvae with associated reef‐scale environmental metrics. Model complexity and overfitting were managed through 10‐fold cross‐validation and iterative variable reduction. The final BRT model was fit with 2700 trees.

3. Results

3.1. Spatial Variation in Thermal Environment

All environmental metrics were derived from long‐term satellite time series extracted at a single pixel per site. The following analyses compare these temporal thermal histories across the four parental colony collection sites, and the same metrics extracted for 32 reef sites across Ningaloo were subsequently used as predictors in the BRT models. Mean annual sea surface temperature (SST) from 1985 to 2022 differed significantly across the four experimental sites: Oyster Stacks (O), Pelican Point (P), Tantabiddi (T) and Bruboodjoo (B) (ANOVA, p < 0.001). Pairwise comparisons of mean SST revealed significant differences among all sites, except between the most northerly sites, Tantabiddi and Oyster Stacks (Figure 1b, Tukey's HDS post hoc, p = 0.661, see Table S5). The warmest site, Tantabiddi, was on average 25.28°C ± 0.02 SE and the coolest, Pelican Point, was 24.2°C ± 0.02 SE.

Over this same period, the frequency of SST anomalies (SSTA) also differed significantly across the four sites (χ ² = 94.229, df = 3, Kruskal–Wallis, p < 0.001). Pairwise comparisons revealed significant differences in SSTA frequency between all sites (Figure 1c, see Table S6), which ranged from 7.85 to 6.46, where Oyster Stacks had the greatest number of anomalies, followed by Tantabiddi and Bruboodjoo.

Thermal stress anomalies (TSA_DHW) from 1985 to 2022 also differed significantly across the four sites (χ ² = 28.6051, df = 3, Kruskal–Wallis, p < 0.001). Pairwise comparisons revealed significant differences between all sites except between Oyster Stacks and Bruboodjoo (Dunn's post hoc p = 0.053) and between Tantabiddi and Pelican Point (Dunn's post hoc, p = 0.385). DTR from 1985 to 2022 varied across the four sites (Figure 1d,e). Bruboodjoo exhibited the highest mean DTR (1.71°C ± 0.59 SD), followed by Pelican Point (1.25°C ± 0.60 SD), Tantabiddi (1.15°C ± 0.47 SD) and Oyster Stacks (0.96°C ± 0.58 SD).

TSA from 2011 to 2022, following the 2011 marine heatwave, (TSA_DHWpost2011) also differed significantly across the four sites (χ ² = 38.0507, df = 3, Kruskal–Wallis, p < 0.001). Pairwise comparisons revealed significant differences between all sites except between Bruboodjoo and Pelican Point (Dunn's post hoc, p = 0.323) and between Oyster Stacks and Tantabiddi (Dunn's post hoc, p = 0.111). Frequency of DHW in 2011 can be seen in Figure 1f.

3.2. Variation in Larval Survival Between Crosses

Survival probability varied significantly across larvae from different family crosses (Figure 2). For Oyster Stacks and Pelican Point, after the final timepoint of 142 h, mean survival of A. tenuis larvae at heat was 58.2% ± 6.58 SE versus 96.9% ± 0.49 SE in the control treatment (Table S7) across all family crosses. For A. tenuis , 22 out of 27 families showed a significant difference in larval survival between heat and control (pairwise one‐sided t‐tests with Benjamini‐Hochberg adjusted p‐values, Table S8). At heat, larvae selectively bred from parents from the same reef (intrapopulation crosses) had a mean endpoint survival of 72.05% ± 3.74% (O × O) and 42.72% ± 4.84% (P × P), while those bred across different reefs (interpopulation crosses) were 65.62% ± 4.11% (O × P) and 52.43% ± 4.03% (P × O) (Table S9), where the first letter denotes maternal and second paternal reef of origin (see Section 2.3 for full notation details). Compared to P × P, survival was 1.7× higher (+68.6%) in O × O larvae, followed by 1.5× higher (+53.6%) in the interpopulation O × P cross and 1.2× higher (+22.7%) in P × O.

FIGURE 2.

(a) Variation in survival rates (mean ± SE) across individual A. tenuis and A. millepora families after 142 h (88 h for B × B and T × T, 92 h for A. millepora ) of exposure to 35.5°C (solid bars) and 27°C (faded bars). Asterisks indicate significant differences (α < 0.05) in survival at high versus ambient temperatures given by pairwise one‐sided t‐tests performed for each unique combination of species and family. p values were adjusted using the Benjamini‐Hochberg (BH) method to control false discovery rate. (b) Effect of parental origin on the survival (mean ± SE) of A. tenuis after 142 h (88 h for B × B and T × T, 92 h for A. millepora ) of exposure to ambient (27°C) and elevated (35.5°C) temperatures.

The final survival timepoint for Bruboodjoo and Tantabiddi was 88 h. At this time, mean A. tenuis survival at heat was 37.2% ± 7.65 SE versus 96.9% ± 0.93 SE in the control (Table S7) across all family crosses. Nine out of 11 families showed a significant difference in larval survival between heat and control treatments (pairwise one‐sided t‐tests with Benjamini–Hochberg adjusted p‐values, Table S8). Intrapopulation crosses at heat had a mean endpoint survival average of 59.47% ± 5.96% (T × T) and 18.58% ± 3.61% (B × B). No interpopulation crosses could be produced. Under heat stress, survival was 3.2× higher (+219.9%) in the intrapopulation T × T cross compared to B × B.

The final timepoint for A. millepora larvae was 92 h for Tantabiddi and Oyster Stacks, in which mean survival at heat was 51.9% ± 10.9 SE versus 96.5% ± 0.9 SE at control across both inter‐ and intrapopulation crosses (Table S7). Here, 11 out of 13 families showed a significant difference in survival between heat and control (pairwise one‐sided t‐tests with Benjamini–Hochberg adjusted p‐values, Table S8). At heat, larvae from intrapopulation crosses had a mean survival of 30.89% ± 4.09% (P × P), while those from interpopulation crosses were 67.42% ± 9.78% (O × P) and 57.33% ± 7.45% (P × O). Compared to P × P, survival was 2.2× higher (+118.3%) in O × P larvae and 1.9× higher (+85.6%) in the interpopulation P × O cross. In both A. tenuis and A. millepora, there were no significant differences between crosses in the control treatment (Table S11).

3.3. Key Environmental Drivers of Larval Survival

BRT were employed to identify key environmental predictors and their relative contribution in explaining variation in larval survival. The final model showed moderate predictive performance, with a mean cross‐validated AUC of 0.67 ± 0.01 (SE) and a cross‐validated correlation between observed and predicted values of 0.31 ± 0.01. The mean residual deviance was 1.28 ± 0.01 consistent with moderate predictive performance. Of the significant predictors, mean SST Anomaly (SSTA) frequency was the most influential (relative influence = 37.2%), followed by post‐2011 TSA (DHWpost‐2011, 22.7%), long‐term mean Daily Temperature Range (DTR, 21.7%), variability in Daily Temperature Range (DTR_stdev, 9.8%) and TSA since 1985 (TSA DHW, 8.5%) (Figure 3a). When all 20 SST predictors were included in the BRT, the top five predictors together accounted for 91% of the total variation in larval survival, while the remaining 15 each contributed < 2% (see Section 2.6).

FIGURE 3.

(a) Relative influence of each predictor variable in the final gradient boosted model, based on their contribution to reducing model deviance. (b) Partial dependence plots showing the marginal effect of individual predictors on the model's output (predicted probability of larval survival), while holding all other variables constant. (c) Three‐dimensional visualisations of the strongest pairwise interactions identified by the model. Left: Interaction between sea surface temperature anomaly (SSTa) frequency and thermal stress anomalies since 2011 (TSA_DHW post 2011); Centre: Interaction between SSTA frequency and mean daily temperature range (DTR); Right: Interaction between DTR and TSA_DHW post 2011.

BRT partial dependence plots of larval survival of both species revealed several non‐linear relationships and thresholds in each of the most influential predictors (Figure 3b,c). The greatest predicted survival occurred at relatively higher SSTA frequencies (~7.5 weeks), compared to 6.96 (range: 6.47–7.85, Figure 3b left panel). This represents a 7.2% increase. The step‐like change between ~6.5 and 7.0 reflects the tree‐based structure of the boosted regression model, which represents relationships as discrete splits rather than smooth functions. This produces apparent thresholds in partial dependence plots, with the increase around ~7.0 indicating a potential transition to higher predicted larval survival. The small peak near the lower bound (~6.5) is likely an artefact of limited data support at the lower end of the range, where model predictions are less reliable. A higher post‐2011 DHW (i.e., cumulative thermal stress since 2011) was positively associated with higher predicted survival up to levels of ~+1.2 DHW, after which the effect plateaued. Reefs with higher predicted larval survival (> 50%) had a mean post‐2011 TSA_DHW of 1.23, compared to 0.96 in the broader study area (range: 0.75–1.59, Figure 3b, centre panel), a 27.6% increase. Higher DTR was negatively associated with larval survival, as reefs with higher predicted survival had a mean DTR of 1.04°C, compared to 1.26°C (range: 0.90°C–1.79°C, Figure 3b, right panel), a 17.5% reduction. Lower DTR stdev was associated with higher predicted survival: 0.54 with high survival versus 0.57 with low survival (range: 0.47–0.64, Figure S1), a 5.3% decrease. TSA_DHW showed a non‐linear relationship with predicted survival. High survival was predicted for larvae that originated from reefs with a mean TSA_DHW of 0.73, compared to 0.64 (range: 0.51–0.96, Figure S1), representing a 14.8% increase. Two‐way interaction plots demonstrate synergistic effects between predictors (Figure 3c).

3.4. Predicting Reef Sites of High Offspring Survival Under Heat Stress

The relationships from the models were used to develop spatially explicit forecasts of larval survival under heat stress across Ningaloo. Predictions were made by applying the trained gradient boosted model to a spatial dataset of environmental conditions encompassing the five most influential predictors identified in the BRT analysis (Table S12). Model projections identified six of 35 reefal areas as having higher survival (> 45%) based on the overlap in predictors (Figure 4). Reefs with predicted larval survival probabilities exceeding 45% were concentrated in the northern and central sections of the Ningaloo Reef, representing 17% of the total reef area analysed. In contrast, the region extending from south of Point Cloates to north of Coral Bay, where the reef crest lies farther offshore and the lagoon is substantially wider, exhibited consistently lower predicted survival (< 30%).

The highest survival reefs were spatially clustered north of Point Cloates; Tantabiddi, Turquoise Bay, Oyster Stacks, Kurrajong, north of One K campground and South Lefroy Bay. On average, high‐survival sites experienced more frequent positive sea surface temperature anomalies (SSTA frequency: 7.21 weeks vs. 6.91 weeks overall mean, +4.3%), positioning them in the upper quartile of thermal anomaly exposure across Ningaloo. High‐survival reefs exhibited mean post‐2011 DHW values of 1.14 compared to 0.96 overall (+18.8%), aligning closely with the optimal threshold (~1.2 DHW) identified in the BRT partial dependence plots where survival benefits plateaued. Importantly, these reefs demonstrated substantially lower daily temperature ranges (mean DTR: 1.02°C vs. overall mean 1.27°C, a drop of −19.9%). Long‐term thermal stress (TSA DHW) was slightly higher at high‐survival sites (0.65 vs. 0.63, +3.5%), consistent with the non‐linear relationship observed in the BRT analysis. Together, these predictors suggest that reefs with higher larval survival under heat stress experience frequent thermal anomalies alongside intermediate cumulative heat stress and thermal variability relative to local conditions, potentially promoting survival under elevated temperatures.

4. Discussion

Here we show that corals' thermal history, including maximum temperatures and daily variability jointly shape coral offspring resilience to acute heat stress in Acropora corals. Specifically, larvae from colonies originating from reef locations with higher frequency of positive thermal anomalies (SSTA frequency), moderate cumulative heat stress relative to local conditions (post‐2011 DHW) and lower DTR showed significantly higher survival at extreme temperatures (35.5°C) compared to those from reefs with other combinations and ranges of temperature metrics. This pattern aligns with growing evidence that exposure to high‐frequency temperature variability can enhance coral heat tolerance (Quigley and van Oppen ²⁰²²) or reduce bleaching probability (Safaie et al. 2018). However, the magnitude of benefit depends on the specific variability regime experienced (Brown, Eyal, et al. ²⁰²³; Brown, Lenz, et al. ²⁰²³; Schoepf et al. ²⁰²²). The high predictive power of thermal history here suggests that local environmental conditions strongly modulate offspring thermal performance, even across relatively small spatial scales (Kler Lago et al. 2025). However, spatial predictions are best interpreted within the context of the thermal variation empirically captured across the four collection sites, which were selected to span the breadth of conditions present along the reef, and predictions at sites approaching the outer limits of this range should be treated with appropriate caution. Increased larval survival under heat in warmer compared to cooler reefs (O × O > O × P > P × O > P × P, where O = Oyster Stacks and P = Pelican Point) follows the expected pattern, consistent with prior evidence that maternal colony thermal history strongly influences offspring survival (Quigley et al. 2016). Indeed, maternal identity strongly influences early life history traits in corals such as larval survival under heat and settlement success (Dixon et al. 2015; Johnston et al. ²⁰²⁴). These findings support the hypothesis that adult colonies residing in historically thermally stressful environments may act as reservoirs of heat‐tolerant offspring. However, divergent survival and transcriptional responses between life stages could constrain the rate of adaptation if heat‐responsive gene expression patterns differ between adults and offspring (Macadam et al. 2025; Ruggeri et al. ²⁰²³). This work collectively underscores the importance of identifying and protecting coral reef regions with the ability to contend with marine heatwaves.

Interestingly, BRT models revealed non‐linear relationships between SST metrics and larval survival, with values of SSTA frequency ~7.5 weeks/year, DTR ~1°C, and cumulative stress TSA_DHW ~0.7–1.2 associated with highest predicted survival. These thresholds suggest the occurrence of an optimal “Goldilocks zone” of thermal stress; enough to prime physiological responses without overwhelming thermal performance capacity (Hazraty‐Kari et al. ²⁰²²; Safaie et al. ²⁰¹⁸). Recent work examining DTR across Indo‐Pacific and Caribbean reefs found inner reef flats had the highest mean DTR at 1.4°C, with outer reef flats and shallow lagoons exceeding 1°C, while most other reef habitats had DTRs between 0.5°C–1°C (Bos and Pinsky 2025). In Moorea's backreefs, annual mean DTR can range between 0.63°C–1.22°C (Cheh 2025), whereas highly heat tolerant coral larvae on the GBR originate in reef regions with ~0.5°C DTR (Quigley and van Oppen ²⁰²²). In comparison, Ningaloo's DTR (0.90°C–1.79°C) lies at the moderate to high end of typical reef variability. The negative relationship between survival and DTR > 1.5°C in our study is consistent with Brown et al. (2024) who found a parabolic response whereby thermal tolerance peaked at intermediate daily variability and declined at both extremes. Our results suggest that intermediate thermal fluctuations may prime tolerance, while high‐amplitude variability (DTR > 1.5°C) imposes increased energetic costs from repeated stress responses. Importantly, what constitutes ‘intermediate’ variability is relative to local conditions and the absolute DTR values at which tolerance peaks will differ between reef systems. This suggests that the relationship between DTR and thermal tolerance follows an optimum curve, where the beneficial priming window is defined by local thermal context rather than a universal threshold. The physiological and molecular mechanisms underlying these thresholds (Guerrero 2024; van Oppen and Oakeshott ²⁰²⁰) are beyond the scope of this study but represent important avenues for investigation.

The emergence of post‐2011 TSA_DHW as a major predictor suggests recent disturbance disproportionately shapes larval survival under heat stress compared to longer term temperature trends at Ningaloo. This suggests that recent events may leave an impact on surviving individuals of these reef ecosystems (“ecological memory,” sensu Hughes et al. ²⁰¹⁹). Alternatively, the 2011 heatwave may have filtered out thermally sensitive genotypes. Taken together, this supports the importance of disturbance history in forecasting future disturbance, recovery and recruitment patterns (Brown and Barott 2022; Hackerott et al. ²⁰²¹; Lenz et al. ²⁰²⁵). Importantly, although these values are below 4°C–weeks, the threshold typically associated with mass bleaching (NOAA Coral Reef Watch), there is a detectable pattern in variation in offspring performance. This underscores that even potentially sublethal stress may act as a selective filter on thermally sensitive genotypes, influencing the distribution of heritable thermal tolerance across reef populations. Finally, the interaction of these environmental predictors highlights that heritable larval survival is not shaped by only single stressors in isolation, but rather by the joint influence of chronic thermal regimes and recent thermal history. It further emphasises the role of environmental variability in shaping resilience. Together, these results suggest that coral populations residing in warm, variable environments with a history of thermal stress likely have on‐average greater heat tolerance.

Reefs in the northern and central regions of Ningaloo showed the highest predicted larval survival. In contrast, southern sites below Point Cloates exhibited lower predicted survival, suggesting they may be less suitable sources of thermally resilient broodstock. Six reefs (17% of the surveyed area) with predicted survival > 45% represent potential sources of heat‐tolerant genotypes. This regional clustering suggests that local oceanographic and geomorphological features may generate persistent thermal regimes that contribute to spatially structured resilience. Across the Ningaloo‐Pilbara region, hydrodynamic and genetic studies indicate pronounced spatial variability in larval settlement, restricted gene flow and signals of local adaptation in reef‐associated taxa (Feng et al. 2016; Thomas et al. ²⁰¹⁷), suggesting that natural spread of heat‐tolerant genotypes from northern to southern reefs may be limited. This underscores the potential value of assisted gene flow and targeted restoration as management tools for this region. Importantly, the extent of coral reef thermal refugia is projected to decline from 84% of reef pixels (1 km resolution SST grid cells) at present to near zero under 1.5°C–2.0°C of global warming (Dixon et al. 2022). This suggests that even cooler than average locations at Ningaloo will face substantial future warming risk, underscoring the critical need to protect and manage this region. Therefore, coral reef management that incorporates future climate scenarios is increasingly essential for long‐term reef resilience planning under climate change (Caruso et al. 2021).

The boosted regression tree model was trained on 3035 individual larval survival outcomes from selectively bred Acropora families, with each observation linked to a full suite of 20 SST‐derived predictors from the maternal colony's reef of origin. Model performance in this framework is driven primarily by the number of response observations rather than the number of sampling locations, as the algorithm learns relationships from the full distribution of individual‐level outcomes (Elith et al. 2008). This approach is supported by evidence that the four source sites span a substantial thermal gradient and collectively bracket the conditions present across all 32 prediction sites, indicating that model predictions are made through interpolation within a defined environmental space rather than extrapolation. The robustness of this framework is further supported by the close agreement between training and cross‐validated performance metrics, suggesting that the model captures a consistent biological signal rather than site‐specific structure.

This BRT model supports current theoretical frameworks whilst also offering a practical tool for reef‐scale spatial planning. By identifying spatial locations and gradients associated with corals capable of producing heat‐tolerant offspring, it enables conservation practitioners to identify priority zones for reef protection, larval seeding or the assisted movement of colonies. Future work could refine these insights by integrating larval dispersal, reef connectivity and fine‐scale oceanographic data, further enhancing spatial prediction. The BRT model captured ~31% of the variation in larval survival, with an AUC of 0.67 ± 0.01 (SE), indicating moderate predictive power. While not highly discriminative at the individual larval level, this performance is consistent with expectations for ecological systems where survival is shaped by a complex interplay of measured and unmeasured factors operating across multiple timescales. Future iterations of this modelling framework would benefit from SHAP‐based interpretability approaches, which offer additionally granular insight into individual variable contributions and are gaining traction in ecological machine learning applications (Cheung et al. 2025).

Satellite‐derived SST data have known limitations, including coarse spatial resolution, measurement of skin (the uppermost ~10–20 μm of the ocean surface) rather than bulk subsurface temperature, and potential biases under cloud cover or in shallow reef environments. Although quality control flagging removed cloud‐contaminated observations in DTR, peak warming events coinciding with persistent cloud cover may still have been underrepresented in our thermal metrics, an inherent limitation of infrared‐based satellite SST products. Nevertheless, our ability to derive consistent seasonal statistics across all seasons suggests data gaps were not extensive or seasonally concentrated enough to substantially bias our results. Furthermore, because all environmental predictors were derived from the same remotely sensed datasets, relative spatial patterns and thermal gradients remain internally consistent and comparable across locations. Additionally, factors not captured by satellite‐derived SST metrics, including symbiont community composition, host genetic variation, parental colony size and condition and local microbiome differences, likely contribute substantially to offspring performance (Howells et al. 2013; Voolstra et al. ²⁰²¹).

The exclusion of geographic coordinates (latitude and longitude) was deliberate to prioritise biophysical predictors, avoid inflating spatial specificity and to maximise transferability of the framework beyond Ningaloo. However, we acknowledge that this choice also means spatial autocorrelation in model residuals was not explicitly assessed and may indicate residual geographic structuring not captured by environmental predictors. Future applications of this framework would benefit from explicit testing for spatial autocorrelation, for example using Moran's I on model residuals, and from incorporating spatial cross‐validation approaches to further validate predictive performance across the reef. The model's strength therefore lies in its ability to identify environmental gradients and interactions to generate spatially explicit predictions of higher survival probabilities. This is especially valuable for spatial forecasting and conservation planning at the reef scale where exhaustive colony screening is not logistically feasible.

In summary, this work identifies locations along Ningaloo World Heritage reef that likely harbour corals with the ability to produce thermally resilient offspring, bridging experimental physiology with spatial prediction. Importantly, the non‐linear relationships detected here emphasise that temperature variation may drive heritable heat tolerance. We note that our data have not addressed whether Ningaloo's thermally tolerant offspring will have greater settlement success, growth rates or competitive ability, which are critical questions for predicting population‐level outcomes. Future work should investigate genetic bases of these performance differences, assess whether tolerance persists across life‐history stages, and model dispersal connectivity between these high and low survival probability reefs. Additionally, the integration of finer‐resolution oceanographic data would likely enhance our spatial predictions. As projected 0.5°C–1.0°C increases in SST in Western Australia by 2040 (CSIRO ²⁰²¹) threaten both vulnerable and resilient coral reef sites beyond their thermal thresholds (Quigley and Baird 2024), protecting coral populations while they persist represents a priority for Ningaloo and similar reef ecosystems worldwide.

Author Contributions

Teresa Baptista Nobre: conceptualization, data curation, formal analysis, investigation, visualization, writing – original draft. Paige J. Maroni: writing – review and editing. Nicole L. Jones: writing – review and editing. Kate M. Quigley: conceptualization, investigation, funding acquisition, writing – review and editing, resources.

Funding

This work was supported by the Government of Western Australia; Australian Research Council, DE230100284; Minderoo Foundation.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: List of explanatory variables used in the boosted regression tree modelling. Predictors with highest relative influence in the final gradient boosted model are highlighted.

Table S2: Total number of unique maternal and paternal colonies used per reef.

Table S3: Total number of families Acropora tenuis and Acropora millepora .

Table S4: Parental colony identifier (Reef.Species.number) and cross produced.

Table S5: Results of Tukey's HSD pairwise comparisons following a one‐way ANOVA for differences in mean annual SST between 1985 and 2022.

Table S6: Results of the Kruskal–Wallis Test (χ²) and Dunn's post hoc test for differences in TSA‐DHW and SSTfreq across locations.

Table S7: Summary statistics for each species by temperature.

Table S8: Results of one‐sided t‐tests comparing larval percent survival between heat (35.5°C) and control (27.1°C) treatments for each A. tenuis and A. millepora family. p‐values were adjusted using the Benjamini–Hochberg (BH) method to control the false discovery rate.

Table S9: Summary statistics for each cross grouped by species and temperature.

Table S10: Results of Tukey's HSD post hoc test following a one‐way ANOVA examining the effect of cross identity on A. tenuis and A. millepora larval percent survival under heat treatment conditions (35.5°C). The analysis tested for pairwise differences in survival between crosses.

Table S11: Results of Tukey's HSD post hoc test following a one‐way ANOVA examining the effect of cross identity on A. tenuis and A. millepora larval percent survival under ambient temperature conditions (27.1°C). The analysis tested for pairwise differences in survival between crosses.

Table S12: Environmental predictor values and predicted larval survival probabilities for reef sites across Ningaloo Reef. Values shown are the five most influential environmental predictors identified in the Boosted Regression Tree analysis: mean sea surface temperature anomaly frequency (SSTa_freq, weeks), post‐2011 thermal stress anomalies (post2011_tsa_dhw, degree heating weeks), mean daily temperature range (DTR, °C), standard deviation of daily temperature range (sd_DTR, °C) and long‐term thermal stress anomalies since 1985 (tsa_dhw, degree heating weeks). Predicted survival probabilities were generated by applying the trained gradient boosted model to the spatial environmental dataset. Sites are ordered from north to south along Ningaloo Reef. Sites with predicted survival > 45% (indicating higher survival under heat stress) are shown in bold.

Figure S1: Generalised Boosted Regression Models (GBMs), specifically focusing on Recursive Feature Elimination (RFE) and predictive deviance with 10‐fold cross‐validation. The solid black curve is the mean, and the dotted curves ±1 standard error, for the changes in predictive deviance (i.e., as measured on the excluded folds of the CV). The red line shows the minimum of the mean, and the green line the number of trees at which that occurs. The final model that is returned in the model object is built on the full data set, using the number of trees identified as optimal.

Figure S2: Partial dependence plots showing the marginal effect of temperature‐related predictors on the probability of larval survival under heat stress. Each plot shows the fitted marginal effect of a single predictor on the response variable (larval survival probability), holding all other variables constant at their average values. The y‐axis represents the predicted probability of survival, while the x‐axis shows the values of the respective predictor. These plots are derived from the final boosted regression tree model used to forecast larval survival outcomes.

Data Availability Statement

The data and code that support the findings of this study are openly available at https://github.com/TeresaNobre/ningaloo‐survival‐sst‐modelling and published at https://doi.org/10.5281/zenodo.21017257.