Research articlesBackground complexity can mitigate poor camouflage
Abstract
Avoiding detection through camouflage is often key to survival. However, an animal's appearance is not the only factor affecting conspicuousness: background complexity also alters detectability. This has been experimentally demonstrated for both artificially patterned backgrounds in the laboratory and natural backgrounds in the wild, but only for targets that already match the background well. Do habitats of high visual complexity provide concealment to even relatively poorly camouflaged animals? Using artificial prey which differed in their degrees of background matching to tree bark, we were able to determine their survival, under bird predation, with respect to the natural complexity of the background. The latter was quantified using low-level vision metrics of feature congestion (or ‘visual clutter’) adapted for bird vision. Higher background orientation clutter (edges with varying orientation) reduced the detectability of all but the poorest background-matching camouflaged treatments; higher background luminance clutter (varying achromatic lightness) reduced average mortality for all treatments. Our results suggest that poorer camouflage can be mitigated by more complex backgrounds, with implications for both camouflage evolution and habitat preferences.
1. Introduction
Remaining undetected is frequently important for a number of reasons, including predator–prey interactions, avoiding social harassment and seeking sneak mating opportunities [1]. Camouflage is the most widespread means of achieving concealment ‘in plain sight’, arguably the most critical factor being the similarity of the object's colour and pattern to that of its immediate background [2–5]. However, a factor extrinsic to the camouflaged object also affects its concealment: background complexity [6–10]. Merilaita [6] argued, based on results from neural network models that the visual complexity of the background is a key determinant of detectability, and that higher background complexity relaxes the requirement for precise background matching. He proposed that this is because more complex backgrounds impose higher information-processing costs, and that predators are limited in their processing capacity; we return to this issue in the Discussion. Analogously, the effect of many highly salient visual features in the background, known as ‘visual clutter’, has been investigated in humans in applied contexts such as visual display design [11,12], and also in a few other species. First, by monitoring predation on artificially patterned backgrounds by birds or fish [7–9] and second, by measuring wild avian predation, and human visual search, for artificial targets against natural backgrounds [10]. Although these studies demonstrate a detrimental effect of background complexity on detection, they do not tell us how important it is relative to matching the background. Somewhat surprisingly, Xiao & Cuthill's [10] experiment suggested that, for birds, background complexity was far more important than matching the immediate background. A key limitation of Xiao & Cuthill [10] is that the effect of background complexity was demonstrated for only a single target colour: that of the average background. This leaves open the question of whether the benefits of background complexity for concealment are independent of background matching, as Merilaita [6] suggested, or whether some level of background matching is required. Murali et al. [13] have addressed this question using humans searching on artificial backgrounds, concluding that background heterogeneity aids concealment, but not when the targets fail to match the background. However, whether such effects apply to non-human predators in the field, and the sort of complexity variation seen in natural backgrounds, need to be addressed. Our present study fills that gap by systematically varying the degree of background matching and establishing the limits of background complexity's ability to impede detection by wild predators searching on natural backgrounds in the field.
Here, we determine the extent to which background complexity can mitigate poor camouflage. Understanding the interaction between conspicuousness and background complexity is important for two main reasons: the first is to understand what pattern of camouflage evolution will be favoured in different habitats [6]; the second is to understand which habitats animals prefer if complexity does indeed decrease detection [9]. Most habitats are heterogeneous in colour and pattern, and many animals move between visually different habitats. So an open question is whether it is better to have coloration that is a compromise between different backgrounds, or specialized to one [6,14–16]. Modelling suggests that a critical factor is a trade-off between improved survival on one background and reduced survival on another [15,17]. Background complexity will affect that trade-off if it mitigates any mismatch of specialist camouflage to alternative backgrounds, and of compromise strategies to all backgrounds. Furthermore, animals benefiting from concealment could potentially select backgrounds with higher complexity [9], and those benefiting from salience (for signalling) could select habitats with lower complexity [18].
By monitoring the survival of artificial prey ‘moths’ in natural woodland, we examined the effect of natural levels of background complexity (as in [10]) on the survival of different degrees of background matching (as in [19]). By recording the frequency of colours across a large sample of European oak tree (Quercus robur) bark within the woodland, we produced treatments which spanned the background luminance frequency distribution. This allowed us to test whether higher background complexity interferes with the detection of all targets regardless of how well they match the background, or whether complexity cannot mitigate poor camouflage. We predicted that high background complexity would only reduce detectability for targets that already match the background well. By manipulating one simple feature—the average luminance or achromatic lightness—that is known to influence the salience of camouflaged objects in our experimental paradigm [20], we sought to determine just how mismatched the target needs to be to the background for complexity to cease to affect detectability. Whether the effect is sudden or continuous is an empirical question that our experiment should help address. To measure background complexity, we used feature congestion [11,12], which is based on features from the early stages of visual processing, namely variation in luminance, colour and edge orientation. It has been shown to predict interference in both human and bird search [10,11]. For avian colour vision, we used Xiao & Cuthill's [10] adaptation of the model of Rosenholtz et al. [11,12].
2. Material and methods
(a) Stimuli
The targets were designed to resemble a non-specific lepidopteran: right-angle triangles at 45 mm wide × 32 mm high. Nine treatments were produced, all having the average hue of oak bark but varying in achromatic lightness. The colour information was derived from 1000 calibrated photographs of oak tree bark, taken in the same woods as the experiment was carried out in (Leigh Woods National Nature Reserve, Somerset, UK, 2°38.6′ W, 51°27.8′ N) the previous year. Photographs were taken at head height, approximately 1 m away from the oak trees, of areas of bark that were free from lichen and not in direct sunlight. The camera was a Nikon D3200 DSLR camera with 35 mm Nikon AF-S DX NIKKOR f/1.8G lens (Nikon Corp., Tokyo, Japan), set at ISO 1600, f8 and automatic integration time. A colour standard, Colorchecker Passport (XRite, Grand Rapids, MI, USA), was pinned to the trees in the bottom left-hand corner of the frame for later calibration (as in [21,22]). These photographs were linearized and normalized to control for variation in light intensity and colour balance, and then mapped to the cone photon-capture colour space of a typical passerine predator, using cone spectral sensitivity data for the blue tit (Cyanistes caeruleus) [23]. The procedures were carried out using custom MATLAB scripts (The MathWorks, Natick, MA, USA), using the same procedures as described in [22,24] (see electronic supplementary material). One hundred random target-sized samples were taken from each photograph (the xy coordinates being pairs of random numbers drawn from a uniform distribution), and the average colour for each was calculated. The measure of lightness was the photon catch of the avian double cones [25], scaled from 0 (black) to 1 (white), and two opponent channels to represent the relevant variation in hue: red-green (the contrast between mediumwave- and longwave-sensitive cones) and blue-yellow (the contrast between shortwave- and the average of mediumwave- and longwave-sensitive cones), both also scaled to lie between 0 and 1 (for further details see [10]). Neither oak bark nor the printed targets reflected ultraviolet, so this component of avian colour could be ignored (for bark reflectance spectra see [26]). The ‘avian luminance’ of the 100 000 samples ranged between 0.07 and 0.85 and had two modes (figure 1). The treatments of 0.05, 0.15, 0.25, 0.35, 0.45, 0.55, 0.65, 0.75 and 0.85 luminance units were chosen to span the range from exceedingly rare and dark to exceedingly rare and light, with treatments also approximating the two peaks and the trough between them (figure 1). Validation of the intended manipulation of target-background contrast is provided in the electronic supplementary material. All treatments had the same red-green and blue-yellow contrasts (0.013 and −0.155, respectively), matching the average of the 100 000 bark samples, so they varied in tone/lightness but not hue. These targets were printed on waterproof paper (Rite-in-the-Rain, J. L. Darling LLC, Tacoma, WA, USA) using a calibrated printer (Canon Imagerunner Advance C5535i; Canon Inc., Tokyo, Japan).
Figure 1. A histogram showing the frequency of the avian luminance of 100 000 oak tree (Quercus robus) bark samples. The red dotted lines show where on the distribution the treatment luminance values fall, with the darkest designated as treatment 1 and the lightest as 9. (Online version in colour.)
(b) Procedure
The experiments were run from October to December 2020. The general experimental protocol followed that of Cuthill et al. [27], with the artificial ‘moths’ pinned to mature oak trees along nonlinear transects with a dead mealworm (Tenebrio molitor) larva frozen at −80°C then thawed underneath the ‘wings’, with a small portion showing. Each transect comprised a block within an overall randomized block design. The transects varied in length from roughly 500 to 1000 m, according to variation in oak tree density within different areas of the woodland. The meandering nature of the transects would make them hard to define in terms of area, but they did not overlap each other. Younger oak trees (trunk circumference at head height less than 0.9 m) were avoided, with no more than one target per tree, pinned at roughly head height, facing away from paths to minimize interference from the public. Once pinned, a photograph was taken of the target and its respective background; four mobile phones were used, two of which were iPhones (iPhone 8 and 11, Apple Inc., Cupertine, CA, USA) and two of which were Samsungs (Samsung SM-A405FN and SM-G970F, Samsung Group, Seoul, South Korea). The known size and reflectance of the target, coupled with calibrations based on photographs of a colour chart (Colorchecker Passport; X-Rite, Grand Rapids, MI, USA), were used to normalize and linearize the photographs, then map them to avian colour space. These photos were then used to extract the same measures of background complexity as in Xiao & Cuthill [10], using Rosenholtz's principles of feature congestion [11,12]. The calculations were carried out using the custom MATLAB scripts described and explained in [24], based on the original MATLAB functions of Rosenholtz and colleagues (https://dspace.mit.edu/handle/1721.1/37593). Rosenholtz et al.'s ‘feature congestion’ can be thought of as a perceptual measure of the variation in three components of a visual scene: luminance, colour and edge orientation. A scene with high levels of local contrast in brightness will score highly on the luminance clutter measure; analogously, spatial variation in colour contributes to the colour clutter metric, and variation in the orientation of edges (lines) contributes to the orientation clutter metric. ‘Local contrast’ is in fact calculated at three spatial resolutions (i.e. capturing variation in each of coarse, medium and fine detail) and summed to provide a single measure of each of what Rosenholtz et al. [11,12] call contrast (luminance), colour and edge orientation ‘clutter’. The feature congestion metric is a weighted sum of the three, based on empirically derived estimates of the contribution of each to perceived differences. Xiao & Cuthill [10] showed that the orientation clutter measure of perceived image complexity, and an equivalent for avian vision, predicted the detectability of triangular (notionally moth-shaped) targets on natural bark backgrounds, for humans and birds respectively. The electronic supplementary material of Xiao & Cuthill [10,24] has a figure that, in a simple intuitive way, demonstrates how the Rosenholtz et al. clutter metrics relate to image features.
Targets were checked at 24, 48, 72 and 96 h, with disappearance of all or most of the mealworm being marked as avian predation, and predation by invertebrates (spiders, slugs and wasps) and ‘survival’ up to 96 h being marked as ‘censored’. Invertebrate predation was determined by either direct observation (one instance of a wasp), a hollowed-out exoskeleton (spiders) or the presence of mucus near the target (slugs). The large sample size that our method allows precludes direct observation of most predation events, so we cannot be certain that birds were responsible for all events scored as bird predation. However, one would expect birds to be the predominant visual predator for such prey in winter in UK woodland, and non-visual predators would only add noise to our data. In each block, 90 targets were placed (10 replicates of each of the nine treatments). Overall, 27 blocks were completed, totalling 2430 targets.
(c) Analysis
Mixed-effects Cox regression was applied using the ‘coxme’ function from the ‘coxme’ R package [28,29]. Block was fitted as a random effect, treatment, and the three metrics of feature congestion of the background were treated as fixed effects. The significance of effects was tested using an analysis of deviance comparing the unexplained variation of models with and without the factor in question, tested against a χ2 distribution. Starting with a maximal model including interactions between treatment and each of the feature congestion metrics, models were step-wise simplified based on the non-significance of terms. Effect sizes are presented as odds ratios with 95% confidence intervals. Treatment 5 was chosen as the baseline for comparison with other treatment levels, as this lay close to the mean of the whole distribution (0.48; see also electronic supplementary material, figure S1) and was also close to the luminance of the single treatment used in Xiao & Cuthill [10].
3. Results
Overall, 27% of targets were censored (8.6% eaten by spiders, 6.6% by slugs, 3.3% lost and 8.7% remained uneaten after 96 h). The main effect of treatment, ignoring background complexity, affects target mortality in a pattern that loosely mirrors the frequency of each luminance in the background (χ2 = 376.41, d.f. = 8, p < 0.001; compare figure 2 with figure 1). Targets which have a more common background shade (treatment 2 to 7) have a lower relative mortality than rarer shades.
Figure 2. More common background shades have enhanced survival. Odds ratio plot for the relative survival of each treatment compared to treatment 5, which lies near to the mean of the whole distribution. Treatments with 95% confidence intervals not overlapping the red dashed line have a lower relative survival (less than 1) than treatment 5. Odds ratios and 95% confidence intervals (bars) were estimated using a mixed-effects Cox regression. (Online version in colour.)
We then examined how the metrics of background complexity altered the survival of the targets; all steps in the statistical modelling can be found in the electronic supplementary material. There was no significant interaction between treatment and colour clutter (χ2 = 5.09, d.f. = 8, p = 0.748), treatment and luminance clutter (χ2 = 10.05, d.f. = 8, p = 0.262), or a main effect of colour clutter (χ2 = 0.00, d.f. = 8, p = 0.979). However, the interaction between treatment and orientation clutter and the main effect of luminance clutter remained in the minimal adequate model (χ2 = 57.04, d.f. = 8, p < 0.001 and χ2 = 22.89, d.f. = 1, p < 0.001, respectively). Survival was higher with greater luminance clutter (odds ratio 0.866, 95% c.i. 0.818 to 0.918). The effect of orientation clutter was also found to boost survival, but only for those treatments with commoner background shades (treatments 2 to 7), with no significant effect for the treatments representing very rare shades, both dark (1) and light (8, 9); see figure 3a,b.
Figure 3. Odds ratios of the effect of (a) luminance clutter and (b) orientation clutter on the relative survival of the treatments. Background complexity only enhances survival for targets that match the background to some degree (treatment 5 is close to the average background luminance, with treatment 1 much darker, and treatments 8 and 9 much lighter, than any background colours). The red dotted line signifies no effect (equal to 1); data above the line has a higher relative survival on more complex bark (greater than 1). Odds ratios and 95% confidence intervals were estimated using a mixed-effects Cox regression. (Online version in colour.)
4. Discussion
Our results support Merilaita's [6] conclusion, based on neural network modelling, that background complexity has an important influence on detectability, and that higher background complexity enhances the benefits of background-matching camouflage. When examining the three visual characteristics of feature congestion (luminance, colour and orientation of edges), we found that two of them had a significant effect on predation rates. A higher background orientation clutter reduced the detectability of all but the rarest background-matching camouflaged treatments (treatment 1, 8 and 9) (figure 3b). With regard to a higher background luminance clutter, there was also a pattern of lower mortality (figure 3a). This effect was similar to, but weaker than, that seen with orientation clutter, but with no detectable treatment-by-background interaction. Therefore, unlike orientation clutter, we cannot confidently conclude that the concealment benefits of high background luminance contrast disappear for rarer background matching samples. The effect of orientation clutter has been previously found in experiments involving humans and wild birds [10]. Although luminance clutter was not significant in that study, we note that our sample size was an order of magnitude greater, so capable of detecting smaller effects. We make no claims that orientation clutter will be the most important factor in all situations; oak bark is characterized by deep linear ridges, and our targets have linear edges, so an effect on the signal-to-noise ratio in the domain of edge detection is expected. Oak bark also has a low chromatic variation (mainly different shades of brown), so it will be interesting to carry out analogous experiments with backgrounds, and targets, with different chromatic and structural characteristics.
A corollary of background complexity aiding concealment is that background complexity mitigates less-than-perfect camouflage [6]. Targets which are matched to at least some of the background are less detectable on visually complex backgrounds than those on backgrounds of lower complexity. These findings have implications for habitat selection and thus animal distributions. If an animal benefits from concealment, all things being equal, it should choose a complex background [9]. Conversely, those benefiting from conspicuousness (e.g. to convey a visual signal) should choose to be seen against a less complex background to maximize their saliency [30]. Habitat choice with respect to habitat complexity could be an effective means of changing the balance between salience and crypsis [31], with different costs and benefits from changing appearance per se. We also found that the rarest background shades were little affected by the complexity of the background. Background complexity does not mitigate a very poor match to the background.
Moving beyond the effect of background complexity, on average the treatments with more common background shades survived better than rarer shades (figure 2). This is expected as, all things being equal, the best camouflage strategy is expected to be the most probable background sample [19]. Settling at random (as in our experiment), a common sample has a higher chance of being against a background that is a similar colour to itself, reducing detectability. In our data, survival generally matched the peaks in background luminance (figures 1 and 2), although there was no detectable dip in survival in treatment 5 compared to 4 and 6, as might be expected from the bimodal luminance distribution (figure 1). This could be a lack of statistical power (although our sample size was large, the benefits of a precise match to the background may be small), or an example of where there is an advantage to a ‘compromise’ strategy intermediate between the two modal background shades [14,15,17,32].
In summary, the experiments of Murali et al. [13], for humans searching on artificial backgrounds, and our findings—using natural backgrounds—suggest that background complexity alters the detectability of background-matched targets. This is true even for those targets which have relatively poor, but not the poorest, background matching. This suggests that visual complexity can play a role in the evolution of camouflage in heterogeneous environments [6] and can mitigate the costs of a poorer match. One caveat is that none of our targets was maximally cryptic (all lacked patterning) and tree bark is relatively homogeneous in comparison with other natural substrates (e.g. leaf litter); it would be of interest to see if similar trends obtain for such environments. Animals could also make habitat choices based on visual clutter, selecting habitats of higher complexity for concealment and lower complexity for signalling. This prediction deserves to be tested. Beyond biology, the results are also relevant to understanding human visual search in natural environments and extending approaches familiar to those in applied psychology and ergonomics (e.g. with regard to visual displays) to more naturalistic tasks.
Ethics
The experiment was approved by the University of Bristol Animal Welfare and Ethical Review Body.
Data accessibility
All data are available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.79cnp5hv8 [33]. The data are provided in the electronic supplementary material [34].
Authors' contributions
Z.W.R.: conceptualization, data curation, formal analysis, methodology, writing-original draft, writing-review and editing; D.J.D.A.: data curation; N.C.: data curation; W.F.: data curation; F.S.: data curation; E.J.W.: data curation; N.E.S.-S.: conceptualization, funding acquisition, methodology, supervision, writing-review and editing; I.C.C.: conceptualization, data curation, formal analysis, funding acquisition, methodology, supervision, writing-review and editing.
All authors gave final approval for publication and agreed to be held accountable for the work performed therein.
Competing interests
We declare we have no competing interests.
Funding
I.C.C. and N.E.S.-S. were supported by grant no. BB/S00873X/1 from the Biotechnology & Biological Sciences Research Council, UK.
Acknowledgements
We are grateful to Sam Green and to three anonymous referees for helping to improve this manuscript.
