The overall relationship between climatic variables and VIS-NIR reflectance
Climate variables were the first two principal components (PCs) from a Principal Components Analysis (PCA) of 9 climatic niche variables. Climate PC1 was higher in species that inhabit in warmer, drier climatic conditions and correlated most strongly with mean temperature (r= 0.98), solar radiation (r = 0.88) and water vapour pressure (r = 0.87) (Fig. 1 left panel). Climate PC2 was higher in species that inhabit areas with higher mean precipitation (r = 0.78) and low precipitation seasonality (r = -0.77) (see Tables S1,2 for the full PCA results).
Climate was a significant predictor of butterfly reflectance for all body regions. In the multivariate phylogenetic regression, mean total VIS-NIR reflectance was predicted by PC1 (F6,334= 12.58, P < 0.001), PC2 (F6,334= 3.66, P = 0.002), and size (F6,334 = 7.44, P < 0.001) (other variables excluded from the model with P > 0.2). Phylogenetic signal was high (λ = 0.76). Post-hoc PGLS models that used the mean reflectance of each body region as a response variable showed that for all body regions, species from colder environments (lower PC1) had lower reflectance, which corresponded to a gradual decrease in the reflectance of butterfly assemblages across latitudinal gradients (Table 1; Fig. 2,3). PC2 was also significant for all body regions (except for the ventral thorax), with species from high precipitation environments having lower reflectance than species from drier environments. Size predicted the reflectance of both dorsal and ventral entire wing: smaller species tended to show higher mean entire wing reflectance than larger species. No other variables remained significant. The results from the models excluding sexually dimorphic males showed the same trends except that PC2 became non-significant for all ventral regions (Supplementary materials, Tables S3,4). Phylogenetic signals were consistently high for all body regions (all λ > 0.72).
The relationship between overall (VIS-NIR) reflectance and climate was evident for both VIS and NIR wavelength ranges but driven more by NIR reflectance for the thorax and basal wing. Results for VIS and NIR reflectance separately were essentially consistent with the overall reflectance models (see Supplementary materials, Tables S5,6, Figs. S3,4 for the full results). PC1 predicted VIS and NIR reflectance for all body regions; however, the relationship was more robust (larger coefficients) for NIR reflectance for the thorax and basal wing, but not the entire wing (dorsal and ventral). PC2 also predicted the reflectance of most body regions except for ventral thorax and basal regions in the VIS range, and the ventral thorax in the NIR range. Size did not predict VIS reflectance of butterflies but predicted NIR reflectance of the entire wing (both dorsal and ventral); smaller butterflies showed higher NIR reflectance than larger butterflies. The direction of this relationship (i.e. whether coefficients had negative or positive values) was always the same as the overall reflectance models for all significant predictors.
We additionally examined whether NIR reflectance show adaptive variation after accounting for its high correlation with VIS reflectance (see methods). Analyses using residuals from the linear regression between NIR and log(VIS) as response variables confirmed the importance of NIR reflectance of the thorax and basal wing in the overall correlations between reflectance and climate. The multivariate phylogenetic regression model showed that PC1 (F6,334 = 4.63,P < 0.001), PC2 (F6,334 range= 2.34, P = 0.03, and size (F6,334 = 7.69,P < 0.001) were significant predictors of NIR residuals (Table 1). Post-hoc PGLS revealed that the residuals of ventral thorax and basal wing regions were higher in species from hotter environments (with higher PC1; Fig. 4). PC2 predicted dorsal thorax, basal wing, and ventral basal wing regions. Size predicted dorsal and ventral entire wing regions. The trends of the significant relationships were the same as the overall reflectance models (Table S7). All other variables were non-significant with Padj> 0.1.