Understanding how communication systems evolve across heterogeneous environments requires examining multiple signalling modalities. In 2017 and 2018, we examined geographic variation in chemical signals of the insular lizard Gallotia galloti across 16 sites on Tenerife (Canary Islands). We analysed interpopulation variation in the lipophilic fraction of male femoral gland secretions in relation to vegetation cover (NDVI), local sex ratio, ectoparasite load, and macroclimatic factors. In 2023, we also analysed visual signals of two lizard phenotypes that occur in contrasting habitats of Tenerife. We modelled the chromatic and achromatic conspicuousness of colour patches under both a conspecific (lacertid) and an avian predator visual system to examine habitat-dependent divergence in signal detectability. Chemical signal diversity decreased with increasing shrub vegetation cover, while males from sites with male-biased sex ratios exhibited greater individual chemical richness, consistent with heightened intrasexual competition. Populations in xeric habitats and with higher mite loads showed relatively higher proportions of α-tocopherol, whereas greener habitats were associated with higher proportions of steroids. In contrast, visual analyses revealed that lizards from densely vegetated habitats exhibited greater chromatic conspicuousness and larger chromatic volumes, particularly under the conspecific visual model. Achromatic contrast showed weaker habitat-dependent differentiation, consistent with the hypothesis that luminance divergence is constrained by predator-mediated selection due to the high luminance sensitivity of avian visual systems. Thus, in open habitats, persistent and chemically diverse secretions may enhance communication efficiency under increased exposure, whereas in vegetated environments chromatic visual signals appear to be amplified, potentially driven by sexual selection while constrained by predation risk. These findings suggest that environmental heterogeneity can promote complementary divergence across signalling modalities, highlighting the importance of integrating sensory ecology and social context to understand the evolution of animal communication.

Statistics Analysis of the Lipophilic Fraction of Male Chemical Signals

Spatial autocorrelation (SAC)

We used Moran's matrix I, where the null hypothesis equals the spatial independence of the data (Legendre and Legendre, 1998). The following variables were spatially independent: MCCR (sd = 0.056, P = 0.46), LCCR (sd = 0.064, P = 0.66), MSDC (sd = 0.060, P = 0.76) and LSDC (sd = 0.060, P = 0.31). However, the PC1 chemicals (sd = 0.055, P = 0.049) and the PC2 _chemicals were spatially autocorrelated (sd = 0.062, P = 0.009). To control SAC in the models analysing PC1_chemicals and PC2_chemicals, we included as predictors an inverse distance-weighted function of neighbouring response (autocovariate) (Dormann et al., 2007).

Model building

Statistical analyses were performed in R version 4.0.4 (R Core Team, 2021). The sampling year (2017 or 2018) and the colour phenotype (i.e., ‘galloti’ or ‘eisentrauti’) were established as factor predictors in a general linear model. We included local sex ratios as continuous predictors (as a proxy of intrasexual competition), median abundance of mites per site in males (as proxies of parasitic pressure), a normalized difference vegetation index (NDVI) (and index that is used to quantify vegetation greenness and approach vegetation density; Weier and Herring, 2020), the human activity index (a proxy of human footprint pressure), and the two bioclimatic/elevation principal factors. We checked the multicollinearity of the predictors in the models using a variance inflation factor (VIF) criterion (Schroeder et al., 1990). We also checked the parametric assumption of model residuals by simulating the residual distribution with the package DHARMa (Hartig, 2024). All models for the chemical composition indices of male lipids in femoral secretions showed a good fit to a Gaussian error distribution. A weighting term, namely, the corresponding sample size of lizards sampled for chemicals per locality, was included in the models.

Model selection and model averaging on male chemical signals

We analysed the geographic variation in the relative abundance of lipids in male femoral secretions performing model selection and model averaging on the Gaussian models with the functions ‘dredge()’ and ‘model.avg()’ of the package ‘MuMIn’ (Bartoń, 2024). This statistical method is recommended for small sample sizes because it removes effects with low influence for the final model (Hegyi and Garamszegi, 2011). For this, the models were considered sufficiently informative when their ΔAICc ≤ 2 (Burnham and Anderson, 2004). We calculated the relative importance of each predictor as the frequency in which it appeared in the equally likely models. Furthermore, a final model was calculated by averaging all equally likely models. We calculate the adjusted standard error of the z-standardized ß coefficient ± adjusted standard error of the predictors in this final model.

Colour measurements of male and female lizards (2023)

In April 2023, we collected spectral data from the colour patches of 100 adult lizards (52 males and 48 females). Specifically, 49 lizards (26 males and 23 females) were sampled in the north of Tenerife for phenotype ‘eisentrauti’, which were clearly differentiated because they have yellow stripes, and 51 lizards (26 males and 25 females) in the centre and southern localities for phenotype ‘galloti’, which only had UV-blue patches (Figure 1c). Each of the colour phenotypes was sampled at three different localities (5-6 individuals by sex and locality; Figure 1e). We collected the spectral reflectance from their first and second eyespots, first and second ventrolateral patches, cheek, yellow patch, and dorsum (Supplementary information). This represents most of the colour patches they have (Figure 1a). The yellow patch as well as the blue cheek is only present in ‘eisentrauti’, so to obtain a proper comparison, we measured reflectance in individuals of phenotype ‘galloti’ in the same body regions. In addition, we sampled a single reflectance spectrum for each of a total of 27 background items (that is, stones, soil, flowers, leaves; Supplementary information) on the same sites of Tenerife where colour of lizards was measured. We chose these items based on their apparent representativeness of the sampled habitats and to be used as a background colouration in the visual models (see below). We measured their reflectance spectra in the field as well as the colour patches of lizards using a Jaz portable spectrophotometer connected to a pulsed xenon lamp (Ocean Optics Inc., Dunedin, FL, USA). We used as reference spectrum one measurement performed on four layers of white-matte Teflon tape mounted on top of a commercial white standard that provided a flat and constant reference white surface. These layers of clean Teflon were replaced by new ones every time they showed signs of dirt. The probe of the spectrophotometer was fitted inside a fully opaque plastic piece that enabled placing the spectrometer probe at constant distance of 3 mm from the measured surface, forming an angle of 90º, and avoiding environmental light to bias the spectral measurements (Megía-Palma et al., 2022). We used an integration time of 10 ms and a boxcar width of 4 units.

We built two visual models based on an ultraviolet sensitive (UVS) lacertid and a violet sensitive (VS) predatory bird and measured the conspicuousness of the coloration of G. galloti against natural backgrounds. To test whether colour patches of lizards are more visible to conspecifics than to predators, conspicuousness was measured as chromatic (i.e., chromaticity) and achromatic (i.e., luminance) distances, expressed as just noticeable differences (JND). JND values over three units were considered discriminable, whereas those between one and three units were considered discriminable under good light conditions (Siddiqi et al., 2004). For the predator model, we focus on the common kestrel, Falco tinnunculus, as G. galloti individuals represent one of the main diet items of this species in Tenerife (Carrillo et al., 2017). To model its vision, we assumed the same parameters as Goedert et al. (2021) for a close relative, Falco sparverius (405, 449, 504 and 567 nm, corresponding to VWS, SWS, MWS and LWS cones in proportions of 1:2:3:3). For the lacertid model, we assumed cone absorbances and abundances from another species in the same family, Podarcis muralis (367, 456, 497 and 562 nm, corresponding to UVWS, SWS, MWS and LWS cones in proportions of 1:1:1:4; Martin et al., 2014). To fit these models, we used the sensmodel function of pavo v.2.7.1 (Maia et al., 2019), which is based on the templates of Govardovskii et al. (2000) and Hart and Vorobyev (2005). For both models, we used the receptor noise limited model (Vorobyev and Osorio, 1998). We used an irradiance spectrum corresponding to daylight conditions ‘D65’ implemented in pavo for both visual models. We used the same irradiance spectrum (instead of irradiance spectra from every site where spectral data was collected) to avoid introducing biases related to the date and climatological conditions associated with each of the sampling events. The Weber fraction value was set to 0.1 for the visual model of Falco (Olsson et al., 2018) and to 0.05 for the lacertid visual model (Martin et al., 2014). To estimate the conspicuousness between the colour patches of G. galloti and the items present in their habitats, we retained all comparisons between colour patches and background elements. Pairwise comparisons between lizard colour patch and field background colour item resulted in 16,200 chromatic/achromatic JND scores per visual model. Given that the conspicuousness between colour patches greatly depends on the background against which they are compared (Caves et al., 2024), we used an averaged back spectrum in the darkest (nearly black) dorsal area as a reference background that was measured in every case avoiding the dorsal yellow stripes in ‘eisentrauti’. The estimates of JND were not averaged per phenotype, instead they were obtained separately for each sex (×2) and colour phenotype (×2) due to the differences they show in their dorsal spectra (Supplementary information). To examine our hypothesis that chromatic contrast would show stronger phenotypic and receiver-dependent differentiation than achromatic contrast, we performed a pairwise within-colour patch analysis with both chromatic and achromatic contrasts simultaneously set as response variables and the three-way interaction between visual model, lizard phenotype and sex, and all surrogate two-way interactions and main effects, as predictors. We inspected the resulting model residuals to ensure that they conformed to assumptions of normality, homoscedasticity, autocorrelation, and kurtosis. We performed a Bonferroni post-hoc test to inspect for significant differences across levels of the three-way interaction.

            Further, we explored whether colour patches showed greater variability under the conspecific visual model, which would suggest that colour may be optimized for intraspecific communication. To do so, we used the ‘voloverlap’ function from pavo v.2.7.1 (Maia et al., 2019) to calculate the chromatic volume (i.e., the spectral variability of a colour as perceived by the receptor’s tetrachromatic vision represented in a 3-dimensional visual space) of each colour patch of lizards from the six sampling sites separately. We did this using the** visual models built using the Falco and Podarcis cone absorbances and proportions (see above). Therefore, this analysis provided a single chromatic volume score per colour patch (resulting from all individuals analysed per locality), receptor visual model, and locality. For this analysis, we used an idealized homogeneous background (i.e., constant illuminance value of 1 across all wavelengths), which also allowed us to obtain the chromatic volume of the dorsum. Then, we fitted a generalized linear mixed model (GLMM) with chromatic volumes as response, lizard colour phenotype, receptor visual model (i.e., conspecific or predator), and their two-way interaction as fixed factors. We also included the body region and locality as random terms in the analysis.