Improved color constancy in honey bees enabled by parallel visual ...

Improved color constancy in honey bees enabled by parallel visual projections from dorsal ocelli Jair E. Garciaa,1 , Yu-Shan Hungb,c,1 , Andrew D. Greentreed , Marcello G. P. Rosae,f,g , John A. Endlerh , and Adrian G. Dyera,e,2 a

Bio-Inspired Digital Sensing Laboratory, School of Media and Communication, RMIT University, Melbourne, VIC 3000, Australia; b National Vision Research Institute, Australian College of Optometry, Carlton, VIC 3053, Australia; c Eccles Institute for Neuroscience, The John Curtin School of Medical Research, The Australian National University, Canberra, ACT 2601, Australia; d Australian Research Council Centre of Excellence for Nanoscale BioPhotonics, School of Science, RMIT University, Melbourne, VIC 3001, Australia; e Department of Physiology, Monash University, Clayton, VIC 3168, Australia; f Biomedicine Discovery Institute, Monash University, Clayton, VIC 3168, Australia; g Australian Research Council, Centre for Excellence for Integrative Brain Function, Monash University, Clayton, VIC 3168, Australia; and h Centre for Integrative Ecology, School of Life & Environmental Sciences, Deakin University, Waurn Ponds, VIC 3216, Australia

How can a pollinator, like the honey bee, perceive the same colors on visited flowers, despite continuous and rapid changes in ambient illumination and background color? A hundred years ago, von Kries proposed an elegant solution to this problem, color constancy, which is currently incorporated in many imaging and technological applications. However, empirical evidence on how this method can operate on animal brains remains tenuous. Our mathematical modeling proposes that the observed spectral tuning of simple ocellar photoreceptors in the honey bee allows for the necessary input for an optimal color constancy solution to most natural light environments. The model is fully supported by our detailed description of a neural pathway allowing for the integration of signals originating from the ocellar photoreceptors to the information processing regions in the bee brain. These findings reveal a neural implementation to the classic color constancy problem that can be easily translated into artificial color imaging systems. daylight | insect | vision | neuron tracing | von Kries

C

olor constancy allows natural and artificial visual systems to solve the challenge of maintaining the color sensation of an object despite changes in the illumination. Color constancy requires the visual system to discount an unknown illumination function, when the only available information to a sensor may be the total response of the different photoreceptors present in the visual system (1). Color constancy by chromatic adaptation is classically modeled by assuming a set of scalar coefficients that adjust the sensitivity of the different sensors responsible for color vision depending on the particular ambient light conditions (2, 3). The chromatic effect produced by variation of the spectral power distribution (SPD) of daylight is explained by differences in the ratio of long to short wavelengths observed at different conditions of daylight (4). This variability is constrained, potentially making possible the recovery of SPD from a limited number of sensor responses and simplifying color constancy solutions (5). Indeed, it is possible to reconstruct the entire SPD for different phases of daylight with just two functions summarizing daylight variability at short (550 nm) wavelengths (6). However, visual targets often present complex shapes and textures and are illuminated by variable intensity, spectrally mixed illumination, potentially requiring the use of contextual information for achieving color constancy (7, 8). Chromatic adaptation is thus thought unlikely to be the sole mechanism enabling constancy (8, 9), and different cortical or other neural processing mechanisms seem important for reliable color vision (10, 11). Considering human vision, color processing involves multistage neural representations from V1 to V4 and the frontal cortex, and although V4 neurons seem critical for automatic color constancy operations (11–14), subsequent neural processing stages have been implicated (15). www.pnas.org/cgi/doi/10.1073/pnas.1703454114

Although this evidence suggests that a highly complex neural system is necessary for enabling color constancy, many animals without such brain structures encounter the same visual processing problems. For example, honey and bumble bees possess a trichromatic color visual system (16), for which it has been shown that color constancy can be maintained despite large changes in the spectral properties of the illumination (17, 18). Color signals captured in the bees’ photoreceptors are sequentially processed along multiple neural pathways, including opponent neurons in the inner layer of the medulla, which seem to exhibit a capacity to adapt to stimulus input (19). Bumble bees are capable of sensing changes in overhead illumination and making complex foraging decisions (20–22), suggesting potential use of contextual information for perceiving variations in ambient illumination. For example, considering changes in the illumination from short to long wavelength-rich radiation, free flying bumblebees (23) and honey bees (17, 24) make reliable color decisions for most dynamic changing conditions. In addition to the two main compound eyes facilitating trichromatic vision, worker honey bees have, as do most insects, three simple lens eyes located on the dorsal surface of the head known as ocelli. Although these organs have been traditionally proposed to mediate flight-oriented Significance Color sensing requires a capacity to discount the changing color of natural light. We present a biologically validated mathematical solution to this classic problem based on honey bee color vision. The observed spectral tuning of two simple ocellar photoreceptors in the honey bee allows for an optimal color constancy solution to different light environments, including standard CIE (Commission Internationale de l’Eclairage) illuminations, natural forest light, sunlight, or shade. The model is fully supported by a neural pathway potentially allowing for the transfer of spectral information originating from the ocellar photoreceptors to the centralized information processing regions in the brain and explains previously observed behavioral results. This solution to color constancy can be implemented into color imaging systems to enable accurate color interpretation. Author contributions: J.E.G., Y.-S.H., A.D.G., M.G.P.R., J.A.E., and A.G.D. designed research; J.E.G., Y.-S.H., A.D.G., and J.A.E. performed research; J.E.G., Y.-S.H., A.D.G., J.A.E., and A.G.D. analyzed data; and J.E.G., Y.-S.H., A.D.G., M.G.P.R., J.A.E., and A.G.D. wrote the paper. The authors declare no conflict of interest. This article is a PNAS Direct Submission. Freely available online through the PNAS open access option. 1

J.E.G. and Y.-S.H. contributed equally to this work.

2

To whom correspondence should be addressed. Email: [email protected].

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10. 1073/pnas.1703454114/-/DCSupplemental.

PNAS | July 18, 2017 | vol. 114 | no. 29 | 7713–7718

NEUROSCIENCE

Edited by John G. Hildebrand, University of Arizona, Tucson, AZ, and approved May 31, 2017 (received for review March 1, 2017)

processes (25), ocelli provide underfocused images (26), have large apertures or field of view, and possess two spectrally different photoreceptor classes containing AmUVop and AmLop2 opsins (27), with peak absorption at 335–360 and 499–500 nm, respectively (28, 29), in their skyward-facing ventral retinas (26). Interestingly, the ocellar long-wavelength opsin differs from the corresponding opsin present in the compound eye (AmLop) (27). This difference suggests independence between the two systems, which would also explain previous behavioral observations that the immediate color corrections in changing illumination conditions do not fit with the time course for physiological chromatic adaptation (1). Furthermore, ocellar ventral retinas are innervated by two small neurons termed S neurons (25, 26). We hypothesized that detecting changes in the ratio of shortto long-wavelength radiation typical in common daylight conditions is a key function of the ocelli. We also reasoned that an efficient biological implementation of the ocellar inputs for attaining color constancy would require the existence of a parallel visual pathway from these structures to higher-order visual processing centers of the bee brain. This pathway would allow the visual system to use a priori knowledge of the general spectral characteristics of the illuminant to achieve color constancy based on either top-down mechanisms or local processing at hierarchically “high” centers. To test our hypothesis, we defined the spectral sensing resolution (SSR) of the ocellar system as the power to resolve changes in the relative contribution of short and long wavelengths for a set of different spectra of daylight. SSR is a function of two variables: (i) the distance (∆λ) between the wavelength of maximum probability of absorption of the two photoreceptors present in the ocelli and (ii) the equivalent color temperature (Tc ) used to describe the perceived color of different spectra of daylight observed at different sites in the Northern Hemisphere under specific conditions (6, 30) (Eq. 2). Therefore, Tc is also a measure of the ratio of longer to shorter wavelengths observed in different daylight spectra. In particular, SSR represents the slope of the surface defined by the ratio of the response of the two ocellar photoreceptors for various Tc values (Eq. 4). Considering the importance of color constancy for reliable color vision, we predicted that the observed tuning of ocellar photoreceptors in the honey bee provides an optimal SSR for the most commonly occurring spectra of daylight measured in the Northern and Southern Hemispheres (3, 4, 6, 31–33). We calculated the value of SSR for Tc values from 4,000 to 10,000 K at 200-K intervals. We also modeled the characteristics of SSR when considering natural light environments typical

A

B

C

of forests, which account for the effect of forest geometry, clouds, and gaps in foliage covering (3); ∆λ values from 0 to 260 nm at 20-nm intervals (Fig. 1) were used for our calculations. The extreme values of ∆λ represent two theoretically extreme tuning schemes: (i) where there is only one effective ocellar photoreceptor and (ii) a system where the two photoreceptors are almost completely separated (SSR modeling and definition are in Materials and Methods). We compare model predictions with previously observed behavior and also show how the model would explain bee choices for real flower colors. Results Modeling Results. Modeling shows that SSR increases with ∆λ,

having the maximum coverage of ideal SSR near 260 nm (Fig. 2C). At this ∆λ, there is minimal receptor overlap (Fig. 1), which consistent with existing theory, would improve independent spectral sampling (34). However, this tuning difference (∆λ = 260 nm) requires a spectral separation involving a longwavelength sensitivity not ever observed in bees (35), and the Tc values at which resolution is maximum are not the most frequent values observed in diurnal conditions (Fig. 2D). Indeed, at ∆λ = 140 nm, a tuning close to the actual ∆λ = 150 nm measured for the honey bee (28), the SSR continues to show optimal resolving power. On the other end, the model predicts a sharp drop in resolution as Tc values diminish and the IL /IS ratio approaches unity (Fig. 2C), collapsing at IL /IS = 1 and low (≈4, 400 K) Tc values. At this point, the model thus suggests a failure of color discrimination (lower left corner in Fig. 2C), consistent with behavioral data for this type of illumination (17). We extended the model to include natural illuminations, such as those typically observed in forest environments, for additional validation (Fig. 3). When using the SPDs of forest shade, woodland shade, small gaps, and open cloudy light environments (3) for modeling SSR as a function of ∆λ, results indicate that, in all cases, SSR increases exponentially with ∆λ (P < 0.001), closely approaching SSRmax at ∆λ = 140 nm (Fig. 3). These results are in agreement with the model based on daylight illumination, thus suggesting a compromise between high SSR values and a biologically plausible spectral tuning. However, considering the high degree of variability that can be found in natural forest lights because of filtering effects (36), there will likely be cases, such as forest shade and open gaps environments, where any model will likely fail. This failure would explain behavioral observations that individual bumblebees show a tendency to prefer foraging within a constant illumination condition when finding colored flowers (20).

D

E

Fig. 1. Example of five ocelli photoreceptor tuning schemes used for the simulations leading to the SSR model. (A) ∆λ = 0, (B) ∆λ = 80, (C) ∆λ = 160, (D) ∆λ = 260, and (E) ∆λ = 320 nm; ∆λ values correspond to the distance between the principal absorption peak (α band) of the short- (magenta lines) and long-wavelength (green lines) photoreceptors present in the ocelli of the honey bee (A. mellifera), and ∆λ = 150 nm corresponds to the spectral tuning measured for dorsal ocelli by means of electrophysiological methods (28).

7714 | www.pnas.org/cgi/doi/10.1073/pnas.1703454114

Garcia et al.

B

C

D

NEUROSCIENCE

A

Fig. 2. The different spectral illumination problem and potential mechanisms to achieve color constancy. (A) Relative total number of photons reflected by Gazania rigens recorded using a calibrated image sensor (53) illuminated by daylight-type radiation (xenon source). Outlined area I shows the effect of long wavelength-rich, low color temperature illumination (∼4,000 K), and outlined area II shows short-wavelength, high color temperature illumination (∼10,000 K); inserted grayscales show comparison considering the 400- to 700-nm spectral range. (B) Relative SPDs of various daylight conditions relevant to honey bee vision from 300 to 650 nm modeled after the work by Judd et al. (6) (solid lines) and reconstructed following the CIE method (hence its characteristic common value at 560 nm) (4). Modeled spectral sensitivity curves of the two photoreceptor classes present in the dorsal ocelli of A. mellifera (dashed lines) peaking at 330 and 470 nm. Data converted to quantum irradiance used for calculations. (C) SSR of the ocelli photoreceptor system (color map) as a function of various spectral tuning positions (∆λ) and Tc values. Up arrows on the x axis indicate ∆λ values equal to (i) 40, (ii) 140, and (iii) 260 nm. The model predicts poor resolution for wavelengths Tc < 4, 000, consistent with behavioral results that honey bee color discrimination is affected (17, 46). (D, Upper) Density histogram of Tc values for the Northern Hemisphere (31, 33) (solid line) and approximate Tc values for natural illuminations (3) (green bar) along with their corresponding maximum SSR values for ∆λ = 40, 140, and 260 nm (vertical lines). (D, Lower) The maximal SSR value for ∆λ = 140 nm matches the most frequently observed Tc values.

How Would Such a System Operate? Improvement of color con-

stancy through ocellar input would require neural structures allowing for information integration between ocelli, color signals sensed by the honey bees’ compound eye trichromatic photoreceptors, and the color processing centers in the bee brain. Indeed, early attempts using cobalt chloride for tracing the S neurons suggested that ocelli neurons may be involved in color processing, but results were inconclusive (37). We used fluorescent dye (dextran-conjugated Texas Red) together with confocal microscopy (Materials and Methods) to establish whether information integration pathways do exist in the bee brain. The reconstructions (Fig. 4A) show the existence of a direct neural pathway between ocelli and color processing centers (19) by means of the S neurons (Fig. 4B). These cells descend their axons posteriorly to the protocerebral bridge, turn outward across the median and lateral protocerebrum, and then, project anteriorly around the lobula and into the inner layer of the medulla, a brain region considered as a key color processing center for adaptation of neural signals to stimulus input (38) (Movie S1). Likewise, the soma of one of the S neurons was found posteriorly to the central body, another brain region related to color signal processing in the honey bee (Fig. 4A). Neuroanatomical observations thus confirm that S neurons do receive inputs from the ventral ocelli

Garcia et al.

retinas (which directly view the sky) and show that neurons can provide information to color processing centers. This input to the higher levels of visual processing in the bee brain allows for implementation of an improved and almost immediate chromatic adaptation as described below. Color Constancy Enabled by Ocellar Input. Information from the

ocellar input may improve color constancy in the honey bee by providing the necessary, instantaneous measurements of the characteristics of the ambient light. This information can then be used to enable adaptation mechanisms for attaining color constancy in a classic von Kries sense (i.e., by varying the sensitivity of three photoreceptors in the compound eyes in accordance to the reconstructed ambient light spectrum) (39). This approach would thus require the reconstruction of the spectral characteristics of the ambient light from just two parameters, namely the response of the two photoreceptor classes in the dorsal ocelli (IS and IL in Eq. 3). Empirical characterization of different conditions of daylight in Northern (6) and Southern Hemispheres (32) plus different forests in both hemispheres (3) shows that the spectral profile of daylight can be estimated by linear models with just a few parameters (36). Furthermore, the spectral signal from natural surfaces

PNAS | July 18, 2017 | vol. 114 | no. 29 | 7715

Fig. 3. SSR as a function of ∆λ when considering light environments typical of vegetation (3). (A) Relative spectral quantum irradiance for forest shade (Tc ≈ 5,160 K; green line), woodland shade (Tc ≈ 6,910 K; blue line), small gaps (Tc ≈ 4,300 K; orange line), and open/cloudy (Tc ≈ 5,190 K; gray line) light environments. (B) SSR as a function of ∆λ for each vegetation light environment. Arrows on the left side of the y axis represent the SSR values predicted for ∆λ = 40,140, and 160 nm. Values on the second y axis correspond to the mean SSR value for the different environments at each respective ∆λ. For ∆λ = 140 nm, which is biologically plausible, SSR values are close to the theoretical maximum SSR value expected at ∆λ = 250 nm.

and illuminations tends to be limited; thus, its spectral profile can be recovered from the output of the receptors by assuming linear models and flat surfaces (40–43). Therefore, it is possible to

reconstruct an approximation of the spectral profile of ambient illumination from the ocellar response by replacing the continuous spectral function by a linear sum of basic functions (5, 44, 45). To assess the potential advantage of using ocelli input from the skyward-facing retina for improving color constancy under various daylight conditions, we calculated the spectral shift predicted from our model against the shift expected from the implementation of classic von Kries chromatic adaptation (Fig. 5). To this end, we initially reconstructed the SPDs of various phases of daylight using Tc values ranging from 4,000 to 10,000 K (Fig. 2B) from their respective ocellar responses (IL , IS ) and subsequently, used the reconstructed spectra for calculating the chromatic adaptation coefficients following standard methods (46, 47). We evaluated the ecological validity of our model using reflectance spectra corresponding to 111 flowers used by ref. 48 and assumed the same SPDs used for calculating the ocellar SSR. The extent of the perceptual difference expected from a honey bee observing each flower under different daylight conditions, assuming the two chromatic adaptation mechanisms, was measured as the length of the vector of the respective loci representing the photon captures expected for each flower and illumination in two color spaces (49): (i) the Maxwell Triangle (MT) and (ii) the Receptor Noise Model (RN) (Fig. 5). Although the MT makes no assumptions about the underlying sensory or neural processing of the color signals (47, 49), the RN predicts thresholds of chromatic discriminability based on the signal to noise ratio imposed by photon noise and internal receptor noise under photopic conditions (50, 51). The use of these two color spaces

Fig. 4. (A) 3D reconstructions of the two honey bee ocellar interneurons (purple and green traces; posterior view) and the color processing centers [medulla (Me), lobula (Lo), lateral ocellus (LOc), oesophageal forman (OF), mushroom bodies (MB), and the central body (CB)] in the honey bee brain. The neurons were traced and reconstructed from serial optical sections of the brain loaded with fluorescent dye (dextran-conjugated Texas Red). All of the neurons innervating the right lateral ocellar retinas were filled; only the two S neurons that project laterally to the optic lobes were traced. Upper shows an enlargement of the boxed area in Lower. The S neurons descend within the ocellar track (red dashed line) posteriorly to the CB and have their axon collaterals in the median and lateral posterior protocerebrum as well as the inner layer of the Me. (Scale bar: 100 µm.) (B) Schematic diagram of the suggested color pathways in the bee brain. The horizontal and vertical arrows are the light stimuli, and the colors of the arrows indicate the different types of spectral sensitivity of the two visual systems (ocelli: purple, 350 nm; cyan, 490 nm; compound eyes: purple, 350 nm; blue, 450 nm; green, 550 nm) in addition to the compound eye color processing pathway. The compound eye pathway is based on the anatomical and electrophysiological data by ref. 19.

7716 | www.pnas.org/cgi/doi/10.1073/pnas.1703454114

Garcia et al.

A

B

C

thus allows for comparing the expected perceivable color shift with different assumptions regarding the neural processing of color signals. Simulation results show that a mechanism enabled by ocellar inputs improves color constancy by diminishing the magnitude of the chromatic shift (DI ), measured as the Euclidean distance between the first and the last points of each locus in the respective color space, expected from changes in illumination compared with the classic von Kries chromatic adaptation model (DvK ) in the two color spaces considered (Fig. 5 ¯ vK = 4.0 ± 2.7 × 10−2 ; D ¯ I = 8.3 ± 6.5 × 10−3 ; A and B) [MT: D Wilcoxon signed rank test (W ) = 6, 900; P < 0.001; RN: ¯ vK = 1.7 ± 1.1; D ¯ I = 2.8 ± 2.1 × 10−1 ; W = 6, 900; P < 0.001]. D Reductions in color shift are predicted to be 79 and 83% by the MT and RN models, respectively. The length (distance) of the locus in a given color space can also be used to establish the minimum color difference (i.e., threshold) required by an observer to accurately identify two color samples as being different from each other. Because of differences in the assumptions underlying the two color spaces, the precise value of the threshold varies for each model. For example, using psychophysics experiments, Dyer and Neumeyer (24) established a distance of 0.04 distance units in the MT as the minimum difference required for trained bees to discriminate between similar yellow or blue hues presented simultaneously to a honey bee (24). Likewise, Vorobyev and Osorio (50) and Vorobyev et al. (51) defined a distance of 1.0 distance unit in the RN space (∆S ) as being the minimum color dissimilarity required for a bee to reliably identify two colors as different. Interestingly, honey bees have been trained to discriminate between similar broadband colors stimuli close to this predicted value (24). Therefore, we can use the threshold values of 0.04 and 1.0 in the MT and RN spaces, respectively, to predict the color shift likely perceived by a honey bee for potentially achieving color constancy by the classic von Kries method or the proposed ocellar-mediated color constancy model. First, the implementation of the SSR model significantly re¯ I = 0.04; P < duces color shift considering either the MT (Ho: D ¯ I = 1.00; P < 0.001) color spaces. Sec0.001) or the RN (Ho: D ond, such a reduction is biologically relevant, because von Kries ¯ vK = 1.00; adaptation is above the threshold in the RN (Ho: D P < 0.001), and SSR is significantly below the threshold in both the MT and RN models. Consequently, honey bees would significantly benefit from attaining color constancy by means of a mechanism that uses ocellar inputs, because this method ensures that the potential color shift produced by changes in illumination is smaller than the color discrimination threshold. This mechanism, which may be implemented by top-down information flow from the medullar or central body or local computations at these strucGarcia et al.

tures, thus ensures the perception of a flower’s color as being constant under the wide range of ambient illuminations more frequently observed in natural conditions (Fig. 2D). The biologically validated mathematical solution that we reveal can be readily implemented into artificial systems; indeed, we were able to validate model predictions considering common natural lighting environments in which a bee may forage. In these scenarios, the mechanism provides SSR values close to the ideal SSR value at the ∆λ value measured for the honey bee ocelli. Because ocelli are common in many insect taxa, our results suggest that the search for analogous mechanisms in other species would be of high value. Materials and Methods Ocellar Photoreceptor Response. The light intensity received by a photoreceptor in the dorsal ocelli of Apis mellifera (I) was defined as

Ii =

Z650 E(λ)Si (λ) dλ,

[1]

300

where E(λ) represents the spectral irradiance of daylight corresponding to a given equivalent color temperature (Tc ) (6), and Si (λ) denotes the spectral sensitivity for each one of the i = 2 photoreceptors present in the dorsal ocelli of the honey bee (28). Tc values describe the color sensation produced by a black body radiator with temperature T (Kelvin), such that (4) " !#−1 θ1 1 1 1 + − , [2] Tc = Tj θ1 + θ2 Tj+1 Tj where Tj and Tj+1 represent two short lines (isolines) crossing the Planckian locus describing the color of a black body as a function of temperature in a CIE (Commission Internationale de l’Eclairage) uniform chromaticity diagram (30), and θ1 and θ2 are the angles between the Tj and Tj+1 isolines. The spectral sensitivity function for each photoreceptor was reconstructed using the nomogram template by Stavenga et al. (52), including the main (α), secondary (β), and tertiary (γ) absorption bands, and assuming a vitamin A1 visual pigment, common in the order hymenoptera (35). The different spectral tuning (∆λ) instances evaluated by the model were simulated by systematically changing the value of maximum absorption wavelength (λmax ) for the α band of one of the ocellar photoreceptors from 340 to 600 nm at 20-nm intervals, while keeping the λmax value for the α band of the second photoreceptor constant at 340 nm. We arbitrarily named the photoreceptor kept constant as “short” (IS ) and the one varying in its maximum absorption peak as “long” (IL ). Absorption peak values for the β and γ bands for the L and S photoreceptors were kept constant at 340 and 275 nm, respectively, as suggested in ref. 52. SSR Modeling. SSR was defined as the resolution attainable by the dorsal ocelli of the honey bee for reconstructing the SPD of the overhead ambient illumination. Therefore, SSR corresponds to the slope of the surface described by the function η(∆λ, Tc ) (Fig. 2C), which expresses the ratio of

PNAS | July 18, 2017 | vol. 114 | no. 29 | 7717

NEUROSCIENCE

Fig. 5. Predicted perceivable color shift for 111 flowers when illuminated by Tc between 4,000 and 10,000 K (A) assuming color constancy by classic von Kries chromatic adaptation (white bars in C) and (B) taking into consideration the ocellar input (cross-hatched bars in C). The chromatic shift for each flower is (A and B) plotted as locus in the Maxwell color space and quantified by its Euclidean distance in both spaces in C. Dashed lines in C indicate the minimum distance in each color space required by two stimuli to be perceived as being dissimilar in color. **Significant differences at α = 0.05.

the response of the two Ii photoreceptors (Eq. 1) to a light intensity corresponding to a given Tc value for a ∆λ position: η(∆λ, Tc ) =

IL . IS

[3]

SSR (Fig. 2) is thus defined as SSR(∆λ, Tc ) =

∂ η(∆λ, Tc ). ∂Tc

[4]

We solved for Eq. 4 numerically using the routine gradient in Matlab v. 8.0 (R2012b; The Mathworks). Neuron Tracing. The morphologies of the S neurons were obtained using fluorescent dye filling and confocal microscopy. S neurons were filled with Texas Red (Invitrogen), and their morphologies were imaged using optical sectioning. Animals were mounted horizontally (dorsal side up) on a metal holder; the head and thorax were secured with a 3:1 mixture of bee wax and violin resin. To access the ocellar retina, the lens was removed, and a very thin layer of Vaseline was applied to the exposed retinas during dye injection. A glass pipette filled with dextran-conjugated Texas Red (Invitrogen) was inserted dorsally close to the ventral retinal area of the lateral

1. Chittka L, Faruq S, Skorupski P, Werner A (2014) Colour constancy in insects. J Comp Physiol A 200:435–448. ¨ 2. von Kries J (1902) Chromatic Adaptation (Festschrift der Albrecht Ludwigs Universitat, Freiberg, Germany). 3. Endler JA (1993) The color of light in forests and its implications. Ecol Monogr 63: 2–27. 4. Wyszecki G, Stiles WS (1982) Color Science Concepts and Methods, Quantitative Data and Formulae (Wiley, New York), 2nd Ed. 5. Maloney LT, Wandell BA (1986) Color constancy: A method for recovering surface spectral reflectance. J Opt Soc Am A 3:29–33. 6. Judd DB, et al. (1964) Spectral distribution of typical daylight as a function of correlated color temperature. J Opt Soc Am 54:1031–1040. 7. Land EH, McCann JJ (1971) Lightness and retinex theory. J Opt Soc Am 61:1–11. 8. Hurlbert A (2007) Colour constancy. Curr Biol 17:R906–R907. 9. Webster MA, Mollon JD (1995) Colour constancy influenced by contrast adaptation. Nature 373:694–698. 10. Shapley R, Hawken MJ (2011) Color in the cortex: Single-and double-opponent cells. Vision Res 51:701–717. 11. Werner A (2014) Spatial and temporal aspects of chromatic adaptation and their functional significance for colour constancy. Vision Res 104:80–89. 12. Conway BR, et al. (2010) Advances in color science: From retina to behavior. J Neuroscience 30:14955–14963. 13. Foster DH (2011) Color constancy. Vision Res 51:674–700. 14. Rinner O, Gegenfurtner KR (2002) Cone contributions to colour constancy. Perception 31:733–746. 15. Conway BR, Moeller S, Tsao DY (2007) Specialized color modules in macaque extrastriate cortex. Neuron 56:560–573. 16. Peitsch D, et al. (1992) The spectral input systems of hymenopteran insects and their receptor-based colour vision. J Comp Physiol A 170:23–40. 17. Neumeyer C (1981) Chromatic adaptation in the honeybee: Successive color contrast and color constancy. J Comp Physiol A 144:543–553. 18. Werner A, Menzel R, Wehrhahn C (1988) Color constancy in the honeybee. J Neurosci 8:156–159. 19. Dyer AG, Paulk AC, Reser DH (2011) Colour processing in complex environments: Insights from the visual system of bees. Proc Biol Sci 278:952–959. 20. Arnold SEJ, Chittka L (2012) Illumination preference, lllumination constancy and colour discrimination by bumblebees in an environment with patchy light. J Exp Biol 215:2173–2180. 21. Lotto RB, Chittka L (2005) Seeing the light: Illumination as a contextual cue to color choice behavior in bumblebees. Proc Natl Acad Sci USA 102:3852–3856. 22. Dyer AG (2006) Bumblebees directly perceive variations in the spectral quality of illumination. J Comp Physiol A 192:333–338. 23. Dyer AG, Chittka L (2004) Biological significance of distinguishing between similar colours inspectrally variable illumination: Bumblebees (Bombus terrestris) as a case study. J Comp Physiol A 190:105–114. 24. Dyer AG, Neumeyer C (2005) Simultaneous and successive colour discrimination in the honeybee (Apis mellifera). J Comp Physiol A 191:547–557. 25. Ribi W, Warrant E, Zeil J (2011) The organization of honeybee ocelli: Regional specializations and rhabdom arrangements. Arthropod Struct Dev 40:509–520. 26. Hung Y-S, Ibbotson MR (2014) Ocellar structure and neural innervation in the honeybee. Front Neuroanat 8:6. 27. Velarde RA, et al. (2005) Pteropsin: A vertebrate-like non-visual opsin expressed in the honey bee brain. Insect Biochem Mol Biol 35:1367–1377.

7718 | www.pnas.org/cgi/doi/10.1073/pnas.1703454114

ocellus. The glass pipette was inserted into the live animal at room temperature for a minimum of 2 h to allow dye uptake and diffusion. After dye injection, the brain was carefully dissected in fixative (3.7% paraformaldehyde in PBS) and then transferred to 3.7% paraformaldehyde in methanol for 45 min. The sample was then dehydrated through an ethanol series and mounted in methyl salicylate (Sigma-Aldrich). Filled neurons were optically sectioned (2-µm thickness) on a confocal microscope (LSMPascal; Zeiss) with a 10× objective lens (N.A. = 0.45; Apochromat; Carl Zeiss). The image series was processed in the Zeiss LSM image viewer and Fiji to produce 2D projections of the bee brain, whereas the 3D model was reconstructed from an image series of the head together with the filled neurons. The images were processed using AMIRA 5.3.3 (Visage Imaging GmbH). Each image was segmented into discrete components by manually tracing outlines of the filled neurons and different brain regions. 3D mesh models of each structure were then generated from the segmented images. ACKNOWLEDGMENTS. Y.-S.H. acknowledges Prof. Michael Ibbotson for discussions and granting access to the confocal microscope. A.D.G. was supported by Australian Research Council Grant FT160100357. M.G.P.R. and A.G.D. were supported by Australian Research Council Grant DP 0878968. J.A.E. was supported by Australian Research Council Grant DP150102817.

28. Goldsmith TH, Ruck PR (1958) The spectral sensitivities of the dorsal ocelli of cockroaches and honeybees an electrophysiological study. J Gen Physiol 41:1171–1185. 29. Ogawa Y, Ribi W, Zeil J, Hemmi JM (2017) Regional differences in the preferred e-vector orientation of honeybee ocellar photoreceptors. J Exp Biol 220:1701–1708. 30. Kelly KL (1963) Lines of constant correlated color temperature based on Macadam’s (u, v) uniform chromaticity transformation of the CIE diagram. J Opt Soc Am 53: 999–1002. 31. Condit HR, Grum F (1964) Spectral energy distribution of daylight. J Opt Soc Am 54:937–944. 32. Dixon ER (1978) Spectral distribution of Australian daylight. J Opt Soc Am 68: 437–450. 33. Henderson ST, Hodgkiss D (1964) The spectral energy distribution of daylight. Br J Appl Phys 15:947–952. 34. Worthey JA, Brill MH (1986) Heuristic analysis of von Kries color constancy. J Opt Soc Am A 3:1708–1712. 35. Briscoe AD, Chittka L (2001) The evolution of color vision in insects. Annu Rev Entomol 46:471–510. 36. Chiao CC, Osorio D, Vorobyev M, Cronin TW (2000) Characterization of natural illuminants in forests and the use of digital video data to reconstruct illuminant spectra. J Opt Soc Am A Opt Image Sci Vis 17:1713–1721. 37. Pan KC, Goodman LJ (1977) Ocellar projections within the central nervous system of the worker honey bee, Apis mellifera. Cell Tissue Res 176:505–527. 38. Paulk AC, Dacks AM, Phillips-Portillo J, Fellous J-M, Gronenberg W (2009) Visual processing in the central bee brain. J Neurosci 29:9987–9999. 39. von Kries J (1970) Chromatic Adaptation (MIT, Cambridge, MA), pp 109–119. 40. Maloney L (1986) Evaluation of linear models of surface spectral reflectance with small numbers of parameters. J Opt Soc Am A 3:1673–1683. 41. Maloney LT, Yang JM, Mausfeld R, Heyer D (2003) The Illuminant Estimation Hypothesis and Surface Colour Perception (Oxford Univ Press, Oxford), pp 335–358. 42. Heikkinen V, et al. (2008) Evaluation and unification of some methods for estimating reflectance spectra from RGB images. J Opt Soc Am A 25:2444–2458. 43. Garcia JE, et al. (2015) Differentiating biological colours with few and many sensors: Spectral reconstruction with RGB and hyperspectral cameras. PLoS One 10:e0125817. 44. Maloney LT (2002) Illuminant estimation as cue combination. J Vis 2:493–504. 45. Maloney LT (1999) Physics-Based Approaches to Modeling Surface Color Perception (Cambridge Univ Press, Cambridge, UK), pp 387–416. 46. Dyer AG (1998) The colour of flowers in spectrally variable illumination and insect pollinator vision. J Comp Physiol A 183:203–212. 47. Kelber A, Vorobyev M, Osorio D (2003) Animal colour vision - behavioural test and physiological concepts. Biol Rev Camb Philos Soc 78:81–118. 48. Dyer AG, et al. (2012) Parallel evolution of angiosperm colour signals: Common evolutionary pressures linked to hymenopteran vision. Proc Biol Sci 279:3606–3615. 49. Kemp DJ, et al. (2015) An integrative framework for the appraisal of coloration in nature. Am Nat 185:705–724. 50. Vorobyev M, Osorio D(1998) Receptor noise as a determinant of colour thresholds. Proc Biol Sci 265:351–358. 51. Vorobyev M, Brandt R, Peitsch D, Laughlin SB, Menzel R (2001) Colour thresholds and receptor noise: Behaviour and physiology compared. Vision Res 41:639–653. 52. Stavenga DG, Smits RP, Hoenders BJ (1993) Simple exponential functions describing the absorbance bands of visual pigments. Vision Res 33:1011–1017. 53. Garcia JE, Greentree AD, Shrestha M, Dorin A, Dyer AG (2014) Flower colours through the lens: Quantitative measurement with visible and ultraviolet digital photography. PLoS One 9:e96646.

Garcia et al.