Fig 1.
Evaluation and prediction of fish aesthetic values.
(a) Workflow of the online survey and deep learning prediction of aesthetic values. (1) Pairs of images were presented to the public during the online survey and scored using the Elo algorithm (see Methods). Left Parma bicolor and right Abudefduf luridus. (2) Once the 345 new images were evaluated online, the values of the 157 images previously evaluated [16] were corrected using the 21 images shared between the 2 surveys. (3) The resulting 481 images with evaluated aesthetic values were used to train a ResNet50 algorithm (see Text E and Fig L in S1 File). Illustration inspired from the PlotNeuralNet [31]. (b) Left: The r2 of the linear relationship between the predicted values averaged across the 5 validation sets and the evaluated values is 0.79 ± SD 0.04 (the color of points indicates the 5 sets used to perform the cross validation). This algorithm was used to predict the aesthetic values of the 4,400 unevaluated images of our dataset. Right: Distribution of the 481 evaluated values in light blue and of the 4,400 predicted aesthetic values in dark blue. The dots at the bottom of the plot indicate the predicted aesthetic values of the images shown in panel (c). Data and code required to generate this Figure can be found in https://github.com/nmouquet/RLS_AESTHE. (c) Examples of fishes representative of the range of predicted aesthetic values. Decreasing aesthetic value from left to right and top to bottom: Holacanthus ciliaris, Aracana aurita, Amphiprion ephippium, Ctenochaetus marginatus, Scarus spinus, Amphiprion bicinctus, Epinephelides armatus, Fusigobius signipinnis, Diplodus annularis, Odontoscion dentex, Nemadactylus bergi, Mendosoma lineatum. See S1 Data for image copyright.
Fig 2.
(a) Regression coefficients (with standard errors) from the final model between the aesthetic value and the 9 significant image features (see Text A in S1 File for a complete description of the image features). The variables have been scaled to visualize how the magnitude of the effects differs between them. Most important variables were: color heterogeneity, color saturation, standard deviation in lightness (SD lightness), pattern repetition, and body elongatedness. (b) Principal Component Analysis (PC1 and PC2) performed with the 9 significant images features (see Fig 2A, Text A, F in S1 File for a description of the image features). Points (fishes) are colored by their aesthetic values, and image feature vectors are projected on the 2 axes. Examples of fishes (chosen on the perimeter of the distribution) are provided for illustration. Clockwise order: Calloplesiops altivelis, Epinephelus ongus, Kyphosus vaigiensis, Epinephelus costae, Jenkinsia lamprotaenia, Phyllogobius platycephalops, Belone belone, Ctenogobiops crocineus, Suezichthys devisi, Opistognathus aurifrons, Pseudanthias ignitus, Pomacentrus auriventris, Mecaenichthys immaculatus, Pomacanthus navarchus, Aracana aurita, Pomacanthus sexstriatus. See S1 Data for image copyright. Data and code required to generate this Figure can be found in https://github.com/nmouquet/RLS_AESTHE.
Fig 3.
Phylogenetic history and ecological originality.
(a) Relationship between the aesthetic value and the age of the species (log transformed) in millions of years (averaged over 100 trees) without (plain line) and after (dashed line) accounting for phylogenetic relatedness. Both models show significant negative slopes (considering phylogenetic relatedness: slope = ‒78.4, p-value < 0.001; not considering phylogenetic relatedness: slope = ‒14.1 ± SD 1.7, p-value < 0.005, over the 100 random trees). (b) Relationship between the aesthetic value and the functional distinctiveness of species without (plain line) and after (dashed line) accounting for phylogenetic relatedness. Both models show significant negative slopes (considering phylogenetic relatedness: slope = ‒1,154.2, p-value < 0.001; not considering phylogenetic relatedness: slope = ‒383.3 ± SD 26.5, p-value < 0.001, over the 100 random trees). On both panels, species’ Evolutionary Distinctiveness (averaged over 100 trees and log transformed) have been used to color the points from low (dark red) to high (dark blue) values. Data and code required to generate this Figure can be found in https://github.com/nmouquet/RLS_AESTHE.
Fig 4.
Aesthetic values across the tree of life.
Phylogenetic tree of the 2,417 fishes. Aesthetic values are mapped over the entire phylogeny with a color gradient obtained by estimating states at internal nodes with maximum likelihood [36]. For illustration, we have highlighted 20 families with contrasted aesthetic values using gray arcs and show examples of fishes for each family. Clockwise order: Cantherhines macrocerus, Pseudobalistes naufragium, Anampses femininus, Scarus spinus, Bodianus unimaculatus, Myripristis jacobus, Gymnothorax annasona, Meiacanthus atrodorsalis, Embiotoca jacksoni, Amphiprion bicinctus, Abudefduf bengalensis, Chromis alpha, Carangoides chrysophrys, Istigobius decoratus, Apogon pacificus, Sarda australis, Pterois miles, Acanthistius ocellatus, Amblycirrhitus pinos, Parequula melbournensis, Pomacanthus navarchus, Chaetodon flavirostris, Diplodus puntazzo, Acanthurus tristis. See S1 Data for image copyright. Data and code required to generate the phylogenetic tree can be found in https://github.com/nmouquet/RLS_AESTHE.
Fig 5.
Fish aesthetic value and conservation status.
(a) Violin plot of the aesthetic value of reef fishes for three groups of conservation status: TH, NE, and LC. Letters indicate significant differences between the groups (Tukey p-values are respectively p < 0.01 between TH and NE, p < 0.001 between LC and TH, and p < 0.001 between LC and NE). (b) Violin plot of the aesthetic value of reef fishes for the 5 groups of fishery importance: “Data deficient” = no data available; “Non commercial” = no interest for fisheries or potential interest or minor interest; “Subsistence fisheries” = importance for subsistence fisheries; “Commercial” = commercial importance for fisheries; “Highly commercial” = high commercial importance for fisheries. Letters indicate significant differences between the groups (all Tukey p-values are < 0.002). Data and code required to generate this Figure can be found in https://github.com/nmouquet/RLS_AESTHE. LC, Least Concern; NE, Not Evaluated; TH, Threatened.