Supplementary methods

Boosted regression trees

Temperature loggers are placed on the metal picket at the start of the first transect on long-term monitoring program (LTMP) survey reefs. Loggers are changed at each visit and the data downloaded into a database within the Australian Institute of Marine Science Data Centre. Temporal records matching the LTMP fish data series were not available at all reefs. For the boosted regression tree (BRT) analysis, we used temperature records where available. When temperature data was missing, we used the average temperature of the month in which surveys were conducted at a given reef as the input for the BRT models. Divers also recorded substrate rugosity at the start of each transect. Substrate rugosity is a measure of the complexity of the underlying reef matrix, without the inclusion of live and dead coral skeletons, and is recorded on a categorical scale between zero (least complex — minimal vertical relief, few holes, crevices and overhangs) and five (most complex — high vertical relief, many holes, crevices and overhangs) This 0–5 scale correlates strongly with a range of other rugosity metrics and has been found to be a good predictor of reef fish diversity and abundance (Wilson et al. 2007). Temperature, substrate rugosity and other benthic and spatial predictors were included as explanatory variables in BRTs to examine the habitat association of predatory fishes of the Great Barrier Reef (De’ath 2007; Elith et al. 2008). Optimal tree complexity, learning rate and bag fraction were determined following Elith et al. (2008). Tree complexity sets the number of divisions for each tree per iteration and was restricted to five to prevent over-fitting models. We tested models with differing learning rates (0.1, 0.01, 0.05, 0.001) and bag fractions (0.5 and 0.8), which respectively alters the effect of the primary tree on the final solution and sets the percentage of the dataset used. We chose the model that kept the optimal number of trees as close to 1000 as possible (to reduce among-model variance) while explaining the greatest deviance between the predicted and observed values.

Analysis of the effects of disturbance

ESM Table S1 Description of disturbances affecting LTMP reefs showing the absolute and relative change (Δ) in hard coral cover (HCC). Disturbances are categorised by impacts using quartiles of the change in relative percentage HCC. Low impacts had a change of between 10.88 and -31.16%, moderate impacts -31.17 to -46.23%, high impacts -46.24 to 66.89% and very high impacts were greater than -66.90%.

Disturbance / Magnitude / Reef / Absolute Δ HCC / Relative Δ HCC
Bleaching / Low / 19131S / -8.74 / -20.99
HASTINGS REEF / -5.13 / -19.0
MARTIN REEF(14123) / -5.75 / -22.80
Moderate / MYRMIDON REEF / -11.02 / -31.87
PANDORA REEF / -18.98 / -32.49
A. planci / Low / SLATE REEF / -5.07 / -11.18
Moderate / GREEN ISLAND REEF / -5.83 / -36.14
MACGILLIVRAY REEF / -9.56 / -36.14
THETFORD REEF / -13.1 / -36.33
High / CHICKEN REEF / -7.76 / -49.65
GREEN ISLAND REEF / -6.94 / -47.08
TURNER REEF / -17.45 / -56.24
Very High / GANNETT CAY REEF / -42.48 / -91.95
HORSESHOE / -41.31 / -79.44
Cyclone / Low / AGINCOURT REEFS (NO 1) / -6.03 / -18.0
OPAL (2) / -5.06 / -19.29
Moderate / MICHAELMAS REEF / -9.45 / -40.77
High / LIZARD ISLAND / -16.41 / -64.84
ST CRISPIN REEF / -13.8 / -46.67
Disease / Low / HORSESHOE / -13.4 / -21.28
LADY MUSGRAVE REEF / -8.16 / -10.88
Moderate / AGINCOURT REEFS (NO 1) / -12.16 / -32.74
Multiple / Low / 21529S / -8.77 / -14.44
HASTINGS REEF / -8.15 / -26.01
LOW ISLANDS REEF / -5.08 / -14.34
LOW ISLANDS REEF / -5.2 / -19.05
Moderate / 20104S / -17.32 / -35.69
THETFORD REEF / -11.43 / -37.11
High / CARTER REEF / -26.79 / -46.78
MACGILLIVRAY REEF / -15.94 / -56.99
NO NAME REEF / -26.65 / -52.61
YONGE REEF / -34.8 / -58.87
Very High / 21529S / -44.78 / -97.45
BROOMFIELD REEF / -25.7 / -67.23
FITZROY ISLAND REEF / -25.11 / -78.62
HAVANNAH REEF / -38.07 / -88.87
JOHN BREWER REEF / -27.33 / -94.47
LINNET REEF / -46.17 / -91.07
MARTIN REEF(14123) / -22.46 / -70.34
NORTH DIRECTION REEF / -27.26 / -75.30
ONE TREE REEF / -73.83 / -96.98
RIB REEF / -59.14 / -89.28
SLATE REEF / -31.89 / -73.16
WRECK ISLAND REEF / -66.68 / -93.36
Storm / Low / CHICKEN REEF / -5.44 / -12.83
DAVIES REEF / -6.12 / -15.06
MYRMIDON REEF / -8.87 / -30.93
PANDORA REEF / -10.84 / -21.78
Moderate / 19131S / -23.41 / -39.02
19138S / -18.9 / -40.04
CARTER REEF / -16.85 / -39.99
DAVIES REEF / -12.73 / -40.16
DIP REEF / -10.56 / -36.13
HAYMAN ISLAND REEF / -15.08 / -32.51
NO NAME REEF / -17.6 / -46.04
High / 19138S / -25.11 / -60.54
CHICKEN REEF / -7.76 / -49.65
DIP REEF / -11.51 / -53.19
HYDE REEF / -11.46 / -65.90
RIB REEF / -7.34 / -46.43
SNAKE (22088) / -26.1 / -48.61
YONGE REEF / -23.2 / -56.22
Very High / GANNETT CAY REEF / -26.86 / -81.08
LADY MUSGRAVE REEF / -76.15 / -98.87
REBE REEF / -15.63 / -86.21
Unknown / Low / NO NAME REEF / -7.42 / -12.54

Observer comparisons

The LTMP has run for more than two decades, and it was unavoidable that multiple observers have been used during this time. To establish that there were no inherent biases due different observers in the counts of reef-associated predatory fishes, we analysed the abundance of individual fish species counted by different observers. Fish abundance was pooled to site to reduce the number of zeros that are inherent in ecological count data and then modelled using Bayesian hierarchical linear mixed models. Each species was modelled individually with the fixed effect of observer and random effects fitted for reef and site to account for spatial autocorrelation. Response variables were modelled against a zero-inflated negative binomial in the Bayesian Regression Models using Stan (brms) package in R

Fig. S1 Comparison of counts of individual species of reef-associated predatory fish on the GBR by different observers. Data are modelled higher posterior distribution median count 250 m–2 and the associated 95% uncertainty intervals (UIs) from Bayesian hierarchical linear mixed models. Statistical differences among observers can be inferred where 95% UIs do not overlap

Fig. S2 Time averaged spatial patterns in the (a, b) density and (c, d) species richness of reef-associated predatory reef fishes on the GBR, only from reefs open to fishing. These models were used to examine the effects of management zone (open and closed to fishing) on the broad-scale spatio-temporal patterns in predatory reef fished and can be compared to Fig. 2 in the main article. Data are modelled higher posterior distribution means and the associated 95% uncertainty intervals (UIs) from Bayesian hierarchical linear mixed models. Statistical differences among sub-regions can be inferred where 95% UIs do not overlap. Latitudinal sectors are CL = Cooktown-Lizard Island, CA = Cairns, TO = Townsville, WH = Whitsunday, SW = Swain, CB = Capricorn-Bunker

Fig. S3 Temporal patterns in the (a, b) density and (c, d) species richness of reef-associated predatory reef fishes on the GBR. These models were used to examine the effects of management zone (open and closed to fishing) on the broad-scale spatio-temporal patterns in predatory reef fished and can be compared to Fig. 2 in the main article. Data are modelled higher posterior distribution means and the associated 95% uncertainty intervals (UIs) from Bayesian hierarchical linear mixed models. Statistical changes through time can be inferred where 95% UIs do not overlap. Data from the Cooktown-Lizard Island outer-shelf sub-region are not included in this model because these reefs changed zone from open to closed in the re-zoning of the GBR 2004. Latitudinal sectors are CL = Cooktown-Lizard Island, CA = Cairns, TO = Townsville, WH = Whitsunday, SW = Swain, CB = Capricorn-Bunker

Table S2 Average lengths (SE) of five species of predatory fishes with inshore restricted distribution. Fish were measured by observers during underwater visual census after 2007. NA = no individuals recorded.

Length (cm) / Max reported length (cm)
Species / Inshore / Mid-shelf / Outer-shelf / Source
Lutjanus sebae / 32.75 (1.28) / NA / NA / 60
110 / Randall et al. 1997
www.fishbase.org
Lethrinus laticaudis / 37.7 (1.45) / NA / NA / 56
56 / Randall et al. 1997
www.fishbase.org
Lujanus vitta / 25.6 (0.31) / NA / NA / 40
40 / Randall et al. 1997
www.fishbase.org
Lutjanus argentimaculatus / 43.1 (2.02) / 44.0 (1.53) / NA / 120
150 / Randall et al. 1997
www.fishbase.org
Lutjanus lemniscatus / 35.4 (1.11) / 35.0 (NA) / NA / 65
65 / Randall et al. 1997
www.fishbase.org

References

De’ath G (2007) Boosted trees for ecological modeling and prediction. Ecology 88:243–251

Elith J, Leathwick JR, Hastie T (2008) A working guide to boosted regression trees. J Anim Ecol 77:802–813

Randall JE, Allen GR, Steene RC (1997) Fishes of the Great Barrier Reef and Coral Sea. University of Hawaii Press, Honolulu

Wilson SK, Graham NAJ, Polunin NVC (2007) Appraisal of visual assessments of habitat complexity and benthic composition on coral reefs. Mar Biol 151:1069–1076