Distance-dependent defensive coloration in the poison frog Dendrobates tinctorius, Dendrobatidae

Significance

Poison dart frogs are well known for their deadly toxins and bright colors; they are a classic example of warning coloration. However, conspicuousness is not the only consideration; defensive coloration must be effective against a diverse predator community with a variety of different visual systems, and variable knowledge of prey defenses and motivation to attack. We found that the bright colors of Dendrobates tinctorius are highly salient at close-range but blend together to match the background when viewed from a distance. D. tinctorius combines aposematism and camouflage without necessarily compromising the efficacy of either strategy, producing bright colors while reducing encounters with predators. These data highlight the importance of incorporating viewing distance and pattern distribution into studies of signal design.

Abstract

Poison dart frogs provide classic examples of warning signals: potent toxins signaled by distinctive, conspicuous coloration. We show that, counterintuitively, the bright yellow and blue-black color of Dendrobates tinctorius (Dendrobatidae) also provides camouflage. Through computational modeling of predator vision, and a screen-based detection experiment presenting frogs at different spatial resolutions, we demonstrate that at close range the frog is highly detectable, but from a distance the colors blend together, forming effective camouflage. This result was corroborated with an in situ experiment, which found survival to be background-dependent, a feature more associated with camouflage than aposematism. Our results suggest that in D. tinctorius the distribution of pattern elements, and the particular colors expressed, act as a highly salient close range aposematic signal, while simultaneously minimizing detectability to distant observers.

Poison dart frogs (Dendrobatidae) are well known for their striking aposematic (warning) signals: distinctive, conspicuous coloration signaling potent toxins (1, 2). Predators learn the association between prey coloration and toxic defense, and bright and highly contrasting color patterns have been demonstrated to increase the speed, accuracy, and longevity of predator-avoidance learning (3–5).

Aposematic species are not, however, immune to predation (6–10). Naïve and specialized predators will ignore warning coloration, and even susceptible predators will actively manage their intake of defended prey in accordance with their nutritional requirements and toxin burden (6–10). For example, birds and snakes have been observed predating toxic dendrobatid (Dendrobates, Phyllobates, and Oophaga spp.) frogs (11–13). Consequently, maximizing detectability is not necessarily the optimal strategy for preventing predation, and defended species may benefit from incorporating aspects of camouflage into their coloration (14–20).

Indeed, rather than being alternative and mutually exclusive forms of defensive coloration, camouflage and aposematism are now frequently considered along a continuum from inconspicuous to highly salient (21). Defended prey trade off the benefits of conspicuous warning signals against a low predator encounter rate, frequently resulting in weaker defenses being associated with smaller and less-saturated aposematic components (20, 22, 23). An alternative, which can maintain color saturation and signal size, is aposematic signals that may also act as context-dependent disruptive camouflage (24) or distance-dependent background matching (14, 21).

Disruptive coloration breaks up the outline of a target into unrecognizable patches, which blend into different background components (25). Because both aposematic and disruptive patterns benefit from highly contrasting colors (3, 5, 26, 27), and disruption does not necessarily require that colors match the background (28), there is potential for cooption of similar pattern components despite the opposing processes (21). However, whereas consistent symmetrical patterns are more easily learned and remembered (29), asymmetric and variable patterns are more effective at concealing prey (26, 30, 31). It has been demonstrated that conspicuous warning signals can appear cryptic in particular microhabitats (24), but it is unknown how widespread this may be in nature.

Distance-dependent patterns, on the other hand, take advantage of limitations in observer visual acuity to appear highly salient at close range, but camouflaged when viewed from a distance (15, 16, 18, 19, 32, 33). This effect can be achieved by either combining patterns of different spatial frequency (size) (18) or by the color components blending together to form a cryptic average color (pattern blending) (19, 32, 33). Previous work using artificial prey suggests that pattern blending can have greater survival than either camouflage or aposematism alone (18, 19). However, although these model systems take inspiration from ecologically relevant patterning, little work has been done into naturally occurring color patterns (15, 16, 32).

Dendrobates tinctorius (Dendrobatidae) is a brightly colored frog found across the Guiana Shield. At the Nouragues Natural Reserve, French Guiana, the frogs are blue-black with a bright yellow ring, which may be broken or joined to form a figure eight (Fig. 1A) (34). However, the presence of asymmetry and variation between individuals does not conform to standard aposematic theory (34, 35), and is reminiscent of cryptic patterning (29). To examine the optical processes involved in the coloration of D. tinctorius, we ran computational models of predator vision, an in situ survival experiment with the frogs’ wild avian predators, and a computer-based detection experiment with human surrogate predators, where we manipulated coloration, patterning, and viewing distance.

Fig. 1.

Avian (VS) visual modeling of D. tinctorius and the leaf litter background (n = 84), the trends are consistent with the other four visual models (SI Appendix). At close-range D. tinctorius and its background are easily discriminated, but at greater viewing distances classification accuracy declines. (A) D. tinctorius photographed in situ. (B) Leaf litter colors in avian visual space. (C) D. tinctorius coloration in avian visual space; the frog contains high-luminance yellows and low-luminance blue-blacks, which are not found in the background. (D) The mean colors of each frog and each background sample are intermixed. (E) Color discrimination (ROC curves) at different spatial resolutions; as resolution decreases (increasing distance), accuracy decreases. (F) Texture discrimination (ROC curves) at different spatial scales; as resolution decreases, accuracy decreases.

Results

Visual Modeling.

We photographed adult D. tinctorius (n = 84) in situ near the Saut Pararé camp in the Nouragues Natural Reserve, French Guiana (Fig. 1A). Detectability at different viewing distances was assessed using models of tetrachromatic bird [both violet- (VS) and ultraviolet (UV) sensitive], trichromatic snake, dichromatic mammal, and trichromatic human visual perception. The avian, snake, and mammalian models are representative of potential visual predators of dendrobatid frogs (11–13, 36), and we included human vision to allow intuitive interpretation of the results.

Support vector machines (SVM) (37) were used as a classifier, and discrimination accuracy (frog vs. leaf litter background) was assessed using the area under the curve (AUC) of receiver-operated characteristic (ROC) curves (38). An AUC of 1.0 represents perfect classification, whereas an AUC of 0.5 indicates random chance. We followed a commonly used grading system interpreting AUC results as: 1.0–0.9 = excellent, 0.9–0.8 = good, 0.8–0.7 = fair, 0.7–0.6 = poor, 0.6–0.5 = fail (39). Data are available in the Dryad Digital Repository.

With all spatiochromatic information, representing close-range viewing, all visual systems were excellent at discriminating frog from background (Table 1 and SI Appendix, Fig. S1). Classification accuracy did, however, change at different spatial resolutions, and the frogs’ color and visual texture converged with that of the background at greater viewing distances (Fig. 1 B–D). For all visual systems, color discrimination accuracy decreased from good at the highest resolution (very high), to fair at low resolution (low), and poor for the mean color, with the dichromatic mammalian model having the largest decline in accuracy as resolution decreased (Table 1 and SI Appendix, Fig. S2). Similarly, the accuracy of visual texture discrimination declined at lower resolutions, where all visual systems were good at the highest resolution but failed to classify effectively at the low resolution (Table 1 and SI Appendix, Fig. S3).

Table 1.

Visual model discrimination accuracy at each spatial resolution

Resolution	Avian VS	Avian UV	Snake	Mammal	Human
All spatiochromatic information
Full resolution	0.98	0.97	0.96	0.96	0.98
Color discrimination
Very high	0.89	0.89	0.88	0.85	0.89
High	0.88	0.88	0.87	0.80	0.88
Medium	0.83	0.83	0.82	0.70	0.83
Low	0.78	0.78	0.75	0.63	0.78
Mean	0.63	0.65	0.62	0.60	0.68
Texture discrimination
Very high	0.89	0.89	0.89	0.89	0.89
High	0.88	0.87	0.87	0.87	0.88
Medium	0.78	0.75	0.72	0.72	0.77
Low	0.60	0.56	0.54	0.54	0.58

AUC of ROC curves for each visual model. All models are excellent at discriminating frogs from backgrounds at high resolutions, but accuracy declines toward random chance (0.5) for all models at low resolution. AUC of 1.0 equals perfect classification and 0.5 indicates random chance: 1.0–0.9 = excellent, 0.9–0.8 = good, 0.8–0.7 = fair, 0.7–0.6 = fair, 0.6–0.5 = fail.

Survival.

Camouflage is largely background-dependent, with even small deviations away from background color and texture leading to significant decreases in survival (40), whereas conspicuous aposematism is more resilient to variation in background coloration (41).

To assess how dependent the survival of D. tinctorius was on the visual characteristics of the natural background, we used plasticine model frogs to record avian predation, and manipulated both frog color pattern and background. Three different frog colors were designed to represent: (i) the natural phenotype (N: yellow-and-black), (ii) aposematism (Y: plain yellow), and (iii) camouflage (C: brown-and-black). These frogs were presented on four backgrounds: the natural leaf litter (NL), two manipulated backgrounds that differed in color and visual texture from the natural background [plain soil (NS) and a homogeneously colored paper square (PA)], and a paper square printed with leaf litter (PL), which acted as a control for the use of paper backgrounds (SI Appendix, Fig. S4).

We found a significant interaction between frog color pattern and background (χ² = 50.67, df = 11, P < 0.001), so each frog color was analyzed separately. There was a significant effect of background on the survival of the brown-and-black frogs (χ² = 29.35, df = 3, P < 0.001). There was no significant difference in survival between the natural background (NL) and the printed leaf litter (PL) background (CNL-CPL: z = 1.45, P = 0.150), but the brown-and-black frogs had significantly higher survival on leaf litter than on both the modified backgrounds (CNL-CNS: z = 3.10, P = 0.002; CNL-CPA: z = 4.40, P < 0.001) (Fig. 2, Left). Conversely, there was no significant effect of background on the survival of plain yellow frogs (χ² = 0.51, df = 3, P = 0.918; all pairwise tests: z < 0.62, P > 0.540) (Fig. 2, Center). These results are consistent with the brown-and-black frog being camouflaged on the leaf litter and the plain yellow being equally detectable on all backgrounds. This conclusion is further supported by visual modeling of the stimuli photographed in situ (SI Appendix, Fig. S5).

Fig. 2.

Relative survival of plasticine frogs in comparison with the natural leaf litter background (NL). Odds ratios with 95% confidence intervals from the model (n = 84 per frog-background combination). (Left, cryptic brown-and-black frogs, C) There was no significant difference between the natural background (NL) and the printed leaf litter (PL), but there was a significantly lower survival for the modified (NS and PA) backgrounds. (Center, plain yellow frogs, Y) There was no significant effect of changing background. (Right, yellow-and-black frogs, N) There was no significant difference between NL and PL, but NL had a significantly higher survival than both NS and PA. The survival of the natural yellow-and-black pattern was dependent on the visual characteristics of the background, suggesting a camouflage component.

We found a significant effect of background on the yellow-and-black frogs, which mimicked the natural phenotype (χ² = 12.10, df = 3, P = 0.007). There was no significant difference between leaf litter backgrounds (NNL-NPL: z = 0.72, P = 0.470), but survival was significantly higher on the natural background than on modified backgrounds (NNL-NNS: z = 2.44, P = 0.015; NNL-NPA: z = 2.90, P = 0.004) (Fig. 2, Right). The survival of the yellow-and-black phenotype, therefore, appears to be dependent on the visual characteristics of the background in a similar manner to the brown-and-black frog, but counter to the plain yellow frog.

Detection.

To investigate further how viewing distance affects the detectability of D. tinctorius, we performed a screen-based detection experiment with human participants (n = 18). We manipulated frog coloration (SI Appendix, Fig. S6), and presented the stimuli on their natural background under conditions representing three viewing distances: near, medium, and far. The human participants were required to click on the frogs, and we recorded both reaction time (Fig. 3 and Table 2) and detection accuracy.

Fig. 3.

Time taken by human observers to detect frogs at the near and far distances. Means with 95% confidence intervals from the model (n = 540 per frog-distance combination): gray lines indicate 95% confidence intervals for treatment A (posterized natural pattern). At close-range (near, black) the natural phenotype (A) groups readily with conspicuous (high yellow) patterns (C, D, and E), which are detected more quickly than cryptic patterns (B, F, G, H, I, J, and K). At the greatest distance (far, orange) the natural phenotype is detected significantly more slowly than conspicuous patterns, and groups more readily with cryptic patterns.

Table 2.

Pairwise comparisons (of a priori interest) for the time taken to detect frogs by human observers

Comparison	Near	Medium	Far
Treatment	χ2 = 3,349.00, df = 11, P < 0.001	χ2 = 3,519.60, df = 11, P < 0.001	χ2 = 2,563.10, df = 11, P < 0.001
A–B	z = −11.43, P < 0.001	z = −15.59, P < 0.001	z = 3.00, P = 0.027
A–C	z = 3.74, P = 0.002	z = 3.49, P = 0.005	z = 29.03, P < 0.001
A–D	z = 3.90, P < 0.001	z = 4.50, P < 0.001	z = 25.56, P < 0.001
A–E	z = 0.46, P = 1.000	z = 2.48, P = 0.114	z = 17.71, P < 0.001
A–F	z = −7.40, P < 0.001	z = −7.40, P < 0.001	z = −0.91, P = 0.976
A–I	z = −12.13, P < 0.001	z = −22.47, P < 0.001	z = −4.38, P < 0.001
I–J	z = −6.87, P < 0.001	z = −2.68, P = 0.067	z = −0.66, P = 0.998
I–K	z = −34.17, P < 0.001	z = −18.99, P < 0.001	z = −0.73, P = 0.995
B–G	z = 3.68, P = 0.002	z = 1.34, P = 0.800	z = 0.71, P = 0.996
J–H	z = 5.74, P < 0.001	z = 5.36, P < 0.001	z = 2.27, P = 0.187
A–L	z = 1.80, P = 0.464	z = 3.16, P = 0.016	z = 9.48, P < 0.001

At close-range (near) the natural pattern (A) groups more readily with conspicuous patterns, whereas at greater distances (far) A is detected in a similar manner to cryptic patterns. A: natural pattern; B: plain blue-black; C: plain yellow; D: reversed color natural pattern; E: yellow pixels grouped into a circular patch; F: yellow highlighting the frog’s edge; G: natural pattern with yellow replaced with the mean color of the background; H: inverse of pattern G; I: mean color of the frog; J: mean color of the background; K: background matching camouflage; and L: unmanipulated frog.

Reaction time.

We found a significant interaction between frog coloration and distance (χ² = 1,670.40, df = 22, P < 0.001), and so treatment effects were analyzed separately for each distance. At close-range (near) we found that the natural yellow and blue-black pattern (A) was detected significantly more slowly than plain yellow (C < A) and reverse color pattern (D < A) frogs, both of which increased the proportion of yellow, but the natural pattern was detected significantly faster than plain blue-black (A < B). At long distances (far), however, it took participants significantly longer to detect the natural pattern compared with the plain yellow (C < A), reversed pattern (D < A), and the plain blue-black (B < A).

The magnitude of these effects shows that at close range the natural pattern grouped more readily with the conspicuous high yellow patterns (C and D) than the more cryptic plain blue-black (B). However, when viewed from a greater distance, the time taken to detect the natural pattern greatly increased, and the natural pattern grouped more readily with cryptic patterns. At the furthest distance the natural pattern provides more effective camouflage than the plain blue-black (B < A) (Fig. 3 and Table 2).

The arrangement of the natural pattern also appears well suited for both short-range detectability and long-range camouflage. At close-range, rearranging the pattern, while maintaining the ratio of color components, could increase but not decrease reaction time (A = E and A < F), whereas at long-range we found the opposite (E < A and A = F). We therefore conclude that the pattern is arranged to be highly salient at close range but cryptic when viewed from a distance.

Moreover, we found no evidence of disruptive camouflage, which would predict that high-contrast patterning would increase reaction time compared with plain colors (26, 30). In contrast, when comparing frogs with brown rather than yellow patterning to plain brown or black frogs, we found that the presence of patterning decreased reaction time at close-range (G < B and H < J) and had no effect at greater distances (G = B and H = J). These data suggest that at close range the pattern itself acts as a salient signal even in the absence of conspicuous coloring.

Furthermore, our data suggest that, when viewed from a distance, the yellow and blue-black components are spatially averaged to produce a mean color, which provides effective camouflage. As distance increases, the time taken to detect the natural pattern (A) converges with that of its mean color (I), and at greater viewing distances the mean color is just as effective as the average color of the background (J) and random-sample background matching (K) at preventing detection.

Detection accuracy.

A similar trend was observed in the detection-accuracy data. There was a significant interaction between treatment and distance (χ² = 297.64, df = 22, P < 0.001), and so each distance was analyzed separately. We found a significant effect of treatment at the near (χ² = 249.26, df = 11, P < 0.001) and medium (χ² = 675.05, df = 11, P < 0.001) distances, but detection accuracy was too high for reliable pairwise tests.

At the furthest distance (far) there was a significant effect of treatment (χ² = 2,769.10, df = 11, P < 0.001) and pairwise tests were possible. Increasing the amount of yellow in the pattern (C and D) increased detection accuracy over the natural pattern (A < C: z = −8.49, P < 0.001; A < D: z = −15.64, P < 0.001), as did rearranging the pattern into a signal yellow circle (A < E: z = −15.40, P < 0.001). Whereas there was only a marginal effect of moving the yellow pixels to highlight the edge of the frog (E < A: z = −2.66, P = 0.076). Furthermore, the natural pattern was detected more accurately than the mean color of the frog (I < A: z = −8.57, P < 0.001) and the unmanipulated frog (L < A: z = −6.93, P < 0.001). No further pairwise tests were significant (z < 2.15, P > 0.265).

Discussion

Aposematic signals are made up of both color and visual texture information. However, whereas color saturation has been studied in detail, patterning has received comparatively little attention. Most research into pattern has focused on close-range signaling properties and suggests that high-contrast patterns can increase the saliency, memorability, and consistency of a signal (3–5). Alternatively, however, it has been suggested that high-contrast patterning may reduce detectability, through either disruptive camouflage (24) or distance-dependent signaling (14–20).

Taken together, our data suggest that D. tinctorius displays a specific ratio and distribution of color components, which combines highly salient aposematic signaling with effective and targeted background-matching camouflage. At close-range the pattern is easily detectable, utilizing bright colors not found in the background to increase color contrast. At greater viewing distances, these highly contrasting colors merge together to form an average that closely matches that of the background, such that the time taken to detect the average color of the frog cannot be distinguished from that of background-matching camouflage.

Moreover, at close-range the pattern of D. tinctorius is distinct from the textural component of the background and appears to act as a salient signal even in the absence of conspicuous coloring. At long-range, in a similar manner to color-blending, textural information converges to match that of background. We therefore found no evidence to support disruptive camouflage but, rather, these data are consistent with distance-dependent pattern blending (19, 32, 33). This result is consistent for avian, snake, mammal, and human visual perception, and translates into a decrease in avian predation on the frog’s natural background.

The bright colors of D. tinctorius have previously been associated with aposematism (42, 43) and sexual signaling (34). Brighter and more conspicuous signals have been demonstrated to improve the efficacy of aposematism (44, 45) and to be favored during mate selection in dendrobatid frogs (46, 47). Under natural conditions, however, variation in predator motivation toward aposematic prey means that incorporating aspects of camouflage may increase survival (6–10, 20). Indeed, even in the absence of aposematic defense, distance-dependent patterning may facilitate the greater color saturation favored for mate attraction without necessarily increasing predation risk, especially where conspecifics and predators operate on different spatial scales.

It has further been suggested that the colors of D. tinctorius may disrupt a predator’s ability to track a moving frog (motion dazzle) (48, 49) and that phenotypic variation may trade off the benefits of salient signaling versus camouflage (35). We propose pattern-blending as an additional (albeit not mutually exclusive) optical mechanism that combines the benefits of both salient signaling and camouflage: reducing predator encounter rates while retaining effective aposematism (18).

Evidence for distance-dependent signaling from natural phenotypes is currently scarce, with most research focusing on predominantly cryptic patterns with small aposematic components (15, 16). In contrast, the yellow and blue-black of D. tinctorius does not appear to share features usually indicative of camouflage, and our data highlight how patterning and background characteristics may influence saliency and detectability.

The particular balance between aposematism and camouflage may be affected by differences in toxin susceptibility between different predators, or temporal shifts in the predators’ motivation to attack. Predator motivation may vary considerably due to seasonal differences in community composition, competition, energetic requirements, and the availability of alternate prey (6, 50), whereas predation risk may also fluctuate as dietary-derived toxicity and the frog’s level of activity may shift with changing environmental conditions (51–53).

Indeed, different components within the coloration of D. tinctorius may fulfill different roles and be individually selected (54). More research is needed to understand how multiple functions interact under different viewing conditions [e.g., lighting conditions (35)/viewing distance and angle], in different contexts [e.g., microhabitats (35)/posture and motion (48, 49)], and to different observers (predators and conspecifics), as well as how color is affected by temporal changes in behavior, toxicity, and the visual environment. Furthermore, intraspecific variation both within (seemingly continuous) (35, 48) and between (largely discrete) (42, 43, 54, 55) populations suggests this balance in selection pressures may vary both geographically and between individuals. However, it is currently unknown whether these differences are the result of natural or sexual selection, or neutral drift within a broad definition of the aposematic signal defined by potential predators. Aposematic patterning, therefore, appears to be a highly complex adaptation, combining different processes at different spatial scales and in different contexts. Future research is needed to understand how these multiple selection pressures interact across wider temporal and spatial dimensions.

Methods

Study Site and Image Collection.

Experiments took place in the rainforest surrounding the Saut Pararé camp of La Station de Recherche en Écologie des Nouragues, French Guiana, between December 2014 and January 2015. We photographed adult D. tinctorius on their natural leaf litter background at a height of 70 cm (n = 84), as well as the leaf litter without frogs at 100 cm (n = 265) and 150 cm (n = 265). In situ photographs were taken with a Nikon D3200 DSLR and AF-S DX NIKKOR 35-mm lens (Nikon Corporation) and contained a ColorChecker Passport (X-Rite Inc. 2009). Ex situ UV photographs were taken of captive D. tinctorius and the plasticine models, using a Nikon D70 DSLR (Nikon) and UV-VIS 105-mm CoastalOpt SLR lens (Jenoptik AG) with human-visible and IR-blocking filters, and each image contained a 15% reflectance Spectralon gray standard (Labsphere).

Visual Modeling.

Photographs of the frogs on their natural background (n = 84) were calibrated and scaled using the ColorChecker Passport (56), and 1,000 pixels from the frog and 1,000 pixels from the background were randomly selected for analysis. We used five different visual models: tetrachromatic avian, both VS [Pavo cristatus (57)] and UV [Sturnus vulgaris (58)] sensitive, trichromatic snake [Masticopis flagellum (59)], dichromatic mammal [Mustela putorius (60)], and trichromatic human (61). As opponent processing is central to human color perception (62), and evidence suggests a similar mechanism for avian color perception (63–65), each visual model used relative cone-capture rates to generate a 3D color space made up of luminance (L) and the opponent channels red-green (rg), and yellow-blue (yb) (19, 20, 66). All visual modeling was performed in MATLAB 2015a (The MathWorks).

To represent increasing viewing distance, we used a log-Gabor filter bank with four spatial scales (wavelength relative to the smallest frog: very high = 1/8, high = 1/4, medium = 1/2, and low = 1) and six orientations (0°–150° in 30° increments) (41, 67, 68). To assess discriminability between frog and background at each spatial scale we used SVMs [R package e1071 (37)] in R 3.1.3 (The R Foundation for Statistical Computing). SVMs act as a nonlinear classifier, projecting the data into a multidimensional space in which a hyperplane can be fitted between groups (39). Data were cross-validated by training the model on one half of the data and testing on the other (39). Classification accuracy was assessed using the AUC of ROC curves [R package pROC (38)] in R 3.1.3 (SI Appendix) (19, 20, 41, 66).

Survival.

We used plasticine model frogs to record avian predation in situ. Plasticine frogs were designed to represent the natural pattern (N, yellow-and-black), aposematism (Y, plain yellow), and camouflage (C, brown-and-black). Model frogs were presented on four backgrounds: the natural leaf litter (NL), a paper square printed with leaf litter (PL), and two manipulated backgrounds that differed from the natural substrate, natural soil (NS) and a homogeneously colored paper square (PA). This created 12 treatments (frog–background pairs).

A randomized block design was used. Twelve blocks containing seven of each treatment (n = 1,008: 84 per treatment) were placed along nonlinear transects through the rainforest. Stimuli were inspected for signs of avian predation at 24, 48, 72, and 96 h. Predation risk was analyzed with a mixed-effects Cox model [R package coxme (69)] in R 3.1.3. Avian predation was included as a full event, whereas nonavian predation, missing or washed out stimuli, and those surviving to 96 h were included as censored values, and block was included as a random factor (SI Appendix).

Detection.

Photographs of 30 different D. tinctorius were randomly selected and the colors of each frog were manipulated to create 12 different treatments (n = 360). To allow for pattern manipulation the colors of each frog were standardized using k-means clustering. The RGB color space was grouped into four clusters, with the centroid with the highest ratio of R+G to B designated yellow, and that with the lowest luminance designated black. This created 12 treatments: A, natural pattern using standardized colors; B, plain blue-black; C, plain yellow; D, reversed color natural pattern; E, all yellow pixels of A grouped into one approximately circular patch; F, all yellow pixels from A moved to the frog’s edge; G, natural pattern with yellow replaced with the mean color of the background; H, inverse of pattern G; I, mean color of the frog; J, mean color of the background; K, background matching camouflage (random sample of leaf litter background); and L, unmanipulated natural frog pattern (SI Appendix, Fig. S6).

Three different viewing distances were created using calibrated photographs of leaf litter taken at different distances (100 cm = 265, 150 cm = 265). Each frog was appropriately scaled and randomly placed onto the leaf litter [as the number of frogs (360) outnumbered the number of backgrounds (265), 95 background images were randomly selected and rotated by 90°]. The 100-cm images created the “near” condition and the 150-cm images created the “medium” distance. The “far” condition was generated by reassigning the frogs to the 150-cm images and applying a 16° Gaussian filter to remove high spatial frequency information.

Human participants (n = 18) searched for frogs on a computer screen. The stimuli were presented to each participant in three sessions (each distance as a separate session), each of which contained all 360 stimuli, split into 10 blocks of 36 images, in an individually randomized sequence (n = 540 per frog-distance combination). We recorded reaction time and detection probability in Psychtoolbox (70) in MATLAB 2015a. Research was approved by the University of Bristol Faculty of Science Research Ethics Committee. All participants gave their informed consent in line with the Declaration of Helsinki.

Detection probability was analyzed with a binomial generalized mixed-effects model and log-transformed reaction time analyzed with a general linear mixed-effects model [R package lme4 (71)] in R 3.1.3. Both models included treatment and distance as fixed effects, and participant number as a random factor. Pairwise comparisons of a priori interest were calculated [R package multcomp (72)] to test three hypotheses: whether the natural pattern is organized to act as (i) maximal conspicuousness, (ii) disruptive camouflage, and (iii) distance-dependent pattern blending (SI Appendix).